Advances in Land Remote Sensing: System, Modeling, Inversion and Application

Advances in Land Remote Sensing Advances in Land Remote Sensing System, Modeling, Inversion and Application Shunlin ...

Author: Shunlin Liang

150 downloads 1743 Views 9MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

Advances in Land Remote Sensing

Advances in Land Remote Sensing System, Modeling, Inversion and Application

Shunlin Liang Editor Department of Geography, University of Maryland, College Park, MD, USA

123

Shunlin Liang University of Maryland College Park, MD USA

ISBN 978-1-4020-6449-4

e-ISBN 978-1-4020-6450-0

Library of Congress Control Number: 2007940919 c 2008 Springer Science+Business Media B.V. ° No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written permission from the Publisher, with the exception of any material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work. Cover illustration: From sensors and platforms, to information extraction and applications (compilation by S. Liang) Printed on acid-free paper 9 8 7 6 5 4 3 2 1 springer.com

Preface

This book is primarily based on presentations in the three reviewing panels of the 9th International Symposium on Physical Measurements and Signatures in Remote Sensing held at the Institute for Geographical Sciences and Natural Resource Research, Chinese Academy of Sciences, China, in October 2005. It presents a collection of review papers on remote sensing sensor systems, radiation modeling and inversion of land surface variables, and remote sensing applications. Each chapter summarizes the progress in the past few years and also identifies the research issues for the near future.

v

Acknowledgements

This symposium series is affiliated with the International Society for Photogrammetry and Remote Sensing (ISPRS) Commission VII/I Working Group on Fundamental Physics and Modeling led by Professor Michael Schaepman (Chair, Wageningen University, the Netherlands), Professor Shunlin Liang (co-Chair, University of Maryland, USA), and Dr. Mathias Kneubuehler (Secretary, University of Zurich, CH, Switzerland) (2004–2012). It was sponsored and/or financially supported by the Institute of Geographical Sciences and Natural Resources Research (IGSNRR) of Chinese Academy of Sciences (CAS), Institute of Remote Sensing Applications (IRSA) of CAS, Chinese 973 Project “Quantitative Remote Sensing of Major Factors for Spatio-temporal Heterogeneity on the Land Surface” undertaken by Beijing Normal University, US National Aeronautics and Space Administration (NASA), International Society for Photogrammetry and Remote Sensing, IEEE Geoscience and Remote Sensing Society, and Scientific Data Center for Resources and Environment, CAS. This symposium would not be successful without scientific leadership by the International Scientific Committee and effective organization by the local Organizing Committee.

vii

International Scientific Committee

Professor Guanhua Xu, Minister of the Chinese Ministry of Science and Technology, China (honorary Chair) Professor Shunlin Liang, University of Maryland, USA (Chair) Dr. Fr´ed´eric Baret, INRA, Avignon, France Professor Mike Barnsley, University of Wales Swansea, UK Professor Marvin Bauer, University of Minnesota, USA Professor Jon Benediktsson, University of Iceland, Iceland Professor Jing Chen, University of Toronto, Canada Professor Peng Gong, University of California at Berkeley, USA Dr. David Goodenough, Pacific Forestry Centre, Natural Resources Canada Dr. Tom Jackson, USDA /ARS at Beltsville, Maryland, USA Dr. David Jupp, CSIRO Earth Observation Centre, Australia Dr. Yann Kerr, CNES/CESBIO, France Dr. Marc Leroy, MEDIAS, France Dr. Philip Lewis, University College London, UK Professor Deren Li, Wuhan University, China Professor Xiaowen Li, Beijing Normal University, China Professor Jiyuan Liu, IGSNRR, CAS, China Dr. John V. Martonchik, Jet Propulsion Laboratory, USA Professor Ranga Myneni, Boston University, USA Professor Ziyuan Ouyang, Institute of Geochemistry, CAS, China Dr. Jeff Privette, NASA /GSFC, USA Dr. Jon Ranson, NASA /GSFC, USA Professor Michael Schaepman, Wageningen University, The Netherlands Professor Jose Sobrino, University of Valencia, Spain Dr. Karl Staenz, Canadian Centre for Remote Sensing, Canada Professor Alan Strahler, Boston University, USA Professor Qingxi Tong, Institute of Remote Sensing Applications, CAS, China Dr. Michel Verstraete, JRC, Ispra, Italy Dr. Charlie Walthall, USDA /ARS at Beltsville, Maryland, USA Dr. Zhengming Wan, University of California at Santa Barbara, USA

ix

x

Organizing Committee Professor Jiyuan Liu, Director of IGSNRR, CAS (co-Chair) Professor Xiaowen Li, Director of Center for Remote Sensing and GIS, Beijing Normal University and Director of IRSA, CAS (co-Chair) Professor Dafang Zhuang, Execute Vice-director of Scientific Data Center for Resources and Environment, CAS (Vice Chair) Professor Mingkui Cao, IGSNRR, CAS Professor Changqing Song, Chinese National Foundation of Sciences Professor Renhua, Zhang, IGSNRR, CAS Professor Lixin Zhang, Beijing Normal University Dr. Keping Du, Beijing Normal University Professor Mengxue Li, National Remote Sensing Center of China Professor Boqin Zhu, Institute of Remote Sensing Applications, CAS Dr. Ronggao Liu, IGSNRR, CAS (General Secretary)

Shunlin Liang University of Maryland, College Park, MD, USA

Reviewers

Each chapter is anonymously reviewed by at least one reviewer. Their valuable comments and suggestions have greatly helped to improve the quality of the volume. Jing M. Chen University of Toronto, Canada Jan Clevers Wageningen University, The Netherlands Ruth DeFries University of Maryland, USA Alan R. Gillespie University of Washington, USA Hongliang Fang University of Maryland, USA Xiuping Jia The University of New South Wales, Australia David L.B. Jupp CSIRO Marine and Atmospheric Research, Australia Yann H. Kerr CNES/CESBIO, France Yuri Knyazikhin Boston University, USA Randy Koster NASA, USA Eric F. Lambin University of Louvain, Belgium Tiit Nilson Tartu Observatory, Estonia

xi

xii

Paolo Pampaloni IFAC-CNR, Italy Bernard Pinty EC Joint Research Centre, Italy Jeff Privette NOAA, USA

Reviewers

Contents

1

Recent Advances in Land Remote Sensing: An Overview . . . . . . . . . . Shunlin Liang

1

Part I Remote Sensing Systems 2

Passive Microwave Remote Sensing for Land Applications . . . . . . . . . Thomas J. Jackson

3

Active Microwave Remote Sensing Systems and Applications to Snow Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Jiancheng Shi

4

Multi-angular Thermal Infrared Observations of Terrestrial Vegetation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 Massimo Menenti, Li Jia, and Zhao-Liang Li

5

Terrestrial Applications of Multiangle Remote Sensing . . . . . . . . . . . . 95 Mark J. Chopping

9

Part II Physical Modeling and Inversion Algorithms 6

Modeling the Spectral Signature of Forests: Application of Remote Sensing Models to Coniferous Canopies . . . . . . . . . . . . . . . . . . . . . . . . . 147 Pauline Stenberg, Matti M˜ottus, and Miina Rautiainen

7

Estimating Canopy Characteristics from Remote Sensing Observations: Review of Methods and Associated Problems . . . . . . . . 173 Fr´ed´eric Baret and Samuel Buis

8

Knowledge Database and Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 Jindi Wang and Xiaowen Li

xiii

xiv

Contents

9

Retrieval of Surface Albedo from Satellite Sensors . . . . . . . . . . . . . . . . 219 Crystal Schaaf, John Martonchik, Bernard Pinty, Yves Govaerts, Feng Gao, Alessio Lattanzio, Jicheng Liu, Alan Strahler, and Malcolm Taberner

10

Modeling and Inversion in Thermal Infrared Remote Sensing over Vegetated Land Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245 Fr´ed´eric Jacob, Thomas Schmugge, Albert Olioso, Andrew French, Dominique Courault, Kenta Ogawa, Francois Petitcolin, Ghani Chehbouni, Ana Pinheiro, and Jeffrey Privette

11

Spectrally Consistent Pansharpening . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 Ari Vesteinsson, Henrik Aanaes, Johannes R. Sveinsson, and Jon Atli Benediktsson

12

Data Assimilation Methods for Land Surface Variable Estimation . . 313 Shunlin Liang and Jun Qin

13

Methodologies for Mapping Land Cover/Land Use and its Change . . 341 Nina Siu-Ngan Lam

14

Methodologies for Mapping Plant Functional Types . . . . . . . . . . . . . . 369 Wanxiao Sun and Shunlin Liang

Part III Remote Sensing Applications 15

Monitoring and Management of Agriculture with Remote Sensing . . 397 Zhongxin Chen, Sen Li, Jianqiang Ren, Pan Gong, Mingwei Zhang, Limin Wang, Shenliang Xiao, and Daohui Jiang

16

Remote Sensing of Terrestrial Primary Production and Carbon Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423 Maosheng Zhao and Steven W. Running

17

Applications of Terrestrial Remote Sensing to Climate Modeling . . . . 445 Robert E. Dickinson

18

Improving the Utilization of Remotely Sensed Data . . . . . . . . . . . . . . . 465 John R. Townshend, Stephen Briggs, Roy Gibson, Michael Hales, Paul Menzel, Brent Smith, Yukio Haruyama, Chu Ishida, John Latham, Jeff Tschirley, Deren Li, Mengxue Li, Liangming Liu, and Gilles Sommeria

19

Emerging Issues in Land Remote Sensing . . . . . . . . . . . . . . . . . . . . . . . 485 Shunlin Liang, Michael Schaepman, Thomas J. Jackson, David Jupp, Xiaowen Li, Jiyuan Liu, Ronggao Liu, Alan Strahler, John R. Townshend, and Diane Wickland

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495 CD-ROM included

Contributors

Henrik Aanaes Informatics and Mathematical Modelling, Technical University of Denmark, Denmark [email protected] Frederic Baret UMR1114, INRA-CSE, 84914 Avignon, France [email protected] Jon Atli Benediktsson Department of Electrical and Computer Engineering, University of Iceland, Hjardarhaga 2-6, 107 Reykjavik, Iceland [email protected] Stephen Briggs European Space Agency, Via Galileo Galilei, 00644 Frascati, Rome, Italy [email protected] Sanuel Buis UMR1114, INRA-CSE, 84914 Avignon, France Ghani Chehbouni Institute of Research for the Development, Center for Spatial Studies of the Biosphere, UMR CESBio, Toulouse, France [email protected] Zhongxin Chen Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry of Agriculture, Beijing 100081, China [email protected] Mark J. Chopping Department of Earth and Environmental Studies, Montclair State University, 1 Normal Ave, Montclair, NJ 07043, USA [email protected]

xv

xvi

Contributors

Dominique Courault National Institute for Agronomical Research, Climate – Soil – Environment Unit, UMR CSE INRA / UAPV, 84914 Avignon, France [email protected] Robert E. Dickinson School of Earth and Atmospheric Sciences, Georgia Institute of Technology, 311 Ferst Drive, Atlanta, GA 30332-0340, USA [email protected] Andrew French United States Department of Agriculture/Agricultural Research Service, US Arid Land Agricultural Research Center, 21881 North Cardon Lane, Maricopa, AZ 85238, USA [email protected] Feng Gao Earth Resources Technology, Inc., 8106 Stayton Dr., Jessup, MD 20794, USA Roy Gibson EUMETSAT [email protected] Pan Gong Institute of Agricultural Resources & Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China Yves Govaerts EUMETSAT, Am Kavalleriesand 31, D-64295 Darmstadt, Germany [email protected] Michael Hales NOAA, Silver Spring, MD 20910, USA [email protected] Yukio Haruyama JAXA, 1-8-10 Harumi, Chuo-ku, Tokyo 104-6023, Japan [email protected] Chu Ishida JAXA, 1-8-10 Harumi, Chuo-ku, Tokyo 104-6023, Japan [email protected] Thomas J. Jackson USDA ARS Hydrology and Remote Sensing Lab, 104 Bldg. 007 BARC-West, Beltsville, MD 20705, [email protected] Fr´ed´eric Jacob Institute of Research for the Development, Laboratory for studies on Interactions between Soils – Agrosystems – Hydrosystems, UMR LISAH SupAgro/INRA/IRD, Montpellier, France [email protected]

Contributors

xvii

Li Jia Alterra Green World Research, Wageningen University and Research Centre, The Netherlands [email protected] Daohui Jiang Institute of Agricultural Resources & Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China David Jupp CSIRO Marine and Atmospheric Research, Canberra ACT 2601 Australia [email protected] Nina Siu-Ngan Lam Department of Environmental Studies, Louisiana State University, Baton Rouge, LA 70808, USA [email protected] John Latham Food and Agriculture Organization, Rome, Italy [email protected] Alessio Lattanzio Makalumedia gmbh, Robert-Bosch Strasse 7, 64296 Darmstadt, Germany Deren Li Wuhan University, 39 Loyu Road, Wuhan, 430070, China [email protected] Mengxue Li NRSCC, 15B, Fuxing Road, Beijing, 100862, China [email protected] Sen Li Institute of Agricultural Resources & Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China Xiaowen Li Research Center for Remote Sensing and GIS, Beijing Normal University, No.19 XieJieKouWaiDaJie Street, Beijing 100875, China [email protected] Zhao-Liang Li Institute of Geographic Sciences and Natural Resources Research, Beijing, China Shunlin Liang Department of Geography, University of Maryland, College Park, USA [email protected] Jicheng Liu Department of Geography and Environment, Boston University, 675 Commonwealth Ave., Boston, MA 02215, USA

xviii

Contributors

Jiyuan Liu Institute for Geographical Sciences and Natural Resource Research, Chinese Academy of Sciences, Beijing, China [email protected] Liangming Liu National Remote Sensing Center of China Ronggao Liu Institute for Geographical Sciences and Natural Resource Research, Chinese Academy of Sciences, Beijing, China [email protected] John Martonchik Jet Propulsion Laboratory, Mail Stop 169-237, 4800 Oak Grove Dr., Pasadena, CA 91109, USA [email protected] Massimo Menenti TRIO/LSIIT, University Louis Pasteur (ULP), Strasbourg, France and Istituto per i Sistemi Agricoli e Forestali del Mediterraneo (ISAFOM), Naples, Italy [email protected] Paul Menzel University of Wisconsin, Space Science and Engineering Center, Madison, WI 53706, USA [email protected] Matti M˜ottus Department of Forest Ecology, FI-00014 University of Helsinki, Finland; Tartu Observatory, 61602 T˜oravere, Tartumaa, Estonia [email protected] Kenta Ogawa Department of Geo-system Engineering, University of Tokyo and Hitachi Ltd, Tokyo, Japan Albert Olioso National Institute for Agronomical Research, Climate – Soil – Environment Unit, UMR CSE INRA/UAPV, Avignon, France [email protected] Francois Petitcolin ACRI-ST, Sophia Antipolis, France [email protected] Ana Pinheiro Biospheric Sciences Branch, NASA’s GSFC, Greenbelt, MD 20771, USA [email protected]

Contributors

xix

Bernard Pinty Global Environment Monitoring Unit, IES, EC Joint Research Centre, TP 440, via E. Fermi, I-21020 Ispra (VA), Italy [email protected] Jeffrey Privette NOAA’s National Climatic Data Center, Asheville, NC 28801-5001, USA [email protected] Jun Qin Institute for Geographical Science and Natural Resource Research, Beijing, China [email protected] Miina Rautiainen Department of Forest Resource Management, FI-00014 University of Helsinki, Finland [email protected] Jianqiang Ren Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry of Agriculture, Beijing 100081, China and Institute of Agricultural Resources & Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China Steven W. Running Numerical Terradynamic Simulation Group, Department of Ecosystem and Conservation Science, University of Montana, Missoula, MT 59812, USA [email protected] Crystal Schaaf Department of Geography and Environment, Boston University, 675 Commonwealth Ave., Boston, MA 02215, USA [email protected] Michael Schaepman Centre for Geo-Information, Wageningen University, Wageningen, The Netherlands [email protected] Thomas Schmugge Gerald Thomas Professor of Water Resources, College of Agriculture New Mexico State University, Las Cruces, NM, USA [email protected] Jiancheng Shi Institute for Computational Earth System Science, University of California, Santa Barbara, CA 93106-3060, USA [email protected] Brent Smith NOAA, Silver Spring, MD 20910, USA [email protected]

xx

Contributors

Gilles Sommeria World Climate Research Programme, WMO, CH-1211 Geneva, Switzerland [email protected] Pauline Stenberg Department of Forest Resource Management, FI-00014 University of Helsinki, Finland [email protected] Alan Strahler Department of Geography and Environment, Boston University, 675 Commonwealth Ave., Boston, MA 02215, USA [email protected] Wanxiao Sun Department of Geography and Planning, Grand Valley State University, USA [email protected] Johannes R. Sveinsson Department of Electrical and Computer Engineering, University of Iceland, Hjardarhaga 2-6, 107 Reykjavik, Iceland [email protected] Malcolm Taberner Global Environment Monitoring Unit, IES, EC Joint Research Centre, TP 440, via E. Fermi, I-21020 Ispra (VA), Italy [email protected] John R. Townshend Department of Geography, University of Maryland, College Park, MD 20742, USA [email protected] Jeff Tschirley Food and Agriculture Organization, 00153 Rome, Italy [email protected] Ari Vesteinsson Department of Electrical and Computer Engineering, University of Iceland, Hjardarhaga 2-6, 107 Reykjavik, Iceland Jindi Wang Research Center for Remote Sensing and GIS, Beijing Normal University, No.19 XieJieKouWaiDaJie Street, Beijing 100875, China [email protected] Limin Wang Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry of Agriculture, Beijing 100081, China Diane Wickland NASA Headquarters, Washington, DC, USA [email protected]

Contributors

Shenliang Xiao Institute of Agricultural Resources & Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China Mingwei Zhang Institute of Agricultural Resources & Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China Maosheng Zhao Numerical Terradynamic Simulation Group, Department of Ecosystem and Conservation Science, University of Montana, Missoula, MT 59812, USA [email protected]

xxi

Chapter 1

Recent Advances in Land Remote Sensing: An Overview Shunlin Liang

Earth’s surface is undergoing rapid changes due to urbanization, industrialization and globalization. Environmental problems such as water shortages, desertification, soil depletion, greenhouse gas emissions warming the atmosphere, deforestation, elevated coastal waterway sediment and nutrient fluxes, among other environmental problems, are increasingly common and troubling consequences of human activities. Policy decisions about the environment rely on accurate and reliable information, especially data and understanding leading to better predictions of natural hazards, epidemics, impacts of energy choices, and climate variations. Comprehensive, systematic Earth observations are key to forecasting Earth system dynamics. Predicting future scenarios of our planet’s habitability requires analysis of what has transpired in the past along with observations of present conditions and processes. Timely, quality long-term global data acquired through remote sensing is essential for the ongoing viability and enhancement of human society on Earth. The field of remote sensing (Earth observation) has developed rapidly. Many publications have documented its progress (e.g., the special issue of Remote Sensing Reviews with a set of papers reviewing the modeling and inversion of surface bidirectional reflectance distribution function (BRDF) (Liang and Strahler, 2000), and the edited or authored books on similar subjects (Liang, 2004; Myneni and Ross, 1991). To systematically summarize the achievements of terrestrial remote sensing in recent years and to set the research agenda for the near future, the 9th International Symposium on Physical Measurements and Signatures in Remote Sensing (ISPMSRS), held in October 2005 in Beijing, organized three review panels. The papers compiled in this book are largely from these panels and are organized into three parts, respectively. Part I of the book (corresponding to the first panel, chaired by Drs. David Jupp and Tom Jackson, with Drs. Ralph Dubayah, Michael Schaepman, Jianchen Shi, Stephen Ungar, and David Le Vine) focuses on remote sensing systems and sensors. As there are many different remote sensing systems (the result of various Shunlin Liang University of Maryland, USA S. Liang (ed.), Advances in Land Remote Sensing, 1–6. c Springer Science + Business Media B.V., 2008

1

2

S. Liang

permutations of passive vs. active, specific domain of the electromagnetic spectrum sampled, and spectral channel bandwidth) currently in operation or planned to be operational in the future, this panel evaluated the capabilities of these systems for estimating key land surface variables and how they can best be improved and combined effectively. The panel presentations covered microwave, thermal-infrared (IR), hyperspectral optical, and Lidar remote sensing. Based on these discussions, four chapters are compiled on this subject, including passive microwave, active microwave, multiangle thermal IR, and multiangle optical remote sensing. In Chapter 2 on passive microwave remote sensing, Jackson identifies six factors in passive microwave sensor design that affect the retrieval of land surface properties: frequency, polarization, view geometry, spatial resolution, temporal coverage, and signal-to-noise ratio. While summarizing the features of three current relevant satellite instruments and two other satellite sensors, he points out that the low frequency observations (<6 GHz) and coarse spatial resolution are obstaclesto-land applications. He further discusses three approaches, currently under investigation, for solving these problems. The first is synthetic aperture radiometry, such as the planned Soil Moisture Ocean Salinity mission. The second is use of large, lightweight antennas and the Hydros satellite mission (recently cancelled) which could have demonstrated this approach. The third is disaggregation by integrating with higher resolution visible to thermal remote sensing data through data fusion. The next decade will see several exploratory missions using new technologies, and innovative approaches to integrating passive microwave with active microwave measurements. Shi summarizes, in Chapter 3 on active microwave remote sensing, the significant developments in mapping snow, inferring snow wetness and snow water equivalence using active microwave sensors (mainly synthetic aperture radar – SAR). Following a brief introduction to different SAR systems, including current and future systems, he discusses three types of SAR measurements: backscattering measurements at a given frequency and polarization, polarization properties obtained by a fully polarimetric instrument, and interferometry from repeat-pass sampling. Shi devotes more space to reviewing various techniques that use these three measurements to map snow properties, followed by a discussion of the need for future SAR systems to map snow properties. Multiangle thermal-IR remote sensing is covered in Chapter 4. Menenti et al. first present experimental evidence of angular signatures of thermal-IR observations, then demonstrate that two soil and foliage temperatures normally can account for most thermal IR observations at the top of the canopy. (further division into sunlit and shadowed components are needed only under extreme conditions). They review various geometric-optics and radiative transfer modeling techniques, and finally describe inversion techniques to estimate component temperatures from multiangle thermal-IR observations. The related subject on estimating land surface skin temperature is given in Chapter 10. In Chapter 5 on multiangle optical remote sensing, Chopping first outlines different sensor systems for acquiring multiangle data, and then reviews a variety of land applications using multiangle observations, such as mapping canopy openness,

1 Recent Advances in Land Remote Sensing: An Overview

3

clumping index, structural scattering index, land cover and community type, snow and ice, and dust emissions. He also outlines different modeling techniques for angular signatures in the optical domain and discusses other factors (e.g., nearsimultaneous and accumulated angular sampling, scale) related to multiangle data acquisition. Although these four chapters cover a variety of sensors and applications, two other remote sensing techniques deserve mention. Lidar has a unique ability to measure vertical and spatial heterogeneity across multiple scales. Many studies have demonstrated that lidar can be used for estimating forest structure variables such as canopy height and profile, canopy cover, growth/successional state, foliar profile, stem density, basal area, biomass, and canopy base height. Except for IceSat, currently there are no spaceborne lidar remote sensing systems available. Hyperspectral remote sensing is another approach that has great potential. A more challenging issue faced by the remote sensing community is how to integrate data from different sensor systems in such a way that derivation of land surface variables is maximized. Part II of the book presents surface radiation modeling and inversion methods corresponding to the second panel chaired by Drs. Alan Strahler and Xiaowen Li, and contributed by Drs. Fr´ed´eric Baret, Fr´ed´eric Jacob, Philip Lewis, Massimo Menenti, Pauline Stenberg, and Jindi Wang. There have been considerable investments on developing physical models to understand the surface radiation regimes. Some of these models have been incorporated into very useful algorithms for estimating land surface variables from satellite observations. This panel assessed the progress of surface radiation modeling and satellite inversion algorithm development in many different aspects. The papers in this part are composed of panel presentations and two additional contributions. Note that Chapters 4 and 5 also contain modeling and inversion mainly for multiangle observations. Stenberg et al. in Chapter 6 describe different geometric-optical models and radiative transfer models for simulating the reflectances of heteorogenerous canopies, particularly forests. In radiative transfer modeling, they review mainly the threedimensional models. They also discuss the structural and spectral characteristics of coniferous forests, the clumping effects and reflectance scaling using the p-theory, and the inputs for simulating the forest reflectance. Baret and Buis in Chapter 7 review various methods for estimating canopy properties from remotely sensed data. They first describe the empirical regression-type algorithms, using either experimental data or radiative transfer simulated data, and different model inversion algorithms, including optimization algorithms, look-up table algorithms, and two stochastic inversion algorithms. They then discuss the nature of ill-posed problem in the inversion of canopy properties, and finally summarize two methods to address the ill-posed issue by using the a priori information or posing additional spatial, temporal and model coupling constraints. Wand and Li in Chapter 8 present a knowledge database and discuss its applications to improving the inversion of land surface information from remotely sensed data. They first discuss the need for a prior knowledge due to the ill-posed problem in remote sensing inversion that is also discussed by Baret and Buis in Chapter 7,

4

S. Liang

then describe the database they have developed and finally demonstrate how it has been used in different inversion studies. They also address the parameter scaling and validation issues. Schaaf et al. in Chapter 9 summarize the current algorithms for estimating land surface broadband albedo and anisotropy from satellite observations. After providing some general background, they present the specific algorithms for three representative sensors: MODIS (multispectral), MISR (multiangle) and Meteosat (geostationary) and also discuss how to integrate historic and current albedo products. In Chapter 10, Jacob et al. comprehensively review various modeling methods of thermal signatures and inversion techniques for estimating land surface skin temperature from thermal-IR remotely sensed data. They start with the definitions of temperatures and emissivities, and analysis of spatial, spectral, angular and temporal signatures in thermal-IR remote sensing. Various modeling techniques and inversion methods are then evaluated. They further examine current research issues, including land surface brightness temperature from atmospheric correction or model simulation, ensemble emissivity and radiometric temperature of mixed surface types from different observation capabilities, aerodynamic temperature, and directional effects of temperature and emissivity. The potential applications of thermal-IR remote sensing are also discussed. Vesteinsson et al. in Chapter 11 present a new method to integrate remote sensing data of different spatial resolutions. After reviewing various image fusion methods in four major types (Frequency, color transformation, statistical, and hierarchical), they present a new data fusion framework. They model the fusion issue as a minimization problem where the objective function is the energy of a Gibbs distribution consisting of three energy terms: spectral consistency constraint, smoothness constraint, and imaging physics constraint. Minimization of the objective function is sought using stochastic optimization. This framework is finally implemented using two IKONOS satellite images by integrating the panchromatic band with multispectral bands. Data assimilation methods for estimating land surface variables are presented by Liang and Qin in Chapter 12. After introducing the basic principles of data assimilation, they identify a series of critical issues, such as data and products to be assimilated, parameters to be estimated, assimilation algorithms to be used, error matrices to be determined, and imperfect numerical models. The latest applications of data assimilation methods to various fields (e.g., hydrology, carbon cycle, agricultural productivity) are also reviewed. Lam in Chapter 13 introduces the use of textural/spatial measures to automated land cover classification and change detection. She first identifies the major criteria for evaluating textural measures and then shows different examples in using them to improve classification accuracy. She also summarizes the existing methods into a framework, and then argues that the textural approach has potential for rapid change detection. She finally demonstrates the use of textural measures solely (not including spectral information) for change detection in New Orleans before and after Hurricane Katrina.

1 Recent Advances in Land Remote Sensing: An Overview

5

In Chapter 14, Sun and Liang review the existing methods for mapping plant functional types (PFTs) at regional to global scale from remote sensing data. Traditionally, land surface models represent vegetation as discrete biomes that are not natural vegetation units but are products of classification. Since most land models are expanding beyond their traditional biogeophysical roots to include biogeochemistry, especially photosynthesis and the carbon cycle, the land modeling community has started using PFTs to represent land surface. There is an urgent need for mapping PFT using remote sensing. The authors finally present a multisource evidential reasoning data fusion framework using high-level land products for improved mapping of PFTs from satellite observations. These chapters in Part II of the book cover a variety of topics related to land surface radiation modeling and inversion. I must emphasize the importance of validation here. The new modeling techniques, new inversion methods, and newly generated products require thorough validation. Though a specific chapter on validation is not presented here, a recently published special issue on this topic (Morisette et al., 2006) provides many relevant details. Part III of the book discusses remote sensing applications corresponding to the last panel chaired by Drs. John Townshend and Jiyuan Liu, and with contributions by Drs. Zhongxin Chen, David Goodenough, and Diane Wickland. Remote sensing science driven by applications will have more widespread use and benefit larger user communities. Significant disconnections between remote sensing development and applications persist. Some products developed by remote sensing scientists have not been widely utilized. Many variables required by surface process models and decision support systems have not been generated. Product accuracy and application requirements may not always be consistent. Even the same variables may be defined differently in remote sensing algorithms and application models. This panel provided a useful dialogue between remote sensing scientists and application experts. This part of the book includes some of the discussions and also presents two more chapters on this topic. Chen et al. in Chapter 15 review the remote sensing applications pertinent to agricultural monitoring and management. They discuss progress in four subjects: crop identification and mapping, crop yield estimation and prediction, crop phenology monitoring, and soil moisture estimation from optical to microwave remote sensing. They demonstrate the need for multisensor data fusion, use of multiple signatures, and integration of data and numerical models. Zhao and Running in Chapter 16 review the historical development and the recent advances in the application of satellite remote sensing data for estimating terrestrial productivity and monitoring carbon cycle-related ecosystem dynamics and changes. They first outline the development of using vegetation index for estimating land surface biophysical variables, then describe the MODIS GPP and NPP products and the findings from the vegetation-related products. They further discuss the application of long-term satellite data to the study of terrestrial ecosystems, including phenology monitoring, changes in regional carbon storage resulting from land use change, carbon flux changes induced by climate change, disturbance detection, and validation of ecosystem models. They also propose a scheme for an integrated study of carbon dynamics.

6

S. Liang

In Chapter 17, Dickinson discusses applications of terrestrial remote sensing to climate modeling. Beginning with an introduction to the formulation of the climate models, he then describes land surface models, their components, and energy and water balance requirements. He reviews the contributions from many land remote sensing products (e.g., LAI, albedo and temperature) in terms of roles of solar radiation in climate model terrestrial system, and summarizes various assessments of the climate models when using these products. He further considers how terrestrial remote sensing can better support climate models and eventually be a component of climate prediction through data assimilation. Townshend et al. in Chapter 18 identify and analyze factors either facilitating or hampering the increased use of satellite data, products and services by compiling 25 case studies. These factors include technical, education and capacity building, financial and policy. They discover that successful adoption of remote sensing products always involves balanced cooperation between space agencies and users. They derive a set of principles that will lead to enhanced usage of space-based data and products and summarize the recommendations to the space agencies. Liang et al. in Chapter 19 summarize the key questions and issues discussed by three review panels and identify some emerging issues in land remote sensing, including sensor networks, modeling complex landscapes, machine learning techniques for inversion, and spatial scaling. Finally, note that two special issues in Photogrammetric Engineering and Remote Sensing (October 2007), and Journal of Remote Sensing (September 2007) from the 9th ISPMSRS add to and complement the review papers in this book. Readers are referred to these special issues for details.

References Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, New York, 534pp Liang S, Strahler A (guest eds) (2000) Land surface bi-directional reflectance distribution function (BRDF): recent advances and future prospects. Special issue of Remote Sens. Rev. 18:83–511 Morisette J, Baret F, Liang S (guest eds) (July 2006) Global land product validation. Special issue of IEEE Trans. Geosci. Remote Sens. 44(7):1695–1937 Myneni RB, Ross J (eds) (1991) Photon-vegetation Interactions: Applications in Optical Remote Sensing and Plant Physiology. Springer, New York

Part I

Remote Sensing Systems

Chapter 2

Passive Microwave Remote Sensing for Land Applications Thomas J. Jackson

Abstract Land applications, in particular soil moisture retrieval, have been hampered by the lack of low frequency passive microwave observations and the coarse spatial resolution of existing sensors. The next decade could see several improved operational and exploratory missions using new technologies as well as innovative disaggregation and data fusion approaches that could lead the way to an order of magnitude improvement in spatial resolution. Keywords: Passive microwave · soil moisture · land surface

2.1 Introduction Passive microwave remote sensing has made major contributions in atmospheric and oceanic sciences. These applications have exploited higher frequencies and used low frequencies to establish background conditions. Land applications have been hampered by the availability of low frequency observations (<6 GHz) and coarse spatial resolution. Conventional technologies and approaches to retrievals have limited spatial resolution to the 50 + km range, which has in turn limited the potential usage to only very large-scale studies. Three factors will affect land applications of passive microwave remote sensing over the next decade: the operational low frequency instruments available, exploratory missions using new technologies, and innovative approaches to disaggregating the coarser resolution passive microwave observations with active microwave measurements. The current satellite sensors, criteria for mission design, and future missions are presented in this chapter. Because soil moisture is one of the most important and challenging problems the discussion will focus on this topic. Thomas J. Jackson USDA ARS Hydrology and Remote Sensing Lab, Beltsville, MD, USA [email protected] S. Liang (ed.), Advances in Land Remote Sensing, 9–18. c Springer Science + Business Media B.V., 2008

9

10

T.J. Jackson

2.2 Satellite Passive Microwave Sensor Systems Low frequency passive microwave remote sensing provides information on the dielectric and temperature properties of the Earth’s atmosphere and surface as well as some characterization of its geometric features. The dielectric properties are dependent upon the water content of the target. As a result, passive microwave remote sensing has been useful in studies of the atmosphere, oceans, snow, ice, and land. The breadth of parameters and variables that can be retrieved is illustrated in Table 2.1 for the Advanced Microwave Scanning Radiometer (AMSR-E) instrument on the National Space and Aeronautics Administration (NASA) Aqua satellite. Land applications of passive microwave remote sensing have included classification, temperature, vegetation characteristics, and soil moisture. Current operational sensors such as the Special Sensor Microwave/Imager (SSM/I) have been used in land cover type classification (Neale et al., 1990), land surface temperatures estimation and in deriving a general surface wetness parameter (Basist et al., 1998). System design for land variable retrieval includes at least six considerations: • • • • • •

Frequency Polarization Viewing geometry Spatial resolution Temporal coverage Signal to Noise

2.2.1 Frequency Selection Lower frequencies provide greater sensitivity to changes in soil moisture through a greater range of vegetation cover conditions. For soil moisture retrieval these are very important considerations (Jackson and Schmugge, 1991). Figure 2.1 illustrates this point using a sensitivity parameter (degrees K/% soil moisture) and identifies several key satellite instruments. For bare soil there is little change in sensitivity as a

Table 2.1 Earth variables derived from AMSR-E Category

Variable

Atmosphere

Total integrated water vapor over the ocean Rainfall over ocean Sea surface temperature Ocean surface wind speed Sea ice concentration Snow depth over sea ice Sea ice temperature Snow-cover water equivalent over land Surface soil wetness

Oceans Sea ice

Snow Land

2 Passive Microwave Remote Sensing for Land Applications High

SMOS

AMSR-E Meteorological Satellites Bare

Sensitivity

Fig. 2.1 Sensitivity of brightness temperature to soil moisture as a function of microwave frequency

11

Vegetated

Low 1

2

3 5 10 20 30 Frequency (GHz)

50

function of frequency (although) there is a decrease in sensing depth with increasing frequency. In the presence of vegetation the impact of frequency is much more significant. Surface roughness can also decrease sensitivity. Another aspect of Fig. 2.1 to note is that research and applications were limited to the lowest frequencies of the meteorological satellites for many years. When AMSR-E and other recent satellite instruments were launched a significant improvement in soil moisture retrieval was anticipated. Another key issue in frequency selection is radio frequency interference. This has become a significant problem with AMSR-E, primarily at 6.9 GHz in the USA. Contamination of the signal has rendered the data unusable and as a result applications have had to adapt to using a less desirable higher frequency (Li et al., 2004; Njoku et al., 2005). As shown in Fig. 2.1, the shift in frequency from 6.9 GHz to 10.7 GHz for soil moisture retrieval results in a reduction in sensitivity and loss of information. It should be noted that the 6.9 GHz band is not a protected radio frequency, however, 10.7 GHz is protected and still is contaminated in some regions of the world.

2.2.2 Polarization It is well known that horizontal polarization is much more sensitive to changes in the soil dielectric constant than vertical polarization measurements. Simple computations using radiative transfer equations can readily verify this fact (Ulaby et al., 1982). However, vertical polarization can be very useful in normalizing the horizontal measurements. Information on the polarization difference has been used to characterize the vegetation in several studies and algorithms (Becker and Choudhury, 1988; Paloscia et al., 2001; Owe et al., 2001; Njoku et al., 2003). With the recent launch of the WindSat instrument (Gaiser et al., 2004) with a fully polarimetric microwave radiometer, investigators are beginning to examine what new information about the land may be contained in the additional channels the sensor provides (Narvekar et al., 2007).

12

T.J. Jackson

2.2.3 Viewing Geometry Nearly all satellite microwave radiometers used for land studies have employed conical scanning. Conical scanning provides constant incidence angle observations, which simplifies the retrieval process. Higher incidence angles result in wider swaths and increased temporal coverage. So, there are good reasons for choosing this design. However, the highest sensitivity to soil moisture will be at low incidence angles. Radiative transfer relationships include the angle in computation. In addition, the path length through the vegetation canopy will increase with angle. At low incidence angles the path attenuation will be smaller. Some algorithm designs actually exploit these changes with incidence angle in their retrievals (Wigneron et al., 2000). In this approach, a correction for vegetation is made in performing the soil moisture retrieval by using several incidence angle observations.

2.2.4 Spatial Resolution Spatial resolution has been the challenge for passive microwave remote sensing of land. Unlike atmospheres and oceans, the land exhibits heterogeneity at several scales that may be significant. With low frequencies it is extremely difficult to design an antenna system necessary for a high resolution observing system. This topic will be covered in depth in a later section.

2.2.5 Temporal Coverage With conical scanning low-resolution sensors, the temporal coverage has not been a major issue for applications. Wide swaths of current and future sensors result in 1–3 day global coverage.

2.2.6 Signal to Noise To date, using conventional antenna technology, signal to noise has not been an issue for land applications (soil moisture) because the range of response and sensitivity to the geophysical parameters are quite large.

2 Passive Microwave Remote Sensing for Land Applications

13

2.3 Selected Conical Scanning Instruments Three current satellite instruments are of interest in passive microwave sensing of land; SSM/I, AMSR-E, and WindSat. Features of each mission are summarized in Table 2.2. There have been several other important satellites for land remote sensing that include the Scanning Multifrequency Microwave Radiometer (SMMR) (Owe et al., 1992) and the Tropical Rainfall Measurement Mission Microwave Imager (TMI) (Bindlish et al., 2003). Worth noting in Table 2.2 is the fact that the SSM/I instrument now has an over 20-year period of record. Interpreting data from the SSM/I to extract surface information requires accounting for atmospheric effects on the measurement. When one considers the atmospheric correction, the significance of vegetation attenuation, and the shallow contributing depth of soil for these high frequencies, it becomes apparent that the data are of limited value for estimating soil water As a result, SSM/I data has been applied to only a limited set of land applications including a soil wetness product (Basist et al., 1998). Some successful attempts have been made at soil moisture retrieval under selected conditions (Jackson, 1997). Aqua was launched in May 2002 and ADEOS-II was also launched in the same year. Each included an AMSR instrument. ADEOS-II was lost after a few months of operation but Aqua is still providing data. As shown in Table 2.1 AMSR-E was to provide significantly lower frequency measurements than the SSM/I with the same level of spatial resolution. AMSR-E holds great promise for estimating soil water content in sparsely vegetated regions and is the best possibility in the near term for mapping soil water. As opposed to previous passive microwave satellite missions, both NASA (Njoku et al., 2003) and the Japanese Aerospace Exploration Agency (JAXA) (Njoku et al., 2000) have included soil moisture as a mission product from AMSR-E. As noted previously, it is not expected that AMSR-E can provide a globally reliable soil moisture product. However, it will work in some regions and the efforts at operational implementation and validation of products are providing valuable lessons for future missions. Table 2.2 Selected satellite conical scanning microwave radiometers Frequency (GHz)

85/91 37 22/23 19 10.8 6.9/6.8 1.4 37 IFOV (km) Observing time Period of record

Satellite Instrument SSM/I

AMSR-E

WindSat

V, H V, H V V, H – – – 27 × 38 Various 1985–

V, H V, H V, H V, H V, H V, H – 8 × 14 1:30 p.m. 2002–

– V, H, U, F V, H V, H, U, F V, H, U, F V, H – 8 × 13 6:00 a.m. 2003–

14

T.J. Jackson

The third satellite instrument is WindSat that launched in 2003 and includes a multifrequency passive microwave radiometer system with a C band channel. This is a risk reduction mission for one component of the next generation of operational polar orbiting satellites that the U.S. will be implementing. Regarding operational sensors, the Advanced Microwave Scanning Radiometer (AMSR-E) on Aqua will transition to a component of the long term Japanese Global Climate Observing Mission (GCOM) and WindSat may evolve to the Conical Microwave Imaging System (CMIS) on National Polar Orbiting Earth Satellite System (NPOESS). The lower frequencies and commitments to both data and algorithms should result in better products for land studies despite the rather coarse spatial resolution. Japan provided the AMSR-E instrument to the Aqua platform and has plans to follow this up with an improved version on the Global Climate Observing MissionsWater (GCOM-W) satellites possibly in 2010. It is uncertain what the USA will be doing as a follow on to WindSat and SSM/I. Until recently plans called for a Conical Microwave Imaging System (CMIS) as part of the National Polar Orbiting Earth Satellite System (NPOESS). CMIS would be similar to WindSat, offering multifrequency (including C and X band) as well as fully polarimetric observations. However, issues with instruments and costs are expected to impact the final design and delay the current launch date. Another source of passive microwave measurements will be satellites involved in the Global Precipitation Mission that follows the successful TMI. This mission is scheduled for 2012.

2.4 New Directions and Missions As noted previously, achieving a spatial resolution that can satisfy land application requirements (or even demonstrate an approach that could open up future options) has been the challenge to passive microwave remote sensing. The problem has been achieving high spatial resolution at lower frequencies requires a large antenna. Getting large and heavy antennas into space is obviously difficult and costly. There are at least three approaches to solving this problem that are currently under investigation: • Synthetic aperture radiometry • Large lightweight antennas • Data fusion-disaggregation using higher resolution remote sensing techniques Synthetic aperture radiometry attempts to solve the spatial resolution dilemma by replacing the traditional large filled array with a sparse array. It can achieve what a very large antenna would, but with a much smaller mass and size. Earlier studies established one dimensional synthetic aperture methods (Le Vine et al., 1994) and demonstrated that this approach could be used for soil moisture retrieval (Jackson et al., 1995). One-dimensional methods use a real aperture along track and synthetic aperture methods across track, essentially a set of long sticks.

2 Passive Microwave Remote Sensing for Land Applications

15

The European Space Agency (ESA) will take synthetic aperture radiometry a step further in the Soil Moisture Ocean Salinity (SMOS) mission that uses a two dimensional design (Kerr et al., 2001). SMOS includes an interferometric radiometer operating at L band (1.4 GHz) and uses a technique based on the cross-correlation of observations from all possible combinations of receiver pairs. SMOS is currently scheduled for launch in late 2008. At an altitude of 763 km, the antenna will view a swath of almost 3,000 km providing a 40 km global soil moisture product every 2–3 days. Another key element of SMOS that is tied to its design is the soil moisture retrieval algorithm. Each SMOS 40 km footprint is the average of 1.2 s. As the satellite moves along its orbital path each pixel is observed at many different viewing angles, which permits the observation of brightness temperature as a function of viewing angle. These multiple angles are used with radiative transfer equations to retrieve soil moisture as well as vegetation information (Wigneron et al., 2000). An alternative solution to the size-mass issues of real aperture antennas is the use of lightweight mesh. This approach is compatible with designs that can be packaged into smaller volumes for launch and also have much lower mass. A large mesh antenna was one of the key features of the Hydros satellite mission that was studied by NASA under its Earth System Science Pathfinder program (Entekhabi et al., 2004) but was recently cancelled by NASA. Using a 6 m mesh antenna operating at 1.4 GHz, Hydros would be able to provide a 40 km brightness temperature/soil moisture product. It would employ conical scanning over a wide swath that results in 2–3 day global coverage. Conical scanning facilitates some aspects of retrieval. However, the spinning of the large antenna would require careful engineering. This concept is currently being re-considered for priority implantation. The passive microwave instruments of SMOS and Hydros would have demonstrated two potential paths to achieving higher spatial resolution using only passive sensors. The 40 km products would satisfy the demands of climatology studies. There are many additional applications that could utilize soil moisture if it was available at higher spatial resolution. If successful, each of these technologies could lead to future missions with resolutions of better than 10 km. Another approach to better spatial resolution is disaggregation. The concept of disaggregation to achieve a higher resolution product is based on the assumption that the passive microwave instrument provides a reliable soil moisture product and that alternative remote sensing measurements while at higher spatial resolutions have lower accuracies. This has been explored in the past using visible-near infrared remote sensing. In regions with uniform soil properties and small amounts of vegetation this might work. However at these wavelengths the high-resolution data only represents an extremely thin surface soil layer that easily disconnects from the lower profile. If there is vegetation this approach will not work unless the vegetation exhibits a response related to the soil moisture state. There is some justification for this; however, it is also compounded by lags in time between changes in soil moisture and changes in canopy visible-near infrared characteristics. Using thermal infrared for disaggregation has also been proposed. It too has limitations. However, it is expected that the lag time between soil moisture changes and

16

T.J. Jackson

temperature changes will be shorter than with physiographic features observed with visible-near infrared. Disaggregation was explored by Chauhan et al. (2003) as a solution for NPOESS. It was proposed that a combination of microwave, visible-near infrared, and thermal infrared could be used. It was demonstrated using 25 km SSM/I products and 1 km Advanced Very High Resolution Radiometer (AVHRR) data. This might work sometimes, however, cloud cover would certainly limit it. Another approach to disaggregation was central to the Hydros mission design. In addition to the passive microwave instrument, Hydros would have included a radar that shared the same antenna. This instrument would provide a 3 km soil moisture product with possibly limited accuracy. However, the real role of this sensor was to serve as a disaggregation and data fusion tool that would be combined with the coarser resolution passive product. It is fairly well established that a radar responds to soil moisture in much the same way as a passive microwave sensor for smooth bare soils. However, roughness and vegetation effects can be more difficult to account for using radar, which leads to a higher uncertainty in radar only soil moisture retrieval algorithms. A potential disaggregation scheme is described in Narayan et al. (2006). Not only would the radar serve in disaggregation it would provide the option of using data fusion of passive and active microwave in soil moisture retrieval. The concept of data fusion of passive and active could be explored using another L band mission being implemented by NASA called Aquarius (Koblinsky et al., 2003). Aquarius would provide (∼2010) both active and passive coarse spatial resolution measurements using a push broom approach. It would provide some useful information for algorithm science but the technology does not provide a pathway to high spatial resolution. It was anticipated that with Hydros on its successor a highly reliable 10 km soil moisture product could result from the integration of passive and active microwave remote sensing. A 10 km soil moisture product would open the doors to a much wider range of applications and the demonstration of the approach could lead to even higher resolutions in the future. There are likely to be limits on the differences in scale that can be used when disaggregating. There has to be some overlap in the dominant processes that control variability at the two resolutions. If a passive technology is demonstrated that can lead to future missions with higher spatial resolutions, then radars with higher resolutions can also be consider that could eventually result in products <1 km. Schemes for disaggregation that could be compatible with SMOS, as well as other missions, are described in Merlin et al. (2005) and Pellenq et al. (2003).

2.5 Summary Land applications, in particular soil moisture retrieval, have been hampered by the lack of low frequency passive microwave observations and the coarse spatial resolution of existing sensors. The next decade will see several exploratory missions using

2 Passive Microwave Remote Sensing for Land Applications

17

new technologies, and innovative approaches to integrating passive microwave observations with active microwave measurements that could lead the way to an order of magnitude improvement in spatial resolution.

References Basist A, Grody NC, Peterson TC, Williams CN (1998) Using the Special Sensor Microwave/ Imager to monitor surface temperatures, wetness, and snow cover. J. Appl. Met. 37:888–911 Becker F, Choudhury, BJ (1988) Relative sensitivity of NDVI and microwave polarization difference index (MPDI) for vegetation and desertification monitoring. Remote Sens. Environ. 24:297–311 Bindlish R, Jackson TJ, Wood E, Gao H, Starks P, Bosch D, Lakshmi V (2003) Soil moisture estimates from TRMM Microwave Imager observations over the Southern United States. Remote Sens. Environ. 85:507–515 Chauhan NS, Miller S, Ardanuy P (2003) Spaceborne soil moisture estimation at high resolution: a microwave-optical/IR synergistic approach. Int. J. Remote Sens. 24:4599–4622 Entekhabi D, Njoku E, Houser P, Spencer M, Doiron T, Belair S, Crow W, Jackson T, Kerr Y, Kimball J, Koster R, McDonald K, O’Neill P, Pultz T, Running S, Shi JC, Wood E, van Zyl J (2004) The hydrosphere state (Hydros) mission concept: an Earth System Pathfinder for global mapping of soil moisture and land freeze/thaw. IEEE Trans. Geosci. Remote Sens. 42:2184–2195 Gaiser PW, St. Germain KM, Twarog EM, Poe GA, Purdy W, Richardson D, Grossman W, Jones WL, Spencer D, Golba G, Cleveland J, Choy L, Bevilacqua RM, Chang PS (2004) The WindSat spaceborne polarimetric microwave radiometer: sensor description and early orbit performance. IEEE Trans. Geosci. Remote Sens. 42:2347–2361 Jackson TJ (1997) Soil moisture estimation using special Satellite Microwave/Imager satellite data over a grassland region. Water Resour. Res. 33:1475–1484 Jackson TJ, Schmugge TJ (1991) Vegetation effects on the microwave emission from soils. Remote Sens. Environ. 36:203–212 Jackson TJ, Levine DM, Swift CT, Schmugge TJ, Schiebe FR (1995) Large-area mMapping of soil-moisture using the ESTAR passive microwave radiometer in Washita92. Remote Sens. Environ. 54:27–37 Kerr YH, Waldteufel P, Wigneron JP, Font J, Berger M (2001) Soil moisture retrieval from space: the Soil Moisture Ocean Salinity (SMOS) mission. IEEE Trans. Geosci. Remote Sens. 39:1729–1735 Koblinsky CJ, Hildebrand P, LeVine D, Pellerano F, Chao Y, Wilson W, Yueh S, Lagerloef G (2003) Sea surface salinity from space: science goals and measurement approach. Radio Sci. 38(4):8064, doi:10.1029/2001RS002584 Le Vine DM, Griffis A, Swift CT, Jackson TJ (1994) ESTAR: a synthetic microwave radiometer for remote sensing applications. Proc. IEEE. 82:1787–1801 Li L, Njoku EG, Im E, Chang P, St. Germain K (2004) A preliminary survey of radiofrequency interference over the U.S. in Aqua AMSR-E data. IEEE Trans. Geosci. Remote Sens. 42:380–390 Merlin O, Chehbouni AG, Kerr YH, Njoku EG, Entekhabi D (2005) A combined modeling and multi-spectral/multi-resolution remote sensing approach for disaggregation of surface soil moisture: application to SMOS configuration. IEEE Trans. Geosci. Remote Sens. 43:2036–2050 Narayan U, Lakshmi V, Jackson TJ (2006) High-resolution change estimation of soil moisture using L-band radiometer and radar observations made during the SMEX02 experiments. IEEE Trans. Geosci. Remote Sens. 44:1545–1554

18

T.J. Jackson

Narvekar PS, Jackson TJ, Bindlish R, Li L, Heygster G, Gaiser P (2007) Observations of land surface passive polarimetry with the WindSat Instrument. IEEE Trans. Geosci. Remote Sens. 45:2019–2028 Neale CMU, McFarland MJ, Chang K (1990) Land-surface type classification using microwave brightness temperatures from the Special Sensor Microwave/Imager. IEEE Trans. Geosci. Remote Sens. 28:829–838 Njoku E, Koike T, Jackson T, Paloscia S (2000) Retrieval of soil moisture from AMSR data. In: P. Pampaloni, S. Paloscia (eds), Microwave Radiometry and Remote Sensing of the Earth’s Surface and Atmosphere. VSP, Zeist, The Netherlands, pp 525–533 Njoku EG, Jackson TJ, Lakshmi V, Chan TK, Ngheim SV (2003) Soil moisture retrieval from AMSR-E. IEEE Trans. Geosci. Remote Sens. 41:215–229 Njoku EG, Ashcroft P, Chan T, Li L (2005) Global survey and statistics of radio frequency interference in AMSR-E land observations. IEEE Trans. Geosci. Remote Sens. 43:938–947 Owe M, van de Griend AA, Chang AT (1992) Surface moisture and satellite microwave observations in semiarid Southern Africa. Water Resour Res. 28:829–839 Owe M, de Jeu R, Walker J (2001) A methodology for surface soil moisture and vegetation optical depth retrieval using the microwave polarization difference index. IEEE Trans. Geosci. Remote Sens. 39:1643–1654 Paloscia S, Macelloni G, Santi E, Koike T (2001) A multifrequency algorithm for the retrieval of soil moisture on a large scale using microwave data from SMMR and SSM/I. IEEE Trans. Geosci. Remote Sens. 39:1655–1661 Pellenq J, Kalma JD, Boulet G, Saulnier G, Wooldridge S, Kerr Y, Chehbouni A (2003) A scheme for disaggregation of soil moisture along topography and soil depth. J. Hydrology. 276:112–127 Ulaby F, Moore R, Fung A (1982) Microwave Remote Sensing: Active and Passive, vol. II, Addison-Wesley, Reading, MA Wigneron JP, Waldteufel P, Chanzy A, Calvet JC, Kerr Y (2000) Two-dimensional microwave interferometer retrieval capabilities over land surfaces (SMOS mission). Remote Sens. Environ. 73:270–282

Chapter 3

Active Microwave Remote Sensing Systems and Applications to Snow Monitoring Jiancheng Shi

Abstract This chapter describes research activities made over terrestrial snow covered areas with active microwave instruments, especially with Synthetic Aperture Radar (SAR). Even though an exhaustive review of all studies could not be considered in the framework of the book, the chapter consists of many significant developments which have been made recently in the space-borne and air-borne based active microwave remote sensing of snow properties including techniques of mapping snow, inferring snow wetness and snow water equivalence (SWE). The abilities and existing problems with the current sensors will be discussed.

3.1 Introduction Seasonal snow cover is one of the most important components in predicting global water- and energy-cycle consequences of Earth-system variability and change. Seasonal snow cover and its subsequent melt can dominate local-to-regional climate and hydrology. It is the major source of fresh water over wide areas of the mid-latitudes. Understanding, characterizing, and predicting snow-related processes across spatial scales in coupled atmospheric and hydrologic models requires improved capability for accurately monitoring spatial and temporal distributions of seasonal snow properties on land, especially snow water equivalence (SWE) and snow wetness. In situ measurements provide direct characterization, but at limited spatial and temporal extent and resolution, and frequently must be acquired under challenging or dangerous field conditions. Furthermore, because of the high spatial and temporal variability of snow cover, snow properties derived from in situ measurements often do not provide the reliable spatial and temporal characterization of distributed snow properties across model domains at the required range of resolutions. Jiancheng Shi Institute for Computational Earth System Science, University of California, Santa Barbara, USA [email protected] S. Liang (ed.), Advances in Land Remote Sensing, 19–49. c Springer Science + Business Media B.V., 2008

19

20

J. Shi

Because visible and near-infrared radiation does not penetrate snow, and because the optical properties of ice and water are similar, snow reflectance in this part of the electromagnetic spectrum is not sensitive to snow depth (except for very shallow snow) or free liquid water in the snowpack. For large, flat regions, passive microwave data at 18 and 37 GHz allow sequential mapping of SWE distribution. In dry snow, the profile of the snow grain sizes also influences the microwave remote sensing signal. A more serious problem for measurement of SWE is that the snow depths are often too large for existing passive microwave remote sensing retrieval algorithms to provide accurate estimates. The signal becomes asymptotic and insensitive to snow water equivalence at and above about 0.5 m at 37 GHz (a commonly used frequency). Moreover, the spatial resolution of the current passive microwave satellite sensors is too coarse to provide useful information at regional and drainage basin scales for hydrological investigations and applications. Active microwave sensors (radars), on the other hand, are sensitive to many snow parameters such as snow density, depth, grain size, free liquid water content, and snow-pack structures that are useful for hydrologic applications. Active microwave sensors, especially Synthetic Aperture Radar (SAR), have been used to estimate snow properties, such as snow wetness and snow water equivalency (SWE), and to discriminate snow with other surfaces. They acquire image measurements of the Earth day or night, in all weather, through cloud cover, smoke and haze, and with high resolution especially useful in monitoring geophysical properties in alpine regions. This chapter summarizes the recent progresses in the technical developments for estimation of snow and vegetation properties using active microwave instruments, mainly synthetic aperture radar (SAR).

3.1.1 Current Available Satellite SAR Data There are several Spaceborne SAR instruments available, in past, current and near future, from different space agencies over the world. The major sensor parameters are summarized in Table 3.1. ERS-1 and ERS-2 SAR instrument is one of the sensors on the first and second European Remote Sensing Satellites that were launched in July 1991 and 1995 by the European Space Agency. They are circling the Earth every 100 min. In 3 days, it can cover the entire planet with exact repeat coverage at 35 days. They have the polar orbit at 780 and 785 km for ERS-1 and ERS-2, respectively. Both ERS-1 and ERS-2 SAR transmit and receive the microwave signals with vertical polarization at frequency 5.3 GHz and a fixed incidence angle of 23◦ . Its primarily applications oriented towards ocean and ice monitoring including those of sea state, sea surface winds, sea surface temperature, ocean circulation and sea and ice level. Due to its all-weather and high resolution microwave imaging capability, its applications have been found over land from the monitoring of crops, tropical deforestation to flooding monitoring. The commonly used products of ERS-1/2 SAR in geophysical

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

21

Table 3.1 Characteristics of current available spaceborne SAR Sensor

Frequency in GHz

Polarization

ERS-1/2 ASAR

5.3 5.33

RADARSAT-1

5.3

SIR-C/X-SAR JERS-1 PALSAR

1.25 and 5.3/9.6 1.27 1.27

RADARSAT-2

5.3

TerraSAR-X

9.6

Fixed at 23◦ Varying with mode HH Varying with mode Fully Varying with polarimetric/VV data takes HH Fixed at 35◦ Fully polarimetric, Varying with dual-polarization, mode and HH Fully polarimetric, Varying with dual-polarization, mode and HH VV or dualVarying with polarization mode VV Dual-polarization

Incidence in degree

Pixel resolution in m

Available time frame

3.8–12.5 3.8–150

Since 1991/1995 Since 2002

10–100

Since 1997

6–27

April and October, 1994 1994–1997 Since December 2005

18 10–100

3–100

2007

1–16

2007

applications are the single look complex product (SLC), the precision image (PRI), and the geocoded product that produced from PRI data after Earth ellipsoid and terrain corrections. The more detailed information about ERS-1 and ERS-2 SAR instrument and data products can be found at http://www.earth.esa.int/ers. In March 2002, the European Space Agency launched Envisat, an advanced polar-orbiting Earth observation satellite which provides measurements of the atmosphere, ocean, land, and ice. An Advanced Synthetic Aperture Radar (ASAR), operating at C-band, ASAR ensures continuity with the image mode (SAR) and the wave mode of the ERS-1/2 AMI. It enhanced capabilities in comparison with ERS SAR include coverage (a ScanSAR mode), range of incidence angles, polarization (dual-polarizations), and modes of operation (with nine different modes). The resulting improvements in image and wave mode beam elevation steerage allow the selection of different swaths, providing the swath coverage of over 400 km wide using ScanSAR techniques. In alternating polarization mode, transmit and receive polarization can be selected allowing scenes to be imaged simultaneously in two polarizations. RADARSAT-1, launched in November 1995, is an operational radar satellite system developed by Canadian Space Agency (CSA) to monitor environmental change and the planet’s natural resources. RADARSAT-1 has the sun-synchronous orbit at 798 km altitude with 6 p.m. ascending node and 6 a.m. descending node. It transmits and receives the microwave signals with horizontal polarization (HH) at C-band (5.3 GHz). It has seven beam modes, 35 beam positions and ScanSAR modes for a wide range of imaging options, varying resolutions from 8 to 100 m, swath widths of

22

J. Shi

50–500 km, and incidence angles from 10◦ to 59◦ , depending on the mode selection. The more detailed information about RADARSAT-1 instrument and data products can be found at http://www.space.gc.ca/asc/eng/satellites/radarsat1. ALOS PALSAR is the Phased Array type L-band Synthetic Aperture Radar (PALSAR) on the Advanced Land Observing Satellite (ALOS) that has been launched on January 24, 2006 and is a follow on the Japanese Earth Resources Satellite-1 (JERS-1). It provides higher performance than the JERS-1’s SAR for land observations. PALSAR has the ScanSAR mode with the swath width of 250– 350 km depending on the number of scans in addition to the fine resolution in a conventional mode. This swath is three to five times wider than conventional SAR images. It has also different polarization selections from single, dual, to fully polarimetric observations and improved orbit controlling capability that allows the more frequent repeat-pass interferometer applications. SIR-C/X-SAR is the Shuttle Imaging Radar-C and X-Band Synthetic Aperture Radar. It was a cooperative experimental SAR mission between the National Aeronautics and Space Administration (NASA), the German Space Agency (DARA), and the Italian Space Agency (ASI) (Evans et al., 1997). The SIR-C/X-SAR system was flown onboard NASA’s Space Shuttle on two 10-day missions in April and October 1994. SIR-C provides radar polarimetric digital images simultaneously at two wavelengths, L-band (24 cm), and C-band (5.6 cm). These polarimetric data allows derivation of the complete scattering matrix on a pixel-by-pixel basis and more detailed information about the geometric structure and dielectric property of a target (Van Zyl et al., 1987; Evans et al., 1988). X-band (3 cm) with VV polarization results in a three-frequency capability. The resolution of this system ranged between 10 and 40 m, and it collected image swaths between 15 and 90 km wide. In addition, there have been two new SAR satellites launched in 2007. RADARSAT-2 is the follow on RADARSAT-1 system co-funded by CSA and MacDonald Dettwiler (MDA). It has been launched in 2007 with 7 years mission duration and will support all RADARSAT-1 imaging modes. It has the same orbit, repeat cycle and ground track as RADARSAT-1. Its improvements in comparison with RADARSAT-1 include the higher resolution (up to 3 m), routine left-looking and right-looking mode for more frequent revisit, selective polarizations with fully-polarimetric imaging modes, on-board GPS receivers to improve the real-time position knowledge, and higher downlink power. The more detailed information about RADARSAT-2 instrument and data products can be found at http://www.radarsat2.info/rs2 satellite. TerraSAR-X is a new German radar satellite with 514 km sun-synchronous orbit at 98◦ Inclination and has been launched in June, 2007 with a 5 year lifetime. It carries a high frequency X-band (9.65 GHz) SAR sensor that can be operated in three different basic modes – the ScanSAR, Stripmap and Spotlight mode at a varying geometrical resolution between 1 and 16 m. TerraSAR-X will provide single or dual polarization data. On an experimental basis additionally quad polarization and along track interferometry are possible.

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

23

3.1.2 SAR Measurement Properties Generally, SAR provides three types of measurements that can be used in study geophysical properties depending on the sensor properties: (1) backscattering measurements at the given frequency and polarization such as ERS1/2, ASAR, and RADARSAT-1; (2) the measurements of the polarization properties by a fully polarimetric instrument (SIR-C, PALSAR, RADARSAT-2, and TerraSAR-X with selected mode); and (3) interferometry from the repeat-pass measurements. Backscattering SAR imagery provides high resolution digital images but with only fixed polarization state of the antenna. With only one or a few intensity measurements per pixel, the applications to monitor snow properties have to rely mainly on the radiometric properties such as snow classification and estimations of snow wetness and SWE. Rather than just measuring amplitude, an imaging radar polarimeter measures the amplitude and relative phase for every polarization state. These complex measurements lead to nine independent real elements of the Stokes’ matrix, which describes how the scattering mechanisms in each pixel transform the illuminating electromagnetic wave back to the receiving antenna. The polarization feature derived from the Stokes’ matrix, is the radar cross section as a function of the antenna polarization state and is useful in interpreting the scattering mechanisms within a resolution cell. The first element of the matrix is the total power or span. The nine cross products of the pixel scattering matrix can be obtained by the linear combinations from the span and the other Stokes’ matrix elements. In addition to the polarization signature, many other features describing the scattering mechanisms within a pixel can be derived from the Stokes’ matrix. These include: the pixel intensity synthesized from the polarization signature for fixed antenna polarization states, the polarization phase difference of the scattering matrix elements, the coefficient of variation, the enhancement factor, the scattering mechanism and the degree of polarization. Thus, polarimetric measurements provide much more information per pixel than the single fixed antenna SAR imagery, and have been shown to be effective in classification of terrain. These measurements have great opportunity to minimize the topographic effects on radar images. In addition to single, multi-polarization, and fully polarimetric radar observations, the repeat-pass interferometric radar measurement also provides the useful information in analyses of land surface geophysical properties. Its techniques for topographic mapping of surfaces promise the high-resolution digital elevation models; but they also permits inference of changes in the surface over the orbit repeat cycle from the correlation properties of the radar echoes. Measurements of interferometer correlation describe processes occurring on the time scales of the orbit repeat time and size scales on the order of a radar wavelength, such as vegetation growth, glacier motion, permafrost freezing and thawing, and soil moisture induced effects. Furthermore, SAR images have two distinct special characteristics: image speckle and a more complicated geometric mapping in contrast to images from optical sensors. The SAR image generation involves a coherent processing carried

24

J. Shi

out on the received signal: fading causes on SAR imagery a grainy appearance referred to as speckle (Ulaby et al., 1982, 1986). This is because a SAR resolution cell, that contains many different scatterers, is very large when compared to the wavelength of the illuminating electromagnetic wave. The returned radar echo results from the coherent summation of all the returns from the amplitudes and phases of the single scatterers. Commonly, it requires processing a multi-look image or applying a speckle filter to reduce its effects and to improve the image quality (Ulaby et al., 1982, 1986; Lee et al., 1998). Furthermore, SAR mapping is mainly an integration of backscattered signals having the same Doppler frequency as along-track measurements and the same distance as across-track measurements. The characteristic radar measurements are range – the distance from the sensor to an object point, and time – the position of the sensor along its flight path where the data are collected. They define the two-dimensional SAR image space. The total received power from a resolution cell is proportional to the radar cross section, and the rest of factors in the radar equation are generally assumed constant. The radar cross section is usually a function of polarization, frequency, viewing geometry, and illuminated area, in addition to factors such as the dielectric properties and geometric structure of the targets. During the data processing, the backscattering coefficient σo is usually presented in terms of a normalized radar cross section to eliminate the dependence on illumination area.

σ (3.1) A where A is illumination area. For a flat area, both illumination area and incidence angle of a pixel can be reasonably estimated so that the backscattering coefficient can be obtained. Topographic effects on radar images can be considered as two aspects: (1) the effects on received radar power, which result from a great variation in illuminated area and incidence angle for a pixel resolution. It can be described as the variation of the received power from an inclined surface compared to the received power from a horizontal surface. This variation is a function of the relative orientation of the surfaces with respect to the illuminating source and their position relative to the sensor. (2) The geometric distortions on image coordinates that are mainly dependent upon the distance between an imaged pixel and the sensor. The effects of rugged terrain on both the received power result from the change in illuminating area and the geometric distortion can be corrected for a resolution cell if a digital elevation model (DEM) with the comparable resolution and the sensor location information are available. They are available only with the geocoded data products. σo =

3.2 Snow Mapping with SAR For climatological and hydrological investigations, the area of snow cover is one of the important parameters. For instance, the snow line on glaciers is an important quantity for hydrological applications and mass balance studies. Visible and

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

25

near-infrared sensors have been used extensively to measure these quantities, but are hampered by cloud cover, which can be pervasive in some regions. In particular, snow cover must be measured on a timely basis to be useful for operational hydrology, and the opportunities for obtaining suitable data from these sensors can be infrequent. Microwave remote sensing is a methodology that is less influenced by cloud conditions, depending on frequency. Especially, Synthetic Aperture Radar (SAR) provides high resolution measurements that are comparable to the scales of the topographical variation in mountain areas and more suitable for mapping snow cover than passive microwave instruments. In the microwave part of the electromagnetic spectrum, ice is almost transparent, and the radar penetration depth, depending on the frequency, can reach tens of meters for dry snow. The major scattering source is the snow–ground interface, and it is difficult to discriminate dry snow cover from bare surfaces or short vegetation with backscattering measurements from a single-polarization radar measurements. Earlier investigators found no significant difference in backscattering from dry snow and snow-free surfaces at either C-band (Rott et al., 1993) or X-band (Fily et al., 1995). However, there is a large dielectric contrast between the solid and liquid phases of water at microwave frequencies. As snow starts melting, even a small amount of liquid water reduces the penetration depth of the radar signal, and thereby changes the dominant scattering source from the snow–ground interface to the snow volume and the air–snow interface. The backscattering coefficients decrease substantially at C-band and X-band (Stiles and Ulaby, 1980; Nagler and Rott, 2000; Shi and Dozier, 1995). As the snow continues melting and the surface ages, backscattering increases and surface scattering at the air–snow interface becomes the dominate scattering ¨ source, as Rott et al. (1993) observed in ERS-1 data over the Otztal, Austria, a snow-covered alpine glacier region. Surface scattering also dominates the signals from bare rock, soil, and glaciers, but these surfaces are usually rougher than wet snow and therefore have stronger backscattering signals. However, the backscattering from wet snow cover may have the similar intensities as that from very smooth bare soils, which can result in a difficulty for separation, as Haefner et al. (1993) showed in ERS-1 data from the Swiss Alps. Since the radiometric quality of SAR images in an alpine region is dependent on flight and imaging parameters (e.g., flight altitude, radar elevation angle) and the topography of the imaged area, the representation of the target materials is likely to be inaccurate. This variation in radar backscatter that is unrelated to the surface cover type is particularly evident for high relief surfaces where a large variation of slope and aspect creates a great variation of local incidence angles and illuminated areas. For example, rock surfaces with greater incidence angle could have lower power return than that from snow or glacier covered areas with smaller incidence angle. The effect of topography, therefore, is another major problem we are facing in using radar to map snow covered areas, especially in alpine regions. The study (Van Zyl et al., 1993) indicates that the topographic effects on radiometric properties measured from spaceborne SAR can be explained by variation in imaged pixel area and in local incidence angle. When topographic information of

26

J. Shi

the study area is available, both spaceborne and airborne SAR image data can be radiometrically calibrated. The remaining problem for target discrimination is the effect of local incidence angle on the received power. If topographic information is not available, we need to consider the effects of variations in both local incidence angle and imaged pixel area for satellite SAR measurements. Therefore, in attempting to use SAR data to map seasonal snow cover over remote and inaccessible areas, we are faced with two major problems: • Compensation must be made for the effects of rugged terrain. • Snow must be distinguished from other surface covers. The currently available techniques for mapping snow cover with SAR imagery can be summarized mainly into three catalogues based on the usage of the SAR measurement properties.

3.2.1 Multi-temporal Single-polarization Techniques Previously, a single-polarization SAR imager provides high-resolution images but with only one available measurement per pixel at a fixed polarization state. With only one intensity measurement per pixel, we have to rely mainly on radiometric properties to distinguish snow covered area from other targets, such as bare ground or vegetation. A practical technique to mapping wet snow cover, that has been developed recently, is the multi-temporal methods. Nagler and Rott (2000) used four ERS-1 repeat-pass images, two from ascending and two from descending orbits to reduce the effect of layover. They compared backscattering ratios of an image with wet snow and a reference image (either snow-free or dry snow) to generate a map of wet snow cover, based on change detection. They found that wet snow cover can be identified by using the time series measurements. The radar observations have shown that the backscattering from wet snow cover could be reduced 3–4 dB at C-band in comparison with eight dry snow cover before melting or bare surface after melting. In addition, they also found that the requirement for accurate local incidence angle was relaxed when they used multi-temporal ERS-1 data. (Rott and Nagler, 1993) since the classification algorithm used the ratio measurements of wet snow cover image to the reference image in which the part of terrain effects – pixel illumination area has been cancelled out if two temporal images are co-registrated accurately. For wet snow mapping in forested regions, Luojus et al. (2006) demonstrated a two step technique to map wet snow cover fraction using multi-temporal ERS-2 measurements for boreal forest region in Finland during the snow melting season. The first step is the forest canopy compensation. This is done by nonlinearly fitting ERS-2 measurements with a semi-empirical forest backscattering model (Pulliainen et al., 2003) with the forest stem volume information to estimate the backscattering signals at ground surface and the volume backscattering signals, the two-way transmittivity of the forest canopy. The second step is the employment of linear interpolation algorithm that uses the reference images and the estimated

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

27

backscattering signals at ground surface to calculate the fraction of wet snow covered area – SCA. o −σo σsurf ground,ref (3.2) SCA = 100% · o o σsnow,ref − σground,ref where σosurf is the estimated backscattering coefficient at the ground surface, σoground,ref is the reference signal from the snow-free ground, and σosnow,ref is the reference signal from the wet snow ground. One reference image describes the signal at the fully wet snow cover situation and the other describes the snow-free at the end of the snow melting season. In addition to two reference images, this method requires the knowledge of the forest stem volume distribution of the study area and suitable to sub image scale applications (many pixels averaged together). This is because the uncertainty produced by radar speckle and the use of forest compensation in which the requested forest stem information is more reliable at coarse resolution than that at ERS-2’s pixel resolution.

3.2.2 Multi-frequency and Multi-polarization Techniques Shi and Dozier (1997) evaluated the characteristics of the backscattering, polarization, and frequency ratios of the targets of the study site near Mammoth Mountain in the Sierra Nevada, U.S using multi-frequency and polarization SIR-C/X-SAR’s measurements. They developed two type of supervised classifiers based on classification tree technique. The first type of the classifier was developed by using intensity measurements, polarization properties, and frequency ratios. It can map dry snow and discriminate dry from wet snow, but it requires topographic information for radiometric terrain correction and to reduce effects of local incidence angle. It is about 79% as accurate as a TM binary classification, but it suffers the same shortcoming – it underestimates total snow cover in regions of mixed pixels, especially forested regions. Its performance on the two data-takes where the snow was dry showed that only a few pixels were misclassified as wet snow. The second type of classifier was developed based on polarization properties and backscattering ratios between different frequencies. Since these measurements can be obtained correctly without radiometric terrain calibration, the classifier does not require topographic information and can be used to map wet snow. Similarly, Shi et al. (1994) also developed a method with multi-polarization C-band airborne SAR to map wet snow and glacier ice without a DEM, using only measurements of the polarization properties. Its accuracy is 77% when compared with TM binary classification, but both underestimate total snow cover.

3.2.3 Repeat-pass Interferometric Technique Except the technique of the first type of classifier as in the multi-frequency and polarization techniques, all above methods are restricted to mapping wet snow-cover

28

J. Shi

since it is difficult to discriminate dry snow cover with bare ground and short vegetation. In the study of using SIR-C/X-SAR data to map snow cover (Shi and Dozier, 1997), it was found that wet snow cover had very similar backscattering intensity and polarization characteristics to smooth bare surface at C-band and X-band. For instance, the backscattering from wet snow-cover is very similar to smooth dry soil, alluvial surfaces, and relatively rough water surfaces. At a large drainage basin or regional scale, where many different targets are within a scene, those techniques might not be reliable. For similar reasons, change detection measurement might be also unreliable since the similar change in backscattering could be caused by different natural environment changes. In order to develop a large-scale snow mapping technique other measurement are required to discriminate between snow and other targets. Interferometric radar techniques for topographic mapping of surfaces promise the high-resolution digital elevation models; but they also permits inference of changes in the surface over the orbit repeat cycle from the correlation properties of the radar echoes. Measurements of interferometer correlation describe processes occurring on the time scales of the orbit repeat time and size scales on the order of a radar wavelength, such as vegetation growth, glacier motion, permafrost freezing and thawing, and soil moisture induced effects. The coherence measurement between two repeatpasses, therefore, provides a useful measurement in addition to backscattering intensities in each scene and their changes between two passes, and makes it possible to develop an algorithm for mapping both dry and wet snow covers over large area. Strozzi et al. (1999) demonstrated that coherence measurements could provide the separation between wet snow cover and bare ground in the cases where the backscatter discrimination failed from analyses of ERS tandem data in Switzerland. The low coherence observed over wet snow cover is mainly caused by the rapid change in scattering properties and geometry as the result of wet snow metamorphism due to the movement of free liquid water content, ice grain growth, displacements of adjacent scatterers, and formation of density heterogeneities (layering, ice-lenses, etc.), which all result in a significant decorrelation. On the other hand, the high coherence is regularly observed over no-forested snow free areas. For forested areas, it can be easily separated with wet snow cover due to their huge difference in backscattering intensity even if its coherence is generally low. They provide the physical bases for separating wet snow with other surfaces. For using C-band measurements, however, it requires the short temporal scale (a few days) between the two repeat-pass measurements in order to avoid the significant temporal decorrelation in other surface targets such as bare or short vegetation surfaces. Shi et al. (1997) evaluated the L-band coherence measurements between two repeat-pass SIR-C image data from its first mission in April (with snow) and second mission in October (without snow), 1994. This measurement indicates that the ground is completely undisturbed between viewings the signals will be highly correlated. Otherwise, the decorrelation will occur. They evaluated the coherence measurements of L-band VV polarization for five land cover types snow, lake, bare surface, short vegetation, and forest. It was found that lake, snow and some forest areas had very low coherence between two data-takes. For the lake, the decorrelation between two data-takes is mainly due to the low S/N ratio and the changes of the

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

29

lake surface roughness characteristics, because of change in wind conditions (speed and direction). For the dry snow case, the dominant scattering at L-band is from the snow-ground interface. In addition to the change of the dielectric contrast from the air-ground to snow-ground, existing dry snow cover will result in large decorrelation. This is due to change in local incidence angle when radar signal passes through snow layer will cause a spatial baseline decorrelation. The radar echoes will be also expected to be close to complete decorrelation when measuring correlation between wet snow-cover and bare ground image passes. This is because the radar signal in snow-covered pass can only penetrate a few centimeters so that the radar senses two different targets. The coherence measurements from forest can have large dynamic range with very low values – similar to those from snow and lake, especially from dense forest. On the other hand, the coherence measurements from the bare surface are significantly higher than those from lake. For the bare surface, a change of soil moisture will result in a decorrelation. However, the amount of decorrelation is expected to be smaller because radar senses a same target with a same scattering mechanism (only change in magnitude). The short vegetation (mainly sagebrush and grass in this study area) has very similar coherence measurement to the bare surface mainly due to the dominant scattering source is from the ground surface at L-band. Therefore, the coherence measurements between one snow covered scene and one without snow provide a very good separation between snow cover and bare surface as well as short vegetation. These two targets are most difficult to discriminate with snow cover. Thus, the correlation measurement provides a significant information to map snow covered area. A pixel-based decision tree classifier (DTC) was developed based on above characteristics. While the coherence measurements to discriminate forest, open water, and snow with short vegetation and bare ground, snow can be separated with forest and open water by using backscattering measurements. In order to verify the classification results, a cloud-free Landsat Thematic Mapper (TM) scene on April 14, 1994 (the SIR-C/X-SAR data-take on April 13, 1994) was acquired. A snow map was generated by the TM data and then projected (or co-registered) to the slant range presentation of the SAR images by using the Space Shuttle ephemeris data and ground control points. In the classification of the TM scene, there are only four target categories: lake, snow, forest, and range land that includes both bare ground and short vegetation. Figure 3.1 shows the SAR classification map (right) and the

Fig. 3.1 Comparison of SAR (right) and TM (left) derived classification maps. Black – Forest and lake, gray – bare surface and short vegetation, white – snow

30

J. Shi

TM classification map (left). The comparison estimation of this two results indicated that at 86% accuracy can be obtained for snow cover area under consideration of the TM classification map as the ground truth.

3.3 Estimates of Snow Wetness from C-Band SAR 3.3.1 Relationships Between Snow Wetness and C-Band SAR Measurements Monitoring spatial and temporal changes of liquid water content in snow is important for hydrological modeling, because the presence of liquid water shows that a particular area of the basin can contribute immediately to runoff. When snow starts melting, there is a great change in scattering mechanism in comparison with the dry snow cover. The radar can not “see” through snow-pack since the penetration depth is about a radar wavelength (a few centimeters) for typical wet snow cover at C-band. Radar backscattering measurements from wet snow are affected by two sets of parameters: (1) sensor parameters, which include the frequency, polarization, and viewing geometry, and (2) snow parameters, which include snow density, liquid water content, particle sizes and shapes of ice and water inclusions, type of the correlation function, and surface roughness. It has been understood that the backscattering from wet snow cover is mainly controlled by snow volume backscattering and the surface backscattering at air–snow interface. • The volume scattering is inversely correlated to snow wetness. The liquid water content mainly causes an increase of snow permittivity because of the high dielectric contrast between ice and water. This results in a significant (1) decrease in transmission at the air–snow interface and (2) high dielectric loss, which greatly increases the absorption coefficient. • The surface scattering is proportional to snow wetness. As the liquid water content increases, the reflectivity at air–snow interface increases greatly so does the surface backscattering. The effects of snow wetness on two major scattering signals are in opposite directions and the actual relationships between radar measurements and snow wetness depend on which scattering component is dominant scattering source. If the volume scattering is dominant scattering source, the radar measurements will show a negative correlation to snow wetness. However, the radar measurements will have a positive correlation to snow wetness if the surface scattering is dominant scattering source. The relationship between backscattering and snow wetness, however, can be either positive or negative, depending on snow characteristics and surface roughness and on incidence angle. Generally, as snow wetness increases the backscattering decreases rapidly for low wetness (<3%) conditions because volume scattering albedo decreases. When wetness is low, the dielectric contrast between air and snow

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

31

is small and volume scattering dominates, so backscattering is not sensitive to surface roughness. As snow wetness further increases, backscattering becomes sensitive to surface roughness. This is because the surface scattering component becomes dominating, resulted from rapidly increasing surface scattering component and decreasing volume scattering component. This complexity of the relationship between the backscattering and snow wetness makes it unrealistic to develop an empirical relation between the radar signal and field measurements.

3.3.2 Snow Wetness Algorithm Using C-Band Polarimetric SAR Measurements An algorithm has been developed for snow wetness retrieval using C-band polarimetric SAR imagery (Shi and Dozier, 1995). This algorithm is developed using a database, which covers the most possible wet snow physical properties including the wide ranges of snow wetness, density, particle size and surface roughness, simulated by the first-order scattering model with both surface and volume scattering components. The major developments in this algorithm were: (1) a simplified surface backscattering model, that describes the relationships between the different polarization measurements for the conditions of most seasonal wet snow covers, to minimize the surface roughness effects with multi-polarization measurements; and (2) the property of the volume scattering ratio in co-polarizations which is only a function of snow permittivity and incidence angle to minimize the volume scattering albedo effects on estimation of snow wetness. The more details on how the algorithm was developed can be found in (Shi and Dozier, 1995). The final inverse model for estimating snow wetness by using three C-band measurements σtvv , σthh , and σtvvhh is given as ∗ ] × (avhx DRS − DT S ) + bvx M2 ] M1 [avx Re[αvv αhh bvx M2 ∗ ]+ − DTV |αvv |2 = M2 avx Re[αvv αhh avhx DRS − DT S

(3.3)

with M1 = σtvvhh − DTV (θi , εs ) × σtvv , DT S =

M2 = σtvv + σthh − DT S (θi , εs ) × σtvvhh

DTV (θi , εs ) + DT H (θi , εs ) |αvv (θi , εs )|2 + |αhh (θi , εs )|2 , DRS = ∗ (θ , ε )] DTV (θi , εs ) × DT H (θi , εs ) Re[αvv (θi , εs )αhh i s

where σtvvhh = Re[Stvv Sthh∗ ] is the real part of the cross product of VV and HH complex scattering elements, αvv and αhh are the polarization amplitudes (Ulaby et al., 1982). avx (θi ), avhx (θi ), and bvhx (θi ) are coefficients derived from statistical analyses and depend only on incidence angle. They are given in (Shi and Dozier, 1995). DTV (θi , εs ) and DT H (θi , εs ) denote the volume backscattering ratios for Re[VVHH∗ ] to VV and Re[VVHH∗ ] to HH which are only depending on the local

32

J. Shi

incidence angle and the dielectric constant of a wet snowpack. This concept is from the first-order volume backscattering model for an inhomogeneous dielectric half space medium: 3 σvpp = ω T2pp · exp −2s2 (k1 cos(θi ) − k2 cos(θi ))2 4

(3.4)

Tvv and Thh are power transmission coefficients for a plane interface for vertical and horizontal polarization. s is the standard deviation of the random surface height. The loss factor exp[−2s2 (k1 cos(θi ) − k2 cos(θi ))2 ] is the rough surface effect on the power transmission coefficient (Ulaby et al., 1986). ω is the volume scattering albedo, which depends on snow density, wetness, particle size, size variation and shape. Under the assumption of spherical grains or randomly orientated particles, the volume scattering albedo is independent of the polarization. Therefore, the ratio for the first-order volume backscattering signals of VV and HH polarizations can be represented as a function of the local incidence angle and dielectric constant: DTH (θi , εs ) =

σvvvhh Re [Tvvhh (θi , εs )]2 = , 2 (θ , ε ) hh σv Thh i s

DTV (θi , εs ) =

σvvvhh Re [Tvvhh (θi , εs )]2 = vvh Tvv2 (θi , εs ) σv

The algorithm derived above requires no information about the volume scattering albedo or the surface roughness parameter. With known local incidence angle, it involves only the calculation of snowpack permittivity, which can be directly related to snow wetness. This algorithm is applicable to the situations of incidence angle from 25◦ to 70◦ , and the snow surface roughness – rms height <0.7 cm and correlation length <25 cm. Figure 3.2 shows the comparison between the field snow wetness measurements (x-axis) and the derived snow wetness from C-band polarimetric SAR data of SIR-C (black squares) and AIRSAR (white triangles) at the study site around Mammoth Mountain, California during the first SIR-C/X-SAR mission in April, 1994. Those included eight snow pits and two transects. The SAR-inferred snow wetness values were obtained from an average value of 3 × 3 windows around the snow pit locations. Snow wetness measurements from transects were also averaged to 100 m scales and the inferred snow wetness from SAR was determined from a mean value of 4 × 2 windows along the transect measurements. Most SAR-derived snow wetness agreed well with estimates of snow wetness. The standard deviation of absolute error was about 1.3% by volume which gives 2.5% error bars at 95% confidence interval for absolute error. The algorithm performed well on both local and regional scales and provided a quantitative estimate of spatial distribution of snow wetness at the top snow layer. As we noticed that the above algorithm requires the C-band polarimetric SAR measurements and can not be directly applied to ASAR image data since it does not provide the measurement of Re[Svv Shh ∗ ]. In addition, the pair of co-polarization measurements VV and HH have little difference at small incidence angle, which results in VV and HH measurements that are almost identical regardless of snow wetness and surface roughness conditions. Therefore, the measurements of VV and

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

33

10

Estimated

8 6 4 Sirc Airsar

2 0

0

2

4

6

8

10

Measured snow wetness in % Fig. 3.2 Comparison between the field measurements and the derived snow wetness using C-band SIR-C’s and NASA/JPL AIRSAR’s polarimetric SAR data that are represented as black squares and white triangles in the plot

HH only provide the opportunity for estimation of snow wetness where the pixels with the moderate to large incidence. Further studies are needed to modify the above algorithm for using only two (ASAR) polarization measurements.

3.4 Estimates of Snow Water Equivalence from SAR Snow water equivalence (SWE), the product of snow density and depth, is the most important parameter for hydrological study because it represents the amount of water potentially available for runoff. Measurement of the amount of water stored in the snowpack and forecasting the rate of melt are thus essential for management of water supply and flood control systems. Because of rough, irregular topography, these attributes exhibit large spatial variability over alpine drainage basins, making it impractical to gather enough in situ measurements. Snow density and depth are generally not highly correlated, so they must be considered as independent variables in field surveys of the spatial distribution of snow and in the study of snow’s microwave backscattering properties. In studies of active microwave remote sensing from SAR on retrieval SWE, Bernier et al. (1998) related the multi-temporal C-band backscattering measurements over frozen agriculture fields to thermal resistance in the snowpack. The relationship between thermal resistance and SWE was found and used to estimate SWE. However, it is commonly known that dry snow is almost transparent at C-band. The snowpack only contributes a small amount of signal so that the sensitivity of C-band measurements to SWE is poor. The more studies may be needed to demonstrate the physics of this technique and its capability to be applied to other areas. This chapter

34

J. Shi

mainly summarizes the other two types of techniques that employ different SAR measurement properties and with strong physical principles: 1. SAR backscattering technique with Multi-frequency (L, C, and X bands) and dual-polarization (VV and HH) measurements (Shi and Dozier, 2000a, b). 2. Repeat-pass SAR interferometric technique: The differential phase shift measurements from repeat-pass SAR interferometry can be directly related to SWE (Rott et al., 2004) or its changes (Guneriussen et al., 2001).

3.4.1 Estimation of SWE with Multi-frequency and Polarization SAR Backscattering Measurements During the past years, theoretical modeling of snowpack microwave backscatter has achieved advanced significantly with respect to understanding snowpack extinction properties and the interactions of the microwave signal with the snowpack volume and surfaces. The dense media theory considers a snowpack as consisting of random, discrete, spherical particles of a single size and permittivity (Tsang et al., 1985), accounting for coherent scattering (near-field effect) within the snowpack, and has been extended to consider multi-size and multi-permittivity randomly distributed spherical particle systems and systems where the particles are not randomly distributed but may have a tendency to form clusters and bonds (Tsang, 1992; Ding et al., 1994). This approach is appropriate for application to the snowpack microstructure because snow metamorphic processes can give rise to clusters or aggregations of ice grains. There has also been significant improvement in the modeling of surface scattering characteristics. The Integral Equation Model (IEM) (Fung, 1994) and the more recent Advanced Integral Equation Model (AIEM) allow characterization of a wider range of the surface roughness conditions than past models. Validations of AIEM (Wu et al., 2001; Li et al., 2002; Chen et al., 2003) have demonstrated significant improvement in modeling surface scattering and emission for microwave remote sensing of land surfaces. These efforts have established a fundamentally-improved understanding of the effects of snow physical parameters and underlying surface dielectric and roughness properties on the microwave measurements of snow-covered terrain, making it possible to characterize the microwave backscatter behaviors more accurately. Radar backscattering coefficient measurements at a given incidence angle θi over seasonal snow covered terrain can, generally, be expressed as a four component model: t a v gv g σ pq ( f ) = σ pq ( f ) + σ pq ( f ) + σ pq ( f ) + Tp · Tq · L p · Lq · σ pq (f)

(3.5)

The total backscattering signals consist of the surface backscattering from the interfaces at the air–snow interface and at the snow–ground interface, direct volume backscattering from snow-pack, the interaction term between snow volume and the snow–ground interface, and are represented by the superscript, t, a, g, v, and gv. The

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

35

f is the radar frequency. The subscript p and q represents polarization status of the observed radar signals. T is the power transmittivity at the air–snow interface. σg pq is the surface backscattering signal at snow–ground interface. Lp = exp[−τp / cos θr ] is the snow-pack attenuation factor. τ = κe d is the optical thickness, which is the product of snow extinction coefficient – κe and snow depth – d. To carry out a forward simulation of a dry snow covered terrain Eq. (3.5) requires a total of 13 surface and snow parameters. The six snowpack parameters include snow depth, density, ice particle size, size variation, stickiness, and temperature. In addition to the dielectric constant of the ground, the six surface roughness parameters include RMS height, correlation length, and the two-parameter correlation functions at the air–snow and the snow–ground interfaces. The importance of each scattering component depends on the sensor’s frequency, polarization, incidence geometry, and snow properties. Based on the frequency dependence of SAR measurements to snow and ground properties, Shi and Dozier (2000a, b) developed a multi-frequency (L, C, and X bands) and dual-polarization (VV and HH) technique to estimate SWE and applied SIR-C/X-SAR image data. This technique uses L-band measurements to estimate snow density and the underground dielectric and roughness properties. The relationship between underground backscattering signals at C-band and X-band can be estimated with the dielectric and roughness properties estimated from L-band measurements. Then, using C-band and X-band measurements with the minimized effects of the underlying backscattering signals estimate snow depth and ice particle size. This technique requires all 3 SIR-C/X-SAR frequency measurements.

3.4.1.1 Estimation of Snow Density with L-Band Dual-polarization SAR Effects of Snow on L-Band SAR Measurements Volume scattering and extinction from dry snow are all very small at L-band. At microwave frequencies, the absorption coefficient (the imaginary part of the dielectric constant) of ice is small, and snow grains are also small compared to an incident L-band wavelength (24 cm). During the past two decades, little attention has been paid to microwave interactions of snowpack at L-band (1.25 GHz) frequencies or lower, while much work has been done at C-band (5.5 GHz) or higher frequencies (Rott and M¨atzler, 1987; Ulaby et al., 1982; Stiles and Ulaby, 1980; Ulaby and Stiles, 1980; Ulaby et al., 1984; Kendra et al., 1998). Assuming there are no large structures including the melting/frozen draining fingers and channels in the snowpack which may occur during snow melting season, snow grains can not generate significant volume scattering at L-band since snow grains are much smaller than incident L-band wavelengths. Under this condition, we can simplify the backscattering model by considering a dry and homogeneous snowpack over a bare soil or rock surface and Eq. (3.5) becomes t g σqp (k0 , θi ) = Tq (θi ) · Tp (θi ) · σqp (k1 , θr )

(3.6)

36

J. Shi

θi , θr , and k1 represent the radar wave incidence angle at the air–snow interface, cosine propagation angle in snowpack, and radar wave propagation number in snowpack for the coherent component, respectively. It indicates that the L-band radar measurements over season dry snow cover are affected by two sets of properties: (1) snow properties through the power transmittivity at the air–snow interface T which is a function of snow density, polarization, and incidence angle and (2) surface backscattering properties at snow–ground interface through surface parameters including the dielectric constant and surface roughness (roughness is generally described by an autocorrelation function and standard deviation of surface roughness height) for a random rough surface with no orientation of features. When the electromagnetic wave passes through the snowpack, versus directly striking the ground, the following differences occur: • Because of refraction within the snow, the incidence angle at the snow–ground interface is smaller. • The incident wavelength at the snow–ground interface is shorter because the snow is dielectrically thicker than air. • The snow layer reduces the dielectric contrast at the snow–ground interface, which in turn reduces the reflectivity at snow–ground interface. • The power loss at the air–snow interface reduces the total energy incident on the snow–ground interface. The first two factors result in a change of the sensor observing parameters. Note that the dielectric contrast εg /εs should be used instead of εg and that k1 should be used g at the snow–ground interface. instead of k0 when calculating the backscattering σqp √ √ The shift in wavenumber k1 = k0 εs (or wavelength λ 0 = λ1 εs ) is a function of snow density. For the range of snow densities considered −100 to 550 kg m−3 – the L-band propagation wavelength in snow ranges from 21 to 16 cm, compared to 24 cm in air. Because the surface roughness effect depends on its size relative to the incident radar wavelength, the shortening of the incident wavelength for higher snow densities will result in the soil surface appearing rougher than it would if the snow were absent. This causes an increase in the surface backscattering signal and is especially strong for a nearly smooth surface. However, this effect caused by the wavelength shift becomes smaller when the surface is rougher or the incident frequency is higher. √ Snell’s law specifies the change in incidence angle: sin(θi ) = sin(θr ) εs . The incidence angle at the snow–ground interface depends only on the incidence angle at the snow surface and the dielectric constant of snow, not on its thickness, thus it is a function of snow density. For a given incidence angle at the snow surface, a greater snow density causes a greater change in the incidence angle at the snow– ground interface. For a given snow density, however, a larger incidence angle at the snow surface results in a greater change in the refractive angle in the snow layer. Therefore, a greater increase in the backscattered power at larger incidence angles is expected than that at smaller incidence angles. Furthermore, the effects of snow density at different co-polarizations HH and VV shows a smaller increase in VV polarization than in HH polarization for the same

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

37

snow density and ground surface properties. The backscattering of VV polarization as a function of incidence angle declines more slowly than that of HH polarization when the surfaces are not too rough. Therefore, the changes in VV polarization are smaller than that of HH polarization. However, for a very rough surface, the angular dependence is smaller than that for a smooth surface, and the difference between VV and HH polarization is smaller. While a dry snowpack does not absorb or scatter the radar signal at low frequencies, it nevertheless affects the magnitude of the backscattering from the underlying rock or soil and the relationship between HH and VV polarization. The magnitude of the effect depends on the radar incidence angle, snow density, roughness and dielectric properties of the soil. Snow is more likely to enhance the backscattering magnitude of a smooth soil than a rough soil. These factors enable development of an algorithm for inferring snow density using L-band SAR measurements. At C-band or higher frequencies, however, snow density affects the magnitude of the volume scattering and the surface scattering properties at the snow–ground interface.

Snow Density Algorithm with L-band Dual-polarization SAR Measurements Based on the above understanding of snow density effects on radar backscattering at L-band, Shi and Dozier (2000a) develop an algorithm to estimate snow density by characterizing the dependence of the surface backscattering on both the incidence angle and the wavelength. It was done by establishing a backscattering σ g hh and σ g vv database using the IEM model (Fung, 1994) over a wide range of incidence angles, dielectric and roughness conditions, and incidence wavenumbers, corresponding to a range of snow densities from 100 to 550 kg m−3 . Then, the relationship of HH and VV backscattering signatures with the wide range of surface dielectric and roughness conditions at each incidence angle and wavenumber were characterized by using regression analysis to find coefficients to parameterize this relationship: g g g g g + c(θr , k1) log10 σhh = a(θr , k1) + b(θr , k1) log10 σhh σhh + σvv + σvv log10 + d(θr , k1 ) log10

g σhh g σvv

+ e(θr , k1 ) log10

g σhh g σvv

2 (3.7)

It represents a relationship between the surface backscattering coefficients σ g hh and σ g vv at a given incidence angle and wavenumber for a wide range of random rough surfaces. The form of the relationship minimizes its sensitivity to the surface dielectric and roughness properties, while maximizing its sensitivity to the incidence angle and wavenumber. The coefficients a, b, c, d and e in Eq. (3.7) depend only on incidence angle and wavenumber at the ground surface. They are given in (Shi and Dozier, 2000a).

38

J. Shi

By placing Eq. (3.6) in Eq. (3.7), the algorithm for estimation of snow density by t and σ t SAR measurements can be derived: using only σhh vv t t σhh σvv + log10 Thh (θi , εs ) Tvv (θi , εs ) t t σhh σvv + 2 (θ , ε ) Tvv2 (θi , εs ) Thh i s t 2 σt σ Tvv (θi , εs ) +c(θr , k1 ) log10 2 hh + d(θr , k1 ) log10 hh t T 2 (θ , ε ) Thh (θi , εs ) σvv hh i s

= a(θr , k1 ) + b(θr , k1 ) log10

+e(θr , k1 ) log10

t T 2 (θ , ε ) σhh vv i s t T 2 (θ , ε ) σvv hh i s

2 (3.8)

In Eq. (3.8), Tpp depends on the polarization pp, the incidence angle θi at the air– snow interface, and the dielectric constant of snowpack εs . εs is the only unknown; θi can be calculated from a combination of the spaceborne orbital data and a digital elevation model. Therefore, for a given L-band SAR measurements of σ t hh and σ t vv , we can estimate εs numerically by varying the coefficients of a, b, c, d, and e to find the root of Eq. (3.8). It does not require a priori knowledge of the dielectric and roughness properties of the soil under the snow. Furthermore, snow density can be estimated from Looyenga’s semiempirical dielectric formula (Looyenga, 1965), which provides a good fit to Polder and van Santen’s physical formula (M¨atzler, 1996). εs = 1.0 + 1.5995ρs + 1.861ρs3 (3.9) Three spatial distributed snow density maps were derived by this technique using three SIR-C/X-SAR’s L-band data-takes from its first mission in April 1994 at Mammoth Mountain study site. The three data-takes were acquired in the early morning around 6 a.m. while the snow pit measurements taken at 11 a.m. showed no signs of liquid water in the snow, even at 2,850 m elevation. Thus the imaged snowpacks were dry. Using the Space Shuttle orbital geometry data and a digital elevation model, the terrain radiometric calibration factor and local incidence angle images corresponding to the SIR-C image data were derived. The terrain radiometric correction factor is sin θ0 / sin θi , where θ0 is the incidence angle used in the initial SAR data processing under a flat surface assumption and θi is the actual local incidence angle (Van Zyl et al., 1993). To reduce the effect of image speckle on estimation of snow density, the Stokes matrixes were first averaged to azimuth (25.1 m) and slant range (13.3 m) pixel spacing to form multi-look imagery. Then, the terrain correction was made. The snow-free pixels were also masked out based on the snow classification results derived from SIR-C/X-SAR’s image data (Shi and Dozier, 1997). Figure 3.3 shows the comparison of the field measurements (x-axis) with SIRC’s L-band image data derived snow density (y-axis). The RMSEs are 42 kg m−3

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

39

Estimated

0.4

0.3

0.2 0.2

0.3

0.4

Measured snow density Fig. 3.3 Compares of the field measurements (x-axis) with SIR-C/X-SAR’s L-band image data (y-axis) derived snow density

and 13% for absolute and relative errors, respectively. It should be noticed that the estimated snow density should represent the mean value from the snowpack’s top and bottom layers. This is because the backscattering signal is mainly controlled by two factors: (1) the snow density at the top layer near the surface mainly affects the power transmissivity at the air–snow interface; (2) snow density at the bottom layer controls the incidence angle and wavenumber at the snow–ground interface. It should be noticed that the snow density algorithm described above is based on a radiative transfer model concept – Eq. (3.6) with an assumption of the wavelength shift when radar wave propagates in snowpack without regarding how thick of snow is. It is understandable if snow depth is too thin, it may not result in the notable wavelength shift. Unfortunately, the current used technique does not allow the evaluation of this issue. The studies of the well controlled laboratory experiment or Monte Carlo Model simulations are needed for further quantitative identification on what fraction of snow thickness to radar wavelength will result in the wavelength shift.

3.4.1.2 Technique on Estimation of Snow Water Equivalency with C- and X-Bands Relationships Between Snow Water Equivalence and SAR Measurements Field experiments using ground scatterometer data have shown different relationships between radar backscattering and SWE. For example, Ulaby and Stiles (1980) showed that backscattering at 8.2 and 17.0 GHz had a positive relation with SWE. Similarly, a positive relationship was also observed by an experiment over a smooth subsurface at 5.3 and 9.5 GHz (Kendra et al., 1998). However, this

40

J. Shi

positive relationship existed only over a frozen subsurface, with no correlation over the unfrozen subsurface (Bernier and Fortin, 1998). On the other hand, negative relationships have been observed at similar frequencies, 5.3 and 9.6 GHz (Strozzi, 1999). In addition, Rott and M¨atzler (1987) observed no significant difference between snow-free and dry snow covered regions at 10.4 GHz. Each field experiment represented particular snow and ground conditions. The existence of both positive and negative relationships between radar backscattering and snow water equivalence indicates that this relationship is quite complex. The confusion may result from varying combinations of snow and ground properties, because backscattered power received by the radar over dry snow depends not only on the total snow mass but also on the snow’s density, grain size, structure, and stratification, along with the dielectric and roughness properties of the underlying surface. Understanding this relationship is essential to the development of a reliable algorithm for estimating snow water equivalence. In order to understand the general relationship between the radar measurements at the frequency range between C-band and Ku-band and snow depth or SWE for dry snow cover, for simplicity, we deploy a simple backscattering model “cloud model” in which the surface scattering term from air–snow interface (the first term in Eq. (3.5)) and the interaction term between the ground surface and the snow volume (the third term in Eq. (3.5)) are ignored. Two remaining scattering terms – the direct volume and ground surface backscattering components (the second and fourth terms in Eq. (3.5)) are dominant backscattering signals. The former can be described as a function of the optical thickness (τ – a product of the snow extinction coefficient κe and snow depth d) and the volume scattering albedo (ω – a function of the snow density, particle size, size variation, stickness, and temperature). It is given in terms of the first-order backscattering model as

3 −2τ (f) v (3.10) σpp (f, θi ) = T2pp (θi , εs ) ω (f) cos(θr ) 1 − exp 4 cos(θr ) In the direct volume backscattering signals, the τ is positively related to SWE. The underground scattering component, however, is the product of the backscattering signals from underground surface and snow-pack attenuation properties through the optical thickness τ, which is negatively related to this scattering component. In other words, the effects of SWE in these two major scattering components are in an opposite way. It results in the major confusion in describing the relationship between SWE and radar observations. To further demonstrate the general relationship, we take the first-order derivative of “cloud model” with respect to snow depth d that represents the slope of the backscattering curves in response to snow depth; it can be written as

t (θ ) ∂ σpp 3 i 2 2κ e g0 = Tpp (3.11) exp(−2κ e d/µr ) · ω µr − σpp ∂d µr 4 The ground surface backscattering signal has a great impact on the relationship between the backscattering and snow depth. The sign of the bracketed term in

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

41

Eq. (3.11) determines whether the relationship between the measured backscattering signal and snow depth is positive or negative. If σ g0 pp > 3/4ω µr , the slope is negative. Physically, this means that the attenuated subsurface scattering signal, after it passes through the snowpack, is larger than the backscattering signal generated by the snowpack. It is possible for the bracketed term to be zero, in which case there is no correlation between backscattering and snow depth. In this case, the attenuated amount of the subsurface scattering signal is the same as that of the backscattering signal generated by snowpack. Even for the same snowpack, a positive relationship to snow depth may be observed over a smooth surface but a negative relationship will be observed over a rough surface. Moreover, the sensitivity of the backscattering signal to snow depth depends not only on the snow properties, but also on the incidence angle, and the magnitude of the subsurface backscattering. From Eq. (3.11), we can also see that as the snow depth increases, the change in backscattering will decrease. Therefore, we expect that the change in the backscattering measurements is more sensitive to shallow snow than to thick snow. The extinction coefficient and volume scattering albedo are positively correlated to the sensitivity of backscattering to snow depth. For a given snow depth, the larger these parameters are, the greater changes will be expected. Similarly, the sensitivity is also proportional to the angle of incidence, which indicates that a large incidence should be more effective than a small incidence for the purpose of monitoring snow depth. Note that the parameters, such as the ground dielectric and roughness properties in the ground–surface and snow–volume interaction terms, also affect the relationships. The combinations of all parameters control the overall result. Therefore, the backscattering signal from a seasonal natural snow cover and the signal’s relationship to snow depth is affected by three sets of parameters: 1. Sensor parameters, which include the frequency, polarization, and incidence geometry 2. Snowpack parameters including snow density, particle size and size variation, free liquid water content, characteristics of particle spatial distribution (stickiness), and stratification 3. Subsurface parameters that include the dielectric and roughness properties at the snow–ground interface These complex relationships make it implausible to characterize the parameters from the limited field experiment measurements and to derive an empirical model for estimating snow depth or water equivalence from SAR measurements. To estimate snow depth and thereby water equivalence, we must separate the varying backscattering signals of the subsurface or minimize the effect of the backscattering signal generated by the snow–ground interface. It is clear that estimating the snow depth requires a physically based inversion model that considers all important scattering terms and a technique to separate the backscattering components of the snowpack itself from those of the air–snow and snow–ground interfaces.

42

J. Shi

Snow Depth Algorithm with C-band and X-band SAR Measurements As shown in (Shi and Dozier, 2000a), the snow density ρs , ground dielectric constant εg , and surface RMS height s can be estimated by L-band SAR measurements. From these parameters, we can obtain εs , θr , k1 , T pp , and the reflectivity Rpp at the snow–ground interface. The remaining unknowns in Eq. (3.5) are the surface g backscattering components σpp ( f ) from the snow–ground interface and σa pp ( f ) from the air–snow interface, the extinction coefficients κe ( f ), the volume scattering albedo ω( f ), and the snow depth d. Through analyses of the simulated data, the techniques for estimating snow depth and ice particle size using SIR-C C-band and X-SAR measurements were developed. They were based on the following approaches (Shi and Dozier, 2000b): Reducing the effect of surface backscattering signal from air–snow interface. In considerations of each backscattering component from dry snow cover in Eq. (3.5) for the SWE retrieval, the backscattering signals from the air–snow interface are generally considered as the “noise” since they are typically small in comparison with the other scattering contributions and have no information on SWE. Its effect on retrieval snow properties can be reduced by using the signal generated by the estimated snow density and typical roughness parameters. Developing semi-empirical models to characterize the snow–ground interaction terms. These models represent the snow–ground interaction components more realistically than formulas developed under the assumption of independent scattering. Natural surfaces in alpine regions are quite rough. Therefore, a significant contribution to the snow–ground interactions from non-coherent components is expected. The semi-empirical models are developed explicitly in terms of snowpack volume scattering albedo, optical thickness, ground reflectivity, and surface RMS height. They are easy to implement in the inverse algorithm. Developing semi-empirical model to characterize the relationships between the ground surface backscattering components at C-band and X-band. With L-band SAR estimates of the ground dielectric constant and RMS height, the relationships between the ground surface backscatterings at C-band and X-band can be well characterized. Thus, the unknowns in the ground surface backscattering formulation can be reduced to one. Parameterizing the relationships between snowpack extinction properties at C-band and X-band. Because the extinction properties are highly correlated, we developed analytical forms for the extinction relationships at C-band and X-band. Using these relationships, the number of unknowns in backscattering components from the snowpack can be reduced to two: the volume scattering albedo ω and the optical thickness τ . From the developments in the above, the number of unknowns has been reduced to three in Eq. (3.5), that is τ, ω and the surface component: σsvv (X). Thus, these three unknowns can be solved numerically at each pixel by using Eq. (3.5) with three SAR measurements: σt vv (C), σt hh (C), and σt vv (X). To estimate snow depth d, the separation between the extinction coefficient κe and d in the estimated τ is needed. In other words, we need to determine the absolute

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

43

value of κe from the estimated parameters, i.e., τ( f ), ω( f ) and snow density ρs at each pixel. This can be done by first estimating the absorption coefficient κa (X):

τ (X) [1 − ω (X)] (3.12) κa (X) = 1.334 + 1.2182 log(Vs ) − 3.4217 log τ (C) [1 − ω (C)] where Vs is the volume fraction of ice and can be derived from the estimated snow density at each pixel by using L-band measurements. τ [1 − ω ] = κa · d = τa is the absorption part of the optical thickness at each frequency after removing scatterω (X)] κa (X) ing effects. ττ(X)[1− (C)[1−ω (C)] = κa (C) gives snowpack temperature information (Shi and Dozier, 2000b). Then, snow depth can be estimated by d=

τ (X) [1 − ω (X)] κa (X)

(3.13)

Figure 3.4 shows the comparison of the field measurements with the depths estimated from the SIR-C/X-SAR image data at Mammoth Mountain study site. The algorithm inferred the overall trend of the snow depths, with an RMSE of 34 cm. During the SIR-C/X-SAR overflights, 19 snow pits and the grain size measurements with detailed vertical profiles at depth intervals of 5–10 cm were obtained in the field measurements. Each individual ice particle size was measured by taking average values from three cross diameters with direction differences of 60◦ . There were more than 200 grains measured in each snow pit. By taking one step further, a formula was also developed for estimating the optically equivalent particle size, defined as a particle size at which the extinction properties from a natural snow volume equal those obtained from an ideal snow volume with a uniform distribution of ice particles (Shi et al., 1993). This optically equivalent particle size can be

300

Estimated

250 200 150 100 50 50

100

150

200

250

300

Measured snow depth in cm Fig. 3.4 Compares of the field measurements (x-axis) with SIR-C/X-SAR’s C- and X-band image data (y-axis) derived snow depth

44

J. Shi 2.0

Estimated

1.5

1.0

0.5

0.5

1.0

1.5

2.0

Measured grain size in mm Fig. 3.5 Compares of the field measurements of snow particle radius with SIR-C/X-SAR’s C- and X-band image data derived grain size

represented by the weighted mean size with respect to the scattering properties for a natural snowpack and can be estimated by r¯s =

0.01κs (X) 2Vs S f (2.8332 + 6.6143Vs )

1 3

(3.14)

where r¯s is in cm. κs (X) can be obtained from the estimated κs (X) and ω (X) at each pixel. Figure 3.5 compares the snow particle radius between the ground measurements and those estimated from the SIR-C/X-SAR image data. The estimated snow particle sizes were obtained from a mean value of 3 × 3 window at corresponding snow pit locations on the SAR images. The algorithm performed quite well. It inferred the overall trend of the snow particle sizes, with an RMSE of 0.27 mm. The current technique described above for estimation of SWE requires five measurements: L-VV and L-HH to estimate snow density and ground dielectric and roughness properties, plus C-VV, C-HH, and X-VV to estimate snow depth and grain size. The sensitivity analysis indicated that the C-band SAR measurements were affected mainly by the ground surface properties. The parts of the signal that comes from a typical snowpack at C-band are about 30% and 15% for HH and VV polarization, respectively. The C-band measurements are expected mainly sensitive to under-ground surface condition. Estimation of snow depth using C-band SAR measurements, therefore, requires an accurate technique to estimate the ground backscattering component. At X-band it about 60% of the signal comes from the snowpack itself. Thus, we expect that the measurement is much more sensitive to snowpack and that the requirement for estimation of the ground backscattering component is less severe for radar measurements at X-band or higher.

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

45

3.4.2 Estimations of SWE and Its Change with SAR Interferometry From last section, it can be seen that retrieving SWE from SAR backscattering measurements requires very accurate inversion models and is a very complex process. The other promising technique for SWE estimation is to use the repeat-pass SAR interferometric measurements. It has been shown that the interferometric phase shift resulting from differences in the propagation due to constant changes in snow properties offers a direct method for estimating SWE (Rott et al., 2004) or relative change in SWE (Guneriussen et al., 2001). This technique has been applied to low frequency interferometric SAR measurements at C-band (Guneriussen et al., 2001; Rott et al., 2004) and L-band (Rott et al., 2004). Considering a layer of dry snow, the radar return is dominated by scattering signal from snow/ground interface at low frequencies (C-band or lower). The repeat-pass interferometric phase Φ consists of following contributions (Rott et al., 2004):

φ = φ f lat + φtopo + φatm + φsnow + φnoise

(3.15)

Φflat and Φtopo are the phase differences due to changes of the relative distance between satellite and target for flat earth and for topography, respectively. Φatm results from changes in atmospheric propagation, and Φnoise is phase noise. Φsnow is the two-way propagation difference in the snow-pack relative to air and is resulted from the refraction of radar wave in dry snowpack. When snowpack volume scattering is neglected, the snow phase term Φsnow for an uniform layer of snow with the depth – d can be written as (Guneriussen et al., 2001):

(3.16) φsnow = −2ki · d · cos θi − εs − sin2 θi For ERS SAR with an incidence angle θi = 23◦ , the phase shift due to a change of SWE can be approximated by a linear relation (Guneriussen et al., 2001):

φsnow = −2ki · 0.87 · ∆SW E

(3.17)

It indicates that at the ERS wavelength one fringe is equivalent to 32.5 mm SWE, and for L-band (λ = 24 cm, θi = 23◦ ) one fringe corresponds to SWE = 138 mm. The coherence is determined by several factors and given by Hoen and Zebker (2000): γtotal = γthermal · γsur f ace · γvolume · γtemporal (3.18)

γthermal depends on the signal-to-noise ratio. Its contribution to decorrelation of dry snow covered ground is usually small since the signals from dry snow ground covered are generally strong. For a dry snow cover, the wave number shifts in slant range and in vertical direction have to be considered, which affects mainly on the terms of γsur f ace and γvolume . Their calculations are given by Hoen and Zebker (2000). The temporal decorrelation is a major effect on in coherence of repeat pass SAR data over dry snow covered areas. It is mainly resulted from changes of the

46

J. Shi

snowpack properties due to snow fall or snow drift. They result in changes of the snowpack structure and the roughness of the snow-surface at sub-pixel scale, and consequently change the propagation path in the snowpack of the radar return from each surface element. Rott et al. (2004) shows the term of the temporal correlation for dry snow cover can be expressed as

γtemporal = exp −2 · ki2 · σz2 cos θi − εs − sin2 θi (3.19) where σz is the standard deviation of the geometric path length through the snowpack. It takes into account of differential phase delay due to non-uniform snow accumulation or erosion. Guneriussen et al. (2001) applied this technique using ERS-1/2 tandem data and found that the repeat-pass interferometric phase measurements were very sensitive to small changes in snow properties that could introduce a significant error in DEM estimation even in the case of high degree of coherence. To avoid the temporal decorrelation effects, Rott et al. (2004) used ERS repeat pass data without snowfall between data acquisitions and derived SWE map over their study site using this technique with the required corrections for all other phase components in Eq. (3.14) and the reference points with zero and known snow accumulation. Both studied have shown that the interferometric phase shift from repeat-pass SAR measurements provides a physically based means for mapping SWE of dry snow. Temporal decorrelation due to differential phase shifts at sub-pixel scale caused by snow fall or wind re-distribution is the major limitation for application of this method. In C-band SAR data, these effects often result to complete decorrelation within a few days. L-band is preferable for this technique on mapping SWE than C-band because it is less affected by temporal decorrelation, better coherence and larger measurement range.

3.5 Need for Future Spaceborne System for Monitoring Snow Properties The current available spaceborne C-band imaging radar systems are either single polarization (ERS-2 with VV polarization and RADARSAT-1 with HH polarization) or dual-polarization ASAR (HV/VV or VH/HH). For monitoring snow properties with backscattering measurements, these current systems are only capable to identify the extent of wet snow. They are not capable of measuring SWE or of quantifying snow wetness. For interferometric measurements, the current and near future satellite systems, however, have the repeat orbit at least 24 days or longer. The presented SWE estimation technique in Section 3.4.2 requires the acquisitions within a few days and is only applicable to ERS 1/2 tandem data. For using interferometric technique to estimate SWE, the low frequency SAR instruments (L-band) is required with a quick repeat time interval. The recent launched PALSAR system has 48 days repeat pass and does not match this requirement. However, it provides the unique tool for mapping both dry and wet snow cover with the combined interferometric

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

47

coherence and backscattering measurements as presented in Section 3.2.3 but do not offer capabilities to estimate SWE and snow wetness over the range of Earth’s environments. Taking into account the capabilities and deficiencies of present remotely sensed snow cover information, its is obvious that improvements are necessary in particular regarding the observation of extent and physical properties of snowpacks – SWE and wetness. Sufficiently high spatial resolution is needed to account for the small scale heterogeneity of snowpack related to topography and land cover. For using backscattering signals, two critical capabilities of the radar instrument are required for monitoring snow water equivalence (SWE) globally. First, the radar signal can penetrate the natural snowpacks so that the snow depth information can be obtained from the measurements. Secondly, the radar signal from snowpack itself has to be large enough to provide the first-order information for snowpack itself. At the frequencies at or lower than C-band, snowpack does not generate the significant volume scattering signals and the attenuation of the ground signals also weak (Shi and Dozier, 2000b). At Ku-band (13–19 GHz), the penetration depths are ranged from 0.5 (extreme large grain) to 10 m (extreme small grain). The studies have shown the one-way penetration depth of dry snow from experimental data which at 17 GHz ranges from about 4 m for coarse grained Antarctic firn (Rott et al., 1993) to 6 m for fine-grained Alpine winter snow (M¨atzler, 1987). Secondly, the direct volume scattering component contributes about 60% radar measurement signals in the co-polarization for a typical dry snow condition at X-band 9.25 GHz (Shi and Dozier, 2000b). The numerical simulations (Shi et al., 2003; Shi, 2004) have indicated that the direct volume contribution will be significantly increased in both co-polarized and cross-polarized signals at Ku-band 17 GHz. The direct volume scattering becomes the dominant scattering source. These numbers indicate that 17 GHz is a good choice for sensing volume properties of dry snow and firn, because at this frequency the snow-pack provides a clear signal but is optically transparent enough to achieve sufficient penetration. Recent studies of the global snow cover with NSCAT, operating at 14 GHz also demonstrated the sensitivity of Ku-band to sense dry and wet snow cover (Nghiem and Tsai, 2001). Therefore, we expect that X and Ku-band instruments have capability to penetrate the most of natural dry snow cover globally and provide much more sensitive to snowpack itself than those by C-band. For future satellite mission on monitoring SWE over range of Earth’s environments, the optimal sensor configuration and corresponding algorithm development are needed.

References Bernier M, Fortin JP (1998) The potential of time series of C-band SAR data to monitor dry and shallow snow cover. IEEE Trans. Geosci. Remote Sens. 36:226–243 Chen KS, Wu TD, Tsang L, Li Q, Shi J, Fung AK (2003) The emission of rough surfaces calculated by the integral equation method with a comparison to a three-dimensional moment method simulations. IEEE Trans. Geosci. Remote Sens. 41(1):90–101

48

J. Shi

Ding K, Zurk LM, Tsang L (1994) Pair distribution functions and attenuation rates for sticky particles in dense media. J. Electromagnet. Waves Appl. 8:1585–1604 Evans D, Farr TG, van Zyl JJ, Zebker HA (1988) Radar polarimetry: analysis tools and applications. IEEE Trans. Geosci. Remote Sens. 26:774–789 Evans D, Plaut J, Stofan E (1997) Overview of the spaceborne imaging radar-c/x-band synthetic aperture radar (SIR-C/X-SAR) missions. Remote Sens. Environ. 59(2):135–140 Fung AK (1994) Microwave scattering and emission models and their applications. Artech House, Norwood, MA Guneriussen T, Høgda KA, Johnsen H, Lauknes I (2001) InSAR for estimation of changes in snow water equivalent of dry snow. IEEE Trans. Geosci. Remote Sens. 39(10):2101–2108 Haefner H, Holecz F, Meier E, Nusch D, Piesbergen J (1993) Capabilities and limitations of ERS-1 SAR data for snow cover determination in mountainous regions. Proc. Second ERS-1 Symposium, ESA PS, pp 353–359 Hoen EW, Zebker HA (2000) Penetration depths inferred from interferometric volume decorrelation observed over the Greenland ice sheet. IEEE Trans. Geosci. Remote Sens. 38(6):2571–2583 Kendra JR, Sarabandi K, Ulaby FT (1998) Radar measurements of snow: experiment and analysis. IEEE Trans. Geosci. Remote Sens. 36:864–879 Lee JS, Papathanassiou KP, Ainsworth TL, Grunes MR, Reigber A (1998) A new technique for noise filtering of SAR interferometric phase images. IEEE Trans. Geosci. Remote Sens. 36(5):1456–1465 Li Q, Shi J, Chen KS (2002) A generalized power law spectrum and its applications to the backscattering of soil surface on the integral equation model. IEEE Trans. Geosci. Remote Sens. 40(2):271–280 Looyenga H (1965) Dielectric constant of heterogeneous mixtures. Physica 21:401–406 Luojus K, Pulliainen J, Metsamaki S, Hallikainen M (2006) Accuracy assessment of SAR databased snow-covered area estimation method. IEEE Trans. Geosci. Remote Sens. 44(2):277–287 M¨atzler C (1987) Applications of the interaction of microwaves with the natural snow cover. Remote Sens. Rev. 2:259–387 M¨atzler C (1996) Microwave permittivity of dry snow. IEEE Trans. Geosci. Remote Sens. 34:573–581 Nagler T, Rott H (2000) Retrieval of wet snow by means of multitemporal SAR data. IEEE Trans. Geosci. Remote Sens. 38(2):754–765 Nghiem SV, Tsai WY (2001) Global snow cover monitoring with spaceborne Ku-band Scatterometer. IEEE Trans. Geosci. Remote Sens. 39(10):2118–2134 Pulliainen J, Engdahl M, Hallikainen M (2003) Feasibility of multitemporal interferometric SAR data for stand-level estimation of boreal forest stem volume. Remote Sens. Environ. 85:397–409 Rott H, M¨atzler C (1987) Possibilities and limits of synthetic aperture radar for snow and glacier surveying. Annals Glaciology 9:195–199 Rott H, Nagler T (1993) Capabilities of ERS-1 SAR for snow and glacier monitoring in alpine areas. Proc. Second ERS-1 Symposium, pp 1–6 Rott H, M¨atzler C, Strobl D, Bruzzi S, Lenhart KB (1988) Study on SAR land applications for snow and glacier monitoring. Technical Report 6618/85/F/FL(SC), European Space Agency Rott H, Nagler T, Scheiber R (2004) Snow mass retrieval by means of SAR interferometry Proc. FRINGE 2003 Workshop, ESA SP-550, article-46, Frascati, Italy Shi J (2004) Estimation of snow water equivalence with two Ku-band dual polarization radar. Proc. IGRASS’04, IEEE No. 04CH37612C Shi J, Dozier J (1995) Inferring snow wetness using C-band data from SIR-C’s polarimetric synthetic aperture radar. IEEE Trans. Geosci. Remote Sens. 33(4):905–914 Shi J, Dozier J (1997) Mapping seasonal snow with SIR-C/X-SAR in mountainous areas. Remote Sens. Environ. 59(2):294–307 Shi J, Dozier J (2000a) Estimation of snow water equivalence using SIR-C/X-SAR, part I: inferring snow density and subsurface properties. IEEE Trans. Geosci. Remote Sens. 38(6):2465–2474

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

49

Shi J, Dozier J (2000b) Estimation of snow water equivalence using SIR-C/X-SAR, part II: inferring snow depth and particle size. IEEE Trans. Geosci. Remote Sens. 38(6):2475–2488 Shi J, Jiang L (2004) Numerical simulations on estimation of snow wetness with dual-frequency and polarization radar. Proc. SPIE, Honolulu 5654:149–156 Shi J, Davis RE, Dozier J (1993) Stereological determination of dry snow parameters for discrete microwave modeling. Annals Glaciology 17:295–299 Shi J, Dozier J, Rott H (1994) Snow mapping in alpine regions with synthetic aperture radar. IEEE Trans. Geosci. Remote Sens. 32(1):152–158 Shi J, Hensley S, Dozier J (1997) Mapping snow cover with repeat pass synthetic aperture radar. Proc. IGARSS ’97, IEEE No. 97CH36042, vol II, pp 628–630 Shi J, Yueh S, Cline D (2003) On estimation of snow water equivalence using L-band and Ku-band radar. Proc. IGRASS’03, IEEE No. 03CH37477C Stiles WH, Ulaby FT (1980) The active and passive microwave response to snow parameters, 1, wetness. J. Geophys. Res. 85:1037–1044 Strozzi T, M¨atzler C (1998) Backscattering measurements of Alpine snowcovers at 5.3 and 35 GHz. IEEE Trans. Geosci. Remote Sens. 36(3):838–848 Strozzi T, Wiesmann A, M¨atzler C (1997) Active microwave signatures of snowcovers at 5.3 and 35 GHz. Radio Sci. 32(29):479–495 Strozzi T, Wegmuller U, Matzler C (1999) Mapping wet snow covers with SAR interferometry. Int. J. Remote Sens. 12(20):2395–2403 Tsang L (1992) Dense media radiative transfer theory for dense discrete random media with particles of multiple sizes and permittivities. Prog. Electromagnetic Res. 6(5):181–225 Tsang L, Kong JA, Shin RT (1985) Theory of microwave remote sensing. Wiley, New York Ulaby FT, Stiles WH (1980) The active and passive microwave response to snow parameters: 2, water equivalent of dry snow. J. Geophys. Res. 85:1045–1049 Ulaby FT, Stiles WH (1981) The active and passive microwave response to snow parameters: 2, water equivalent of dry snow. J. Geophys. Res. 85(C2):1045–1049 Ulaby FT, Moore RK, Fung AK (1982) Microwave remote sensing: active and passive, 2, radar remote sensing and surface scattering and emission theory. Addison-Wesley, Reading, MA Ulaby FT, Stiles WH, Abdelrazik M (1984) Snowcover influence on backscattering from terrain. IEEE Trans. Geosci. Remote Sens. GE-22:126–132 Ulaby FT, Moore RJ, Fung AK (1986) Microwave remote sensing, active and passive. vol III. Artech House, Durham, MA. Van Zyl JJ, Zebker HA, Elachi C (1987) Image radar polarization signatures: theory and observations. Radio Sci. 22(4):529–543 Van Zyl JJ, Chapman BD, Dubois P, Shi J (1993) The effect of topography on SAR calibration. IEEE Trans. Geosci. Remote Sens. 31(5):1036–1043 Wu TD, Chen KS, Shi J, Fung AK (2001) A transition model for the reflection coefficient in surface scattering. IEEE Trans. Geosci. Remote Sens. 39(9):2040–2050

Chapter 4

Multi-angular Thermal Infrared Observations of Terrestrial Vegetation Massimo Menenti, Li Jia, and Zhao-Liang Li

Abstract This chapter reviews the experimental evidence on the anisotropy of emittance by the soilvegetation system and describes the interpretation of this signal in terms of the thermal heterogeneity and geometry of the canopy space. Observations of the dependence of exitance on view angle by means of ground-based goniometers, airborne and space-borne imaging radiometers are reviewed first to conclude that under most conditions a two-components, i.e., soil and foliage, model of observed Top Of Canopy (TOC) brightness temperature is adequate to interpret observations. Particularly, airborne observations by means of the Airborne Multi-angle TIR/VNIR Imaging System (AMTIS) and space-borne observations by means of the Along Track Scanning Radiometers (ATSR-s) are described and examples presented. Modeling approaches to describe radiative transfer in the soil– vegetation–atmosphere system, with emphasis on the thermal infrared region, are reviewed. Given the dependence of observed TOC brightness temperature on leaflevel radiation and heat balance, energy and water transfer in the soil–vegetation– atmosphere system must be included to construct a realistic model of exitance by soil–vegetation systems. A detailed modeling approach of radiation, heat and water transfer is first described then applied to generate realistic, multi-angular image data of terrestrial landscapes. Finally, a generic algorithm to retrieve soil and foliage component temperatures from Top Of Atmosphere (TOA) radiometric data is described. Column water vapor and aerosols optical depth are estimated first, to obtain TOC radiometric data from the TOA multi-angular and multi-spectral observations. M. Menenti TRIO/LSIIT, University Louis Pasteur (ULP), Strasbourg, France Istituto per i Sistemi Agricoli e Forestali del Mediterraneo (ISAFOM), Naples, Italy [email protected] L. Jia Alterra, Wageningen University and Research Centre, The Netherlands Z.-L. Li TRIO/LSIIT, University Louis Pasteur (ULP), Strasbourg, France Institute of Geographic Sciences and Natural Resources Research, Beijing, China S. Liang (ed.), Advances in Land Remote Sensing, 51–93. c Springer Science + Business Media B.V., 2008

51

52

M. Menenti et al.

Then vegetation fractional cover and soil and foliage component temperatures are determined by inverting a simple two-components mixture model. The accuracy of all elements of the algorithm is evaluated by using a combination of actual measurements and synthetic radiometric data. Although applicable to multi-angular radiometric data irrespective of spatial resolution, the approach presented would be particularly relevant if space-borne observations with a footprint of 100 × 100 m or better would be available. Observing systems, presently at the design stage, with this capability are briefly described.

4.1 Introduction 4.1.1 Vegetation–Atmosphere Exchanges of Energy and Water The exchange of energy between the land surface and the atmosphere and within terrestrial vegetation canopies is a significant determinant of processes in the atmospheric boundary layer and in terrestrial ecosystems. In these processes it is crucial to determine accurately the partitioning of available energy into sensible heat flux density H (heating or cooling of the surface) and latent heat flux density λ E (evaporation from surface) over a wide range of spatial and temporal scales. Observation and modeling of turbulent heat fluxes at the land surface has been a very active area of research at least since the work of Bowen (1926) on the relative magnitude of heat transfer over dry and wet surfaces (Monteith, 1965; Feddes, 1971; Verma et al., 1976; Hall et al., 1979; Price, 1982; De Bruin and Jacobs, 1989; Beljaars and Holtslag, 1991; Lhomme et al., 1994; Sellers et al., 1995, 1996). Most conventional techniques that employ point measurements to estimate the components of energy balance are representative only of local scales and cannot be extended to large areas because of the heterogeneity of the land surface, of the dynamic nature and of the spatial distribution of heat transfer. Remote sensing is one of the few techniques to provide representative measurements, e.g., surface temperature and albedo, at regional and global scales. Methods using remote sensing techniques to estimate heat exchange at the landatmosphere interface fall into two main categories: (1) use surface radiometric temperature Trad to calculate H then obtain λ E as the residual of the energy balance equation (Blad and Rosenberg, 1976; Seguin et al., 1989; Hatfield et al., 1984); (2) use Trad to estimate the Crop Water Stress Index or the evaporative fraction (the ratio of evapotranspiration to the available energy) (Jackson et al., 1981; Menenti and Choudhury, 1993; Moran et al., 1994). The former category can be further subdivided into single-source, dual-source and multisource models corresponding with a single-, dual- or multi-layer schematization of the surface respectively. Successful estimations of heat fluxes have been achieved over horizontal homogeneous surfaces, such as a surface fully covered by vegetation, open water and bare soil (Jackson, 1985; Huband and Monteith, 1986; Choudhury et al., 1986). Large deviations from these conditions occur at partial canopies which are geometrically and

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

53

thermally heterogeneous. Recent years have seen increasing evidence of specific difficulties inherent to the heterogeneous nature of terrestrial vegetation. For instance, in many semi-arid environments where the surfaces are partially covered by vegetation, both the soil surface and cooler foliage determine the heat exchanges. This leads to the challenge of relating the separate contributions from these elements to the turbulent transport of heat across the land-atmosphere interface. Exchange of water and CO2 between land surface and atmosphere determines to a significant extent the dynamics of the Convective Boundary Layer (CBL) (Mahfouf et al., 1987; Segal et al., 1988; Hatfield, 1991; Xue and Shukla, 1993). Over homogeneous land surfaces the controlling factor is the partition of net radiation into sensible, latent and soil heat flux. The partition of net radiation is determined by the presence and functioning of vegetation and by available soil moisture. Heterogeneous land surfaces compound the complexity of these processes, since the spatial pattern of land surface properties determines CBL motion at small length scales (Hatfield, 1991; Wang and Mitsuta, 1992). These studies brought the attention of a wide scientific community to the need for significant improvements in models of such land surface processes and of the interactions of land surfaces with the atmosphere (e.g., Avissar and Pielke, 1989; Pielke et al., 1991). High resolution atmospheric models may be used to interpret observations in complex landscapes as for example done by Peng and Cheng (1993) and Yan (1999), in the framework of the Hei He basin International Field Experiment (HEIFE). Observations of the anisotropic emittance of land cover provide unique access to the thermal heterogeneity of soil–vegetation systems. All land surfaces are anisotropic and in the thermal-infrared domain, the directional variation of emitted fluxes (described by the so-called brightness temperature) is mainly determined by the distribution of temperature and emissivity between the elements of the canopy, and by the structure of the vegetation (see Balick et al., 1987; Kimes and Kirchner, 1983). Similar to the solar domain, the distribution of shadowed and illuminated parts, as well as the amount of soil and vegetation observable from a particular direction, are the main drivers of the anisotropy models that have been developed to describe the directional variations in thermal infrared spectral domain (e.g., Norman and Chen, 1990; Ottermann, 1990; Smith and Goltz, 1995; Franc¸ois et al., 1997). Determination of the soil and foliage component temperature requires inverse modeling of observed emittance, which is a challenge (Jia et al., 2003; Jia, 2004; Franc¸ois et al., 1997; Li et al., 2001; Franc¸ois, 2002; Liu, 2002), given the strength of the signals and the accuracy which can be achieved. On the other hand it provides the only opportunity to overcome a major shortcoming of current parameterizations of heat fluxes at the land–atmosphere interface (Jia, 2004). Observations of foliage and soil temperature provide also a reliable indicator of crop water stress, either directly as soil-foliage temperature difference or indirectly through modeling of soil evaporation and plant transpiration (Franc¸ois, 2002; Luquet et al., 2003).

54

M. Menenti et al.

4.1.2 Photosynthesis: Light Use Efficiency of Sun-lit and Shadowed Leaves Observations of leaf and soil temperature are also quite relevant for understanding and modeling carbon exchange between terrestrial vegetation and the atmosphere. The rate of photosynthesis depends on many factors, including carbon dioxide concentration in the atmosphere, leaf temperature, or mineral deficiencies (in particular nitrogen) in the soil. The nitrogen content of leaves is strongly related to their chlorophyll content (Field and Mooney, 1986). Foliage temperature. Autotrophic respiration is the process by which some of the chemical energy stored by photosynthesis is used by the plants themselves to grow and develop. This process is critical to the carbon cycle because it results in the rapid release of a large fraction of the carbon initially stored through photosynthesis back to the atmosphere. Autotrophic respiration depends on foliage temperature, growth rates and total biomass, as well as on the biochemical composition of the products formed in the plants (Medlyn et al., 2002). Soil temperature. Heterotrophic respiration is the process by which some of the carbon stored in organic soil components is released. The soil carbon reservoir can be very large compared to the above-ground biomass. Understanding the fluxes of carbon to (senescence, mortality) and from (respiration or mineralization) this soil reservoir becomes a major issue when closing the carbon cycle at the local scale. Heterotrophic respiration is very dependent on soil temperature, and the availability of water and nutrients, particularly nitrogen. Apart from nitrogen fertilization or deposition, symbiotic fixation of atmospheric nitrogen, and leakage or volatilization, the nitrogen cycle is intimately linked to the carbon cycle within the soil via the biotic activity. Evaluation of heterotrophic respiration is a major challenge in the description and modeling of NEP (Valentini et al., 2000; Matteucci et al., 2000). Due to the strong temperature dependence (Fig. 4.1) of both leaf photosynthesis and soil respiration, observations of foliage and soil temperature are useful to

Soil respiration [umol CO2 m−2 s−1 ]

7

An(Normalisedto 1 at 25ºC)

1.6

An

1.4 1.2 1 Betula pendula 0.8

Quercus petraea

0.6 Acer pseudoplatanus 0.4 0.2

Gossypium hirsutum

0 0

5

10

15

20

25

30

Leaf temperature ( ⬚C)

35

40

6 5

Col., Q10=2.21

4 3 2 Wal., Q10=2.16

1 Mdm., Q10=2.09

0

−5

0

5

10

15

20

Soil temperature at 5 cm depth [⬚C]

Fig. 4.1 Carbon fluxes and component soil and foliage temperatures: (left) relative net photosynthesis, An (normalized to the observed value at 25 ◦C), vs. foliage temperature (Medlyn et al., 2002); (right) soil respiration vs. soil temperature. (After Matteucci et al., 2000.)

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

55

understand and model these processes. As regards field experiments, the use of very high resolution TIR images is a widely used solution to obtain direct measurements of foliage and soil temperature (see Section 4.5). As regards observations from space, the only feasible solution is multi-angular observations of Top Of Canopy (TOC) brightness temperature and subsequent inverse modeling to determine foliage and soil temperature from observed anisotropy of emittance. A minimum of two spectral bands is required to establish the brightness temperature of the surface, taking atmospheric influences into account.

4.2 Nature of Anisotropic Emittance of Land Surface 4.2.1 Anisotropic Emissivity of Leaves and Soils The observed anisotropy of land emittance is (see Section 4.3) caused by a combination of two different processes: the inherent anisotropic emissivity of terrestrial materials and the thermal heterogeneity of complex, three-dimensional (3D) land targets. The latter is due to the interaction of radiative and convective energy transfer with the 3D structure of land targets These processes are briefly reviewed in this section. The accuracy of radiometric measurements of land surface temperature depends significantly on accurate knowledge of land emissivity. A 1% uncertainty on surface emissivity can cause about 0.6◦ error on land surface in temperature (Becker, 1987). An angular variation of emissivity has been observed by a number of scientists either in the field or in the laboratory (Becker et al., 1985; Labed and Stoll, 1991; Xu et al., 2000). For example, when sea surface wind produces wavelets, an angular variation of surface emissivity is observed (Masuda et al., 1988; Franc¸ois and Ottl`e, 1994), and an angular variation of the “effective” surface emissivity has also been observed with satellite data (Nerry et al., 1998; Petitcolin et al., 2002). Many efforts have been devoted to measure the directional emissivity of soil, leaves and other natural surfaces. Commonly, emissivity of natural surfaces decreases with increased zenith angle of observation. Becker et al. (1985) measured in the thermal infrared band the bi-directional reflectivity (BDR) of different types of bare soils including quartz sands, agricultural soil, and well-calibrated powders, a large variability of the BDR for different samples with no apparent systematic behavior apart from the backscattering peak was observed and the roughness has large impact on the BDR which means that the same material with different grain sizes may exhibit different angular distributions. As an example, Fig. 4.2 gives the measured angular variation of emissivity for water, sand, clay, slime, and gravel (Sobrino and Cuenca, 1999).

56

M. Menenti et al.

Fig. 4.2 Angular variation of surface emissivity for several natural surfaces. (After Sobrino and Cuenca, 1999, Fig. 3.4.)

Fig. 4.3 Anisotropy change resulting from changing the single-scattering reflectance from 0.3 to 0.2 in the Hapke model. The plot is in the 30◦ azimuth plane with the angle of incidence at 32◦ . Positive zenith values represent backscattering. (After Snyder et al., 1997.)

To describe the anisotropic emissivity of sands and soils, Snyder et al. (1997) defined a so-called anisotropy factor, a: a=

πf 1−ε

(4.1)

where f is the bi-directional reflectance distribution function (BRDF) of the surface, with the same meaning of the bidirectional reflectivity, ε the emissivity. To illustrate the magnitude of anisotropy, a simulation with Hapke BRDF model was done by varying single-scattering reflectance and zenith angle (Fig. 4.3).

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

57

This figure shows that the difference between the two curves for different singlescattering reflectance is relatively small with respect to the overall variation. The simulation by Snyder et al. (1997) demonstrates that for typical materials, over a realistic range of azimuth and incident and reflected zenith angles, the upper bounds of the impact of this difference is 4.8% (RMS) and 9.5% (maximum error) respectively. Efforts to evaluate the anisotropy of foliage emissivity have been limited, because it is very close to unity and has small angular variation. Therefore, for most studies, leaves are assumed to be Lambertian. The ray-tracing method has been applied to estimate the directional reflectance of canopy in VIS/NIR. An extension of this method to the TIR region may be useful to simulate the angular variation of canopy thermal infrared reflectance.

4.2.2 The 3D Structure of Vegetation Canopies The architecture of most vegetation canopies leads to a complex three-dimensional distribution of absorbed radiant energy and, therefore, of the local balance of energy within the canopy space (Fig. 4.4). On the one hand, within the canopy space the surface temperature of foliage and soil varies significantly. On the other hand, the

20

Frequency (%)

11 April, 13:00, s50 15 10 5

0 10

15

20

25

30

35

Brightness temperature (⬚C) 20

Frequency (%)

11 April, 10:30, s50

15 10 5 0 10

15

20

25

30

35

Brightness temperature (⬚C)

Fig. 4.4 Thermal anisotropy of soil–vegetation system resulting from canopy geometry and variability of absorbed solar radiation by different components: wheat leaves (blue) are significantly colder than soil (red). (After Jia, 2004.)

58

M. Menenti et al.

nadir view

off - nadir view θv

cool vegetation

warm soil

Fig. 4.5 Illustration of observed TIR radiance as a function of canopy geometry, foliage and soil component temperatures and the zenith view angle θv . The light gray bar indicates the fraction of foliage in the instantaneous field of view (IFOV) of the sensor, the dark gray bar indicates the fraction of soil in the IFOV. (Jia 2004.)

vertical distribution of foliage temperature is also variable with the solar elevation, the density of leaves and the angle distribution of leaves. The thermal heterogeneity within a vegetation canopy leads to the fact that Trad measured by thermal infrared (TIR) sensors is a function of canopy geometry, vertical distribution of foliage temperature Tf , soil temperature Ts , sensor view angle (θv , ϕv ) and incoming radiation (Kimes, 1980; Franc¸ois et al., 1997) (Fig. 4.5).

4.2.3 Radiation and Convection in the Canopy Space For incomplete canopies, a frequent case in nature, the interaction between the canopy and the atmosphere becomes complex due to the canopy geometry in terms of the size and spacing between plants, the leaf density and the leaf angle distribution. Figure 4.6 illustrates how the elements of a sparse canopy are interacting with their environment. The complex canopy geometry determines the distribution of absorbed solar radiation in the canopy, thereafter inducing spatial variability of sources and sinks of heat and water vapor in the canopy space. A large spacing between plants or lower leaf density, for instance, makes the exposed soil to play an important role in the land–atmosphere interaction. Canopy geometry has also influence on the airflow in the canopy space and the boundary layer resistance of leaves and soil, thus changing the source/sink strength. The interaction between thermodynamic and dynamic processes will lead to thermal heterogeneity, which will in turn give rise to the anisotropy in the exitance of canopy. (See, e.g., Albertson et al., 2001; Wilson et al., 2002)

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

59 Horizontal Divergence

Canopy Height

NET SENSIBLE RADIATION HEAT

WATER VAPOR

Water Vapor

Water Vapor Sensible Heat

Horizontal Divergence

dz dy dx

Storage terms in volume: ∑ Temp. change of crop ∑ Temp. change of moist air ∑ Absolute humidity change

Sensible Heat

Soil surface SOIL HEAT

Horizontal Divergence

Fig. 4.6 Schematic illustration of the 3D structure of a vegetation canopy and of the interactions between canopy elements and the canopy environment.

4.2.4 Thermal Heterogeneity of Vegetation Canopies The radiative and thermal state of elements in a soil-vegetation canopy system is strongly dependent on its geometric structure (radiative transfer) and on its environmental situation (convection and conduction processes). The interactions between the land surface and the atmosphere consist of the interactions between foliage and soil surface, foliage and air in the canopy space, soil surface and air in the canopy space, and between the soil surface and deeper soil layers, i.e., root zone. A realistic model requires describing the processes involved at each spatial point. However, it may not be possible or necessary to do so. Adequate simplification is necessary to redefine the canopy, which should retain the dominant aspects of 3D radiative, heat and mass transfer (see Section 4.3). Commonly, the soil temperature is much higher than foliage temperature because of the different thermal properties between soil and foliage. Thus, to simulate energy exchange between canopy and natural environment, vegetative canopy is commonly represented by two-source model: the sensible heat flux between soil and air, and the one between foliage and air. As the penetration of downwelling solar shortwave radiation and long-wave sky radiation, the foliage temperature changes with the depth of canopy. The temperature profile of vegetative canopy with the height of canopy has been observed by several scientists, e.g., by Paw et al. (1989) who measured the variation of leaf temperature with height within the canopy space by means of a Teletemp Infrared Thermometers (Fig. 4.7). The temperature profile of canopy layers result from the energy balance for each layer. As seen from Fig. 4.7 both foliage at top layers and at bottom layer have

60

M. Menenti et al.

HEIGHT(cm)

200 150 110 50

HEIGHT(cm)

0 150 200 50 0 19

20

21

22

23

24

25

26

27

28

29

30

31

TEMPERATURE(⬚C)

Fig. 4.7 Leaf temperature profile in the canopy, at 1,200–1,300 h, for all days. The top graph shows each individual leaf temperature measurement. The bottom figure shows a 10-point moving average of leaf temperature. (After Paw et al., 1989, Fig. 7.)

higher temperature. The higher temperature of top layer is due to the relatively large incident short-wave radiation which may decrease with depth within the canopy. The bottom layer of vegetative canopy will receive more long-wave radiation than the middle canopy layer from the soil which temperature is usually higher than foliage, with convective cooling being less efficient than at the top of the canopy. Lei (2004) made a detailed temperature measurements of sunlit and shaded leaves of Blackbrush and sunlit soil from March through December 2001 at the Clark Mountain (roughly 35.7◦ N and 115.5◦ W; elevation 1.475 m). Figure 4.8 shows the temperature difference between sunlit and shaded foliage and the temperature difference between sunlit soil and foliage.

4.3 Observed Magnitude of Anisotropy 4.3.1 Ground Observations of Tb (θ ) at TOC The anisotropy in canopy exitance implies that brightness temperature (Tb0 ) at the Top Of the Canopy (TOC) changes with view zenith angle θv and azimuth angle ϕv as shown by field measurements over a range of canopies, especially over sparse

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

3

Soil-leaf temperature difference (⬚C)

Leaf temperature difference (⬚C)

4 Spring Summer Winter

2

1

0

6

9 12 15 Hour of day

18

61

20

15

Spring Summer Winter

10

5

0

6

9 12 15 Hour of day

18

Fig. 4.8 Temperature difference between: sunlit and shaded leaves (left); sunlit soil surface and sunlit leaves (right); Clark Mountain of southeastern California. (After Lei, 2004, Figs. 4 and 5.)

Fig. 4.9 Observation of TOC brightness temperature Tb0 at different view angles. The circles represent the footprints of IFOV at the Top Of Canopy (TOC) and at the bottom of the canopy for different view angles. The components in the volume between TOC and the bottom are observed by a radiometer located above the canopy. (Jia, 2004.)

canopies. The dependence of observed Top Of Canopy (TOC) brightness temperature Tb0 (θv , ϕv ) on the view angle is best documented by ground measurements with a goniometer-mounted radiometer (Fig. 4.9, After Jia, 2004). Tb0 (θv , ϕv ) is the temperature measured by a radiometer at zenith angle θv and azimuth angle ϕv and is simply derived from the radiance Rλ measured by a radiometer by inverting the Planck’s function. Therefore, the observation of Tb0 (θv , ϕv ) and the observation of TIR radiance are equivalent. In this chapter, for simplification the symbol “R” denotes only TIR radiance and the subscript TIR will be neglected. TIR radiance of

62

M. Menenti et al.

a target is usually referred to as “exitance”, i.e., the sum of emitted and reflected TIR radiance by the target concerned. Nielsen et al. (1984) have shown that it is common to have large (up to 20K or more) differences between sun-lit soil and shadowed leaf surfaces, particularly when the top soil is dry. [127] found differences between bare soil and air temperature as large as 27◦ C. For a soybean canopy with 35% ground cover, the soil temperature exceeded the canopy temperature by 11◦ C and was 15◦ C higher than air temperature (Kimes 1980). Usually, Tb0 (θv , ϕv ) is measured by a radiometer in a specific spectral range (centered at some wavelength) and in a particular direction (θv , ϕv ), within an instantaneous field-of-view (IFOV) ΩIFOV . The portions of canopy components with different surface temperatures in the IFOV will change with the view angle (Fig. 4.9). As a consequence, strong anisotropy in exitance, i.e., a significant variation in Tb0 (θv , ϕv ) with the direction of observation, can be observed over thermally heterogeneous systems like sparse canopies. For instance, Qualls and Yates (2001) observed in a cotton field that the difference in Tb0 (θv , ϕv ) between the 0◦ (mixture of vegetation and soil) and the 80◦ (vegetation only) zenith view angles was 16.2◦ C at noon, while the difference was only 0.9◦ C in the early morning. Lagouarde et al. (1995) observed a difference of up to 3.5K for a corn canopy and 1.5K for grass (20 cm high) with a view zenith angle between 0◦ and 60◦ around solar noon. A goniometer (Fig. 4.10a) was designed specifically for canopy directional Tb0 (θv , ϕv ) measurements (Zhang et al., 2002; Jia, 2004; Li et al., 2004). Two arms are connected perpendicularly to each other with the longer one being fixed onto a

(a)

(b)

N

0

30

330 300

60

90

270

(c)

330

0 60

30

40 Target

300

60

20

270

90

120

240 210

180

150

Fig. 4.10 Goniometric system used to measure Tb0 (θv , ϕv ): (a) overview of the goniometer system and of the footprint of the TIR radiometer; (b) zenithal and azimuth sampling scheme; (c) detail of the azimuth sampling scheme. (Liu, 2002; Jia, 2004.)

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

63

circular track, which is set up on the ground, and the shorter one is kept horizontal on top of which radiometers can be mounted. The longer arm can move along the track to change the azimuth position. At each specific azimuth position ϕv the longer arm sways over a range of zenith angles θv (maximum 60◦ ). Such movements are designed and performed to measure the TIR radiance of the same target on the ground (Fig. 4.10b) within a desired range of azimuth and zenith angles. The diameter of the footprint, however, increases with increasing θv because of slant viewing. A 10◦ interval is normally used for the zenith angle change, and 15◦ interval for azimuth angle change (Fig. 4.10c). Two radiometers were used (Jia 2004) to measure Tb0 (θv , ϕv ). One radiometer was set up on the top of the short horizontal arm to measure the radiance of the canopy at each azimuth and zenith angle. Distance to the target was the same at any position of the arm so that the radiometer footprint included the same target at all positions as shown in Fig. 4.10b. The other radiometer was mounted on a mast observing continuously the canopy at nadir. The second radiometer provided the continuous measurements needed to correct for the temporal change in the measurements of Tb0 (θv , ϕv ) during a complete goniometer scan. The latter usually took about 20 min during which the surface temperature may change significantly due to the variation of the solar radiation and windspeed. Due to technical problems, different radiometers were used during the experiment (Table 4.1). A thermal camera (AGEMA THV 900 LW), mounted on top of another goniometer, was used to obtain images of surface temperature of the wheat crop Tb0 (θv , ϕv ) for prescribed azimuth and zenith directions. The AGEMA thermal camera has a scanning HgCdTe detector and a Stirling cooler with the single channel covering the spectral range between 8–12 µm, the frame rate is 15 Hz for 136 × 272 pixels, and the nominal sensitivity is 80 mK at 30◦ C. The camera was equipped with a lens having a FOV of 5◦ × 10◦ . Figure 4.11 shows the change in Tb0 from nadir to off-nadir view zenith angles at each view azimuth angle at different hours during the two selected days: (a) 11 April and (b) 21 April. Only the measurements across two perpendicular planes are shown: one in the N–S direction along the canopy rows and one in the E–W direction across the row. At each azimuth direction, measurements were made twice, e.g., at both 0◦ and 90◦ azimuth the first observation started from θv = +60◦ through nadir to θv = −60◦ (denoted as “go” in Fig. 4.11) and the second observation went back Table 4.1 The characteristics of the radiometers used to measure Tb0 (θv , ϕv ) during the field campaign of QRSLSP (Jia 2004) Instrument

Wavelength (µm)

Radiometer 1 Radiometer 2 Radiometer 3 Radiometer (single channel) Raytek radiometer

8–11, 10.4–14 8–11, 10.4–14 8–11, 10.4–14 8–14 8–14

FOV (◦ ) 4.7 8.6 8.6 15 7

64

M. Menenti et al.

(a) 0-180 go 0-180 back 90-270 go 90-270 back

3

2

1

0 −60

−40

−20

0

20

40

11 April 11 : 00

3

2

1

0 −60

60

−40

20

40

3

2

1

0 −60

−40

−20

0

20

40

3

2

1

0 −60

−60

−40

−20

0

20

40

1

0 −40

−20

0

20

40

40

60

0-180 go 0-180 back 90-270 go 90-270 back

2

1

0 −40

−20

0

20

40

60

Zenith view angle (⬚)

0-180 go 0-180 back 90-270 go 90-270 back

2

1

0 −60

−40

−20

0

20

40

60

Zenith view angle (⬚)

Zenith view angle (⬚) 4

20

4 11 April 14:00

3

60

0

11 April 13:00

−60

Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

2

Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

0-180 go 0-180 back 90-270 go 90-270 back

−20

3

60

4

−60

−40

Zenith view angle (⬚)

4

3

0

Zenith view angle (⬚)

0-180 go 0-180 back 90-270 go 90-270 back

Zenith view angle (⬚)

11 April 13:30

1

4 11 April 12:30

60

0-180 go 0-180 back 90-270 go 90-270 back

2

60

Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

0-180 go 0-180 back 90-270 go 90-270 back

11 April 12 : 00

Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

Tb0 (nadir) Tb0 (offnadir) (⬚C)

0

4

4

Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

−20

11 April 11:30

3

Zenith view angle (⬚)

Zenith view angle (⬚)

11 April 14:30

0-180 go 0-180 back 90-270 go 90-270 back

3

2

1

0 −60

−40

−20

0

20

40

60

Zenith view angle (⬚)

4 0-180 go 0-180 back 90-270 go 90-270 back

11 April 15:00

3

b0

2

T (nadir) - T (off-nadir) (⬚C)

1

3

0-180 go 0-180 back 90-270 go 90-270 back

11 April 16:00

2 1 0

b0

Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

0-180 go 0-180 back 90-270 go 90-270 back

Tb0 (nadir) Tb0 (offnadir) (⬚C)

11 April 10:30

Tb0 (nadir) - Tb0 (offnadir) (⬚C )

Tb0 (nadir) - Tb0 (off-nadir) (⬚C)

4

4

4

0

−1 −60

−40

−20

0

20

Zenith view angle (⬚)

40

60

−60

−40

−20

0

20

40

60

Zenith view angle (⬚)

Fig. 4.11 Measurements of the directional variation of TOC brightness temperature difference Tb0 (nadir) – Tb0 (off – nadir) with zenith view angle: (a) on 11 April and (b) on 21 April at different times of the day at the QRSLSP site (Liu et al. 2002). Tb0 (nadir) is Tb0 at θv = 0◦ , Tb0 (off – nadir) is Tb0 measured at θv = 0◦ . The positive zenith view angle correspond to the azimuth 0◦ and 90◦ planes, the negative zenith view angle correspond to the 180◦ and 270◦ azimuth planes. (Jia 2004.)

from view zenith θv = −60◦ through nadir to θv = +60◦ (denoted as “back” in Fig. 4.9). All the measurements shown in Fig. 4.11 have been corrected for temporal change taking into account the measurements provided by the second nadir looking radiometer. Changes in Tb0 with θv are significant and show different trends at different hours during a day. Around noon, the observed near-nadir Tb0 is always higher than that at off-nadir positions. On the contrary, in the early morning and the late afternoon,

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

65

near-nadir Tb0 tends to be lower than the off-nadir values. Maximum difference between near-nadir Tb0 and those at off-nadir is about 2.8◦ C on 11 April, while 4.4◦ C is observed on 21 April, both around noon time. The row structure of the winter wheat also plays an important role in determining the angular change of Tb0 . Such structure effects are evident when comparing the shape of the curves in the along-row direction (the plane from 0◦ to 180◦ ) and the curves in the across-row direction (the plane from 90◦ to 270◦ ) in Fig. 4.11. The latter shows a steeper slope, particularly in the position opposite to the sun, e.g., at 270◦ plane (negative zenith view angle) at 10:30, 11:00 and 11:30 on 11 April 2001 when the sun was located between 90◦ and 180◦ planes, and asymmetric than the former. For the proper interpretation of multi-angular measurements the geometry of observations needs to be taken into account, because of significant differences in footprint size, position and shape at different view angles. The change in the diameter of the radiometer footprint when the radiometer observes the target at different zenith view angles implies significant differences in the canopy elements captured by observations.

4.3.2 Airborne Observations To our knowledge the only airborne imaging system having the capability to perform multi-angular thermal infrared observations is the Airborne Multi-angle TIR/VNIR Imaging System (AMTIS) developed by Wang et al. (2002). This is a prototype three-channel multi-angle imaging system. It provides high-resolution data in visible/near infrared and thermal infrared spectral bands for use in deriving bidirectional reflectance factors. The precision of the viewing angle is determined by the pointing precision of the AMTIS and the knowledge of the attitude of the airplane (pitch, roll and heading angles). The AMTIS consists of one visible CCD camera, one near-infrared CCD camera, one thermal video system (TVS), camera bench, swing assembly, motor driver and controller, exposing synchronization signal generator, data grabber, real-time display and recorder. Two CCD cameras and the TVS are mounted on the camera bench. A stepper motor rotates the camera bench. The instantaneous field of view (IFOV) of each pixel has been fixed at 0.3 mrad for the VNIR bands, 1 mrad for the TIR band. The total field of view of each camera is about 20◦ . The altitude range is roughly from 500 m to 10,000 m which thus translates into a spatial resolution ranging from 0.15 m to 3 m for VNIR, 0.5 m to 10 m for TIR. Normally, AMTIS is operated at an altitude of 3,000 m with a ground speed of 250 km/h, and swing along-track through discrete nine programmable viewing angles within a range of ±45◦ . A step motor through a gearbox drives the cameras bench. The accuracy of the pitch angle is 0.3◦ . The maximum swing speed is 200 steps/s, or 60◦ /s. Normally, the swing angular velocity is about 10◦ per second when AMTIS takes photograph, then swings back at the angular velocity of about 30◦ per second.

66

M. Menenti et al.

4.3.3 Space Observations of Tb (θ ) at TOA and TOC The measurements of thermal emission at top of atmosphere include the contributions of atmosphere (through atmospheric upwelling radiation and scattering of surface thermal emission) and those of the surface. To perform nearly simultaneous observations of Tb0 (θv , ϕv ) from space at multiple view angles two technical solutions may be used: (a) multiple optics or (b) accurate along-track pointing. The Along Track Scanning Radiometer 1 (ATSR-1) flown on ERS-1 was based on solution (a). ATSR-1 was followed by ATSR-2 in 1995 and the Advanced ATSR (AATSR) on ENVISAT in 2001. The characteristics of these systems (Mutlow et al., 1994) are summarized below. All ATSR sensors acquire dual-view angle data (approximately 0◦ and 53◦ at surface) in four channels for ATSR-1 and seven channels for ATSR-2 and AATSR. The nominal noise equivalent temperature difference (NEδT) of ATSR-2 for IRT channels is 0.04K. The use of the along track scanning technique (Fig. 4.13) makes it possible to observe the same point on the earth’s surface at two view angles through two different atmospheric path within a short period of time. The first view is at a view angle of 55◦ (approximately 53◦ at the earth surface) along the direction of the orbit track when the satellite is flying toward the target point, which is referred to as forward observation in this Chapter. Within 21/2 min the nadir (0◦ ) view observation is made over the same point, which will be referred to as nadir observation later on. Figure 4.12 shows the viewing geometry of ATSR-2 observations. The swath width of ATSR-2 is 500 km, which provides 555 pixels across the nadir (0◦ zenith angle) swath and 371 pixels across the forward (55◦ zenith angle) swath. The nominal pixel size of ATSR-2 is 1 × 1 km at the centre of the nadir swath and 1.5 × 2 km at the centre of the forward swath. The Along-Track Scanning Radiometer (ATSR) instruments are imaging radiometers which are currently the only observing system able to provide from

Frequency (%)

8

6

4

2

0 −6

−4

−2

0

2

4

6

8

10

12

Tb(0) - Tb(45)

Fig. 4.12 Difference of surface brightness temperature Tb between nadir view and forward view (45◦ ); observations by AMTIS in Shunyi of Beijing on 11 April 2001.

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

67

ATSR Instrument

Sub-satellite Track

55⬚

Nadir view swath (555 nadir pixels 1 km resolution)

Flight direction Forward view swath (371 along track pixels 1.5 km x 2 km resolution)

Fig. 4.13 Illustration of Along Track Scanning Radiometer (ATSR) observation. (After Mutlow et al., 1999.)

space quasi-simultaneous bi-angular radiance measurements of the earth’s surface in the TIR and SWIR spectrum regions (in addition to VIS/NIR channels). ATSR1 onboard the first European Remote Sensing satellite (ERS-1) was launched in July 1991 and operated until June 1996. ATSR-1 had four channels – one short wave infrared (SWIR) channel located at 1.6 µm and three TIR channels centred at 3.7 µm, 11 µm and 12 µm. ATSR-1 was designed particularly for providing data over the sea. ATSR-2 onboard the ERS-2 satellite was launched in April 1995 and is currently providing data both over land and over sea. In addition to one SWIR channel and three TIR channels as on ATSR-1, ATSR-2 and AATSR have three narrowband visible-near infrared channels in the blue, green and red spectrum located at 0.55 µm, 0.67 µm and 0.87 µm respectively for vegetation monitoring. Table 4.2 gives information on channel spectral characteristics. (see the World Wide Web site at www.atsr.rl.ac.uk/software.html for details). The ATSR-2 data analysed below pertain to images acquired on June 6th 1999 over Spain at 10:30 a.m. local solar time. The major part of this image is over a cultivated zone, the rest being bare soils. This image is considered cloud-free, inasmuch as ATSR-2 data do not indicate the presence of clouds. The four images have been corrected for atmospheric effects: channel 1 (12 µm) nadir and forward views; channel 2 (11 µm) nadir and forward views. Since ATSR has a limited dynamical range of brightness temperatures, and radiometric saturation is supposed to occur at 312◦ K, all pixels having Tb ≥ 312◦ K have been masked. Atmospheric corrections have been performed using the split-window method given

68

M. Menenti et al.

Table 4.2 Central wavelength and bandwidth of ATSR-2 spectral channels Channel No.

Central wavelength (µm)

Full width at half maximum (µm)

12.0 11.0 3.7 1.6 0.87 0.65 0.55

11.60–12.50 10.52–11.33 3.47–3.90 1.575–1.642 0.853–0.875 0.647–0.669 0.543–0.565

1 2 3 4 5 6 7

TOC(nd)-TOA11(nd)

45

TOC(nd)-TOA12(nd)

40

TOC(fw)-TOA11(fw) TOC(fw)-TOA12(fw)

Frequency (%)

35 30 25 20 15 10 5 0 −2

0

2

4

6

8

10

12

14

Tb(TOC) - Tb(TOA)

Fig. 4.14 Atmospheric impact on surface brightness temperature for two TIR channels at both nadir and forward views. Data are from ATSR-2 on June 6th 1999 over Spain. (nd = nadir; fw = forward.)

by Eq. (4.7) (Li et al., 2003). The difference in brightness temperature at a given wavelength and a given view angle between the TOC and TOA levels (Fig. 4.14) indicates the magnitude and the dependence on view angle of atmospheric effects. The temperature at TOC is always higher than the temperature at the TOA level and this difference is larger at the 45◦ view angle because of the larger atmospheric optical depth. For a given view angle, the difference is larger for channel 1 (12 µm) than for channel 2 (11 µm) due to the stronger atmospheric absorption in channel 1. The surface anisotropy signature is observable after atmospheric correction (Fig. 4.15). Most pixels show a lower temperature in the forward image than in the nadir image. The histogram peaks around 3.5 K at TOC level and around 5.0 K and 5.2 K at TOA level for channels 2 and 1, respectively. This figure shows also that the angular variation observed at TOA comes from both the angular variation at TOC and the

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

69

TOC(nd) - TOC(fw)

25

TOA11(nd) - TOA11(fw) TOA12(nd) - TOA12(fw)

Frequency (%)

20

15

10

5

0

−5

0

5

10

15

Tn - Tf

Fig. 4.15 Difference in surface brightness temperature between nadir and forward view at TOC and TOA. Data are from ATSR-2 on June 6th 1999 over Spain

angular dependency of the atmospheric effects, and the effects of atmosphere generally increases the TOA anisotropy signature as compared to the angular variation observed at TOC. The experimental evidence reviewed in this section leads to conclude that the thermal heterogeneity of soil–vegetation systems is very significant (see, e.g., Figs. 4.4, 4.8 and 4.11). On the other hand, the same observations suggest that soil– foliage temperature differences are much larger than the difference between sunlit and shadow elements, both foliage and soil. In other words, available data suggest that a 2-components conceptual model of soil–canopy systems may be adequate in most cases, with 4 components (sunlit, shadow; foliage, soil) necessary under extreme radiative forcing. This conclusion leads to the modeling approach presented below.

4.4 Modeling the Anisotropic Exitance of Soil-Canopy Systems Several models have been proposed and developed to describe and handle the anistropic exitance of soil–canopy systems. They can be loosely classified in three categories: 1. Simple geometric (deterministic) model of the system: this approach applies to structured vegetation (row crops, tree lines, patches) inasmuch geometry is known and the system can be described with a small number of known parameters (e.g., dimensions, soil emissivity, vegetation emissivity, ...). (Sutherland and

70

M. Menenti et al.

Bartholic, 1977; Kimes et al., 1981; Kimes, 1983; Sobrino and Caselles, 1990; Caselles et al., 1992). Attempts to incorporate a coupling with the atmospheric radiation (down welling) have been scarce (Colton 1996). Such an approach is quite useful for sensitivity studies or to assess the feasibility of the retrieval coupled to the atmospheric correction requirements. Except when the geometry is accurately known, or can be well inferred from other measurements (also from satellite), usefulness of this type of model is however limited since it cannot deal with the physical processes within the system, and model inversion is very sensitive to uncertainties in parameters. 2. Radiative transfer within the canopy: this approach applies to dense (or less dense) systems that can be described statistically and using biome characteristics. Models in this domain solve radiative transfer in the canopy with atmosphere and soil as boundary conditions, assuming plant type, distribution, plant architecture, LAI (total, horizontally/vertically projected), LIDF, etc. (Ottermann, 1990; Ottermann et al., 1992, 1995, 1999; Balick et al., 1987; McGuire et al., 1989; Kimes, 1981; Norman and Chen, 1990; Smith and Goltz, 1995; Smith et al., 1996). Soil temperature, leaf temperature, temperature gradient within the canopy may either be assumed, or simultaneously solved for. Observed TIR anisotropy may reveal whether there exists a temperature gradient within the canopy or not. However, interpretation of directional radiance implies that all parameters of the system are known or can be accurately retrieved from other measurements (from satellite in the visible, NIR and SWIR domains). Since the fluxes within the canopy are coupled to the flux above the canopy, the micrometeorological parameters have to be known. It turns out that the surface TIR directional effect is quite sensitive to ambient conditions. Thus, for a given biome, the TIR emitted radiance may reverse the sign of its angular variation with zenith angle (i.e., decrease or increase with increasing zenith angle), or even show no variation at all. This category of models may not be best for inverse modeling of observations. Nevertheless, such models are extremely useful for: (1) for evaluating the order of magnitude of the angular effect that can be expected and (2) for comparing what is observed with outputs of models. It is worth noting that the modeled anisotropy is in no case large, no more than a few K only if large (>60◦ ), zenith angle can be used. The Geometric-Optics method was initially proposed to model the radiative transfer through conifer canopy in near infrared and visible domains (Li and Strahler 1985). After taking into account the component emittances at the pixel scale, the geometricoptics method has been extended to the thermal infrared domain to model the angular variation of thermal emission from vegetative canopies at the local scale (Su et al. 2003). The core idea behind geometric optics method is that, a vegetation canopy is assumed to be represented by different component elements with different shapes (such as cone, elliptical, or sphere), the parameters for these shapes (such as height, radius, and apex angle) are specified, the spatial distribution for these canopies in pixel plane is previously determined (random, regular, or row), and the position of solar and sensor is used as input. After accounting for the mutual shading between canopies, the fractions of four components (sunlit foliage, sunlit

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

71

soil, shaded foliage and shaded soil) are computed through geometrical optical rules. Then the contributions of each component to the reflected or emitted radiance are combined to compute the reflected or emitted radiation at top of canopy level or pixel level. Multi-scattering between different components is also included. Some improvements on the original geometric optic method have been implemented to take into account the spatial heterogeneity including clumping index of vegetative canopy which describe the non-random distribution of canopy elements including leaves, branches, and the non-random spatial distribution of vegetative canopies in pixel scale (Chen and Leblanc 1997). A limitation of the geometric optic method is that the mutual shading, gap distribution inside vegetative canopy, multi-reflection, and scattering cannot be accounted for. To overcome this shortcoming, a so-called hybrid geometric-optical radiative transfer method was proposed to improve the original geometric-optical method. By using this geometric-optical radiative transfer method, Yu et al. (2004) simulated the directional brightness temperature over a maize canopy with a row structure. The gap probability between rows was computed with geometric-optical rules, while the gap inside rows was calculated with radiative transfer theory. Row and canopy structure parameters including row width, canopy height and width between rows are needed to initialize their model. The comparison between the simulations with their model and in situ measurements showed that, the angular variation of brightness temperature can be precisely captured with this type of model. Radiosity (Gerstl et al., 1991; Gerstl and Borel, 1992) is a computational algorithm to describe the scattering of light between ideally diffuse (Lambertian) surfaces. Although this method has been devoted to modeling bi-directional reflectance distribution function (BRDF) in VIS/NIR optical remote sensing for a long time (Borel et al., 1991), there are few reports on its application to thermal infrared. 3. Inhomogeneous thick vegetation layer: that can be statistically described by an angle dependent “gap fraction” or “gap frequency” (Nilson, 1971). This approach represents an intermediate situation between 1 and 2. It allows the directional TIR radiance to be described as a linear combination of the foliage and soil radiance contributions, weighted by the gap fraction. Inverting directional TIR radiance may be used to retrieve vegetation temperature and soil temperature, assuming, for instance, an angle dependent canopy emissivity. A detailed and comprehensive discussion of direct and inverse modeling is found in (Franc¸ois et al., 1997). The advantage of this type of models is that the gap fraction can be phenomenologically correlated with measurements in the visible-NIR domain through appropriate vegetation indices (Baret et al., 1995). Hence, combination of measurements in the solar reflected domain and in the thermal infrared domain may be of great value, if not yet enough, to solve the problem (i.e., retrieve soil and foliage temperatures). A further step can be made by exploiting detailed BRDF (including hot spot as much as possible) measurements from satellites in the visible – NIR domain to assess, e.g., LIDF type, that would help improving the gap fraction estimate.

72

M. Menenti et al.

4.4.1 Detailed 3D Models of Radiative Transfer in Vegetation Canopies Realistic simulation of the canopy structure and remote sensing scenario has been developed to model the radiative transfer of downwelling solar radiance and atmospheric radiance reflected by vegetative canopy and soil (Kimes and Kirchner, 1982) and of thermal emission from leaves and soil by Guillevic et al. (2002). The so-called DART (Discrete Anisotropic Radiative Transfer) model has been developed (Guillevic et al., 2002; Gastellu-Etchegorry et al., 2004; CESBIO, 2006) to simulate pixel scene and radiation energy budget in this scene. The whole scene is specified by the locations of different objects (including tree, lake, soil, lake, and building) and combined by a 3-D matrix of parallelepiped cells of various sizes. The “atmosphere” part is simulated with “air” cells, while the “terrestrial” part is made up of “air” cells. Cells are associated with 4 types of elements (soil, vegetation, wall, water), with or without relief (DEM, Digital Elevation Model). More details of this model can be found in Guillevic et al. (2002) and Gastellu-Etchegorry et al. (2004). A realistic representation (Fig. 4.16) of canopy geometry and of the spatial heterogeneity of radiative and convective processes can be constructed using a discrete 3D grid (Welles and Norman, 1991). The grid points are the basic unit in the 3D model dealing with the radiation, heat and water vapor fluxes calculations. The components of radiative environment, such as the direct beam radiation from the sun, the diffuse radiation from the sky and soil, and the diffuse radiation scattered by foliage at any grid point are calculated. This description of geometry may also be used to model convection of heat and water vapor. One should note that since we have assumed wind speed, air temperature and vapor pressure are homogeneously distributed along the horizontal directions, the grid points at the same vertical level have identical values of these variables wherever they are located in the horizontal directions.

4.4.2 Simpler Models: Linear vs. Non-linear, 2 vs. 4 Components Mixture Models Typically, the pixel or IFOV (Instantaneous Field Of View) of an imaging radiometer includes both foliage and soil where soil is commonly much warmer than foliage. The experimental evidence reviewed in Section 4.3 suggests that the thermal heterogeneity of a soil-canopy system may be represented under most environmental conditions using foliage and soil components only. Under extreme radiative forcing the difference between sunlit and shaded canopy elements becomes larger. Simple linear and non-linear models have been proposed to simulate radiance or brightness temperature measured at TOC. A 2-component mixture model can be used to simulate the directional distribution of thermal radiation as (Ottermann et al., 1992, 1995):

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

73

Reference height in atmospheric surface layer

TOC

F Canopy space R, u , S, C Soil surface

Root zone Fig. 4.16 Simplified schematic illustration of interactions between points (either containing foliage or soil) in a three-dimensional canopy (soil + vegetation) and in the atmosphere just above the TOC, with all possible physical, chemical and biophysical processes included. TOC is represented by “green plane” while soil surface is represented by “orange plane”. The green blocks represent sub-canopies. Points are symbolized by symbol “•” interaction between points are represented by implies that the interactions are 3D (vertical and horizontal exchanges). “lines”. The symbol Each point is characterized by absorbed radiation flux density R, windspeed u, concentration C of a scalar (i.e., temperature, moisture, CO2 , etc.), source/sink S of a scalar. F represents the flux density of a scalar between points. (Jia, 2004.)

Bλ (Tb0 ) = f εv Bλ (Tv ) + (1 − f )εg Bλ (Tg )

(4.2)

where, Tb0 is the brightness temperature at TOC. εv and εg are emissivities of foliage and soil, respectively. f is the fractional coverage of vegetation. Tv and Tg are temperatures of foliage and soil, respectively. Bλ denotes the Planck function. Previous study showed that the directional thermal radiation is not sensitive to the uncertainty of soil and leaf emissivities (Franc¸ois et al., 1997). This model implies that both the multiple reflection between vegetation and soil, and the temperature differences between sunlit and shaded canopy elements are neglected. It should be noted that Eq. (4.2) is a linear representation of directional thermal radiation at top of canopy without taking into account the atmospheric downwelling radiation and the temperature differences between sunlit and shaded elements inside canopy. As indicated, under extreme radiative forcing, the sunlit elements are much warmer than the shaded ones. The latter implies that accurate modeling of TOC emittance requires taking into account leaf level radiation balance and 3D canopy

74

M. Menenti et al.

structure, either by explicit modeling of leaf level radiative processes or by parameterization. Thus, the sunlit and shaded elements need to be treated differently as, 4

Bλ (Tb0 ) = ∑ fi εi Bλ (Ti )

(4.3)

i=1

Where: fi , εi , and Ti are the fraction, emissivity, temperature of component element i, respectively. Totally, four elements, namely sunlit foliage, shaded foliage, sunlit soil and shaded soil are included. The equations for a four component system are derived in a similar way as shown here for two components (Jia, 2004). The Eqs. (4.2) and (4.3) do not account for the multiple interactions between canopy elements and soil which is the so-called cavity effect (Sutherland and Bartholic, 1977). To simulate such a cavity effect, a non-linear item is introduced as (Li et al., 2001; Menenti et al., 2001): B(Tb0 (θ )) = P(θ )εs B(Ts ) + [1 − P(θ )] εv B(Tv ) + (1 − Ph )εv B(Tv )P(θ )(1 − εs ) +α (1 − εv )εv B(Tv ) [1 − P(θ )] + β (1 − εv )εs B(Ts ) [1 − P(θ )] (4.4) +(1 − εc )Ratm↓ where P(θ ) is the ground fractional cover viewed at angle P(θ ) = 1 − f (θ ), εs and Ts are the soil emissivity and temperature, εv and Tv are leaf emissivity and temperature, Ph is the hemispheric gap frequency defined as the ratio of the radiation travelling through the canopy and reaching the soil to the incident radiation into the canopy over the hemisphere, α and β are respectively the probability of the radiation emitted by a leaf and reflected by other leaves in the canopy and the probability of the radiation emitted by soil and reflected by the leaves above it, εc is the canopy emissivity and Ratm↓ is the downward hemispheric atmospheric radiance divided by π . The first term represents the proportion of the soil radiation that reaches the top of the canopy in the direction θ. The second term is the upward emitted radiation from the vegetation in the direction. The third term represents the downward radiation emitted by the vegetation and reflected by the soil and subsequently traveling upwards through the vegetation in the view direction (vegetation–soil interaction). The fourth term is the contribution of the radiation emitted by the leaves towards other leaves into the canopy and reflected by these leaves towards outside the canopy in the view direction (vegetation–vegetation interaction). The fifth term is the contribution of the radiation emitted by soil towards the leaves and reflected by these leaves towards outside the canopy in the view direction (soil–vegetation interaction), The last term is the downward hemispheric atmospheric radiance reflected by the canopy system. If we define the effective emissivity of soil and vegetation, Es and Ev , as: Es = εs + (1 − Ph )εv (1 − εs )B(Tv )/B(Ts ) Ev = εv + α (1 − εv )εv + β (1 − εv )εs B(Ts )/B(Tv )

(4.5a) (4.5b)

Eq. (4.4) is reduced to Eq. (4.2) with effective emissivity instead of actual emissivity.

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

75

4.5 Modeling of Observations and Sensor Design 4.5.1 Modeling Approach The relationship between canopy geometry, leaf and soil properties, radiative and convective processes is rather complex (see previous Sections) leading to significant thermal heterogeneity, while radiometric measurements capture the overall effect of such heterogeneity only. To understand the nature and information conveyed by radiometric measurements it is useful, therefore, to apply the modeling approach outlined in Section 4.4 to model observations and, particularly, their dependence on canopy (e.g., LAI) and soil (e.g., water content) properties Jia et al., 2005). Simulation of radiometric data or full images can be done considering realistic canopy and environmental conditions by combining a soil–vegetation–atmosphere transfer (SVAT) model and atmospheric radiative transfer (RT) model (Jia, 2004). Some TOC biophysical variables needed as input to the SVAT model are simulated using the RT models PROSPECT and GeoSAIL (Verhoef and Bach, 2003). The TOC Tb (θ) was simulated using the SVAT model Cupid. Radiative transfer in the atmosphere is simulated using the RT model MODTRAN 4.0 taking into account the sensor specifications, particularly spectral coverage, spectral sampling and channel spectral response. The simulated images of directional radiance in two TIR channels (11 and 12 µm) at top-of-atmosphere (TOA) are obtained by adding the atmospheric effects to the top-of-canopy radiance (or images). The SVAT model Cupid is a one-dimensional, multi-layer model that simulates various plant-environment interactions (Norman, 1979; Norman and Campbell, 1983; Wilson et al., 2003). The essential processes simulated in the Cupid model are divided as above-ground processes and below-ground processes stratified by canopy layers (see Fig. 4.16). Above-ground processes are dominated by vegetation including the transport of energy, mass and momentum between plants and their environment. Above-ground canopy is stratified by horizontal layers by identical increment of leaf area index and each horizontal vegetation layer is stratified by leaf angle class. Interactions between leaves in each individual leaf angle class in each horizontal layer and their local environment are first formulated. Collective effect of all the leaves in each horizontal layer is integrated to obtain the response of the layer. Canopy-level responses are simulated by numerical integration over all canopy layers where soil layers are also included. The below-ground processes describe the transport of heat and mass between the roots and their soil environment and between the soil layers defined by identical increment of depth. Unlike many other SVAT models that take soil surface as lower boundary, Cupid has been developed to have plant root zone as the lower boundary. The soil lower boundary conditions consist of soil temperature and soil water content near the bottom of the root zone. All the processes occurring at the soil surface layer are simulated rather than input.

76

M. Menenti et al.

Leaf optical (reflectance, transmittance and emissivity) and physiological properties (photosynthetic rate, respiration rate and stomatal conductance) and leaf position and orientation arrangement determine energy and mass exchanges at leaf level, while canopy structure (leaf density distribution, total leaf area, and canopy height) determines the distribution of absorbed solar radiation in canopy layers thereafter determines mass and heat transfer between layers. The forcing of all the processes described in Cupid model is ambient atmospheric conditions above canopy and soil boundary conditions at the bottom of the root zone. Ambient atmospheric conditions are defined by solar radiation, air temperature, humidity, wind speed, and precipitation measured at some reference height above the canopy. Equations describing heat and mass throughout the entire canopy (leaf energy budget for all leaves and the vertical flux-gradient equations of soil and vegetation layers) are solved simultaneously. Among others canopy layer temperature profile (including soil surface layer) and thermal radiation flux profile are obtained from the solution of Cupid which can then be used to simulate canopy directional radiance in the concerned wavelength. This is done by assuming that the contributions of various leaf layers and of the soil layer are appropriately weighted by the fraction of each layer viewed by the radiometer in a particular view direction. An overview of the simulation procedure is given in Fig. 4.17.

4.5.2 Generation of Synthetic Multi-angular Images The at-sensor radiance is a combination of the surface-emitted radiance and the atmospheric contributions. The anisotropy of at-sensor radiance is due to atmospheric scattering, absorption and emission in addition to the inherent anisotropy of TOC radiance. The most significant interaction of TIR radiance with the atmosphere is attributed to atmospheric absorption primarily due to ozone and water vapor in the atmosphere. The optical path length is greater in the off-nadir views where the water vapor path increases, thus contributing higher upwelling atmospheric radiance, while atmospheric transmittance is smaller. To demonstrate the approach described above a simulated data has been generated (Fig. 4.17; Jia et al., 2005) for three European sites with a heterogeneous and detailed spatial structure to represent major, rather different biomes, such as evergreen forest, deciduous forest, savannah, semi-grasslands and agricultural land. Result presented here relate to two sites briefly described below: 1. Alpine Foreland site. The land surface is classified by 22 land use classes with grassland, mixed forest, shrubs, coniferous forest, deciduous forest, and various crops as dominant land cover types. Four dates were selected for image simulation: 22/042009, 17/06/2009, 19/07/2008, 18/09/2008. The central coordinates of these images are about 47.9◦ N and 11.1◦ E.

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

77

Broadband emissivity of leaf and soil

Other inputs

Cupid model

Sensor geometry

Atmospheric profiles

- canopy temperature profile (including soil surface layer) - thermal radiation flux profile

MODTRAN Canopy RT model

τ(λ,θ, ϕ), Ratm↑(λ,θ, ϕ), Ratm↓(λ) εc(λ ,θ, ϕ)

Trad(λ ,θ, ϕ)

Tb0(λ ,θ, ϕ)TOC

Atmospheric RT model

Sensor channel response function

Tb(11, θ, ϕ) TOA Tb(12, θ, ϕ) TOA

Spatial model

SPECTRA images of Tb(11, q, j) TOA Tb(12, q, j) TOA

Fig. 4.17 The major steps and applied models in the TIR image simulation. (Jia et al., 2005.)

2. Boreal forest site. Boreal forest site is located in Sodankyla of Finland with typical boreal forest covered mainly by deciduous forest (ca. 42%) and coniferous forest (ca. 40%) and dark litter (ca. 12%). The soil in Sodankyla site is mainly sandy type. After obtaining the atmospheric path radiance and transmittance using MODTRAN 4.0 in combining the atmospheric profiles at the two sites, the TOA TIR images were generated by applying these variables to the TOC TIR images. Figure 4.18 shows the images at the Sodankyla site as an example.

78

M. Menenti et al.

Fig. 4.18 Synthetic multi-angular images of TOC Tb (θ ); Boreal forest at Sodankyla, Finland: θ = 0◦ (left) and θ = +60◦ (right).

4.6 Retrieval Algorithms 4.6.1 Introduction An operational algorithm was described by Jia et al. (2003) to retrieve soil and foliage component temperatures over heterogeneous land surface based on the analysis of bi-angular multi-spectral observations made by ATSR-2. This algorithm is a good illustration of the synergistic use of multiangular observations in the VNIRSWIR region and in the TIR region to retrieve simultaneously both land surface and atmospheric variables (see also Verhoef and Menenti, 1998). This algorithm is described in this section. On the basis of the radiative transfer theory in a canopy, a model is developed to infer the two component temperatures using six channels of ATSR-2. Four visible, near-infrared and short wave infrared channels are used to estimate the fractional vegetation cover within a pixel. A split-window method is developed to eliminate the atmospheric effects on the two thermal channels. An advanced method using all four visible, near-infrared and short wave channel measurements at two view angles is developed to perform atmospheric corrections in those channels allowing simultaneous retrieval of aerosol optical depth and land surface bi-directional reflectance (Fig. 4.19). This general approach has been applied to image data acquired with the airborne AMTIS (see Section 4.3). The algorithm was applied to retrieve foliage and soil temperatures of a winter wheat canopy (Liu, 2002). since detailed ground measurements of component temperatures and of directional emittance were only available for this land cover type. Validation could, therefore, be done only for retrievals over winter wheat. All other land cover types were left out of consideration (light blue area in Fig. 4.20).

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

79

(TOA) ATSR data :Tb (nadir), Tb (forward), ρ (nadir),

Cloud

Retrieval of Aerosol Optical Depth

Retrieval of Water Vapor

Atmospheric correction for VIS / NIR / SWIR channels

Fractional vegetation cover ff (nadir), ff (forward)

Atmospheric correction for TIR channels

(BOA) Surface brightness temperature Tb0 (nadir), Tb0 (forward)

Spatial Smoothing

Inversion model

Component temperatures Tf , Ts

Fig. 4.19 Scheme of the operational algorithm for retrieval of Tf and Ts from ATSR multi-spectral and dual-angular measurements. Tb is the brightness surface temperature at TOA measured by TIR channels of ATSR-2; ρ is reflectance at TOA measured by the VIS/NIR/SWIR channels of ATSR2. (After Jia et al., 2003.)

4.6.2 Retrieval of Tb (θ ) at TOC (Top of Canopy) from Tb (θ ) at TOA (Top of Atmosphere) Since the satellite measures the TOA brightness temperature (T ), and the inversion of separate soil and vegetation temperature model needs the TOC brightness temperature (Tb0 ), atmospheric corrections have to be performed. Moreover, since directional ground radiance (equivalent to ground brightness temperature) is likely to result from radiometric 3-D heterogeneity of the surface or the surface cover, neither a kinetic surface temperature nor an emissivity can be simply and uniquely defined

80

M. Menenti et al.

Fig. 4.20 Foliage (left) and soil (right) component temperatures determined from AMTIS multiangular measurements of exitance at 4,200 m height; Shunyi experiment, China, April 2001. (After Liu, 2002.)

(Becker and Li, 1995) or even a value or angular behaviour assumed. Any separation between temperature and emissivity would rely on a priori assumptions or “definitions” of such variables. Thus, the first step in looking for directional effects is to consider the ground radiance as a whole. The consequence of that is that accurate atmospheric corrections have to be applied to TOA radiances (or brightness temperatures) in a given channel, preferably less affected by atmospheric perturbations, or, to say things differently, a common Split-Window method (SW) is not adequate (nor is a double angle method with (A)ATSR nadir/forward views). Indeed, a SW or similar regression algorithm for atmospheric corrections is designed to give the surface kinetic temperature, based on several assumptions regarding spectral and angular emissivity. Single channel atmospheric correction is “just” inverting the radiative transfer equation integrated over the channel (i) bandwidth: ↑ Bi (Ti ) = Bi (Tb0i )τi + Ratm i

(4.6)

in order to get the ground radiance Bi (Tb0i ), where Tb0i is the ground brightness temperature, Ti is the TOA measured brightness temperature, τi is the atmospheric ↑ is the atmospheric upward radiance. All quantities refer to transmission and Ratm i a particular view direction, defined by zenith angle θ at ground level. The accuracy on Bi (Tb0i (θ)) is determined by that on the atmospheric quantities, which in turn depends on the goodness of the radiative transfer code and the description of the vertical structure of the atmosphere. The second condition is by far the most important source of error. There is no simple parameterisation allowing for determination ↑. Atmospheric PTU profiles are needed. of both τi and Ratm i Alternatively, if the ground brightness temperature is, or can be assumed, independent of the channels used to measure it, the Split Window (SW ) (Sobrino et al., 1994; Becker and Li, 1995) method can be used to get Tb0 . Following the procedure developed by Becker and Li (1990, 1995) a general SW algorithm is derived for ATSR-2 nadir and forward views using simulated radiometric data (Li et al., 2003):

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

Tb0 (θ ) = [a(θ ) + b(θ )W ] + [c(θ ) + d(θ )W ]T11 (θ ) +[e(θ ) + f (θ )W ][T11 (θ ) − T12 (θ )]

81

(4.7)

where θ is view angle, W is the total column water vapor in the atmosphere. For the large range of surface parameters and atmospheric conditions (W ≤ 4.5 g/cm2 , air temperature at surface Ta , 272 K ≤ Ta ≤ 311 K and −5 K ≤ Tg − Ta ≤ 15 K), the coefficients a- f have been generated for ATSR-2 nadir and forward views. They are a = −4.89, b = 3.74, c = 1.0205, d = −0.0151, e = 0.916, f = 0.509 the rms residual retrieval error σ = 0.10K for the nadir image, and a = −14.41, b = 8.51, c = 1.0582, d = −0.0343, e = 0.565, f = 0.857σ = 0.24K for the forward view.

4.6.3 Retrieval of Water Vapor from ATSR-2 Split-window Channel Data over Land Water vapor content in the atmosphere is required to improve the accuracy of the remotely sensed surface parameters (Sobrino et al., 1994, Francois and Ottle, 1996). Nowadays, a number of different satellite approaches have been proposed and developed over the past two decades to measure atmospheric water vapor. According to the wavelength used, these approaches may be grouped into three categories: Near-infrared techniques (Frouin et al., 1990, Kaufman and Gao, 1992); Passive microwave techniques (Prabhakara et al., 1985, Alishouse et al., 1990, Schulz et al., 1993); and Thermal infrared techniques (Chesters et al., 1983, Susskind et al., 1984, Kleespies and McMillin, 1990, Jedlovec, 1990, Iwasaki, 1994, Ottl´e et al., 1997, Sobrino et al., 1999). Because the near-infrared technique is based on detecting the absorption by water vapor of the reflected solar radiation as it is transferred down to the surface and up through the atmosphere, use of this technique needs to have at least one channel in the water absorption band (0.94 µm), and one nearby channel in the atmospheric windows (0.86 µm, 1.05 µm and 1.24 µm). Since ATSR-2 (Along-Track Scanner Radiometer) on board ERS-2 (European Remote Sensing) has only four channels in the visible and near infrared domain (0.55 µm, 0.65 µm, 0.87 µm, 1.60 µm) and three channels in the thermal infrared domain (3.7 µm, 11 µm and 12 µm), no channel in the water absorption band is available, the near-infrared technique cannot be applied to ATSR2 data, the only one applicable technique is thermal infrared technique. Up to now, there have been several attempts to derive water vapor using two splitwindow channels (11 µm and 12 µm). For instance, Kleespies and McMillin (1990) proposed a method based on the ratio of split-window channel brightness temperature differences assuming that the atmosphere and surface emissivities in the split-window channels are invariant. Jedlovec (1990) proposed an extension of this concept and showed that the water vapor content can be derived using the ratio of the spatial variance of the channel brightness temperature. On the basis of these methods, Iwasaki (1994) developed a new algorithm to reduce the non-linear effect of air temperature and unresolved cloud effect on the estimation of water vapor

82

M. Menenti et al.

content using the split-window data. Sobrino et al. (1994) improved Jedlovec (1990) method by the use of Split-Window Covariance-Variance Ratio (SWCVR). It has been shown that all these split-window methods are sensitive to instrument noise and are difficult to be applied to satellite data such as AVHRR in an operational manner (Sobrino et al., 1994, 1999). Under the condition that the atmosphere and directional surface emissivity are constant or the effects of their spatial variations are not larger than the combined effects of both instrument noise over the N neighboring pixels, Li et al. (2003) presented a new algorithm to determine quantitatively column water vapor content (W) directly from (A)ATSR Split–Window radiance measurements using the following formulae: (a) For ATSR2 nadir view: W = 13.73 − 13.662τ j /τi

(4.8)

where the subscripts i, j denote respectively channel 11 µm and channel 12 µm of ATSR. (b) For ATSR2 forward view (θ ∼ = 53◦ ): W = 10.02 − 9.971τ j /τi with

(4.9)

N

τj εi = R ji with R ji = τi εj

∑ (Ti,k − T i )(T j,k − T j )

k=1

N

,

(4.10)

∑ (Ti,k − T i )2

k=1

in which the subscript k denotes pixel k, Ti,k and T j,k are the brightness temperatures of pixel k in channels i and j measured at satellite level, respectively, T i and T j are the mean (or the median) brightness temperatures of the N neighboring pixels considered, respectively. This method was developed and applied to several ATSR2 data sets. The water vapor contents retrieved using ATSR2 data from SGP’97 (USA), Barrax (Spain) and Cabauw (The Netherlands) are in good agreement with those measured by the quasisimultaneous radiosonde. The mean and the standard deviation of their difference are respectively 0.04 and 0.22 g/cm2 . It is shown that water vapor content derived from ATSR2 data using the proposed algorithm is accurate enough in most cases for surface temperature determination with split-window technique using ATSR2 data and for atmospheric corrections in visible and near-infrared channels of ATSR2. A more detailed description of this method and its applicability can be found in Li et al. (2003).

4.6.4 Retrieval of Aerosol Optical Depth from ATSR-2 Data for Atmospheric Corrections Atmospheric perturbations (mainly due to absorption and scattering processes) are responsible for substantial modifications of the surface spectral reflectance

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

83

measured by satellite instruments. It is therefore necessary to correct the atmospheric effects to retrieve the surface reflectance. Methods of atmospheric corrections are generally concerned with the estimation of the atmospheric effects associated with molecular absorption, molecular and aerosol scattering. Current methods for the estimation of the atmospheric effects employ a radiative transfer model (Vermote et al., 1997, Beck et al., 1999) whose inputs are generally the vertically integrated gaseous contents, aerosol optical properties and geometric conditions. If ρi∗ is the reflectance measured in channel i at the top of atmosphere (TOA), from radiative transfer theory, the surface reflectance in channel i, ρi , can be expressed as (Rahman and Dedieu, 1994; Vermote et al., 1997)

ρiac (θs , θυ , ∆φ ) 1 + Si ρiac (θs , θυ , ∆φ )

ρi (θs , θυ , ∆φ ) =

(4.11)

with

ρi∗ (θs , θυ , ∆φ ) ρiac =

tgi (θs , θυ )−ρi (θs ,θυ ,∆φ ) a

ti (θs )ti (θυ )

where θs and θυ are solar and viewing zenith angles, respectively. ∆φ is the relative azimuth between sun and satellite direction. Si is the spherical albedo of atmosphere in channel i. tgi is the total gaseous transmission in channel i associated with gaseous absorption along the sun–target–sensor atmospheric path. ρia (θs , θυ , ∆φ ) is the atmospheric reflectance. ti (θs ) and ti (θυ ) are the total atmospheric scattering transmittance along the sun-target and target-sensor atmospheric paths, respectively. In general, the independent measurements of atmospheric composition and aerosol optical properties are not available; it is therefore desirable to derive them directly from satellite data. The most important gases in atmospheric corrections in visible and near infrared channels are water vapor and ozone. Water vapor content in the atmosphere may be derived from the two split-window channel measurements as shown above, and ozone content is taken from climatological data. As for the determination of the aerosol optical properties, if the surface reflectance may be considered isotropic, then the difference in surface reflectance retrieved from multi-angle directions using Eq. (4.1) may be used to derive the atmospheric optical thickness if aerosol type is assumed. However, most land surfaces are far from Lambertian (Hapke, 1981) With multi-angle measurements, it is imperative to consider non-Lambertian reflectances. Several multi-look aerosol retrieval schemes for ATSR-2 have been proposed (Flowerdew and Haigh, 1997; Mackay and Steven, 1998; North et al., 1999). The iteration of a two step-process proposed by North et al. (1999) can be used. The first step is to derive using Eq. (4.1) eight land surface reflectances ρi (θs , θv , ∆φ ) from the TOA reflectance ρ ∗ made at four channels (0.55 µm, 0.65 µm, 0.87 µm, 1.60 µm) and two view angles (nadir and forward views), an initial estimate of the atmospheric aerosol and optical depth agiven . The second step is to fit land surface bi-directional reflectance at 550 nm τ550

84

M. Menenti et al.

model to eight retrieved surface reflectances by the minimization of the error metric function 2

E=

4

∑ ∑ [ρi (θs , θv , ∆φ ) − ρim (θs , θv , ∆φ )]2

(4.12)

θv =1 i=1

where ρim (θs , θv , ∆φ ) is the land reflectance in channel i predicted by the reflectance model. Since there are maximum eight land surface reflectance measurements, land surface bi-directional model must have maximum seven free model parameters so that there is at least 1◦ of freedom available a for atmospheric parameter retrieval, . Considering the land surface to be for instance, the aerosol optical depth τ550 composed of opaque facets, each with Lambertian reflectance ωi , and separating parameters relating to the wavelength invariant three-dimensional structure of the surface from wavelength dependent parameters describing the component spectra, North et al. (1999) developed a seven free parameter (P nadir, P forward, γ and ωi (i = 1, 4)) model as

ρi (θs , θv , ∆φ ) = (1 − Di )P(θs , θv , ∆φ )ωi +

γωi [Di + (1 − γ )ωi (1 − Di )] 1 − (1 − γ )ωi (4.13)

where Di is the incident diffuse fraction which excludes the radiation scattered close to the solar beam direction. Di can be estimated by radiative transfer model for solar direction and aerosol optical depth. P(θs , θv , ∆φ ) is the geometric parameter dependent only on view and illumination directions, γ denotes the mean hemispherically integrated probability of escape of light without further interaction, after a scattering event at the land surface. In case where there are no four channels available, an alternative scheme can be used to retrieve the aerosol optical depth by assuming that the functional shape of the bidirectional effects is invariant with respect to the wavelength within the visible and near-infrared region (Flowerdew and Haigh, 1997; Mackay and Steven, 1998), namely: ρi (θs , θυ 1 , ∆φ ) ρ j (θs , θυ 1 , ∆φ ) = (4.14) ρi (θs , θυ 2 , ∆φ ) ρ j (θs , θυ 2 , ∆φ ) This relationship gives a constraint for atmospheric correction by forcing the retrieved bidirectional reflectance to have a consistent angular variation, even thought the magnitude of the reflectance may vary greatly. The aerosol optical thickness is therefore obtained through the minimization of the error metric function n n ρi (θs , θυ 1 , ∆φ ) ρ j (θs , θυ 1 , ∆φ ) 2 − (4.15) E = ∑∑ ρ j (θs , θυ 2 , ∆φ ) i=1 j>i ρi (θs , θυ 2 , ∆φ ) where n is the total number of channels available, i and j are channel numbers. The details of these methods can be found in North et al. (1999).

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

85

4.6.5 Retrieval of Vegetation Fractional Cover To retrieve soil and foliage temperatures from Eqs. (4.2) or (4.4), the viewing angledependent vegetation fraction needs to be determined independently from additional observations, such as the TOC reflectances in the VNIR and SWIR spectral channels of ATSR-2 and AATSR. The approach presented here is an example, while for a review on retrieval of fractional cover see, e.g., Weiss et al. (1999). The approach to estimate the fractional vegetation cover uses visible and near infrared data. A stepwise multiple linear regression were applied in this study to estimate the fractional vegetation cover ff (θv ) using TOC reflectances ρi (θs , θv , ∆ϕ ). The stepwise multiple linear regression is written n

ff (θv ) = a0 (θs , θv ) + ∑ ai (θs , θv )ρi (θs , θv , ∆ϕ )

(4.16)

i=1

where n is the number of channels used. The model OSCAR (Verhoef, 1998) has been applied to generate surface reflectances for the viewing and illumination conditions applying to the ATSRobservations used in this study and to an ensemble of canopy and atmospheric conditions. This synthetic data base, where f f is known for any given canopy condition, has been used to determine the coefficients ai in Eq. 4.16. A different set of coefficients is obtained for each viewing geometry (see also Jia, 2004).

4.7 Conclusions Vegetation fractional cover. The regression model Eq. (4.16) works rather well when vegetation is green, i.e., corn and alfalfa, while slightly larger errors appear for vegetation with some fraction of brown leaves, like barley at this time of the year (DOY = 179). Table 4.3 gives the summary of the model performance represented by RMSE and Absolute Difference (AD) for each case and for the total data set.

Table 4.3 The performance of the regression equation for estimating fractional vegetation cover Eq. (4.15) as presented by RMSE and AD at nadir and forward 53◦ for each vegetation type. “Total” indicates the overall performance over the entire simulation dataset Vegetation

Corn Barely Alfalfa Total

θv = 0◦

θv = 53◦

RMSE

AD

RMSE

AD

0.0339 0.1086 0.0327 0.0456

0.025 0.109 0.030 0.046

0.1038 0.0745 0.0515 0.0785

0.103 0.073 0.049 0.076

86

M. Menenti et al.

Table 4.4 RMSE between retrieval and simulation of Tf and Ts for corn, barley and alfalfa crops

RMSE of Tf RMSE of Ts

Corn

Barley

Alfalfa

0.381905 0.669429

1.164064 0.867213

0.590805 0.772624

Retrieval of soil and foliage component temperatures. The validation of the retrieved component temperatures using ATSR-2 bi-angular measurements is challenging due to the difficulty of obtaining observations of temperatures of soil and foliage in situ at ATSR-2 spatial resolution (i.e., 1.5 × 2.5 km for the forward view. As done to evaluate the algorithm for ff (θv ), the same method, i.e., using synthetic radiometric data generated with detailed modeling of radiative transfer in the soil–vegetation–atmosphere system and inversion using a simplified algorithm (suitable for operational processing of actual data), can be used to evaluate the retrievals of Tf . The reference Tf and Ts were simulated using the complete model CUPID and the TOC radiance was calculated (for each image pixel) with the procedure outlined in Section 4.5. The Tb0 (θv , ϕv ) at θv = 0◦ and θv = 53◦ obtained in this way were then treated as observations in Eq. 4.4 to retrieve Tf and Ts . The comparisons between the retrievals and the simulations of Tf and Ts for the three crops are encouraging (Table 4.4).

4.8 Conclusions: Limitations of Current Systems and Perspectives Over the last 25 years measurements of emittance by goniometer-mounted radiometers has been the main source of observations to document and understand the anisotropic emittance of vegetation canopies. This experimental body of knowledge led to the development of detailed models of the complex processes which determine the significant thermal heterogeneity of the canopy space. The latter is of particular relevance towards a better understanding of the relation between canopy architecture and foliage–atmosphere exchanges of radiation, water and carbon. On the other hand, the capability of goniometer-mounted radiometers is severely limited both as regards the size of targets and sampling of the spatial variability. Actual exploitation of the unique information conveyed by anisotropic emittance requires imaging radiometers to sample sufficiently large terrestrial targets at a spatial resolution sufficiently large to integrate emitted radiance over an area representative of the target, but sufficiently small to provide independent observations of different land cover types. The ATSR series of instruments is the only current observing system providing directional (at two view angles) observations of emittance and these observations have been used to demonstrate the relevance of directional measurements in the

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

87

TIR region. (see, e.g., Jia et al., 2001; Menenti et al., 2001; Hurk et al., 2002). On the other hand, the large footprint of (A)ATSR and even more the large difference in footprint between the nadir and forward view, make validation of data products very challenging and restrict applications to large scale modeling. For example, the difference between foliage and soil temperature might be used as a measure of drought or crop water stress, but given the low resolution is difficult to relate to the spatial scale of crop and water management practices. The European Space Agency had been studying a high spatial resolution mission for multi-angular land observations (SPECTRA, see Menenti et al., 2005), but this project has been abandoned. The Italian (ASI) and Canadian (CSA) Space Agencies are developing an Earth Observation mission with similar characteristics (Galeazzi et al.,2006) with expected launch in 2010. The latter would provide simultaneous VNIR – SWIR hyper-spectral and multi-spectral observations at high spatial resolution and is being designed specifically for land observations. In general terms, multi-angular measurements of TIR emittance over land provide access to radiative and convective processes in the canopy space which concur to determine the response of terrestrial vegetation to environmental forcing. Acknowledgements Work by the Authors summarized in this chapter has been funded by the European Space Agency (Contracts 13177/98NL/GD; 15164/01/NL/SF; 17169/03/NL/GS; 17179/03/NL/GS), the European Commission (Contracts FP6 GMES no. 502057, FP5 MIND EVK2-CT-2002-00158), the Netherlands Users Support Program (GO-2 Contract SRON EO049.) and the Italian Space Agency (ASI Contract ASI I/R/073/01). The Shunyi field campaign was part of the “Quantitative of Remote Sensing theory and application for Land Surface Parameters (QRSLSP)” project funded by China’s Special Funds for Major State Basic Research (project No. G2000077900, led by Prof. X. Li). Dr. Qiang Liu is thanked for the data processing of goniometer and AMTIS measurements.

References Albertson J, Katul G, Wiberg P (2001) Relative importance of local and regional controls on coupled water, carbon, and energy fluxes. Adv. Water Resour. 24:1103–1118 Alishouse JC, Snyder SA, Vongsathorn J, Ferraro RR (1990) Determination of oceanic total precipitable water from the SSM/I. IEEE Trans. Geosci. Remote Sens. 28:811–816 Avissar R, Pielke RA (1989) A parameterization of heterogeneous land-surface for atmospheric numerical models and its impact on regional meteorology. Month. Weather Rev. 117:2113– 2136 Balick LK, Hutchison BA, Smith JA, McGuire MJ (1987) Directional thermal exitance distributions of a deciduous forest in summer. IEEE Trans. Geosci. Remote Sens. 25:410–412 Baret F, Clevers JPW, Stevens MD (1995) The robustness of canopy gap fraction estimates from the red and near infrared reflectance: a comparison between approaches for sugar beet canopies. Remote Sens. Environ. 54:141–151 Beck A, Anderson GP, Acharya, PK, et al. (1999) MODTRAN4 User’s Manual, Air Force Research Laboratory, Hanscom AFB, MA Becker F, Li Z-L (1990) Towards a local split window method over land surfaces. Int. J. Remote Sens. 11(3):369–393

88

M. Menenti et al.

Becker F, Li Z-L (1995) Surface temperature and emissivity at various scales: definition, measurement and related problems. Remote Sens. Rev. 12:225–253 Becker F (1987) The impact of spectral emissivity on the measurement of land surface temperature from a satellite. Int. J. Remote Sens. 8:1509–1522 Becker F, Ramanantsizehena P, Stoll M-P (1985) Angular variation of the bidirectional reflectance of bare soils in the thermal infrared band. Appl. Opt. 24(3):363–375 Beljaars ACM, Holtslag AAM (1991) Flux parameterization over land surface for atmospheric models. J. Appl. Meteorol. 30:327–341 Blad BL, Rosenberg NJ (1976) Evaluation of resistance and mass transport evapotranspiration models requiring canopy temperature data. Agron. J. 68:764–769 Borel CC, Gerstl SA, Powers BJ (1991) The radiosity method in optical remote sensing of structured 3-D surfaces. Remote Sens. Environ. 36:13–44 Bowen IS (1926) The ratio of heat losses by conduction and by evaporation from any water surface. Phys. Rev. 27:779–787 Caselles V, Sobrino JA, Coll C (1992) A physical model for interpreting the land surface temperature obtained by remote sensors over incomplete canopies. Remote Sens. Environ. 39:203–211 CESBIO, 2006, DART User Manual v1.1. Chen JM, Leblanc S (1997) A 4-scale bidirectional reflection model based on canopy architecture. IEEE Trans. Geosci. Remote Sens. 35:1316–1337 Chesters D, Uccellini LW, Robinson P (1983) Low-level water vapor fields from the VISSR atmospheric sounder (VAS) split-window channels. J. Appl. Meteorol. 22:725–743 Choudhury BJ, Reginato RJ, Idso SB (1986) An analysis of infrared temperature observations over wheat and calculation of the latent heat flux. Agric. Forest Meteorol. 37:75–88 Colton AL (1996) Effective thermal parameters for a heterogeneous land surface. Remote Sens. Environ. 57:143–160 De Bruin HAR, Jacobs CMJ (1989) Forest and regional scale processes. Phil. Trans. R. Met Soc. B (324):393–406 Feddes RA (1971) Water, heat and crop growth. Ph.D. thesis, Wageningen, Agricultural University, The Netherlands. H. Veenman, Zonen, Wageningen, The Netherlands, 184pp Field CB, Mooney HA (1986) The photosynthesis – nitrogen relationship in wild plants. In: TJ Givnish (ed) On the economy of plant form and function. Cambridge University Press, New York, pp 25–55 Flowerdew RJ, Haigh JD (1997) Retrieving land surface reflectances using the ATSR-2: a theoretical study. J. Geophys. Res. 102(D14):17163–17171 Francois C (2002) The potential of directional radiometric temperatures for monitoring soil and leaf temperature and soil moisture status. Remote Sens. Environ. 80:122–133 Franc¸ois C, Ottl`e C (1994) Estimation of the angular variation of the sea surface emissivity with the ATSR/ERS-1 data, Remote Sensing of Environment, 48:302–308 Francois C, Ottle C (1996) Atmospheric corrections in the thermal infrared: global and water vapor dependent split-window algorithms-applications to ATSR and AVHRR data. IEEE Trans. Geosci. Remote Sens. 34(2):457–470 Francois C, Ottle C, Prevot L (1997) Analytical parameterization of canopy directional emissivity and directional radiance in thermal infrared. Application on the retrieval of soil and foliage temperatures using two directional measurements. Int. J. Remote Sens. 18(12):2587–2612 Frouin R, Deschamps P-Y, Lecomte P (1990) Determination from space of atmospheric total water vapor amounts by differential absorption near 940 nm: theory and airborne verification. J. Appl. Meteorol. 29:448–460 Galeazzi C, Gamache M, Hollinger A, De Cosmo V, Menenti M, Staenz K, Beck J, Bergeron M, Ananasso C, Loizzo R, Maszkiewicz M (2006) ASI-CSA joint development phase for a hyperspectral mission. In: I. Reusen and J. Cools (eds.) Imaging Spectroscopy: Innovation in Environmental Research. EARSEL, Paris:1–11 Gastellu-Etchegorry JP, Martin E, Gascon F (2004) DART: a 3D model for simulating satellite images and studying surface radiation budget. Int. J. Remote Sens. 25(2):73–96

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

89

Gerstl SAW, Borel CC (1992) Principles of the radiosity method versus radiative transfer for canopy reflectance modelling. IEEE Trans. Geosci. Remote Sens. 30(2):271–275 Gerstl SAW, Borel CC, Powers BJ (1991) The radiosity method in optical remote sensing of structured 3D surfaces. Remote Sens. Environ. 36:13–44 Guillevic P, Gastellu-Etchegorry JP, Demarty J, Pr´evot L (2002)Thermal infrared radiative transfer within three-dimensional vegetation. J. Geophys. Res. (Atmos.) 108(D8):ACL 6-1, CiteID 4248, doi:10.1029/2002JD002247 Hall AE, Canell GH, Lawton HW (1979) Agriculture in semi-arid environments. Springer, Berlin– Heidelberg, New York, 340pp Hapke B (1981) Bidirectional reflectance spectroscopy. 1. Theory. J. Geophys. Res. 86:3039–3054 Hatfield JL, Reginato RJ, Idso SB (1984) Evaluation of canopy temperature – evapotranspiration models over various crops. Agric. Meteorol. 32:41–53 Hatfield JL (1991) Development of agricultural crops and practices adapted to arid land conditions, 1991. Proc. SCS Multiconference on AI simulation: Agricultural Crops and Practices Adapted to Arid Lands. University of New Mexico Press, Albuquerque, pp 39–60 Huband NDS, Monteith JL (1986) Radiative surface temperature and energy balance of a wheat canopy. Bound.-Lay. Meteorol. 36:1–17 Hurk BJJM, Jia L, Jacobs C, Menenti M, Li Z-L (2002) Assimilation of land surface temperature data from ATSR in an NWP environment – a case study. Int. J. Remote Sens. 23(24):5193–5210 Iwasaki H (1994) Estimation of precipitable water over land using the split-window data from the NOAA satellite. J. Meteorol. Soc. Jpn. 72(2):223–233 Jedlovec GJ (1990) Precipitable water estimation from high-resolution split window radiance measurements. J. Appl. Meteorol. 29:863–876 Jia L, Menenti M, Su ZB, Djepa V, Li Z-L, Wang J (2000) Modeling sensible heat flux using estimates of soil and vegetation temperatures: the HEIFE and IMGRASS experiments. In: M Beniston, M Verstraete (eds), Satellite remote sensing and climate simulations. Kluwer, Dordrecht, The Netherlands Jia L, Menenti M, Su ZB, Djepa V, Li Z-L, Wang J (2001) Modeling sensible heat flux using estimates of soil and foliage temperatures: the HEIFE and IMGRASS experiments, in “Remote sensing and climate modeling: Synergies and Limitations”, M Beniston and M Verstraete (eds) Kluwer, Dordrecht, 23–49 Jia L, Li Z-L, Menenti M, Su Z-B, Verhoef W, Wan Z-M (2003) A practical algorithm to infer soil and foliage component temperatures from bi-angular ATSR-2 data. Int. J. Remote Sens. 24(23):4739–4760 Jia L (2004) Modeling heat exchanges at the land-atmosphere interface using multi-angular thermal infrared measurements. Wageningen University, The Netherlands, ISBN 90-8504-041-8, 199pp Jia L, Li Z-L, Menenti M (2005) Simulation of directional thermal infrared radiance images using soil-vegetation-atmosphere heat and radiation transfer models. Proc. 9th Int. Symp. on Physical Measurements and Signatures in Remote Sens. (ISPMSRS), ISSN 1682–1750 ISPRS, XXXVI (7/W20):331–333 Jackson RD, Idso SB (1975) Surface albedo and desertification. Science 189:1012–1013 Jackson RD, Idso SB, Reginato RJ, Pinter PJ (1981) Canopy temperature as a crop water stress indicator. Water Resour. Res. 17(4):1133–1138 Jackson RD (1985) Evaluating evapotranspiration at local and regional scales. Proc. IEEE 73:1086–1095 Kaufman YJ, Gao B-C (1992) Remote sensing of water vapor in the near IR from EOS/MODIS. IEEE Trans. Geosci. Remote Sens. 30(5):871–884 Kimes DS (1980) Effects of vegetation canopy structure on remotely sensed canopy temperatures. Remote Sens. Environ. 10:165–174 Kimes DS (1981) Remote sensing of temperature profiles in vegetation canopies using multiple view angles and inversion techniques. IEEE Trans. Geosci. Remote Sens. 19(2):85–90 Kimes DS (1983) Remote sensing of row crop structure and component temperatures using directional radiometric temperatures and inversion techniques. Remote Sens. Environ. 13:33–55

90

M. Menenti et al.

Kimes DS, Smith JA, Link LE (1981) Thermal IR exitance model of a plant canopy. Appl. Opt. 20(4):623–632 Kimes DS, Kirchner JA (1982) Radiative transfer model for heterogeneous 3-D scenes. Appl. Opt. 21(22):4119–4129 Kimes DS, Kirchner JA (1983) Directional radiometric measurements of row-crop temperatures. Int. J. Remote Sens. 4(2):299–311 Kleespies TJ, McMillin LM (1990) Retrieval of precipitable water from observations in the split window over varying surface temperature. J. Appl. Meteorol. 29:851–862 Labed J, Stoll MP (1991) Spatial variability of land surface emissivity in the thermal infrared band: spectral signature and effective surface temperature. Remote Sens. Environ. 38:1–17 Lagouarde JP, Kerr YH, Brunet Y (1995) An experimental study of angular effects on surface temperature for various plant canopies and bare soils. Agric. Forest Meteorol. 77:167–190 Lei SA (2004) Leaf temperatures of Blackbrush: relationships with air and soil surface temperatures. USDA Forest Service Proc. RMRS-P-31 Li Z-L, Jia L, Su Z, Wan Z, Zhang R-H (2003) A new approach for retrieving precipitable water vapour from ATSR2 split-window channel data over land area. Int. J. Remote Sens. 24:5095– 5117 Li X-W, Strahler AH (1985) Geometric-optical modeling of a conifer forest canopy. IEEE Trans. Geosci. Remote Sens. GE-23(5):705–721 Li Z-L, Stoll MP, Zhang RH, Jia L, Su Z (2001) On the separate retrieval of soil and vegetation temperatures from ATSR2 data. Science in China, Series D 44(2):97–111 Li Z-L, Zhang R-H, Sun X-M, Su H-B, Tang X-Z, Zhu Z-L, Sobrino JA (2004) Experiment system for the study of the directional thermal emission of natural surfaces. Int. J. Remote Sens. 25(1):195–204 Liu Q (2002) Study on Component Temperature Inversion Algorithm and the Scale Structure for Remote Sensing Pixel. Ph.D. dissertation of Chinese Academy of Sciences, 111p Liu QH, Li XW, Chen LF (2002). Field campaign for quantitative remote sensing. In: Beijing. Proc. IEEE 2002 Int. Geosci. Remote Sens. Symp. (IGARSS’02) vol VI, pp 3133–3135 Lhomme J-P, Monteny B, Chehbouni A, Troufleau D (1994) Determination of sensible heat flux over Sahelian fallow savannah using infra-red thermometry. Agric. Forest Meteorol. 68:93–105 Luquet D, B´egu´e A, Vidal A, Clouvel A, Dauzat J, Olioso A, Gu X-F, Tao Y (2003) Using multidirectional thermography to characterize water status of cotton. Remote Sens. Environ. 84:411– 421 Mackay G, Steven MD (1998) An atmospheric correction procedure for the ATSR-2 visible and near infrared land surface data. Int. J. Remote Sens. 19(15):2949–2968 Mahfouf JF, Richard E, Mascart P (1987) The influence of soil and vegetation on the development of mesoscale circulations. J. Clim. Appl. Meteor. 26:1483–1495 Masuda K, Takashima T, Takayama Y (1988) Emissivity of pure and sea waters for the model sea surface in the infrared window regions. Remote Sens. Environ. 313–329 Matteucci G, Dore S, Stivanello S, Rebmann C, Buchmann N (2000) Soil respiration in beech and spruce forests in Europe: trends, controlling factors, annual budgets and implications for the ecosystem carbon balance. In: ED Schulze (ed), Carbon and Nitrogen Cycling in European Forest Ecosystems, vol 142. Springer, Berlin, pp 217–236 McGuire MJ, Smith JA, Balick LK, Hutchison BA (1989) Modeling directional thermal radiance from a forest canopy. Remote Sens. Environ. 27:169–186 Medlyn BE, Dreyer E, Ellsworth DS, Forstreuter M, Harley PC, Kirshbaum MUF, Le Roux X, Montpied P, Strassemeyer J, Walcroft A, Wang K, Loustau D (2002) Temperature response of parameters of a biochemically-based model of photosynthesis. II. A review of experimental data. Plant, Cell Environ. 25:1167–1179 Menenti M, Choudhury BJ (1993) Parameterization of land surface evapotranspiration using a location-dependent potential evapotranspiration and surface temperature range. In: HJ Bolle, et al. (eds), Exchange processes at the land surface for a range of space and time scales, IAHS, Oxfordshire, vol 212, pp 561–568

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

91

Menenti M, Jia L, Li Z-L, Djepa V, Wang J, Stoll MP, Su ZB, Rast M (2001) Estimation of soil and vegetation temperatures from directional thermal infrared observations: the HEIHE, SGP’97, IMGRASS experiments. J. Geophys. Res. 106:11997–12010 Menenti M, Rast M, Baret F, Hurk B, Knorr W, Mauser W, Miller J, Moreno J, Schaepman M, Verstraete M (2005) Understanding vegetation response to climate variability from space: recent advances towards the SPECTRA Mission. Riv. Ital. Teleril. Special Issue Hyperspectral Observation of Terrestrial Environments. 33/34:193–206 Monteith JL (1965) Evaporation and environment. In: GE Fogg (ed), The state and movement of water in living organisms, Sympos. Soc. Exp. Biol. Academic, New York, 19:205–234 Moran MS, Clarke TR, Inoue Y, Vidal A (1994) Estimating crop water deficit using the relation between surface air temperature and spectral vegetation index. Remote Sens. Environ. 49(3):246 263 Mutlow CT, Z´avody AM, Barton IJ, Llewellyn-Jones DT (1994) Sea surface temperature measurements by the along-track scanning radiometer on the ERS 1 satellite: early results. J. Geophys. Res. 99(C11):22575–22588 Mutlow C, Murray J, Bailey P, Birks A, Smith D (1999) ATSR-1/2 User Guide, http://www.atsr.rl. ac.uk/documentation/docs/userguide/atsr user guide rev 3.pdf Nerry F, Labed J, Stoll M-P (1988) Emissivity signatures in the thermal IR band for remote sensing: calibration procedure and method of measurement. Appl. Opt. 27(4):758–764 Nielsen DC, Clawson KL, Blad BL (1984) Effect of solar azimuth and infrared thermometer view direction on measured soybean canopy temperature. Agron. J. 76:607–610 Nilson T (1971) A theoretical analysis of the frequency of gaps in plant stands. Agric. Meteorol. 8:25–38 Norman JM (1979) Modeling the complete crop canopy. In: BJ Barfield, JF Gerber (eds) Modification of the Aerial Environment of Plants. ASAE Monogr. Am. Soc. Agric. Engr., St. Joseph, MI, pp 249–277 Norman JM, Campbell G (1983) Application of a plant-environment model to problems in irrigation. In: D Hillel (ed) Advanced in irrigation, vol 2. Academic, New York, pp 155–188 Norman JM, Chen JL (1990) Thermal emissivity and infrared temperatures dependence on plant canopy architecture and view angle. In: Proc. Int. Geosci. Remote Sens. Symp. (IGARSS’90, IEEE Geosci. Remote Sens. Soc., New York), pp 1747–1750 North PRJ, Briggs SA, Plummer SE, Settle JJ (1999) Retrieval of land surface bi-directional reflectance and aerosol opacity from ATSR-2 multiangle imagery. IEEE Trans. Geosci. Remote Sens. 37(1):526–537 Otterman J (1990) Inferring parameters for canopies non uniform in azimuth by model inversion. Remote Sens. Environ. 33:41–53 Otterman J, Brakke TW, Susskind J (1992) A model for inferring canopy and underlying soil temperatures from multi-directional measurements. Bound.-Lay. Meteorol. 61:81–97 Otterman J, Susskind J, Brakke T, Kimes D, Pielke R, Lee TJ (1995) Inferring the thermal-infrared hemispheric emission from a sparsely-vegetation surface by directional measurement. Bound.Lay. Meteorol. 74:163–180 Ottermann J, Brakke TW, Fuchs M, Lakshmi V, Cadeddu M (1999) Longwave emission from a plant/soil surface as a function of the view direction: dependence on the canopy architecture. Int. J. Remote Sens. 20(11):2195–2201 Ottl´e C, Outalha S, Francois C, Le Maguer S (1997) Estimation of total atmospheric water vapor content from split-window radiance measurements. Remote Sens. Environ. 61:410–418 Paw U KT, Ustin SL, Zhang CA (1989) Anisotropy of thermal infrared exitance in sunflower canopies. Agric. Forest Meteorol. 48:45–58 Peng XD, Cheng LS (1993) Numerical simulations of the PBL mean structure and fluxed in HEIHE region. Proc. Int. Symp. HEIFE. Kyoto University, Kyoto, Japan, pp 437–442 Petitcolin F, Nerry F, Stoll MP (2002) Mapping temperature independent spectral indice of emissivity and directional emissivity in AVHRR channels 4 and 5. Int. J. Remote Sens. 23(17):3473– 3491

92

M. Menenti et al.

Pielke RA, Dalu G, Snook JS, et al. (1991) Nonlinear influence of mesoscale land use on weather and climate. J. Climate 4:1053–1069 Prabhakara C, Chang HD, Chang ATC (1985) Remote sensing of precipitable water over the oceans from NIMBUS 7 microwave measurements. J. Appl. Meteorol. 21:59–68 Price JC (1982) On the use of satellite data to infer surface fluxes at meteorological scales. J. Appl. Meteorol. 21:1111–1122 Qualls RJ, Yates DN (2001) Directional radiometric temperature profiles within a grass canopy. Adv. Water Resour. 24:145–155 Rahman H, Dedieu G (1994) SMAC: a simplified method for the atmospheric correction of satellite measurements in the solar spectrum. Int. J. Remote Sens. 15:123–143 Schulz J, Schluessel P, Grassl H (1993) Water vapor in the atmospheric boundary layer over oceans from SSM/I measurements. Int. J. Remote Sens. 14:2773–2789 Segal R, Avissar M, McCumber C, Pielke RA (1988) Evaluations of vegetation effects on the generation and modification of mesoscale circulations. J. Atmos. Sci. 45:2268–2292 Seguin B, Assad E, Freteaud JP, Imbernon J, Kerr YH, Lagouarde JP (1989) Use of meteorological satellite for water balance monitoring in Sahelian regions. Int. J. Remote Sens. 10:1001–1017 Sellers PJ, Heiser MD, Hall FG, Goetz SJ, Strebel DE, Verma SB, Desjardins RL, Schuepp PM, McPherson JI (1995) Effects of spatial variability in topography, vegetation cover and soil moisture on area-averaged surface fluxes: A case study using the FIFE 1989 data. J. Geophys. Res. Vol. 100 (D12): 25607–25630 Sellers PJ, Randall DA, Collatz GJ, Berry JA, Field CB, Dazlich DA, Zhang C, Colleli GD, Bounoua L (1996) A revised land surface parameterization (SIB2) for atmospheric GCMS. Part 1: model formulation. J. Climate 9(4):676–705 Smith JA, Goltz SM (1995) A thermal exitance and energy balance model for forest canopies. IEEE Trans. Geosci. Remote Sens. 32(5):1060–1066 Smith JA, Chauchan NS, Ballard J (1996) Remote sensing of land surface temperature: the directional viewing effect. IEEE Trans. Geosci. Remote Sens. 13(4):2146–2148 Snyder WC, Wan Z-M, Zhang Y, Feng Y-Z (1997) Thermal Infrared (3–14 µ m) bidirectional reflectance measurements of sands and soils. Remote Sens. Environ. 60(1):101–109 Sobrino JA, Caselles V (1990) Thermal infrared model for interpreting the directional radiometric temperature of a vegetative surface. Remote Sens. Environ. 33:193–199 Sobrino J, Li Z-L, Stoll MP, and Becker F, 1994. Improvements in the split-window technique for land surface temperature determination. IEEE Trans. Geosc. and Rem. Sens. Vol. 32(2): 243–253 Sobrino J, Li Z-L, Stoll MP, Becker F (1994) Improvements in the split-window technique for land surface temperature determination. IEEE Trans. Geosci. Remote Sens. 32(2):243–253 Sobrino JA, Raissouni N, Simarro J, Nerry F, Petitcolin F (1999) Atmospheric water vapor content over land surfaces derived from AVHRR data: application to the Iberian Peninsula. IEEE Trans. Geosci. Remote Sens. 37(3):1425–1434 Sobrino JC, Cuenca J (1999) Angular variation of thermal infrared emissivity for some natural surfaces from experimental measurements. Appl. Opt. 38(18):3931–3936 Su H-B, Zhang R-H, Tang X-Z, Sun X-M, Zhu Z-L, Li Z-L (2003) An alternative method to compute the component fractions in the geometrical optical model: visual computing method. IEEE Trans. Geosci. Remote Sens. 41(3):719–724 Susskind J, Rosenfield J, Renter D, et al. (1984) Remote sensing of weather and climate parameters from HIRS2/MSU on TIROS-N. J. Geophys. Res. 89:4677–4697 Sutherland RA, Bartholic JF (1977) Significance of vegetation in interpreting thermal radiation from a terrestrial surface. J. Appl. Meteorol. 18:759–763 Valentini R, Dolman AJ, Ciais P, Schulze ED, Freibauer A, Schimel DS, Heimann M (2000) Accounting for carbon sinks in the biosphere, European perspective. CARBOEUROPE European Office. Max-Planck Institute for Biogeochemistry, P.O. Box 100164. 07700 Jena, Germany, 17p Verhoef W, Menenti M (1998) Final Report of Spatial and Spectral Scales of Spaceborne Imaging Spectro-radiometers (SASSSIS). Contract Report 12072/96/NL/CN, ESA. Final Report, NLRCR-98213, National Aerospace Lab. NLR, Amsterdam, The Netherlands, 314pp

4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation

93

Verhoef W (1998) Theory of radiative transfer models applied in optical remote sensing of vegetation canopies. Ph.D. thesis, Wageningen University, The Netherlands, 311pp Verhoef W, Bach H (2003) Simulation of hyperspectral and directional radiance images using coupled biophysical and atmospheric radiative transfer models. Remote Sens. Environ. 87:23–41 Verma SB, Rosenberg NJ, Blad BL, Baradas NW (1976) Resistance-energy balance method for predicting evapotranspiration, determination of boundary layer resistance, and evaluation of error effects. Agron. J. 68:776–782 Vermote EF, Tanre D, Deuze JL, Herman M, Morcrette JJ (1997) Second simulation of the satellite signal in the solar spectrum, 6S: an overview. IEEE Trans. Geosci. Remote Sens. 35:675–686 Wang J, Mitsuta Y (1992) An observation study of turbulent study of turbulent structure and transfer characteristics in Heihe oasis. J. Meteor. Soc. Jpn. 70:1147–1154 Wang J, Yan F, Xiao J, Wei T, Wang J (2002) Development of an airborne multi-angle TIR/VNIR imaging system. IEEE 2002 Int. Geosci. Remote Sens. Symp. (IGARSS’02), vol IV, pp 2245– 2248 Weiss M, Baret F, Myneni R, Pragn`ere A, Knyazikhin Y (1999) Investigation of a model inversion technique for the estimation of crop characteristics from spectral and directional reflectance data. Agronomie 20:3–22 Welles JM, Norman JM (1991) Photon transport in discontinuous canopies: a weighted random approach. In: RB Myneni, J Ross (eds), Photon-vegetation interactions. applications in optical remote sensing and plant ecology. Springer, Heidelberg, Germany, pp 389–414 Wilson K, Goldstein A, Falge E, Aubinet M, Baldocchi D, Berbigier P, Bernhofer C, Ceulemans R, Dolman H, Field C, Grelle A, Ibrom A, Law BE, Kowalski A, Meyers T, Moncrieff J, Monson R, Oechel W, Tenhunen J, Valentini R,Verma S (2002) Energy balance closure at FLUXNET sites. Agric. Forest Meteorol. 113(1–4):223–243 Wilson TB, Norman JM, Bland WL, Kucharik CJ (2003) Evaluation of the importance of Lagrangian canopy turbulence formulations in a soil–plant–atmosphere model. Agric. Forest Meteorol. 115:51–69 Xu X, Chen L, Zhuang J (2000) The passive measurements of object’s directional emissivity in laboratory. Science in China 43:55–61 Xue Y, Shukla J (1993) The influence of land surface properties on Sahel climate. Part I: desertification. J. Climate 6:2232–2245 Yan Y (1999) Numerical simulation of land surface process on heterogeneous surface. Ph.D. dissertation. Lanzhou Institute of Plateau Atmospheric Physics. Chinese Academy of Sciences, Lanzhou, China, 90pp Yu T, Gu X-F, Tian G-L, Legrand M, Baret F, Hanocq J-F, Bosseno R, Zhang Y (2004) Modeling directional brightness temperature over a maize canopy in row structure. IEEE Trans. Geosci. Remote Sens. 42(10):2290–2304 Zhang R-H, Sun X-M, Su H-B, Zhu Z-L, Tang X-Z (2002) A new design for measuring directional radiant temperature and data analysis. Proc. IEEE 2002 Int. Geosci. Remote Sens. Symp. (IGARSS’02), vol V, pp 2768–2770

Chapter 5

Terrestrial Applications of Multiangle Remote Sensing Mark J. Chopping

Abstract Multiangle remote sensing is an emerging technology that enables important applications of terrestrial (land) remote sensing, in ecology and land cover mapping as well as in a variety of disciplines in the Earth Sciences. Advances have been realized in three major areas: the measurement or characterization of canopy structure and surface roughness; the separation of the contributions of the upper canopy and the background in forest and shrub-dominated environments; and improvements in the accuracy of classifications of land cover in environments as dissimilar as deserts and ice sheets. The focus of the chapter is on land surface applications of solar wavelength multiangular data acquired from the air and space; it avoids discussion of methods for retrieval of essentially radiometric quantities such as shortwave albedo, bidirectional and hemispheric reflectance factors (Chapter 9), retrieval of land surface temperature (Chapter 4), theoretical radiative transfer modeling studies, and those based uniquely on field measurements. The chapter introduces existing instruments providing multiangle data; a review of work performed under the broad headings Empirical and Synergistic Approaches, Radiative Transfer, The RAMI Exercise, Canopy Openness, Clumping Index, Structural Scattering Index, Geometric-Optical and Hybrid Models, Direction and Wavelength, The Background in Canopy Reflectance Modeling, Land Cover and Community Type Mapping, Snow and Ice, and Dust Emissions; and finally, brief discussions under the headings Near-Simultaneous and Accumulated Sampling, Angular Sampling, and Scale and Multiangle Observation. The emphasis is on the added value that existing solar wavelength multiangular data can provide to applications. There has been a wide array of approaches, all of which have resulted in studies demonstrating progress and many of which show the advantages possible over the use of purely nadir-spectral techniques, particularly for accessing measures of canopy structure.

Mark J. Chopping Department of Earth and Environmental Studies, Montclair State University, USA [email protected] S. Liang (ed.), Advances in Land Remote Sensing, 95–144. c Springer Science + Business Media B.V., 2008

95

96

M.J. Chopping

5.1 Introduction The exploitation of data obtained by observing the Earth’s surface at different viewing and illumination angles in the solar wavelengths is a rapidly growing field. An excellent summary of work in this field is given in Liang et al. (2000) with a recent update by Diner et al. (2005). Multiangle remote sensing seeks to exploit the reflectance anisotropy of the planet’s terrestrial surface that is described by the bidirectional reflectance distribution function (BRDF) that quantifies the angular distribution of spectral radiance that is scattered by a surface (Nicodemus et al., 1977). The BRDF is thus a fundamental description of the terrestrial surface from the photon’s point of view. A key concept is that multiangle measurement provides a unique way to infer surface information within the observed element (“pixel”) and is thus useful for mapping land surface parameters over large areas. Remote sensing instruments acquiring imagery at varying angles explicitly are designed to collect multiple looks over as short a timeframe as possible and – in the case of those flown on satellites – taking into account the constraints imposed by the orbits of their platforms. Multiangle data sets may also be constructed by accumulation of observations over periods from days to weeks, following the longstanding practice of filtering images from many swaths to construct quasi-cloud-free composites. A limiting factor of this method is the temporal displacement of the measurements, since either the atmosphere or the surface, or both, may vary over short timescales (hours to days). The BRDF gives the reflectance of a target as a function of illumination geometry and viewing geometry; it depends on wavelength and is determined by the structural and optical properties of the surface, such as shadow-casting, multiple scattering, mutual shadowing, transmission, reflection, and absorption by surface elements, facet orientation distribution and facet (e.g., leaf, soil particle) density. The BRDF of a surface (unit: sr−1 ) is defined for all illumination and viewing directions over the upper hemisphere and is not itself measurable as the light incident on the surface is partly diffuse and the measurements involve finite solid angles (Liang, 2004; Bacour and Br´eon, 2005). Researchers have sometimes ignored the correct nomenclature and taken multiangle observations to provide a (usually sparse) sampling that is approximately bidirectional, accepting the assumption that the size/distance ratio of the illuminating source and the observing sensor is zero. However there are now efforts under way to ensure that the terminology in use is more accurate and consistent (Martonchik et al., 2000; Schaepman-Strub et al., 2006). The radiometric data products available from multiangle instruments include bidirectional reflectance factor (BRF; BRDF normalized by the equivalent reflectance of a Lambertian, or diffuse scattering surface) and/or hemisphericaldirectional reflectance factor (HDRF; the single integral of BRDF on the incoming directions; i.e., including direct and diffuse illumination), directional-hemispherical reflectance (black-sky albedo; single integral of BRDF on the outgoing directions) and bihemispherical reflectance (white-sky albedo; the double integral of BRDF over all viewing and illumination directions). These quantities must be derived via modeling since infinitesimal elements of solid angle do not include measurable amounts of radiant flux, and small light sources and unlimited sensor fields-of-view

5 Terrestrial Applications of Multiangle Remote Sensing

97

do not exist; all measurable quantities of reflectance are therefore performed in the conical or hemispherical domain (Schaepman-Strub et al., 2006). The most accurate description of the spectro-directional radiance measurements made by remote sensing instruments is therefore hemispherical-conical reflectance factor, even though solar illumination is usually highly directional, especially under clear skies. The most important physical phenomena underlying the BRDF are surface scattering (geometric effects, shadowing and shadow-hiding; Camacho-de Coca et al., 2004); and volume scattering by facets (leaves, soil particles). It is an intrinsic surface property. Multiangle remote sensing seeks to exploit fundamental aspects of the remote sensing problem and leads to important improvements in the accuracy with which vegetation and other surface parameters can be retrieved. The information returned is unique in that the data are sensitive to both the optical (spectral reflectance, transmittance, and absorption) and structural properties of surfaces (e.g., canopy architecture: horizontal and vertical heterogeneity), that is, from both the angular and spectral domains (Diner et al., 1999, 2005). Since most multiangle work pursues the one-dimensional (1-D) approach that seeks to derive “sub-pixel” information via variation in radiance with solar and/or viewing angles and in which each “pixel” is treated independently, the range of scales over which multiangle sensing can play a role is large (the term “pixel” is used here since in remote sensing it is commonly understood to correspond to the area under the sensor ground-projected field-of-view, even though this usage is incorrect: “pixel” is a contraction of “picture element”). Much work has gone into developing algorithms for retrieving well-understood physical parameters with straightforward interpretations via 1-D model inversions (Chopping et al., 2003; Widlowski et al., 2004; Koetz et al., 2005a), although researchers must grapple with the problem of non-uniqueness of solutions (different combinations of model parameters or state variables can result in similar patterns of observations), which can make this difficult (Weiss et al., 2000; Gobron et al., 2000; Combal et al., 2002). Nevertheless, using simulations of top-of-atmosphere radiance observations in 201 spectral bands and seven look angles, Verhoef (2005) demonstrated that: • Multi-angular hyperspectral measurements could in some cases provide dramatically improved retrievals of canopy structure, leaf properties, soil moisture, and radiative variables (e.g., fraction of photosynthetically-active radiation, albedo) than monodirectional measurements. • The information content of the spectral and angular domains is highly complementary, i.e., the additional information from multiangle remote sensing is not redundant. • For remotely estimating canopy architecture variables, multiangle information is more important than hyperspectral information (Fig. 5.1). Although the results above were obtained in the absence of real-world constraints, it can be asserted from first principles that canopy leaf foliar chemistry and ecosystem function are better assessed using hyperspectral methods, while canopy physical structure (in the sense of physiognomy, or architecture) is better assessed using multiangle methods (Schaepman, 2006). These two approaches are not mutually

Retrieval accuracy (dB)

98

M.J. Chopping 30

20

Hyperspectral Monodir. Hyperspectral Multidir. Multispectral Multidir. Multispectral Monodir.

10

0

s en fd e ar fb

ta ze % Cv ss Di fB t ho Fb D LI a DF LI I LA

b N_ _b Cs _b m Cd _b Cw _b b Ca

g N_ _g Cs _g m Cd _g Cw _g b Ca % SM do be Al R PA fA ver o fC

Geo-biophysical variable

Fig. 5.1 Relative retrieval accuracies expressed in dB above the a priori uncertainty levels, for low sensor noise (0.1 W/m2 /µm/sr) and four mission types. The parameters estimated were fractional cover (fCover); fraction of absorbed PAR (fAPAR); albedo; soil moisture (SM%); green leaf chlorophyll (Cab g), water (Cw g), dry matter (Cdm g), senescence pigments (Cs g) and the N parameter representing leaf mesophyll structure (CN g); brown leaf parameters (the same but identified with the suffix b), canopy leaf area index (LAI), leaf inclination distribution function (LIDF in 2 parameters); a hot spot parameter (hot), the fraction of brown leaves (fB), dissociation factor (Diss), crown cover fraction (Cv%), and tree shape factor (zeta); and the fractions of dense vegetation and bare soil in the immediate environment of the observation (fDENS and fBAR), accounting for adjacency effects. (Courtesy of Wout Verhoef).

exclusive but to date no studies have demonstrated the joint benefits of a spectrodirectional system with regular and broad geographic coverage; however, case studies over limited regions have demonstrated the considerable potential afforded by exploiting both domains simultaneously (Garc´ıa-Haro et al., 2006).

5.2 Major Multiangle Instruments Under a restricted definition for Earth Observation, including only instruments that provide near-simultaneous acquisition of >2 angular looks of the same area on the Earth’s surface, and excluding the thermal or longer wavelengths, only three such instruments have ever been placed in orbit: the NASA Multiangle Imaging Spectro-Radiometer (MISR) flown on the NASA Earth Observing System Terra; the French Space Agency’s (CNES) POLarization and Directionality of the Earth’s Reflectance (POLDER), flown initially on the Japanese ADEOS satellite series and currently on the CNES Parasol platform; and Sira (UK)’s Compact High Resolution Imaging Spectrometer (CHRIS), flown on the European Space Agency (ESA) Proba-1 satellite. Terra was launched in December 1999, Proba in 2001, and Parasol in 2004. Unprecedented efforts have also been made in the last 15 years properly to treat data from across-track scanning instruments such as the NOAA Advanced Very High Resolution Radiometer (AVHRR) and the MODerate resolution Imaging Spectroradiometer (MODIS) flying on NASA’s Terra and Aqua EOS satellites as

5 Terrestrial Applications of Multiangle Remote Sensing

99

intrinsically dependent on surface BRDF. Although for the latter instruments multiangular data sets are acquired sequentially rather than near-simultaneously, work with these accumulating, or “sequential” systems is also making important contributions to multiangle remote sensing. For this reason, reference is made here to MODIS as a “multiangle” instrument, even though it was not designed to acquire multi-angle imagery and the variation in viewing and illumination angles is an accident of the across-track acquisition method. Indeed, under this paradigm many other moderate resolution sensor systems, including those on geostationary platforms, could also be deemed to be “multiangle”. It is preferable to think of these systems in this way rather than to insist on the narrower definition because angular artifacts in data from these instruments might otherwise be overlooked. This chapter will concentrate on applications using data from orbiting instruments providing more than two views; this excludes Japan Aerospace Exploration Agency’s ASTER on Terra, which exploits dual views via a “geometric strategy” rather than the “radiometric strategy” necessary with data from sensors with a larger instantaneous field-of-view (IFOV) and the Advanced Along-track Scanning Radiometer (AATSR), an important European instrument whose dual views have proven valuable for aerosol characterization (King et al., 1999). The geometric strategy involves real or apparent differences in the shape or location of observed objects, resulting from changes in perspective (for targets with 3-D structure), stereoscopic parallax (displacement dependent upon distance from the observer), or actual motion of the target during the time interval between views; the radiometric approach refers to changes in the brightness, color, contrast, or other radiance-related quantities as a function of view and/or illumination angles (Diner et al., 2005, Table 5.1).

5.2.1 MISR on Terra The Multi-angle Imaging SpectroRadiometer (MISR) was designed and built at NASA’s Jet Propulsion Laboratory (JPL), in Pasadena, California and flies on NASA’s Terra satellite. MISR measures Earth’s spectral radiance in four spectral bands (blue, green, red and near-infrared), at each of nine look angles in the forward and aft directions along the flight path (Fig. 5.2). One camera views at nadir (close to 0◦ view zenith angle), while the others view at ±26.1◦ , ±45.6◦ , ±60.0◦ , and ±70.5◦ (Fig. 5.3). Spatial samples are acquired every 275 m. In the global mode, all nine cameras report the red band and the nadir camera reports all other bands in at this resolution; for the off-nadir cameras the other bands are downsampled to 1.1 km via contiguous averaging of four cross-track by four along-track line samples within the instrument before data transmission. Over a period of 7 min, a ∼400 km wide swath of Earth comes into view at all nine angles. Global coverage is achieved every 9 days, with repeat coverage between 2 and 9 days depending on latitude (the repeat time is shorter at higher latitudes). In local mode, all four spectral bands are retained at the acquisition resolution of 275 m in all nine cameras. Special attention has

100

Table 5.1 Major satellite instruments with multiangle capability (Adapted from Diner et al., 2005.) POLDER (ESA)

CHRIS/Proba

MODIS (NASA)

CERES (NASA)

AATSR (ESA)

Description

Multiangle pushbroom imager

Wide Field-of-View pushbroom radiometer

Across-track imaging scanner

Scanners

Circular scanner

Launch Date

December 1999, on EOS Terra

December 1999, on EOS Terra

December 1999, on EOS Terra

March 1, 2002, on ENVISAT

Angular coverage

0◦ , 26◦ , 46◦ , 60◦ , 70◦ forward and backward of nadir in along-track direction

August 1996, on ADEOS 1; December 02, on ADEOS II; December 04, on Parasol Comprehensive (over several days)

Tiltable pushbroom spectrometer October 2001, on Proba-1

Five looks in the along-track direction

0–55◦ continuously

Limb to limb

Multiangle strategies employed Spectral coverage

Radiometric, geometric

Radiometric

Radiometric

Radiometric

446, 558, 672, 866 nm

444.5, 444.9, 492.2, 564.5, 670.2, 763.3, 763.1, 907.7, 860.8 nm

Radiometric, geometric Programmable up to 63 bands, 0.4–1.0 µm Bandwidth: <11 nm

Two looks, 55◦ forward and close to nadir in the along-track direction. Radiometric

36 bands: 0.4–14.5 µm

Shortwave: 0.3–5 µm Thermal window: 8–12 µm Total: 0.3–200 µm

Spatial resolution

275 m or 1.1 km, depending on channel

6 × 7 km nominal, 6 × 6 after processing, 2,400 km swath

250 m − 1 km at nadir, depending on channel

20 km at nadir

Scene dimensions

Pole-to-pole × 400 km

Global coverage in

9 days

Pole-to-pole 2, 400 × 1, 800 km 1 day

18–34 m depending on the node selected, 13 km swath 13 × 13 km approximately NA

0.555, 0.659, 0.865, and 1.61 0.555, 0.659, 0.865, and 1.61 µm (VIS-IR); 3.7, 10.8, and 123.7, 10.8, and 12 µm (TIR) 1 km at nadir

Pole-to-pole × 2, 300 km 2 days

Pole-to-pole × limb-to-limb 1 day

Pole-to-pole × 512 km 6 days

M.J. Chopping

MISR (NASA) (CNES)

5 Terrestrial Applications of Multiangle Remote Sensing

101

Fig. 5.2 Computer-generated image showing the Terra spacecraft and MISR instrument operation. The direction of flight is toward the lower left. The locations imaged by the 9 MISR cameras, each with 4 spectral bands, are illustrated here with translucent surfaces (Courtesy of Shigeru Suzuki and Eric M. De Jong, Solar System Visualization Project. NASA/JPL-Caltech, image P-49081.)

Fig. 5.3 MISR images of Zambia and Botswana, Africa, acquired on August 25, 2000. The left image is a “true” color composite from the red, green, and blue vertical-viewing (nadir) camera. The middle image combines data from the green, red, and near-infrared bands. The right image contains red band data only but as a composite of imagery from the nadir, 70.5◦ forward, and 70.5◦ aftward cameras. The color variations in the multi-angle composite arise because light is reflected at different angles. (Courtesy of the MISR Science Team.)

been paid to providing highly accurate absolute calibration, using on-board hardware consisting of deployable solar diffuser plates and several types of photodiodes (Diner et al., 1998; Bruegge et al., 2002). The land surface reflectance products currently generated at the NASA Langley Atmospheric Sciences Data Center include the spectral hemispherical–directional reflectance factors (HDRF) at the nine MISR view angles and the associated bihemispherical reflectances (BHR). The HDRF and the BHR characterize the surface reflectance under direct and diffuse illumination.

102

M.J. Chopping

An algorithm for the generation of vegetation green leaf area index (LAI) and the fraction of photosynthetically active radiation absorbed by vegetation (fPAR) from MISR BHR and bidirectional reflectance factor (BRF) was implemented for operational processing (Knyazikhin et al., 1998c). Unlike single-angle approaches, the MISR LAI/fPAR algorithm does not rely on prescribed vegetation type maps as it uses the multiangular information to constrain the retrievals.

5.2.2 MODIS on Terra The MODerate resolution Imaging Spectroradiometer (MODIS) instrument flying on NASA’s Earth Observing System (EOS) Terra and Aqua satellites provides 12 bit imagery in 36 spectral bands ranging in wavelength from 0.4 to 14.4 µm. Two bands (red and NIR) are imaged at a nominal resolution of 250 m at nadir, with five bands at 500 m, and the remaining 29 bands at 1 km. MODIS is an across-track scanning radiometer providing observation along a scanline that observes the surface at zenith angles of ±55◦ . At the EOS orbit of 705 km, it achieves a 2,330 km swath and provides global coverage every 1–2 days. MODIS-AM was placed into orbit with the launch of Terra on December 18, 1999; the second instrument, MODIS-PM, is flying on the Aqua spacecraft launched on May 4, 2002. Thanks to the very wide swath, MODIS provides multiangular measurements of spectral radiance from the same location on the surface only through the accumulation of looks over a period of several days (sometimes called the “sequential” method); it is not capable of nearsimultaneous multiangle observation. While global coverage is provided in a period of 48 h or less, cloud cover means than data from several overpasses are almost always required in order to map clear-to-surface surface-leaving spectral radiances. The across-track scanning method provides angular observations closer to the solar principal plane (the plane intersected by the sensor, target, and the Sun) than along-track instruments at mid-latitudes, although this is highly dependent on latitude. The major disadvantage of this method is that the surface may change during the accumulation period; however the variation in the signal owing to accumulated sampling is generally far lower than that owing to the BRDF. Since MODIS data are highly dependent on viewing and illumination geometry and surface BRDF, major efforts have been expended to properly account for BRDF effects, and to exploit the directional signal to retrieve global land albedo (see Chapter 9).

5.2.3 CHRIS on Proba-1 The CHRIS sensor was developed by Sira Electro-Optics (UK). Since April 2006 Sira space and imaging operations have been a part of Surrey Satellite Technology Ltd., Guildford, UK, a spin-off company of the University of Surrey. CHRIS produces imagery in up to 62 spectral channels in the range 415–1,050 nm with a

5 Terrestrial Applications of Multiangle Remote Sensing

103

Fig. 5.4 (a) Artist’s impression of the Proba satellite (b) CHRIS instrument multi-angle acquisition via tilting of the platform (Courtesy of ESA and Mike Barnsley, University of Wales at Swansea.)

spectral resolution of 5–12 nm. It is remotely configurable in terms of both spectral channels and spatial resolution (European Space Agency, 1999), with a nominal ground resolution of between 17 and 34 m. The nominal swath is 13 km. CHRIS was launched on the Proba-1 satellite (Verhaert, Belgium) on October 21, 2001 (Fig. 5.4). Proba-1 is a small platform, weighing approximately 100 kg and measuring approximately 60 × 60 × 80 cm. Thanks to its four reaction wheels the platform is highly maneuverable: along-track pointing allows a given site to be imaged five times during a single overpass, while across-track pointing ensures that the revisit time for a site of interest is potentially less than a week (Barnsley et al., 2004). Proba was not placed into its intended orbit because of an anomaly with the final stage of the Indian Polar Satellite Launch Vehicle (PSLV), resulting in a much lower and more elliptical orbit than intended. This caused a number of operational problems that necessitated on-ground and on-board software changes and delayed the start of the science program. It took more than 12 months to rewrite the software required for precise orbital navigation, so that the first useable overlapping hyperspectral imagery was available to most users only in early 2003 (albeit often with four partially overlapping looks, rather than the five intended). Since then, further improvements have been made so that five overlapping images are generally available and the overlap area is larger. Both the satellite and the CHRIS instrument continue to function well at the time of writing (April 2006), although striping is evident in some images owing to detector anomalies; however, third-party algorithms have been written to correct for this. The CHRIS/Proba system has several unique features: it employs tilting of the platform to gain a series of angular views in the along-track direction; the platform nods in order to acquire images; and the CHRIS sensor is programmable for spatial resolution and spectral coverage. There are well over 60 groups around the world engaged in research with CHRIS in categories including calibration/validation, land and environment, hydrology, atmosphere, agriculture, forestry, hazards, renewable resources, coastal zones, topographic mapping, and geology (European Space Agency, 2006). It is

104

M.J. Chopping

worth noting that not all projects seek to exploit the multiangle feature of the instrument; many projects are focusing on the exploitation of the hyperspectral capability.

5.2.4 POLDER on ADEOS / Parasol The Polarization and Directionality of the Earth’s Reflectance (POLDER) instrument developed by France was initially launched on the Japanese ADEOS-1 (ADvanced Earth Observation Satellite) platform in August 1996, with the satellite failing in June 1997 (a solar panel failure terminated operations after 8 months of continuous acquisition); a second identical instrument, was launched on ADEOS-2 in December 2002, with that satellite failing in October 2003 (another solar panel failure terminated operations with 7 months of continuous acquisition). This version of POLDER provided a swath of ∼2, 400 km with a footprint of 6 × 7 km2 . The instrument was a two-dimensional CCD detector array with wide field of view telecentric optics and a rotating wheel carrying spectral and polarization filters. There were 15 filters and polarizers in the visible and the near infrared range providing nine spectral bands, three of which (443, 670 and 865 nm) were associated with polarized filters. Three bands were dedicated to the observation of ocean color (443, 490, 565 nm) but POLDER-1 was designed to provide dedicated measurements of clouds and atmospheric aerosols. Its applications in terrestrial remote sensing have therefore taken a secondary role, although they are not insignificant (Maignan et al., 2004; Br´eon et al., 2002; Camacho-de Coca et al., 2002). While POLDER-1 acquired spectral radiance data at multiple angles for the same location on a nearsimultaneous basis in the along-track direction, its full angular sampling potential was only realized after a number of successive orbits has been accomplished, typically within 15 days. This means that POLDER-1 was a hybrid system and has some features in common with accumulating instruments. The third POLDER instrument – hereafter called POLDER-2 as it differs from the first version, POLDER-1 – is being flown on the CNES Parasol platform, which was launched from Europe’s spaceport in Kourou, French Guiana in December 2004 (Fig. 5.5). Parasol is a micro-satellite that flies in formation with the NASA EOS “A Train”. The POLDER-2 instrument has provided near-continuous acquisition since March 2005. It observes in eight spectral bands from 440 (blue) to 910 nm (NIR) using a bi-dimensional CCD matrix providing a spatial resolution of ∼6 km; three of the bands are polarized: 440, 670, 865 nm. As with the POLDER instrument on the ADEOS-2 platform, the Parasol-borne POLDER-2 provides its most dense angular sampling via accumulation over a period of a few days. Over successive days, the instrument is at different places in the sky, which provides another set of angular measurements: after a few days, the directional space within 60◦ of view zenith angles is fully covered with a good sampling of the principal and perpendicular planes, including the hot spot around which the solar and viewing directions coincide.

5 Terrestrial Applications of Multiangle Remote Sensing

(a)

105

(c) 50 45

[%]

40 35 30

Ref_443 Ref_565 Ref_670

25 20 −80

−60

−40

−20

0

20

Phase Angle

(b)

(d) 10 8

[%]

6 4 2

RefPol_443 RefPol_370 RefPol_865

0 −2 −80

−60

−40

−20

0

20

Phase Angle

Fig. 5.5 POLDER-2 images over desert target in the Sahara (a) 3 color composites in total light 443–670–865 nm (b) 3 color composites in polarized light 443–670–865 nm (c) directional signatures in total light (d) directional signatures in polarized light (Courtesy of F-M Br´eon, Laboratoire des Sciences du Climat et de l’Environnement.)

5.3 Multiangle Applications: Vegetation Canopies Satellite remote sensing would ideally be able to provide information on both the structure and function of ecosystems and the types of plant communities from which they are formed, at regular intervals, and for the entire terrestrial surface. While remote sensing of vegetation function is an established fundamental and very important application of remote sensing – spectral measures have proved very effective in mapping vegetation productivity – less has been achieved with respect to remotely measuring vegetation structure. The optical properties of vegetation have allowed the exploitation of reflected light in the solar spectral wavelengths via spectral vegetation indices, spectral unmixing techniques and in the temporal domain, phenology metrics, that have proved useful in obtaining global maps of vegetation parameters, such as the MODIS Vegetation Continuous Fields (VCF) % Tree Cover product (Hansen et al., 2002). In spite of these successes, it is difficult to access canopy structure information using single-angle sensing since structural surface properties are largely confounded in nadir-spectral measures.

106

M.J. Chopping

Canopy structure is defined as the vertical and spatial distribution, orientation and density of foliage and its supporting structures. Vegetation structure determines the pattern of light attenuation and the distribution of photosynthesis, respiration, transpiration, and nutrient cycling in the canopy (Widlowski et al., 2004). Net primary production of woody vegetation is always dependent on vegetation structure, since the maximum rate of carbon uptake is constrained by total leaf area, and the rate of carbon loss is constrained by total woody biomass (Widlowski et al., 2004). Measures of canopy structure are required in order to estimate biomass, to assess geographically shifting abundances of woody plant material (shrubs and trees), to estimate carbon emissions to the atmosphere from wildfires, and to examine where forest regrowth and secondary succession are occurring, and at what rates (Liu and Kafatos, 2007). Canopy architecture and spatial heterogeneity are important not only in forests but also affect the available carbon storage, vegetation dynamics and productivity of grasslands undergoing woody encroachment, shrublands, and savannas: different degrees of structural heterogeneity lead to a wide range of estimates for these environments (Asner, 2000; Chopping et al., 2003). It is also necessary for mapping fire fuel loads; and for estimating surface roughness, with the implications this has for fluxes between the surface and the atmosphere. The surface BRDF is dependent on the structural as well as optical characteristics of surfaces and it should thus be possible to exploit angular measurements to retrieve structural canopy parameters. Although this is not straightforward, in the last few years multiangle reflectance data from satellites have been used to develop a variety of approaches (Pinty et al., 2002; Bacour et al., 2002; Gao et al., 2003; Hu et al., 2003; Widlowski et al., 2004; Jenkins et al., 2004; Nolin, 2004; Koetz et al., 2005b; Chen et al., 2005; Strahler et al., 2005; Diner et al., 2005; Kimes et al., 2006; Laurent et al., 2005; Bach et al., 2005; Chopping et al., 2006c; Disney et al., 2006; Garc´ıa-Haro et al., 2006; Heiskanen, 2006; Chopping et al., 2007). The approaches taken are very diverse and include 3-D radiative transfer modeling, geometric-optical canopy reflectance modeling, exploitation of the structural information available in the parameters returned via inversion of physical and semi-empirical models, purely empirical methods, and heuristic and data mining algorithms. All have demonstrated some degree of success.

5.3.1 Empirical and Synergistic Approaches For some years Earth observation scientists have been considering the range and importance of synergies that may be realized between multiangle instruments and other types of sensor. The focus has been on synthetic aperture radar (SAR) and – more recently – lidar, both active systems, although good synergies with high resolution stereoscopic imagers can also be found for some applications (e.g., cloud top heights). Disney et al. (2006) demonstrated that a combined structural and radiometric modeling approach provides a flexible and powerful method for simulating the remotely sensed signal of a forest canopy in the solar and microwave

5 Terrestrial Applications of Multiangle Remote Sensing

107

domains, useful for exploring the impact of canopy structure on the resulting signal and for combined retrievals of forest structural parameters. They developed a detailed 3-D structural model of a conifer forest canopy in order to simulate the observed reflectance (optical) and backscatter (microwave) signals and showed that it is feasible to model forest canopy scattering using detailed 3-D models of tree structure (including the location and orientation of individual needles). A structural growth model of Scots pine was modified to simulate observed growth stages of a Scots pine canopy from age 5 to 50 years. Model canopies were generated by adding needles to tree structural models and these were used to drive optical and microwave models of canopy scattering. Simulated canopy radiometric response was compared with airborne hyperspectral reflectance data from the HyMAP instrument (Integrated Spectronics, Australia) and airborne synthetic aperture radar backscatter data. Model simulations agreed well with observations, particularly at solar wavelengths. The choice of needle shape and phyllotaxy was shown to have a significant impact on multiple scattering behavior at the branch scale. In the microwave domain, simulated backscatter values agreed reasonably well with observations at L-band and less so at X-band. L-band simulated backscatter significantly underestimated observed backscatter at younger canopy ages, probably as a result of inappropriate modeling of the soil/understory background. Successful and convincing empirical applications of multiangle reflectance data have been rare until recently, when the exploration of synergies with lidar data became feasible. Researchers at the NASA Goddard Space Flight Center, Greenbelt, Maryland, performed work in which forest structural data from the airborne Laser Vegetation Imaging Sensor (LVIS) were used with multivariate regression and neural networks to allow the estimation of forest vertical structure from multiangle reflectance data from the Airborne Multi-angle Imaging SpectroRadiometer (AirMISR) (Kimes et al., 2006). The AirMISR data were acquired on 28 August 2003 over a 7,000 ha experimental forest near Howland, Maine, consisting of small plantations, multi-generation clearings and large natural forest stands (predominantly boreal-northern hardwood transitional forest with spruce-hemlock-fir, aspen-birch, and hemlock-hardwood associations). Twenty-eight AirMISR multiangle spectral radiance values were selected from the 36 possible combinations (nine angles and four spectral bands). Data acquired at nadir viewing and at ±26.1◦ , ±45.6◦ , and ±60.0◦ in the flying direction were used and resampled into 15 × 15 m pixels. The LVIS data were acquired 3 days later, with a total of ∼160, 000 LVIS shots collected in the study area. The LVIS data set provided a relatively direct measure of forest vertical structure at a fine scale (20 m diameter footprints). For each laser shot, the following were recorded: location (long, lat), surface height (m), and the heights of the 25%, 50%, 75% and 100% of waveform energy data (m). These data were gridded into 15×15m pixels using inverse-distance weighted average of the four nearest shots. The best model accurately predicted the maximum canopy height as measured by LVIS using AirMISR data with a root mean square error (RMSE) of 0.92 m and a R2 = 0.89 (Fig. 5.6). This result is all the more remarkable because these high accuracies were achieved over a study site with an elaborate patchwork of forest communities with exceptional diversity in forest structure. The study concluded that models

108

M.J. Chopping

Predicated LVIS H100 Values (m)

25

a

20

15

10

5 10

5

15

20

25

75

100

True LVIS H100 Values (m) 25 True/Predicted Forest Height (m)

b Predicted Undisturbed Stand True Undisturbed Stand Predicted Selective Cut True Selective Cut

20

15

10

5

0 0

25

50

LVIS Energy Quartile Heights (%)

Fig. 5.6 (a) LVIS forest canopy heights predicted by AirMISR versus heights from LVIS for 17 × 17 pixel windows. The model was applied to 1,000 random test points. (b) Predicted versus true LVIS energy for canopy height (H100) and quartile heights (H75, H50, H25) for the logged and unlogged areas. The predictive models used only AirMISR data (Reprinted from Kimes et al., 2006. Copyright, Elsevier. With permission.)

using multiangular data are capable of accurately predicting the vertical structure of forest canopies. Further work using data from the Geoscience Laser Altimeter System (GLAS) on the Ice, Cloud, and Land Elevation Satellite (IceSAT) platform to train MISR-derived estimates of canopy and extend them continuously over large areas has provided encouraging but somewhat uneven results. This is most likely to be at least partly owing to the azimuthal orientation of MISR data away from the principal plane – that sometimes can be advantageous, e.g., for the retrieval of albedo and fPAR (B. Pinty, personal communication) – since the AirMISR data were acquired close to this plane (K.J. Ranson, personal communication).

5 Terrestrial Applications of Multiangle Remote Sensing

109

Fig. 5.7 Predicting stand basal area with AirMISR (non-linear multiple regression) using an artificial neural network with 5 nodes. Training R2 = 0.68, RMS Error = 8 m2 /ha (Courtesy of Rob Braswell and Julian Jenkins, Institute for the Study of Earth, Oceans and Space (EOS), University of New Hampshire.)

Another study using independent field biometric measurements to train multiangle data was carried out over the Bartlett Experimental Forest in New Hampshire. An artificial neural network with five nodes was used to map basal stand area over an area of 19.1 km2 . A high correlation was obtained between optical multiangle reflectance signature measured with AirMISR and field observations of stand basal area, with an RMSE of 8 m2 /ha (Jenkins et al., 2004; Fig. 5.7). These studies show that multiangle data can be used to leverage lidar or surface assets to provide wall-to-wall coverage using purely empirical methods, rather than explicit modeling of light scattering. Further advances might be expected using lidar synergistically with multiangle data in modeling approaches as well as in an empirical training mode (Hese et al., 2005; see Geometric-Optical and Hybrid Models, below).

5.3.2 Radiative Transfer Three-dimensional radiative transfer (RT) modeling is used as part of MISR surface product retrievals. A look-up table (LUT) is used to rapidly determine the matching modeled reflectances and the associated values of leaf area index (LAI) and fraction

110

M.J. Chopping

of photosynthetically active radiation (fPAR): all necessary radiative transfer parameters are pre-computed and stored in a Canopy Architecture Radiative Transfer (CART) file. The MISR LAI retrieval algorithm does not depend on a particular canopy radiation model as the elements of the CART are components of various forms of the energy conservation law. The MISR leaf area index (LAI) product has been shown to be accurate to within 0.5 LAI in herbaceous vegetation and savannas and is an overestimate by about 1.0 in broadleaf forests (Diner et al., 2005). When using single angle data, uncertainty in biome classification can lead to large errors. However, the use of multiangle data minimizes the impact of biome misidentification on LAI retrievals; that is, despite errors in biome classification, the use of multiangle information leads to a similar error rate – even when the biome is misclassified – as is obtained with single angle approaches and a prescribed biome map. The multiangle method therefore permits LAI retrieval without the use of a prescribed biome map. In a case study in Africa, 80% of the LAI values were retrieved using an incorrect biome type – but with a probability of about 70%, uncertainties in LAI retrievals due to biome misclassification did not exceed uncertainties in the observations (Hu et al., 2003). A parameter closely related to LAI, the canopy “recollision probability” (defined as the probability with which a photon scattered in the canopy interacts with a phytoelement again) has been shown to describe canopy spectral absorption and scattering; this is a new and developing area of multiangle imaging research. Knyazikhin et al. (1998a, b) proposed that – to a good approximation – the amount of radiation absorbed by a canopy should depend only on the wavelength and a canopy structural parameter (p), which is wavelength independent. Knowing the recollision probability value of a canopy, the scattering coefficient of the canopy at any wavelength can be predicted from the leaf scattering coefficient at the same wavelength. Knyazikhin et al. (1998a, b) also introduced a similar parameter (pt ) relating canopy transmittances at two different wavelengths to the leaf scattering coefficients at these wavelengths. Given the absorption (p value) and transmission (pt value), total reflectance (the upward scattered part of the incident radiation) is also known. It is recognized that LAI may not depend on only these parameters alone but may also vary with leaf orientation and the degree of foliage clumping. A recent study by Smolander and Stenberg (2005) confirmed that the spectral absorption and scattering of structurally simple uniform canopies can be well described by the canopy p value, which furthermore showed close relationship with the LAI but insensitivity to solar zenith angle.

5.3.3 The Radiation Transfer Model Intercomparison (RAMI) Exercise Radiative transfer model accuracy and robustness are key issues for improving remotely sensed surface variables that may depend heavily on the angular distribution of samples. The Radiation Transfer Model Intercomparison (RAMI) exercise

5 Terrestrial Applications of Multiangle Remote Sensing

111

(Pinty et al., 2001, 2004a) has been ongoing since 1999 with the most recent and third phase completed in 2005. RAMI provides a framework within which spectral bidirectional reflectance models designed to simulate the transfer of radiation within plant canopies and over bare soil surfaces can be benchmarked. This is useful because the interpretation of remote sensing data generated by Earth Observing satellites hinges on the exploitation of such models: it is clearly very important that they are reliable, accurate, and fit for purpose. RAMI approaches this endeavor by providing a series of test cases of different degrees of complexity, broadly divided into homogeneous and heterogeneous cases (Fig. 5.8). Participants perform simulations based on these scenarios and deliver modeling results to the Institute for Environment and Sustainability at the European Commission Joint Research Center in Ispra, Italy, that coordinates the exercise. The subsequent publication of the findings provides an objective way to evaluate the performance and limits of applicability of the models. In addition, new models may be developed and existing ones improved by comparing their output with the available simulation results on the RAMI web pages (http://rami-benchmark.jrc.it/). The overall benefit to the remote sensing community is a demonstration of model maturity and a better understanding of how the models may be used in the interpretation of remote sensing data. The latest results, that include 5 1-D models and

Fig. 5.8 Selected scenarios in the Phase 3 RAMI set of radiative transfer modeling exercises: (a) homogeneous turbid cases in the solar domain: leaves treated as a turbid medium (b) homogeneous discrete cases in the solar domain: leaves represented as a large number of non overlapping discshaped objects (c) heterogeneous turbid cases in the solar domain: large number of non overlapping spherical objects with crowns represented as a turbid medium (d) discrete floating spheres in the solar domain: large number of non overlapping spherical objects representing the individual plant crowns that contain randomly distributed finite size disc-shaped scatterers, with the orientation of the scatterers following a uniform distribution function (e) Heterogeneous Scene-Based Experiments: large number of disc-shaped scatterers contained within a series of non-overlapping spherical and cylindrical volumes representing identically sized plant crowns (f) Boreal Birch Stand Scene: large number of randomly located non-overlapping tree-like entities of differing sizes with individual objects represented by ellipsoidal crowns with randomly distributed finite sized foliage. See: http://rami-benchmark.jrc.it/HTML/RAMI3/RAMI3.php

112

M.J. Chopping

13 3-D models, are reported in Widlowski et al. (2006). It should be noted that model accuracy is an important but not the only limiting factor in the application of reflectance models in retrieving surface information useful to applications using multiangle remote sensing data: inversion algorithms (Chapters 6–8), model paradigm and suitability for purpose, noise and contamination in data sets, and angular and spectral sampling also play important roles.

5.3.4 Canopy Openness One approach to estimating canopy openness makes use of parametric fits of red wavelength MISR or AirMISR data to the Rahman-Pinty-Verstraete (RPV) model (Rahman et al., 1993) or the modified version of this (MRPV, Engelsen et al., 1996). The MRPV model parameters from inversion against MISR data are included in the operational MISR Land product at level 2. The RPV and MRPV models are based on a consideration of the main aspects of BRDF shapes and are threeparameter models. The first parameter is a factor describing the overall amplitude of reflectance; the second describes the steepness of the BRDF bowl or bell shape; and the third controls the relative importance of back- and forward-scattering via a Henyey-Greenstein function (for the RPV model) or an exponential function in scattering angle (for the MRPV model). The equations for the MRPV model are given in Eq. (5.1): BRFRPV = ρs (θs , θv , φ ; ρ0 , b, k) = ρ0 M F H

(5.1)

M = [cos θs cos θv (cos θs + cos θv )]k−1 F = exp(−b cos(g)) H = 1 + [(1 − ρ )/(1 + G)]

(5.2) (5.3) (5.4)

where:

where ρ is the average measured reflectance factor and G = [tan2 θs + tan2 θv , −2 tan θs tan θv cos φ ]1/2 cos g = cos θs cos θv + sin θs sin θv cos φ

(5.5) (5.6)

The first parameter, ρ0 , is a factor describing the overall amplitude of reflectance; the second, b, controls the relative importance of back- and forward-scattering via a Henyey-Greenstein function in the original RPV formulation, replaced by an exponential term in the MRPV model Eq. (5.2) and in the MISR BRDF/albedo product. The rationale for this modification is that except for the hotspot term the coefficients of the MRPV model can be rapidly solved by linear least-squares by taking logarithms of the expression; but note that a recently-developed algorithm allows efficient, reliable and accurate inversion of the original RPV model and additionally

5 Terrestrial Applications of Multiangle Remote Sensing

113

Fig. 5.9 Typical BRF anisotropy in the red spectral domain for radiatively homogeneous (left) and heterogeneous (right) vegetation canopies. The BRFs of the heterogeneous surface were generated using a 3-D RT model and typified by a bell-shape (k > 1). The BRFs of the homogeneous surface covers were generated using a 3D RT model and a 1D IPA approach, and are generally bowl-shaped (k < 1) (Reproduced from Widlowski et al., 2001. Copyright, AGU.)

provides a measure of uncertainty on the retrieved parameters (Lavergne et al., 2007). The modified Minnaert function parameter, k, is used to quantify the degree by which the observed bidirectional reflectance factor data resemble a bowl- or bellshaped pattern in azimuthal planes (Fig. 5.9). The presence of a very sparse or very dense and quasi-homogeneous vegetation layer within the sensor IFOV is characterized by a bowl-shaped reflectance anisotropy shape (k < 1), while vertically elongated foliage clumps of medium-to-high densities uniquely generate bell-shaped shapes (k > 1) (Widlowski et al., 2001, 2004; Pinty et al., 2002; Fig. 5.9). This broad relationship holds well for snow-covered areas, with homogeneous bare or non-forested areas exhibiting a bowl-shaped pattern (k < 1), indicating that subpixel homogeneity (sparse or extremely dense vegetation cover) will result in k < 1. A dependency on vegetation density was demonstrated for the first time in Gobron et al. (2002). The k parameter has a limited physical meaning (Nolin, 2004); in order to translate these relationships to measurable canopy structural parameters, it is possible to interpret k values as a ratio between mean effective scene height and the ratio of mean tree density and mean nearest-tree distance (Widlowski et al., 2004). Effective scene height is defined as the mean height of all structures within the IFOV weighted by their fractional surface coverage. The relationship can be seen in kred results obtained by these authors using radiation transfer simulations performed with an illumination zenith angle of 30◦ , a soil reflectance of 0.126, a bark reflectance of 0.251 and a leaf reflectance (transmittance) of 0.018 (0.021) (Fig. 5.10). The data points generated differ only because of their structural characteristics (stem and foliage density). Researchers using multiangle data from the European CHRIS instrument found that the results of RPV model inversion enabled discrimination

114

M.J. Chopping

Fig. 5.10 The organization of the kred values in a plane defined by the “mean effective scene height”, i.e., the height (m) of all structures within the IFOV weighted by their fractional surface coverage and – on a log scale – the ratio between average tree density (stems/m2 ) and the mean nearest tree distance (m). The data points show the characterizations of ∼200 structurally diverse boreal forest scenes (Reproduced from Widlowski et al., 2004. Copyright, Springer. With kind permission of Springer Science and Business Media.)

between different surface types based on the their inherent reflectance anisotropy and that the k parameter was successfully linked to lidar measurements representing the 3-D structure of the canopy (Koetz et al., 2005a).

5.3.5 Clumping Index Foliage clumping is an important forest structural canopy attribute: it affects both the gap fraction for the same LAI, radiation interception and distribution within the canopy, which in turn affects photosynthesis. It can be quantified in a Clumping

5 Terrestrial Applications of Multiangle Remote Sensing

115

Legend 0.0 - 0.3 0.3 - 0.4 0.4 - 0.5 0.5 - 0.6 0.6 - 0.65 0.65 - 0.7 0.7 - 0.8 0.8 - 1.0 No Value Water Excluded

Fig. 5.11 Global clumping index map derived from POLDER 1 data using the normalized difference between interpolated hotspot and darkspot NIR reflectance and applied to vegetated land cover. Clumping increases with decreasing values of the index (Reprinted from Chen et al., 2005. Copyright, Elsevier. With permission.)

Index (CI) based on a vegetation dispersion parameter Ω obtained via a modification to Beer’s law. A physical modeling approach to obtaining CI has been taken by two collaborating groups in Canada at the University of Toronto and the Canada Centre for Remote Sensing, Ottawa (Chen et al., 2005). These authors have produced the first global maps of foliage clumping with multiangle POLDER data, assisted by a geometrical optical model (Fig. 5.11). They used the Normalized Difference between Hotspot and Darkspot (NDHD), an angular index that can be obtained from multiangle remote sensing that characterizes reflectance anisotropy and has been successfully related to ground measurements of CI (Chen et al., 2003). Using the relative magnitude of the darkspot to the hotspot minimizes the dependence of the index on foliage optical properties that are important determinants of bidirectional reflectance. This work built on previous research using data from POLDER (Lacaze et al., 2002).

5.3.6 Structural Scattering Index A semi-empirical approach applied to MODIS data is the use of the Structural Scattering Index (SSI) developed to exploit the kernel weights derived by inversion of the LiSparse-RossThick kernel-driven BRDF model (Gao et al., 2003). The MODIS BRDF/Albedo (MOD43) product is generated on a regular 16-day cycle using this model. The model depends on three parameters (kernel weights) describing the interaction of light with the surface. These parameters are used in a forward version of the model to reconstruct the surface anisotropic effects and to correct MODIS reflectances to a common view geometry (the MOD43B4 Nadir BRDF-Adjusted

116

M.J. Chopping

Reflectances – or NBAR – product). The model can also be used to compute integrated black-sky albedo at some solar zenith angle as well as white-sky albedo MOD43B3. The model is a linear, semi-empirical, kernel-driven (LiSK) model formulated as a superposition of weighted kernels Eq. (5.7): R(θs , θv , ϕ ) = fiso + ( fgeo × kgeo (θs , θv , ϕ )) + ( fvol × kvol (θs , θv , ϕ ))

(5.7)

where R(θs , θv , ϕ ) is the modeled bidirectional reflectance for the solar zenith, view zenith, and relative azimuth angles θs , θv , and ϕ , respectively. The model kernels kgeo and kvol are analytical functions of the solar and viewing angles derived via simplifying physical terms (Roujean et al., 1992; Wanner et al., 1995): kgeo accounts for geometric scattering and shadowing, while kvol accounts for volume scattering from a discrete medium of randomly located facets, assuming the single scattering approximation, an isotropic facet distribution function, and an optically thick medium. The kernel weights fiso , fgeo and fvol are theoretically dependent on the three dimensional structure and optical properties of a canopy and its background; fiso should represent the bidirectional reflectance viewing a surface at nadir with the overhead Sun, while fgeo and fvol measure the relative contributions of the geometric/shadowing and volume scattering components. Models of this kind are also often referred to generically as Li-Ross models, since the terms describing anisotropic scattering are obtained from functions derived by Xiaowen Li and Juhan Ross (see Wanner et al., 1995). The SSI is calculated as ln( fvol nir / fgeo red ), where fvol nir is the volume scattering kernel weight in the NIR band and fgeo red is the geometric kernel weight in the red band. It is based on the principle that surfaces with a higher (lower) and a relatively homogeneous (heterogeneous) vegetation cover will exhibit BRDFs that exhibit a behavior closer to (further from) that of an ideal turbid medium and will therefore generally have a higher (lower) volume scattering kernel weights in the near-infrared (NIR) wavelengths; while in the red wavelengths sparser, more clumped vegetation will exhibit higher kernel weights (Fig. 5.12). Since SSI depends on the vegetation cover fraction as well as vertical structure, a relative structural scattering index has been defined that partially removes the effects of cover by estimating a linear relationship between SSI and a spectral vegetation index. This approach has been found to be useful in exploiting canopy structure as translated through a kernel-driven model and should provide superior results for mapping land cover.

5.3.7 Geometric-optical and Hybrid Models The use of discrete object models such as geometric-optical (GO) and hybrid geometric-optical/radiative transfer (GORT) models represents another approach to exploiting multiangle data (Strahler et al., 2005). Geometric-optical (GO) models treat the surface as an assemblage of discrete, identical, and large objects defined by

5 Terrestrial Applications of Multiangle Remote Sensing

117

Structural Scattering Index

5 4 3 2 1 0 −1 0

0.2

0.4

0.6

0.8

1

Nadir View NDVI (SZN = 30) All Samples Open Shrubland (Jan-Mar) Barren/Sparsely Veg (Nov-Mar)

Evergreen Needleleaf Forest Open Shrubland (Apr-Oct) Barren/Sparsely Veg (Apr-Oct)

Fig. 5.12 Structural Scattering Index (SSI) vs nadir view NDVI from MODIS. The solid trend line represents the change of SSI as the function of nadir view NDVI. Samples above the trend line show the stronger volumetric scattering; those below show the stronger geometric scattering. SSI is also dependent on temporal changes in cover types (Reprinted from Gao et al., 2003. Copyright, Elsevier. With permission.)

geometric primitives (e.g., spheroids, cones, or cylinders) placed in a Poisson distribution above an underlying surface (Li and Strahler, 1985, 1992) which is frequently considered Lambertian (Scarth and Phinn, 2000). The remotely sensed observation is modeled as a linear combination of contributions from viewed sunlit and shaded crown and background components, with each contribution a product of component reflectance and the fraction of the sensor’s field of view occupied by the component Eq. (5.8): (5.8) R = C × k C + G × kG + T × kT + Z × kZ where C, G, T , and Z are the (assumed Lambertian) spectral reflectances or “component signatures” of the sunlit crown and background and shaded crown and background, respectively; and the kn are the component fractions. GO models vary considerably in complexity, with some including terms that allow for volumetric scattering within crowns rather than simple signatures (Ni and Li, 2000; Chen et al., 2003, 2005); and in the way the background contribution is represented (Ni and Li, 2000; Chopping et al., 2005). They hold the potential for obtaining upper canopy information such as stand density, openness, mean object size and canopy height, and crown morphology, which are useful in multiple disciplines (ecological modeling, forestry, hydrology, radiation budget; Garc´ıa-Haro et al., 2006). GO models have been used in the calculation of Clumping Index (see above);

118

M.J. Chopping

some results from applying a simple GO model to estimating woody plant cover in semi-arid environments in the southwestern USA are provided in the sections entitled The Background in Canopy Reflectance Modeling. Using lidar data obtained over forest in the boreal ecosystem-atmosphere study (BOREAS) sites in central Canada, Ni-Meister (2005) showed that lidar waveforms can be simulated accurately and with precision using a modified version of the Geometric-Optical/Radiative Transfer (GORT) canopy reflectance model (Li et al., 1995), driven with canopy structure inputs. This model is dependent on canopy geometry parameters such as tree size, shape, and tree density; and on the spectral reflectance properties of leaves and the BRDF of the background. The GORT model was designed for use with passive remote sensing data and is appropriate for the interpretation of multiangle data. Since it can be used to relate canopy parameters in both the active and passive domains it can be used as a bridge between large footprint lidar and moderate resolution multiangle data. This is a particularly promising direction.

5.3.8 Direction and Wavelength The future will certainly see greater use made of the directional signal in many wavelengths as the advantages of spectro-directional remote sensing become more apparent (Garc´ıa-Haro et al., 2006; Schaepman, 2006). However, most use is still made of red and NIR band data, mainly because it is in these regions that the dominant signatures in both spectral and angular domains are see; partly because these have long been used in remote sensing of vegetation; partly because MODIS and MISR both provide moderate resolution data (250–275 m) in these bands; and partly because interactions with the canopy are better understood. As seen above, several groups working with MISR data have adopted modeling approaches in which only the red band reflectance data are used. This is predicated on the following arguments: • The single scattering regime is dominant in the red spectral domain. • It exploits the reflectance and absorption contrasts between vertically clumped vegetation and the background. • The linear mixture assumption underlying geometric-optical models is more valid for the red than near infra-red wavelengths. • BRDF model inversion experiments using numerical methods show that there are generally fewer problems such as trapping at local minima in the red compared to the near-infrared (Gemmell, 2000). • It keeps the modeling problem (relatively) simple. On the other hand, there have been attempts to investigate and exploit the spectral and directional domains simultaneously (Weiss et al., 2000; Gobron et al., 2000, 2002; Bacour et al., 2002; Casa and Jones, 2005; Bach et al., 2005; Garc´ıa-Haro et al., 2006). Weiss et al. (2000) addressed the question of a potentially optimal set of

5 Terrestrial Applications of Multiangle Remote Sensing

119

spectral bands and directions for retrieval of biophysical parameters using the Scattering by Abitrarily Inclined Leaves (SAIL) radiative transfer model and the PARCINOPY nested radiosity/turbid medium model code. These were used to simulate multiangle remote sensing data in the principal and perpendicular planes (32 directions) for a solar zenith angle of 45◦ and for a wide range of canopy conditions. They found that for estimating the primary variables LAI (leaf area index) and Cab (chlorophyll a + b content), the optimal number of bands is approximately six and the optimal number of directions for these bands is between four and seven, with six directions or less required for reliable retrievals of the primary variables plus fractional vegetation cover and fraction of absorbed photosynthetically active radiation (fAPAR). Additional directions were found to contain redundant information and so their use would merely induce noise in the retrieval process. The ideal directions would be located in the principal plane close to the hot spot direction and in the forward scattering direction for large zenith angles. Note that these results assume idealized conditions (e.g., no cloud cover). Bacour et al. (2002) sought to utilize both the directional and the spectral information of the images acquired in 16 flights by the AirPOLDER instrument (eight bands in the visible-NIR) over the Alpilles test site (January–October 1997), as part of the Alpilles-ReSeDA campaign. Estimation of biophysical variables was executed by inversion of three one-dimensional radiative transfer models, SAIL, KUUSK and IAPI, coupled with the PROSPECT leaf optical properties model. They focused on the capability of model inversion to retrieve the leaf area index (LAI) of wheat, maize, sunflower and alfalfa crops. A quasi-Newton inversion algorithm was used to estimate the Cab , LAI, the mean leaf inclination angle (θl ), the hot spot parameter (sl ) and a multiplicative soil parameter (asoil ). The three models were shown to accurately estimate the LAI compared to planimeter measurements but tended to underestimate LAI for values above 2.3. Compensation effects between LAI and Cab emerged from the spatial analyses of these variables, although the uncertainties were low. Casa and Jones (2005) investigated an unconventional approach for the estimation of leaf area index (LAI) and leaf angle distribution (LAD), based inversion of a canopy ray tracing model against multiangle data from a ground-based multispectral camera. The model was developed using the Persistence of Vision Raytracer (POVRAY) and inversion was carried out using a look-up-table approach. Tests using an extensive data set gathered on a potato crop during experimental trials carried out at Viterbo (Italy) over 3 years showed that LAI was estimated with a RMSE varying from 0.29 to 0.75 in the different years. Bach et al. (2005) used data from the CHRIS instrument on the Proba platform to invert the four-stream Soil-Leaf-Canopy radiative transfer model SLC, an extension of the canopy reflectance models SAILH and GeoSAIL, producing maps of LAI, leaf chlorophyll, and leaf inclination distribution for agricultural fields in the Upper Rhine Valley test site along the German/French border for July 2003 (Fig. 5.13). The input parameters include structural and physiological information on the canopy, soil optical properties and the observation geometry. The Hapke model is used to model a non-Lambertian soil BRDF, including its variation with moisture content. The canopy itself is modeled with a two-layer version of the model SAILH

120

M.J. Chopping

>2.0 2.0-2.5 2.5.3.0 3.0-3.5 3.5-4.0 4.0-4.5 4.5-5.0 5.0-5.5 5.5-6.0

>35 35-40 40-45 45-50 50-55 55-60 60-65 65-70 70-75

Angle LIDF a 56⬚ -0.30 52⬚ -0.20 49⬚ -0.10

Fig. 5.13 LAI, chlorophyll per leaf area (µg/cm2 ) and average leaf angle maps for a 3 × 4 km area of maize, retrieved by adjusting the SLC model against data from CHRIS for the Upper Rhine Valley test-site along the German/French border (July 2003)

and transmittance of green and brown leaves is calculated using the PROSPECT sub-model. The leaf angle distribution is approximated by parameters that describe the average leaf slope and the “bimodality” of the distribution. The spatial distributions of all three retrieved parameters delineated fields and the additional information obtained from the directional data showed that canopy structure was crop-specific but also changed with phenological development. The field pattern in the average leaf slope map (Fig. 5.13c) values was thought to result from different varieties of maize. When the average leaf angle was retrieved for the same site 2 weeks after the initial CHRIS acquisition, the overall average angle retrieved changed by about 10◦ to a more vertical distribution. This can be interpreted as maturation of maize where leaves become more vertical with development. The study showed that crop parameters retrieved from multiangle CHRIS data can provide input parameters essential for crop growth models. Garc´ıa-Haro et al. (2006) developed an approach called directional spectral mixture analysis (DISMA) for retrieving vegetation parameters with the focus on fractional cover and leaf area index. This seeks to combine a consideration of the spectral signatures of soil and vegetation components with an analytical approximation of the radiative transfer equation, resulting in a fast, invertible model suitable for use with discontinuous canopies. Data from the AirPOLDER and HyMap instruments were used to test the model and its inversion using a lookup table (LUT). Retrievals of LAI corresponded well to ground measurements of LAI, with an RMSE of 0.5–0.6 and an R2 of the fitting of around 0.92. The spatial distribution matched that obtained by inversion of the physically-based New Advanced Discrete Model (NADIM) radiative transfer code, also known as the Semi-discrete model (Gobron et al., 1997) (Fig. 5.14).

5 Terrestrial Applications of Multiangle Remote Sensing

(a)

121

(b)

0.0 0.5 1.0 1.5 2.0 2.5 3.0 >3.5

Fig. 5.14 (a) Green LAI obtained via DISMA using AirPOLDER (b) the corresponding results from inversion of the physically-based NADIM radiative transfer code. The maps shows a 5 × 5 km area 28 km from Albacete, Spain; 1% and 7% of the pixels exceed the saturation value of 3.5 (From Garc´ıa-Haro et al., 2006. Copyright, IEEE. With permission.)

5.3.9 The Background in Canopy Reflectance Modeling Many researchers have recognized that for successful modeling of heterogeneous, clumped, or non-closed canopies it is important to account adequately for the contribution of the background of soil and understory (Gemmell, 2000; Ni and Li, 2000; Chen et al., 2005; Koetz et al., 2005b; Bach et al., 2005; Chopping et al., 2006a). In all canopies except closed – especially where the understory is heterogeneous – attempts to exploit canopy reflectance models can be confounded by spatial variation in background reflectance magnitude and anisotropy. In the last few years modelers have moved from the assumption of a Lambertian background, through imposing a single, static background BRDF, to attempts to estimate a spatially dynamic background BRDF. Recent work has shown that the soil/understory reflectance can be obtained for both coniferous and deciduous forests using MISR data, with the retrieved values following seasonal trajectories similar to those of adjacent grasslands, a partial validation of the approach (Canisius and Chen, 2007). Recent progress in dynamically estimating the anisotropic soil-understory contribution for discontinuous open shrub canopies in desert grasslands exhibiting a wide range of canopy-background configurations – including young, small honey mesquite shrubs over a dark grassland matrix and older, larger ones over bright, sandy soils – has been made using the simplified geometric-optical (SGM) model incorporating a kernel weighting approach with MISR data. Estimating the background contribution from MISR-derived LiSparse-RossThin model kernel weights while fixing number shrub density, shape and height at typical values and using numerical optimization to retrieve mean shrub radius has allowed the mapping of fractional shrub cover at landscape scales in the United States Department of

122

M.J. Chopping

Agriculture (USDA), Agricultural Research Service (ARS) Jornada Experimental Range in southern New Mexico. A root mean square error of 0.03 in fractional cover with respect to values estimated from high resolution IKONOS panchromatic imagery was obtained (measured fractional cover range: 0.03–0.19); however only a moderate proportion of the variation in the measured data was explained by the estimated data set, with an overall R2 of 0.19 (Chopping et al., 2006b). When the same algorithm was used to predict fractional shrub cover for pastures a few kilometers from the study area for which independent estimates were available (from image segmentation on a 0.6 m QuickBird panchromatic image), the distributions exhibited an improved spatial correlation (Fig. 5.15) and an R2 of 0.47 was obtained with two model variants (Fig. 5.16). When the algorithm was applied with MISR data over much larger areas – for the Jornada Experimental Range (∼783 km2 ) and the Sevilleta National Wildlife Refuge (∼1, 000 km2 ) and their environs – the maps include trees on the San Andres mountains, in other elevated areas (e.g., Summerford Mountain in the Jornada), and in the riparian environments of the Rio Grande (southwest quadrant of the Jornada map and center of the Sevilleta map), in addition to shrubs. The resulting distributions compare well with those of trees in the MODIS Vegetation

Fig. 5.15 (a) QuickBird shrub map for pasture 12 in the Jornada, red = shrub, white = background (b) shrub cover aggregated to 250 m cells; brighter = greater shrub cover (c) retrieved using MISR red band data to invert the SGM GO model (d) MISR/SGM shrub cover map for pastures 8/9 in the Jornada (e) the corresponding QuickBird map (From Chopping et al., 2006c. Copyright, American Geophysical Union.)

0.30

y = 0.4358x + 0.0208 R2 = 0.4676

0.20

0.10

0.00 0.00

++ + + + ++ + + + ++ + + + + + + + + ++ + + +++ + +++ + + ++ ++ + ++ +++ ++++ + ++ ++ + + + + + ++ ++ +

0.10

0.20

0.30

Fractional Shrub Cover (QB-NIH)

Fractional Shrub Cover (MISR/SGM)

Fractional Shrub Cover (MISR/SGM)

5 Terrestrial Applications of Multiangle Remote Sensing

123

0.30

y = 0.5427x + 0.0777 R2 = 0.4824

0.20

+

0.10 + + ++ +

0.00 0.00

++ + + + + + + + ++ ++ ++ + + + +++ ++ + + + + + + ++ + + + + + + ++ + +++ ++ + ++++ + + + + +++ ++ + + + ++ ++ ++ + ++ +

0.10

0.20

0.30

Fractional Shrub Cover (QB-NIH)

Fig. 5.16 Retrieved vs. measured shrub cover (a) SGM (b) SGM with sunlit crown unweighted by the fraction viewed (i.e., kC = 1.00) (Reproduced from Chopping et al., 2006c. Copyright, American Geophysical Union.)

Continuous Fields (VCF) percent tree cover maps (Fig. 5.17). Note that the MISRderived map includes all large woody plants: it is based on exploiting the canopy structure information encapsulated in the multiangle reflectance data (Chopping, 2006). Subsequent inversions of the SGM that allowed the crown shape parameter to vary indicated that it is possible to obtain regional maps of canopy height as well as crown cover, allowing estimates of aboveground woody biomass. Retrievals of cover, canopy height, and biomass showed good matches with US Forest Service maps, with coefficients of determination of 0.78, 0.69, and 0.81, and absolute mean errors of 0.10, 2.2 m, and 10.1 Mg/ha, respectively, after filtering for high model fitting error, the effects of topographic shading, and a small number of outliers (Chopping et al., 2007; http://csam.montclair.edu/∼chopping/wood/).

5.3.10 Land Cover and Community Type Mapping Many studies have shown that there is much potential for improving the accuracy of land cover classification if patterns in the angular domain as well as the spectral domain can be exploited (Abuelgasim et al., 1996; Hyman and Barnsley, 1997; Bicheron et al., 1997; Sandmeier and Deering, 1999). Chopping et al. (2002) showed that the kernel weights of Li-Ross LiSK models obtained by adjusting a LiSparse-RossThin variant against accumulated multiangle data from the NOAA AVHRR provide contingency tests for community types in semi-arid environments in Inner Mongolia Autonomous Region (IMAR; 10 types) and New Mexico (NM; 19 types) that are superior to those obtained using maximum-value compositing using the Normalized Difference Vegetation Index (NDVI) as the criterion (denominated MVC), and in particular perform better in the worst case: the kernel weights

124

M.J. Chopping

Fig. 5.17 Woody plant fractional crown cover map obtained by adjusting the SGM geometricoptical model against MISR red band data (a) in the Sevilleta National Wildlife Refuge in central New Mexico (b) % tree cover from the corresponding MODIS VCF product (c) in the Jornada Experimental Range in southern New Mexico, dotted line is the tree-shrub boundary from the VCF map (d) % tree cover from the corresponding VCF map product. Solid lines indicate the boundaries of the Sevilleta and fencelines in the Jornada. (Reproduced from Chopping et al., 2006c. Copyright 2006 American Geophysical Union.)

provided minimum reliabilities of 83% and 69% for IMAR and NM experiments, respectively, compared to only 11% and 26%, respectively, with the MVC data set. These performances were reflected in Kappa Index values of 0.93 and 0.91 for the IMAR and NM spectro-directional data sets against 0.74 and 0.46 for the MVC data sets, respectively (the Kappa Index is a means to test two data sets to determine the extent to which their similarities or differences are due to chance). When directional

5 Terrestrial Applications of Multiangle Remote Sensing

125

information was incorporated into the signatures, errors of omission or commission fell from ∼24% and ∼51% obtained with MVC to only ∼6% and ∼8% for the IMAR and NM experiments, respectively. More robust cross-validated tests of classifications of multiangle remote sensing metrics from MISR and nadir-spectral data from the An (nadir) camera have demonstrated that using multi-angular data and anisotropy patterns raises the overall classification accuracy importantly. Su et al. (2007) studied maximum likelihood and support vector machine (SVM) algorithms for mapping community types in the Jornada Experimental Range and the Sevilleta National Wildlife Refuge, New Mexico (19 classes). Half of the samples were randomly selected as the training set and the other half retained as the testing set. A total of 66 classifications were performed with various combinations of data sets: the ρ0 , k, and b parameters of the MRPV model (see Canopy Openness, above); the isotropic, geometric and volume scattering kernel weights of a Li-Ross BRDF model; the structural scattering index; and MISR surface reflectance estimates. Using multiangle data raised the overall classification accuracy from 45.4% obtained with nadir observations only to 60.9%, and with surface anisotropy patterns derived from MRPV and LiSparseRossThin BRDF models (separately) an overall accuracy of 67.5% can be obtained with a maximum likelihood classifier. Using a non-parametric SVM algorithm the classification accuracy was increased to 76.7%. Note that the classes in these experiments are community types that often differ very subtly in terms of their spectral signatures, rather than broad land cover types (Fig. 5.18).

Fig. 5.18 Community type maps for the Sevilleta National Wildlife Refuge. (a) LTER vegetation map (b) MISR maximum-likelihood classification (c) MISR support vector machine classification (Reproduced from Su et al., 2007. Copyright, Elsevier. With permission.)

126

M.J. Chopping

5.4 Other Applications Most applications – or prototype applications – of multiangle remote sensing of land have been concerned with vegetation canopies. However, there are other important Earth Science applications that are often tangentially related to vegetation, including estimating land albedo and the Earth’s radiant energy budget (see Chapter 9); snow and ice; and mapping dust emission sources to simulate the likelihood of large emissions.

5.4.1 Snow and Ice Multiangle remote sensing data are able to provide important additional information on Earth surfaces covered with snow and ice. For example, differentiating between clouds and snow or ice surfaces using spaceborne detectors is difficult because the surface may often be as bright and as cold as the overlying clouds, and because polar atmospheric temperature inversions sometimes mean that clouds can be warmer than the underlying surface. Mega-sastrugi ice fields in East Antarctica, with dunelike features as high as 4 m and separated by 2–5 km – a result of unusual snow accumulation and redistribution processes influenced by the prevailing winds and climate conditions – appear more like cloud formations: they exhibit a rippled appearance. However, the Angular Signature Cloud Mask (ASCM), a MISR product (Di Girolamo and Wilson, 2003) is able to detect clouds over snow and ice as well as over ocean and land (Fig. 5.19). MISR imagery indicated that these mega-sastrugi were stationary surface features between 2002 and 2004. (a)

708 backward viewing camera

(b)

confidence high low low high cloud

surface

Fig. 5.19 (a) December 16, 2004 MISR image over Antarctica, showing sastrugi (b) colorcoded image showing the Angular Signature Cloud Mask results (Courtesy of the MISR Team, NASA/JPL/Caltech, and L. Di Girolamo and M.J. Wilson, University of Illinois.)

5 Terrestrial Applications of Multiangle Remote Sensing

127

Hendriks and Pellikka (2004) presented a study of the multiangular reflectance of a glacier surface. They used multiple ALTM (Optech, Ontario, Canada) digital metric camera frames acquired from the air on August 12, 2003, to examine the angular spectral reflectance response of snow, firn, and ice surfaces on the Hintereisferner glacier in Austria. The camera was operated in color infrared (CIR) mode, resulting in images in three spectral bands: green (510–600 nm), red (600–720 nm) and near-infrared (720–800 nm); the pixel size varied between 0.25–0.30 m as a result of topography. They found important differences in the angular signatures of the three surface types and compared their results to MISR imagery (local mode: nominal 275 m footprints at all angles and in all four bands). Their results showed that MISR data reveal the backscattering of dirty ice, firn and old snow; BRF increases going from the nadir image in the backward direction and at the viewing camera Ca (60◦ ) backwards, BRF is 30% higher relative to the value in the nadir viewing camera (An) (Fig. 5.20). Multiangle views provide unique sub-pixel resolution information about the ice sheet surface that can be used to improve the characterization of climate and ice dynamics processes. Nolin et al. (2002) had previously shown that MISR data can be used as a proxy for surface roughness. They developed a normalized difference angular index (NDAI) using a combination of forward and backward scattered radiation in the MISR red band at the 60◦ fore and aft viewing angles of the MISR instrument. A positive (negative) NDAI value indicates that backward (forward) scattering exceeds forward (backward) scattering and that the surface is rough

0.52 0.47 0.42

Firn/Old snow

BRF

0.37 0.32 0.27

Dirty ice

0.22 0.17 0.12

Cf

Bf

Af

An

Aa

Ba

Ca

60

45.6

26.1

0

26.1

45.6

60

View angle (degrees)

Fig. 5.20 MISR BRFs for dirty ice and firn/old snow. The green, red and NIR bands are shown by the green, red and blue lines, respectively, for the Hintereisferner glacier in Austria, August 14, 2003 (Courtesy of J.P.M. Hendriks and P. Pellikka.)

128

M.J. Chopping

Fig. 5.21 Ice sheet surface classes for 2000–2005 derived from ISODATA classification of summer NIR albedo and NDAI from MISR images acquired in the vicinity of the Jakobshavn glacier and its upland drainage basin. Contour lines of elevation are at 250 m intervals (Reprinted from Nolin and Payne, 2006. Copyright, Elsevier. With permission.)

(smooth). The NDAI was shown to be highly correlated with surface roughness derived from an airborne laser altimeter. Nolin and Payne (2006) used the ISODATA clustering technique with NIR albedo and NDAI values from surface hemisphericaldirectional reflectance for consecutive MISR summer images from 2000–2006 over an area in the vicinity of the Jakobshaven glacier in Greenland. The classified maps (Fig. 5.21) demonstrated good spatial and temporal consistency for seven ice sheet classes for all 6 years; moreover, the classes roughly correspond with glacier facies mapped previously by other researchers. The classes differ in albedo, roughness (including the presence of crevasses), wetness, and the age of the snow at the surface. Nolin (2004) performed a study using data from MISR to demonstrate how the angular pattern of reflectance from vegetation over snow can provide information on forest cover density. This is important in snow studies as vegetation structure and density affect the dynamics of snow accumulation and ablation and affects the ability to estimate snow-covered area accurately from satellite-based sensors. The study area was located in north-central Colorado. MISR red band level 1B2 (top-of-atmosphere radiometrically and geometrically calibrated spectral radiances) data from 15 February 2002 were converted to top-of-atmosphere bidirectional reflectance factors – no atmospheric correction was applied. The Rahman– Pinty–Verstraete (RPV) semi-empirical parametric model was successfully used to simulate the angular patterns of reflectance. The model’s k parameter was used to characterize the angular signatures of selected pixels. In the RPV model this parameter is used to quantify the degree by which the observed bi-directional reflectance

5 Terrestrial Applications of Multiangle Remote Sensing

129

2 71%

1.5

93%

Normalized BRF

1

99%

0.5 0 −0.5 −1 −1.5 −2 −2.5 −80

−60

−40

−20

0

20

40

80

60

MISR camera angle Fig. 5.22 MISR angular signatures for three forest cover densities of fir–spruce forest. Note the change in shape with increasing tree cover; this is captured in the RPV model’s k parameter (Reprinted from Nolin, 2004. Copyright, Wiley. With permission.)

factor data resemble a bowl- or bell-shaped pattern. The results showed distinct patterns in the retrieved k parameter values, with a marked dependence on density and cover type. Non-forested areas exhibited a bowl-shaped pattern (k < 1.0) of reflectance versus viewing angle. Low-density deciduous forests also exhibited this bowl-shaped reflectance pattern, changing as density increases. Other forest cover types show transitional patterns between bowl and bell shapes and distinct bellshaped patterns (k > 1.0) for higher densities (Fig. 5.22). The relationship between k and density does not hold for forest cover densities that approach 100%. For a density of 99%, the fir – spruce forest cover type has a distinct bowl shape and a k value of only 0.69. This is in agreement with previous work indicating that sub-pixel homogeneity (whether because of sparse vegetation cover or extremely dense vegetation cover) will result in k < 1.0. This study indicated from a qualitative standpoint that multiangle reflectance data captures information on forest cover density at the sub-pixel scale.

5.4.2 Dust Emissions Laurent et al. (2005) used POLDER multiangle data with a surface roughness parameter estimated from the Roujean kernel-driven model (Roujean et al., 1992). They used the geometric kernel weight normalized by the diffuse (isotropic) scattering kernel weight (i.e., geo/iso) to simulate dust emissions from deserts in East Asia

130

M.J. Chopping

(N. China and Mongolia). The approach was derived from work previously accomplished by Roujean et al. (1997) and employed an empirical relationship established by Marticorena et al. (2004). The composite surface roughness map they developed describes the spatial variability of the erosion threshold (10 m wind velocities) on dust emission frequency. The map was found to be consistent with geomorphologic interpretations from Landsat imagery and with soil properties described in the literature. The retrieved roughness lengths are in agreement with the roughness lengths experimentally determined over similar surface types in other deserts of the world. The authors computed dust emission frequencies for 3 years (1997–1998–1999) by combining the 10 m erosion threshold wind velocities, the European Centre for Medium-Range Weather Forecasts (ECMWF) surface wind fields, the snow depth and the soil moisture computed using the Food and Agriculture Organization of the United Nations (FAO) soil texture profiles, and ECMWF meteorological data. The simulated frequencies of significant dust emissions were compared to the frequencies of occurrence of Total Ozone Mapping Spectrometer (TOMS) Absorbing Aerosol Index (AAI) higher than 0.7. Both the location and the relative intensity of the highest dust emission frequencies identified from the simulations were in agreement with the observations (Fig. 5.23). 1

Simulated dust event frequency

0.8

0.6

0.4

0.2

0 0

0.2

0.4

0.6

0.8

1

Frequency of TOMS AAI > 0.7

Fig. 5.23 Monthly frequencies of significant simulated dust emissions (flux > 10−10 g cm−2 s−1 ) as a function of the monthly frequencies of TOMS AAI > 0.7 over the Taklimakan desert for the 3 years 1997, 1998, and 1999. Small dots represent individual data; circles represent the averaged frequency of simulated dust emissions for classes (5% width) of frequency of TOMS Absorbing Aerosol Index (AAI) > 0.7; the solid line represent the linear fit of the averaged data (without accounting for the last class which is not representative) (Reproduced from Laurent et al., 2005. Copyright, American Geophysical Union.)

5 Terrestrial Applications of Multiangle Remote Sensing

131

5.5 Some Considerations in Multiangle Acquisition 5.5.1 Near-simultaneous and Accumulated Sampling Near-simultaneous acquisition of multiple observations of the same area on the Earth’s surface is possible by viewing in the along-track direction. From low Earth orbit, multiple looks of the same area can be acquired over a time span of less than 10 minutes. This strategy may be contrasted with accumulated sampling, in which repeat multiangular coverage of a particular area is obtained with a wide cross-track swath and observations are accumulated during multiple orbits. Accumulation of multiangle data in this manner takes many days. This may degrade the angular data set to some degree, depending on how rapidly surface conditions change and – to some extent – on changes in the atmosphere, since it is more difficult to correct extreme conditions. For along-track systems, there will be delay of only a few minutes between the first and last acquisitions, whether the instrument employs multiple sensors (MISR), is tilted (CHRIS/Proba), has a conical scan (AATSR), or has a very wide field-of-view (POLDER). This lag is unavoidable unless a constellation of geostationary platforms is available – an unlikely scenario for the near future. Most Earth observing instruments for which the effects of the surface BRDF have been studied and – to a far lesser extent – exploited are across-track instruments on polar-orbiting instruments such as the AVHRR and MODIS. These systems are designed to view large swaths, covering the globe at least once per 24 h period. They are intrinsically off-nadir viewing devices but to be used as multiangle instruments, observations of the same land areas must be accumulated from multiple overpasses. In addition, sensors such as Spinning Enhanced Visible and Infra-Red Imager (SEVIRI) on Meteosat Second Generation (MSG) geostationary satellites provide angular sampling through diurnal variation in illumination angles. The motivation for BRDF studies with accumulating instruments has been twofold: the need to adjust observations to a common Sun-target-sensor geometry for consistency (a much-touted but actually rather rare quality in Earth Observation in general); and to obtain more accurate estimates of terrestrial albedo than could possibly be obtained from a single view, since sunlight is never scattered in a perfectly diffuse manner at the surface. In the case of the MODIS BRDF/Albedo algorithm, an additional product is Nadir-equivalent BRDF Adjusted Reflectance (NBAR), defined as the best estimate of reflectance at nadir viewing at the mean solar zenith angle of the observations used to invert the Li-Ross BRDF model (in this case LiSparseMODIS-RossThick). NBAR was quickly found to provide superior performance in applications such as land cover mapping (Zhang et al., 2000). The possibility of obtaining additional surface information by exploiting the directional signal encapsulated in these data was something of an afterthought until the 2000s; however, in the last 5 years this has changed, with several studies addressing the retrieval of vegetation structural parameters such as LAI and fractional vegetation coverage (Camacho-de Coca et al., 2002; Chen et al., 2005).

132

M.J. Chopping

Accumulating instruments have one major advantage over those providing nearsimultaneous multiangular acquisitions: they provide greater spatial and temporal coverage. This is particularly important for global applications where cloud and cloud shadow contamination present a problem for instruments with revisit periods of less than a few days. However, near-simultaneous multiangle imaging provides some important advantages over accumulated sampling: • It greatly helps to solve the aerosol scattering problem over land and especially over bright surfaces – including deserts – where spectral information alone currently appears to be inappropriate. • Directional data sets acquired by accumulation may afford serious limitations for applications under those conditions where either the atmosphere or the surface – or both – change more rapidly than the accumulation period. This advantage is more applicable in some circumstances than in others and is probably more dependent on the development of the surface rather than that of the atmosphere. For example, changes owing to snowfall, snowmelt, flood, and fire may not be resolvable by accumulated sampling as they happen much more rapidly than the sampling rate. Near-instantaneous imaging is therefore the only way to assess surface directional properties under these circumstances. In other cases, for example, where changes in surface conditions are slow and/or sparse and intermittent, accumulated sampling is more acceptable. • There is the potential to assess information on canopy structure and retrieve LAI and fPAR with greater accuracy (Hu et al., 2003).

5.5.2 Angular Sampling The angular sampling of both multiangle imaging and accumulating instruments is determined by sensor design (field-of-view/swath), the platform on which the sensor is flown, and by the constraints imposed by the orbit into which the platform is injected. The first and last of these are straightforward. However, that the platform itself can impact the angular sampling regime of a sensor has a meaning restricted to satellites that are able to re-orient themselves in order to acquire multiple looks. Currently this definition applies to only one system: CHRIS on the Proba-1 satellite. In the future, increasing use of constellations of small, agile satellites such as Proba might be expected to provide greater flexibility in angular sampling. One such system concept developed at NASA/GSFC, dubbed Leonardo-BRDF has not yet been pursued (Wiscombe, 2000). The angular sampling provided depends on whether the instrument is an across-track scanner (using a revolving or oscillating mirror; e.g., AVHRR, MODIS, the forthcoming US National Polar-orbiting Operational Environmental Satellite System (NPOESS) Visible/Infrared Imager/Radiometer Suite, or VIIRS); an along-track scanner ((A)ATSR(−2)); multiple along-track moderate swath CCD arrays (MISR); or an along-track wide swath CCD array (POLDER/Parasol). Important issues with respect to angular sampling include:

5 Terrestrial Applications of Multiangle Remote Sensing

133

• The range of viewing and illumination zenith angles • The azimuthal coverage of the samples (how well the hemisphere is covered, particularly with respect to the principal and perpendicular planes) • The angular sampling density, particularly around the hot spot (rarely achieved) • The spatial resolution, determined by the size of the solid angle of the individual observing element, which is somewhat large for POLDER and much smaller for MISR and CHRIS • The interval over which the sampling is acquired (minutes for true multiangle imagers such as MISR, CHRIS, and POLDER; up to weeks for accumulators) • The frequency with which a given sampling may be repeated Since most instruments are flown on polar-orbiting platforms with sun-synchronous orbits having orbital inclinations only somewhat greater than 90◦ and acquisitions are made either side of local noon, across-track scanners tend to provide a sampling that is closer to the principal plane – where variation in radiance with viewing and illumination zenith is greatest – for mid-latitudes. Along-track imagers on platforms with sun-synchronous orbits such as MISR provide a sampling that is often far from the principal plane but approaches it with increasing latitude; and for any given latitude the sampling remains consistent. Owing to the Proba-1 orbit and the narrow field-of-view, CHRIS provides a less regular sampling from site to site and it is only occasionally able to image a target area from directly overhead view; Proba-1 has to be tilted at some – usually small – angle in the across-track direction so that the target area is viewed. The platform acquires images of the target when the zenith angle of the platform is one of the following: ±55◦ , ±36◦ and 0◦ . This means that the angles at which images are acquired vary from site-to-site, depending on their positions with respect to the orbital track. The pattern with which any site is accessible to the platform varies at roughly 8-day intervals, but with some change in sampling because the orbit does not repeat exactly (Barnsley et al., 2004). The POLDER design allows the most comprehensive sampling of the Earth’s radiation field from space to date. The very wide field-of-view (2, 400 × 1, 800 km2 ) allows observation of the same target under many different angular configurations (between 10 and 15 observations for each passage of the satellite), with view zenith angles up to ∼60◦ for the full azimuth range (Laurent et al., 2005).

5.5.3 Scale and Multiangle Observation The spatial scale at which multiangle observations over land are made has an important impact on the kinds of applications in which they might be used, as well as on data transmission rates and geographic coverage. Progress in remote sensing has often been measured in terms of increasing the sampling rate in the spatial domain, so that the smaller the ground sampling distance and the higher the spatial resolution, the better (NASA, 2000). Indeed, this is often postulated to be the major limitation of moderate resolution remote sensing devices: while wall-to-wall coverage is provided, heterogeneity in surface features cannot be resolved, limiting many applications,

134

M.J. Chopping

including cover type classification and the assessment of small-scale but widespread fragmentation in a variety of ecosystems (Phinn et al., 1996). However, this paradigm is flawed in certain respects: although high resolution imagers show the details of land features, this comes at the price of high data volumes monetary expense, and restricted coverage. Although multi-angle observation can only provide statistical metrics on sub-pixel structure, this is often sufficient to address the problem of interest. Moreover, data volumes are smaller and global coverage is achievable. For example, an hypothetical 10 m resolution single-angle imager generates ten times more data than a 100 m multiangle imager having ten angles, for the same swath; it can therefore cover a swath ten times wider for the same data volume. Note too that with multiangle remote sensing a small ground-projected instantaneous field-of-view (GIFOV) can be a disadvantage; specifically, for the radiometric, or 1-D, approach that seeks to derive “sub-pixel” information via variation in radiance with Sun and/or viewing angles and in which each “pixel” is treated independently. In this approach, if the GIFOV is small in relation to surface features such as trees or shrubs, there is a higher probability that an observation will include a sampling of elements that is unrepresentative of the surface as a whole; furthermore, it may also include important contributions from partial features at the extremities. These may result in noise-like fluctuations and adjacency effects in the angular signal that will impact on models that treat the surface as either a semi-infinite, homogenous, plane–parallel medium, or as a set of identical objects distributed spatially in a Poisson distribution (e.g., geometric-optical models). In both cases, an unrepresentative sampling of features and/or a large relative contribution from partial features at the edge of the GIFOV will obscure the angular signal. If the GIFOV is large in relation to surface features then the sampling of surface elements will be more representative and the periphery will make a relatively small contribution. Other considerations are that a large GIFOV may result in greater proportions of observations with unavoidable cloud and cloud shadow contamination and with mixtures of too many surface types. On the other hand, a very small GIFOV may also suffer from multiple scattering to/from adjacent footprints (“pixel cross talking”). Clearly, for any given landscape there is an optimum sampling resolution: if the GIFOV is much larger than the typical length scales then spatial trends in BRDF and parameters derived from model inversions will be less well-defined. There is also a sampling resolution that is the global optimum; i.e., over all surface types of interest. Following Pinty et al. (2002, 2004), Widlowski et al. (2005) addressed the question of the degree to which 1-D radiative transfer models can explain as well as describe the reflectance fields of 3-D forest targets of varying composition and complexity, over a range of spatial sampling scales (sensor ground sampling resolutions). Explaining these reflectance fields requires that the model’s internal parameters (state variables) match and are consistent with those of the 3-D target; it is not sufficient that the modeled anisotropy matches that of the target. Both conditions must hold if any model is to provide surface parameters on inversion which have meaning and utility in remote sensing applications; if they are not, there is considerable doubt on the reliability of surface parameters retrieved. The study tackled this problem through addressing two important questions:

5 Terrestrial Applications of Multiangle Remote Sensing

135

• Can 1-D models represent the anisotropy of a reflectance field generated by a more complex 3-D model at multiple spatial resolutions? and • Will the inversion of a 1-D model (at low spatial resolution where they are more appropriate) lead to systematic (and significant) biases in the retrieval of surface properties? These questions were pursued by using data sets for a level Scots pine landscape simulated by the Rayspread 3-D Monte Carlo raytracing model, using allometric relationships for a range of canopy parameter configurations in the red and near infra-red wavelengths. A 1-D model known as the Semi-discrete model (Gobron et al., 1997) was used to generate a large series of lookup tables for the comparisons. The study examined the impact of changing sensor GIFOV and foliage clumping on the match between the 3-D and 1-D models, as well as differences in interception, absorption and transmittance of solar radiation through and by the canopy arising from the 1-D and 3-D modeling. The conclusions were that 1-D models can provide good matches to reflectance fields generated by 3-D models in terms of both magnitude and directionality but that it does not follow that the 1-D model’s internal parameters and radiation budget match those of the 3-D target. It was shown that differences in the shape of the reflectance anisotropy between a 3-D target and its 1-D homologue – featuring identical spectral and structural properties with the exception of foliage clumping – lead to the retrieval of BRF-equivalent 1-D solutions with parameter values that diverge from those of the 3-D target. Part of this problem is understood as the problem of lack of uniqueness of solutions. A key finding is that for conifer forest canopies, an observation spatial scale of less than 100 m is more likely to introduce large discrepancies between the reflectance fields generated by 3-D and 1-D models for a wide range of conditions.

5.6 Conclusions This chapter has reviewed recent work towards exploiting solar wavelength remote sensing data acquired at multiple viewing angles from the air and space in applications in forestry, ecology, land cover mapping and land cover change, agriculture, hydrology, and glaciology, with important implications for enhancing ecological, crop growth, biogeochemical, hydrologic, and energy balance models. Even when excluding the numerous studies concerned with goniometric measurements, multiangle observations in the thermal wavelengths, and model development and simulation work, there has clearly been a great deal of activity over the last few years. Advances have been seen in a diversity of approaches, with notable gains for synergistic methods that seek to use multiangle data together with other kinds of remote sensing data and/or high resolution inventory data. The synergistic use of multiangle data with those from lidar instruments has provided an especially promising direction for the future, especially if the capability of models such as GORT to link multiangle and lidar data can be further developed. While canopy reflectance models have become more accurate, often incorporating a wider range of

136

M.J. Chopping

spectral measures, the hindrances to the use of geometric-optical models that have prevented widespread adoption in applications are being removed through methods for isolating the contribution of the soil/understory background from that of the upper canopy. Studies directed at assessing the gains from adopting the multiangle approach over nadir-spectral sensing, as well as obtaining meaningful and consistent metrics that characterize surface and canopy conditions, have universally found multiangle data to be valuable. Although efforts to exploit the additional, unique information available through multiangle remote sensing have been ongoing for many years, progress towards applications is accelerating, with high quality data sets available from MISR and MODIS on NASA’s EOS satellites, a third POLDER in orbit on Parasol, innovative new experimental systems such as CHRIS on Proba, greater demonstrated potential for satisfying multiple user groups in diverse disciplines (e.g., atmospheric science, ecology, glaciology, land management, and climate modeling), the realization of important synergies with active instruments, and a greater number of researchers engaged. This is borne out by the tenfold increase in the number of peer-reviewed publications using data from MISR, POLDER or CHRIS in the 10 years to 2005 (Fig. 5.24). Canopy reflectance modeling work with a variety of model types and improved reference data from high resolution imaging and lidar continues to shed new light on the constraints to robust retrievals of biophysical parameters and on important issues such as optimal scales of observation. Efforts continue to be made to engage remote sensing scientists with user groups (e.g., Chopping and Diner, 2005), although this is an ongoing task. Continued improvements in knowledge and understanding, together with greater experience with both near-simultaneous and

70 POLDER-1 + POLDER-2 + PARASOL

60

MISR

Number of publications

CHRIS

50 40 30 20 10 0 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005

Year

Fig. 5.24 Growth in peer-reviewed publications focused on exploitation of POLDER, MISR, and CHRIS, 1990–2005 (Courtesy of David J. Diner, MISR scientist, NASA/JPL.)

5 Terrestrial Applications of Multiangle Remote Sensing

137

accumulating multiangle instruments, will drive the design of future Earth observation systems. The examples presented in this chapter provide a good, albeit not exhaustive, indication of the range and depth of work that has been accomplished over recent years in just one area of multiangle remote sensing (terrestrial applications in the solar wavelengths). They reflect the progress achieved in the ten years since the NASA-sponsored Workshop on Multi-angular Remote Sensing for Environmental Applications (Privette et al., 1997) and point the way to further opportunities for making better use of the unique information provided by multiangle data over land. Acknowledgements Special thanks are owed to David J. Diner for his very valuable suggestions on the manuscript as well as the many others who helped me in compiling this review; and in particular Daniel Kimes, Alan Strahler, Bernard Pinty, Jean-Louis Roujean, Franc¸ois-Marie Br´eon, David Jupp, Jon Ranson, Jing Chen, John V. Martonchik, Anne Nolin, Rob Braswell, Julian Jenkins, Mike Barnsley, Wout Verhoef, Wenge Ni-Meister, Andres Kuusk, Soeren Hese, Hamlyn Jones, Raffaele Casa, F. Javier Garc´ıa Haro, Johan Hendriks, Petri Pellikka, Janne Heiskanen, Larry Di Girolamo, Mathias Disney, Wolfgang Lucht, Sampo Smolander, Gunar Fedosejevs, Mike Cutter, Gabriela Schaepman, and Michael Schaepman. I thank also the participants in the NASA/MISR Workshop on Multiangle Remote Sensing in Ecological Modeling not mentioned above. I must also acknowledge the many people not named here who sent me their recent and often unpublished work, and especially those whose work I could not incorporate here: I am deeply grateful. Any errors or omissions are uniquely mine.

Glossary 1-D 3-D AAI AATSR ADEOS ARS ASCM AVHRR BHR BOREAS BRDF BRF Cab CART CCD CHRIS CI CIR DISMA ECMWF EOS

One-dimensional Three-dimensional Absorbing Aerosol Index Advanced Along-Track Scanning Radiometer Advanced Earth Observation Satellite (Japan) Agricultural Research Service (USDA) Angular Signature Cloud Mask Advanced Very High Resolution Radiometer (NOAA) Bihemispherical Reflectance Boreal Ecosystem-Atmosphere Study (NASA) Bidirectional Reflectance Distribution Function Bidirectional Reflectance Factor Chlorophyll a + b content (leaves in a given canopy) Canopy Architecture Radiative Transfer (MISR LAI algorithm) Charge Coupled Device Compact High Resolution Imaging Spectrometer Clumping Index Color Infrared Directional Spectral Mixture Analysis European Center for Medium range Weather Forecasting Earth Observing System (NASA)

138

FAO fAPAR GIFOV GLAS GO GORT HDRF ICESat IFOV ISRO LAD LAI Lidar LiSK LVIS MISR MODIS MRPV MSG MVC NADIM NBAR NDAI NDVI NIR NDHD NPOESS POLDER POVRAY PSLV RAMI RMSE RPV RT SAIL SAR SEVIRI SGM SSI TOMS UK USDA VCF VIIRS

M.J. Chopping

Food and Agriculture Organization of the United Nations fPAR Fraction of Absorbed Photosynthetically-Active Radiation Ground-projected (sensor) Instantaneous Field-Of-View Geoscience Laser Altimeter System Geometric-Optical (model) Geometic-Optical/Radiative Transfer (model) Hemispherical-Directional Reflectance Factor Ice, Cloud, and Land Elevation Satellite (NASA EOS) (sensor) Instantaneous Field-Of-View Indian Space Research Organization Leaf Angle Distribution Leaf Area Index (one-sided leaf area per unit ground area) Light Detection and Ranging (sensor) Linear, Semi-empirical, Kernel-driven Laser Vegetation Imaging Sensor Multiangle Imaging Spectro-Radiometer (NASA/JPL EOS) MODerate resolution Imaging Spectroradiometer (NASA EOS) Modified Rahman-Pinty-Verstraete (model) Meteosat Second Generation Maximum-Value Compositing (on NDVI) New Advanced Discrete Model (model) Nadir BRDF-Adjusted Reflectances (EOS MOD43B4 product) Normalized Difference Angular Index Normalized Difference Vegetation Index Near Infra-Red (sometimes written near-infrared) Normalized Difference between Hotspot and Darkspot National Polar-orbiting Operational Environmental Satellite System Polarization and Directionality of the Earth’s Reflectance Persistence of Vision Raytracer Polar Satellite Launch Vehicle (ISRO; India) Radiation Transfer Model Intercomparison (RAMI) Exercise Root Mean Square Error Rahman-Pinty-Verstraete (model) Radiative Transfer Scattering by Abitrarily Inclined Leaves (model) Synthetic Aperture Radar Spinning Enhanced Visible and Infra-Red Imager Simplified Geometric-optical Model Structural Scattering Index Total Ozone Mapping Spectrometer United Kingdom of Great Britain and Northern Ireland United States Department of Agriculture Vegetation Continuous Fields (MODIS product) Visible/Infrared Imager/Radiometer Suite (NPOESS)

5 Terrestrial Applications of Multiangle Remote Sensing

139

References Abuelgasim AA, Gopal S, Irons JR, Strahler AH (1996) Classification of ASAS multi-angle and multispectral measurements using artificial neural networks. Remote Sens. Environ. 57(2):79–87 Asner GP (2000) Contributions of multi-view angle remote sensing to land surface and biogeochemical research. Remote Sens. Rev. 18:137–162 Bach H, Begiebing S, Verhoef W (2005) Analyses of hyperspectral and directional CHRIS data for monitoring of agricultural crops using the canopy reflectance model SLC. Proc. 9th Int. Symp. on Physical Measurements and Signature in Remote Sens. (ISPMSRS 2005), vol. I, Int. Soc. Photogrammetry and Remote Sens., 53–55 Bacour C, Br´eon F-M (2005) Variability of biome reflectance directional signatures as seen by POLDER. Remote Sens. Environ. 98:80–95 Bacour C, Jacquemoud S, Leroy M, Hautecoeur O, Weiss M, Pr´evot L, Bruguier N, Chauki H (2002) Reliability of the estimation of vegetation characteristics by inversion of three canopy reflectance models on airborne POLDER data. Agronomie 22:555–565 Barnsley MJ, Settle JJ, Cutter MA, Lobb DR, Teston F (2004) The PROBA/CHRIS mission: a lowcost smallsat for hyperspectral multiangle observations of the Earth surface and atmosphere. IEEE Trans. Geosci. Remote Sens. 42(7):1512–1520 Bicheron P, Leroy M, Hautecoeur O, Br´eon F-M (1997) Enhanced discrimination of boreal forest covers with directional reflectances from the airborne POLDER instrument. J. Geophys. Res. 102(D24):29517–29528 Br´eon F-M, Maignan F, Leroy M, Grant I (2002) Analysis of hot spot directional signatures measured from space. J. Geophys. Res. 107(D16):10.1029/2001JD001094 Bruegge CJ, Chrien NL, Ando RR, Diner DJ, Abdou WA, Helmlinger MC, Pilorz SH, Thome KJ (2002) Early validation of the Multi-angle Imaging SpectroRadiometer (MISR) radiometric scale. IEEE Trans. Geosci. Remote Sens. 40(7):1477–1492 Camacho-de Coca F, Garc´ıa-Haro FJ, Mart´ınez B, Meli´a J (2002) An operational strategy for retrieval of vegetation products from SEVIRI and AVHRR-3 data. Proc. LSA SAF Training Workshop, Ed. EUMETSAT, Darmstadt, Lisbon, July 8–10, pp 1–10 Camacho-de Coca F, Br´eon FM, Leroy M, Garcia-Haro FJ (2004), Airborne measurement of hot spot reflectance signatures. Remote Sens. Environ. 90:63–75 Canisius F, and Chen JM (2007) Retrieving vegetation background reflectance from Multi-angle Imaging SpectroRadiometer (MISR) data. Remote Sens. Environ. 107(1–2):312–321 Casa R, Jones HG (2005) LAI retrieval from multiangular image classification and inversion of a ray tracing model. Remote Sens. Environ. 98:414–428 Chen JM, Liu J, Leblanc SG, Lacaze R, Roujean J-L (2003) Multi-angular optical remote sensing for assessing vegetation structure and carbon absorption. Remote Sens. Environ. 84:516–525 Chen JM, Menges CH, Leblanc SG (2005) Global mapping of foliage clumping index using multiangular satellite data. Remote Sens. Environ. 97:447–457 Chopping MJ (2006) Progress in retrieving canopy structure parameters from NASA multi-angle remote sensing. Proc. Int. Geosci. Remote Sens. Symp. 2006 and 27th Canadian Remote Sens. Symp., Denver, CO, July 31–August 4, 2006 Chopping M, Diner D (2005) Workshop on Ecological Modeling using NASA Multi-angle Remote Sensing. The Earth Observer, NASA/GSFC, November–December 2005, 17(6):32–34 Chopping MJ, Rango A, Ritchie, JC (2002) Improved semi-arid community type differentiation with the NOAA AVHRR via exploitation of the directional signal. IEEE Trans. Geosci. Remote Sens. 40(5):1132–1149 Chopping MJ, Rango A, Havstad KM, Schiebe FR, Ritchie JC, Schmugge TJ, French A, Su L, McKee L, Davis RM (2003) Canopy attributes of Chihuahuan Desert grassland and transition communities derived from multi-angular airborne imagery. Remote Sens. Environ. 85(3):339–354

140

M.J. Chopping

Chopping M, Martonchik JVM, Rango A, Peters DPC, Su L, Laliberte A (2005) Geometric-optical modeling of desert grassland canopy structure with MISR. Proc. 9th Int. Symp. on Physical Measurements and Signature in Remote Sens. (ISPMSRS 2005), vol. I, Int. Soc. Photogrammetry and Remote Sens. 141–143 Chopping M, Su L, Rango A, Martonchik JV, Peters DPC, Laliberte A (2006a) Remote sensing of woody shrub cover in desert grasslands using MISR with a geometric-optical canopy reflectance model, ISPMSRS’05. Special Issue of Remote Sens. Environ. Chopping M, Su L, Laliberte A, Rango A, Peters DPC, Kollikathara N (2006b) Mapping shrub abundance in desert grasslands using geometric-optical modeling and multiangle remote sensing with CHRIS/Proba. Remote Sens. Environ. 104(1):62–73 Chopping, MJ, Su L, Laliberte A, Rango A, Peters DPC, Martonchik JV (2006c) Mapping woody plant cover in desert grasslands using canopy reflectance modeling and MISR data. Geophys. Res. Lett., 33, L17402, doi:10.1029/2006GL027148 Chopping M, Moisen G, Su L, Laliberte A, Rango A, Martonchik JV, and Peters DPC (2007) Large area mapping of southwestern forest crown cover, canopy height, and biomass using the NASA Multiangle Imaging SpectroRadiometer, Remote Sens. Environ, in press. Combal B, Baret F, Weiss M (2002) Improving canopy variables estimation from remote sensing data by exploiting ancillary information. Case study on sugar beet canopies. Agronomie 22:205–215 Di Girolamo L, Wilson MJ (2003) A first look at band-differenced angular signatures for cloud detection from MISR. IEEE Trans. Geosci. Remote Sens. 41(7):1730–1734 Diner DJ, Beckert JC, Reilly TH, Bruegge CJ, Conel JE, Kahn RA, Martonchik JV, Ackerman TP, Davies R, Gerstl SAW, Gordon HR, Muller J-P, Myneni RB, Sellers PJ, Pinty B, Verstraete MM (1998) Multi-angle Imaging SpectroRadiometer (MISR) instrument description and experiment overview. IEEE Trans. Geosci. Remote Sens. 36(4):1072–1087 Diner DJ, Asner GP, Davies R, Knyazikhin Y, Muller J-P, Nolin AW, Pinty B, Schaaf CB, Stroeve J (1999) New directions in Earth observing: scientific applications of multi-angle remote sensing. Bull. Am. Meteorol. Soc. 80(11):2209–2229 Diner DJ, Braswell BH, Davies R, Gobron N, Hu J, Jin Y, Kahn RA, Knyazikhin Y, Loeb N, Muller J-P, Nolin AW, Pinty B, Schaaf CB, Seiz G, Stroeve J (2005) The value of multiangle measurements for retrieving structurally and radiatively consistent properties of clouds, aerosols, and surfaces. Remote Sens. Environ. 97:495–518 Disney M, Lewis P, Saich P (2006) 3D modelling of forest canopy structure for remote sensing simulations in the optical and microwave domains. Remote Sens. Environ. 100:114–132 Engelsen O, Pinty B, Verstraete MM, Martonchik JV (1996) Parametric bidirectional reflectance factor models: evaluation improvements and applications. Report EUR16426EN European Commission Joint Research Centre Space Applications Institute, Ispra, Italy, 120pp European Space Agency (1999) Exploitation of CHRIS data from the Proba mission for science and applications. Experimenters’ Handbook Issue 4: Baseline Programme, Scientific Campaign Unit, ESA-ESTEC European Space Agency (2006) Achievements of the Proba mission: projects and publications at: http://earth.esa.int/pub/ESA DOC/PROBA/ProjectsandPublications.html Gao F, Schaaf CB, Strahler AH, Jin Y, Li X (2003) Detecting vegetation structure using a kernelbased BRDF model. Remote Sens. Environ. 86:198–205 Garc´ıa-Haro FJ, Camacho-de Coca F, Meli´a J (2006) A directional spectral mixture analysis method. Application to multiangular airborne measurements. IEEE Trans. Geosci. Remote Sens. 44:365–377 Gemmell F (2000) Testing the utility of multi-angle spectral data for reducing the effects of background spectral variations in forest reflectance model inversion. Remote Sens. Environ. 72:46–63 Gobron N, Pinty B, Verstraete MM, Govaerts Y (1997) A semi-discrete model for the scattering of light by vegetation. J. Geophys. Res. 102:9431–9446

5 Terrestrial Applications of Multiangle Remote Sensing

141

Gobron N, Pinty B, Verstraete MM, Martonchik JV, Knyazikhin Y, Diner DJ (2000) Potential of multiangular spectral measurements to characterize land surfaces: conceptual approach and exploratory application. J. Geophys. Res. 105(D13):17539–17549, 10.1029/2000JD900154 Gobron N, Pinty B, Verstraete MM, Widlowski J-L, Diner DJ (2002) Uniqueness of multiangular measurements part 2: joint retrieval of vegetation structure and photosynthetic activity from MISR. IEEE Trans. Geosci. Remote Sens. 40:1574–1592 Hansen MC, DeFries RS, Townshend JRG, Sohlberg RA, Dimiceli C, Carroll M (2002) Towards an operational MODIS continuous field of percent tree cover algorithm: examples using AVHRR and MODIS data. Remote Sens. Environ. 83:303–319 Heiskanen J (2006) Tree cover and height estimation in the Fennoscandian tundra–taiga transition zone using multiangular MISR data. Remote Sens. Environ. doi:10.1016/j.rse.2006.03.015 Hendriks JPM, Pellikka P (2004) Multiangular reflectance of a glacier surface. Presentation at the 2004 Joint Assembly (CGU, AGU, SEG & EEGS), Montr´eal, Canada, May 17–23, 2004 Hese S, Lucht W, Schmullius C, Barnsley M, Dubayah R, Knorr D, Neumann K, Riedel T, Schr¨oter K (2005) Global biomass mapping for an improved understanding of the CO2 balance – the Earth observation mission Carbon-3D. Remote Sens. Environ. 94:94–104 Hu J, Tan B, Shabanov N, Crean KA, Martonchik JV, Diner DJ, Knyazikhin Y, Myneni RB (2003) Performance of the MISR LAI and FPAR algorithm: a case study in Africa. Remote Sens. Environ. 88:324–340 Hyman AH, and Barnsley MJ (1997) On the potential for land cover mapping from multiple view angle (MVA) remotely-sensed images. In: Observations and Interactions: Proc. 23rd Annu. Conf. Exhibition Remote Sens. Soc., 135–140 Jenkins J, Ollinger S, Braswell R, Martin M, Plourde L, Smith M-L, et al. (2004) Detecting patterns of canopy structure and carbon uptake with multi-angle remote sensing. Presentation at MISR Science Team Meeting, Pasadena, CA, December 9, 2004 Kimes DS, Ranson KJ, Sun G, Blair JB (2006) Predicting lidar measured forest vertical structure from multi-angle spectral data. Remote Sens. Environ. 100(4):503–511 King MD, Kaufman YJ, Tanr´e D, Nakajima T (1999) Remote sensing of tropospheric aerosols from space: past, present, and future. Bull. Am. Meteorol. Soc. 80(11):2229–2260 Knyazikhin Y, Kranigk J, Myneni RB, Panferov O, Gravenhorst G (1998a). Influence of smallscale structure on radiative transfer and photosynthesis in vegetation cover. J. Geophys. Res., D: Atmospheres 103:6133–6144 Knyazikhin Y, Martonchik JV, Myneni RB, Diner DJ, Running SW (1998b) Synergistic algorithm for estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation from MODIS and MISR data. J. Geophys. Res. 103(D24):32257–32276 Knyazikhin Y, Martonchik JV, Diner DJ, Myneni RB, Verstraete M, Pinty B, Gobron N (1998c) Estimation of vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation from atmosphere-corrected MISR data. J. Geophys. Res. 103(D24):32239–32256 Koetz B, Kneub¨uhler M, Widlowski J-L, Morsdorf F, Schaepman M, Itten K (2005a) Assessment of canopy structure and heterogeneity from multi-angular CHRIS-PROBA data. 9th Int. Symp.on Physical Measurements and Signatures in Remote Sens. (ISPMSRS), Beijing, China, 2005, 76–78 Koetz B, Kneub¨uhler M, Schopfer J, Schaepman M, Itten K (2005b) Influence of changing background on CHRIS/Proba data over an heterogeneous canopy. Proc. EARSeL 4th Workshop on Imaging Spectroscopy, Warsaw, Poland, 1–9 Lacaze R, Chen JM, Roujean J-L, Leblanc SG (2002) Retrieval of vegetation clumping index using hot spot signatures measured by POLDER instrument. Remote Sens. Environ. 79:84–95 Laurent B, Marticorena B, Bergametti G, Chazette P, Maignan F, Schmechtig C (2005) Simulation of the mineral dust emission frequencies from desert areas of China and Mongolia using an aerodynamic roughness length map derived from the POLDER/ADEOS 1 surface products. J. Geophys. Res. 110(D18):D18S04, doi:10.1029/2004JD005013 Lavergne T, Kaminski T, Pinty B, Taberner M, Gobron N, Verstraete M, Vossbeck M, Widlowski J-L, Giering R (2007) Application to MISR land products of an RPV model inversion package using adjoint and Hessian codes. Remote Sens. Environ. 107(1–2):362–375

142

M.J. Chopping

Li X, Strahler AH (1985) Geometric-optical modeling of a conifer forest canopy. IEEE Trans. Geosci. Remote Sens. 23:705–721 Li X, Strahler AH (1992) Geometric-optical bidirectional reflectance modeling of the discrete crown vegetation canopy: effect of crown shape and mutual shadowing. IEEE Trans. Geosci. Remote Sens. 30(2):276–291 Li X, Strahler AH, Woodcock CE (1995) A hybrid geometric optical-radiative transfer approach for modeling albedo and directional reflectance of discontinuous canopies. IEEE Trans. Geosci. Remote Sens. 33:466–480 Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, Hoboken, NJ, 534 pp Liang S, Strahler AH, Barnsley MJ, Borel CC, Gerstl SAW, Diner DJ, Prata AJ, Walthall CL (2000) Multiangle remote sensing: past, present and future. In:, Liang S and Strahler AH (eds), The Special Issue on Land Surface Bidirectional Reflectance Distribution Function (BRDF); Recent Advances and Future Prospects, Remote Sens. Rev. 18:83–102 Liu X, Kafatos M (2007) MISR multi-angular spectral remote sensing for temperate forest mapping at 1.1-km resolution. International Journal of Remote Sensing 28(2):459–464 Maignan F, Br´eon F-M, and Lacaze R (2004) Bidirectional reflectance of Earth targets: evaluation of analytical models using a large set of spaceborne measurements with emphasis on the hot spot. Remote Sens. Environ. 90:210–220 Marticorena B, Chazette P, Bergametti G, Dulac F, Legrand M (2004) Mapping the aerodynamic roughness length of desert surfaces from the POLDER/ADEOS bi-directional reflectance product. Int. J. Remote Sens. 25:603–626 Martonchik JV, Bruegge CJ, Strahler A (2000) A review of reflectance nomenclature used in remote sensing. Remote Sens. Rev. 19:9–20 NASA (2000) NASA Earth Science Enterprise Research Strategy for 2000–2010 Ni W, Li X (2000) A coupled vegetation – soil bidirectional reflectance model for a semi-arid landscape. Remote Sens. Environ. 74:113–124 Ni-Meister W (2005) 3D vegetation structure extraction from lidar remote sensing. Proc. 9th Int. Symp. on Physical Measurements and Signature in Remote Sens. (ISPMSRS 2005), vol. I, Int. Soc. Photogrammetry and Remote Sens. 47–49 Nicodemus FE, Richmond JC, Hsia JJ, Ginsberg IW, Limperis T (1977) Geometrical Considerations and Nomenclature for Reflectance. In: National Bureau of Standards, US Department of Commerce, Washington, DC Nolin AW (2004) Towards retrieval of forest cover density over snow from the Multi-angle Imaging SpectroRadiometer (MISR). Hydrol. Process. 18:3623–3636 Nolin AW, Payne M (2006) Classification of glacier zones in Western Greenland using albedo and surface roughness from the Multi-angle Imaging SpectroRadiometer (MISR). Remote Sens. Environ. 107(1–2):264–275 Nolin AW, Fetterer FM, Scambos TA (2002) Surface roughness characterizations of sea ice and ice sheets: case studies with MISR data. IEEE Trans. Geosci. Remote Sens. 40(7):1605–1615 Phinn S, Franklin J, Hope A, Stow D, Huenneke L (1996) Biomass distribution mapping using airborne digital video imagery and spatial statistics in a semi-arid environment. J. Environ. Manage. 47:139–164 Pinty B, Gobron N, Widlowski J-L, Gerstl SAW, Verstraete MM, Antunes M, Bacour C, Gascon F, Gastellu J-P, Goel N, Jacquemoud S, North P, Qin W, Thompson R (2001) Radiation Transfer Model Intercomparison (RAMI) exercise. J. Geophys. Res. 106:11937–11956 Pinty B, Widlowski J-L, Gobron N, Verstraete MM, Diner DJ (2002) Uniqueness of multiangular measurements – Part I: an indicator of subpixel surface heterogeneity from MISR. IEEE Trans. Geosci. Remote Sens. 40(7):1560–1573 Pinty B, Widlowski J-L, Taberner M, Gobron N, Verstraete MM, Disney M, Gascon F, Gastellu J-P, Jiang L, Kuusk A, Lewis P, Li X, Ni-Meister W, Nilson T, North P, Qin W, Su L, Tang S, Thompson R, Verhoef W, Wang H, Wang J, Yan G, Zang H (2004a) RAdiation transfer Model Intercomparison (RAMI) exercise: results from the second phase. J. Geophys. Res. 109, D06210 101029/2003JD004252

5 Terrestrial Applications of Multiangle Remote Sensing

143

Pinty B, Gobron N, Widlowski J-L, Lavergne T, Verstraete MM (2004b) Synergy between 1-D and 3-D radiation transfer models to retrieve vegetation canopy properties from remote sensing data. J. Geophys. Res. 109, D21205 Privette JL, Deering D, Wickland D (1997) Report on the workshop on multiangular remote sensing for environmental applications. NASA Technical Memorandum 113202, 54 pp. Rahman H, Pinty B, Verstraete MM (1993) Coupled surface-atmosphere reflectance (CSAR) model. 2. Semiempirical surface model usable with NOAA Advanced Very High Resolution Radiometer data. J. Geophys. Res. 98:20791–20801 Roujean J-L, Leroy M, Deschamps P-Y (1992) A bidirectional reflectance model of the Earth’s surface for the correction of remote sensing data. J. Geophys. Res. 97:20455–20468 Roujean J-L, Tanr´e D, Br´eon F-M, Deuz´e J-L (1997) Retrieval of land surface parameters from airborne POLDER bidirectional reflectance distribution function during HAPEX-Sahel. J. Geophys. Res. 102:11201–11218 Sandmeier S, Deering DW (1999) Structure analysis and classification of boreal forests using airborne hyperspectral BRDF data from ASAS. Remote Sens. Environ. 69(3):281–295 Scarth P, Phinn S (2000) Determining forest structural attributes using an inverted geometricoptical model in mixed eucalypt forests, Southeast Queensland, Australia. Remote Sens. Environ. 71:141–157 Schaepman M (2006) Spectrodirectional remote sensing: from pixels to processes. Int. J. Appl. Earth Obs. Geoinformation 6(3–4):271–282 Schaepman-Strub G, Schaepman M, Painter T, Dangel S, Martonchik J (2006) Reflectance quantities in optical remote sensing – definitions and case studies. Remote Sens. Environ. 103(1):27–42 Smolander S, Stenberg P (2005) Simple parameterizations of the radiation budget of uniform broadleaved and coniferous canopies. Remote Sens. Environ. 94(3):355–363 Strahler AH, Jupp DLB, Woodcock CE, Li X (2005) The discrete-object scene model and its application in remote sensing. Proc. 9th International Symp. on Physical Measurements and Signatures in Remote Sens. 1:166–168 Su L, Chopping MJ, Rango A, Martonchik JV, Peters DPC (2007) Support vector machines for recognition of semi-arid vegetation types using MISR multi-angle imagery. Remote Sens. Environ. 107(1–2):299–311 Verhoef W (2005) Earth observation model sensitivity analysis to assess mission performances in terms of geo-biophysical variable retrieval accuracies. Proc. 9th Int. Symp. on Physical Measurements and Signature in Remote Sens. (ISPMSRS 2005), Beijing, October 17–19, 2005, vol. I, Int. Soc. Photogrammetry and Remote Sens. 324–327 Wanner W, Li X, Strahler AH (1995) On the derivation of kernels for kernel-driven models of bidirectional reflectance. J. Geophys. Res. 100:21077–21090 Weiss M, Baret F, Myneni R, Pragnere A, Knyazikhin Y (2000) Investigation of a model inversion technique for the estimation of crop characteristics from spectral and directional reflectance data. Agronomie 20:3–22 Widlowski J-L, Pinty B, Gobron N, Verstraete MM, Davis AB (2001) Characterization of surface heterogeneity detected at the MISR/TERRA subpixel scale. Geophys. Res. Lett. 28(24):4639–4642 Widlowski J-L, Pinty B, Gobron N, Verstraete M, Diner DJ, Davis B (2004) Canopy structure parameters derived from multi-angular remote sensing data for terrestrial carbon studies. Climatic Change 67:403–415 Widlowski J-L, Pinty B, Lavergne T, Verstraete MM, Gobron N (2005) Using 1-D models to interpret the reflectance anisotropy of 3-D canopy targets: issues and caveats. IEEE Trans. Geosci. Remote Sens. 43:2008–2017, doi10.1109/TGRS.2005.853718 Widlowski J-L, Taberner M, Pinty B, Bruniquel-Pinel V, Disney M, Fernandes R, GastelluEtchegorry J-P, Gobron N, Kuusk A, Lavergne T, Leblanc S, Lewis P, Martin E, Mottus M, North PJR, Qin W, Robustelli M, Rochdi N, Ruiloba R, Soler C, Thompson R, Verhoef W, Verstraete MM, Xie D (2006) The third RAdiation transfer Model Intercomparison (RAMI) exercise: documenting progress in canopy reflectance modelling. J. Geophys. Res. 112(D9):D09111. doi:10.1029/2006JD007821

144

M.J. Chopping

Wiscombe W (2000) Formation Flying, Leonardo, Balloon-ERB and Triana: Forays to the Frontier, NASA/NIAC workshop presentation, November 7, 2000, online at http://www.niac.usra. edu/files/library/meetings/misc/nov00 wks/wwiscombe.pdf, last access 06/14/06 Zhang X, Schaaf CB, Gao F, Friedl MA, Strahler AH, Hodges JCF (2000) Mapping land cover and green vegetation abundance using MODIS-like data: a case study of New England. Proc. Int. Geosci. Remote Sens. Symp. 2000: vol 5, pp 2005–2007

Part II

Physical Modeling and Inversion Algorithms

Chapter 6

Modeling the Spectral Signature of Forests: Application of Remote Sensing Models to Coniferous Canopies Pauline Stenberg, Matti M˜ottus, and Miina Rautiainen

Abstract The vertical and horizontal structure of forest canopies is one of the most important driving factors of various ecosystem processes and has received increasing attention during the past 20 years and served as an impetus for earth observation missions. In the remote sensing community, the variables which describe canopy structure are called biophysical variables, and are directly coupled with the fundamental physical problem behind remote sensing: radiative transfer in vegetation. There are basically three different approaches to interpreting biophysical variables from remotely sensed data: (1) empirical, (2) physically based, and (3) various combinations of them. The physical approach builds upon an understanding of the physical laws governing the transfer of solar radiation in vegetative canopies, and formulates it mathematically by canopy reflectance models which relate the spectral signal to biophysical properties of the vegetation. In this chapter, we will first outline the basic principles and existing physically based model types for simulating the spectral signature of forests. After this, the focus is on the specific issues related to applying these models to the complex 3D structure of coniferous canopies.

6.1 Introduction The assessment of many fundamental ecological questions at global scale is possible only through remote sensing, since integrated analyses of the biosphere, atmosphere and hydrosphere require simultaneous measurements over large areas. Pauline Stenberg Department of Forest Resource Management, University of Helsinki, Finland [email protected] Miina Rautiainen Department of Forest Resource Management, University of Helsinki, Finland Matti M˜ottus Tartu Observatory, T˜oravere, Tartumaa, Estonia S. Liang (ed.), Advances in Land Remote Sensing, 147–171. c Springer Science + Business Media B.V., 2008

147

148

P. Stenberg et al.

The three-dimensional (3D) structure of forest canopies is amongst the most important driving factors of various physiological and ecological processes. By analyzing the 3D structure of canopies, it is possible to detect growth patterns and phenological cycles of forests, as well as other changes such as insect or fire outbreaks and damages, illegal deforestation, and environmental stresses caused by drought or pollution. In the remote sensing community, the variables which describe canopy structure are called biophysical variables. Biophysical variables are coupled with the radiative transfer problem (Chandrasekhar, 1960) in vegetation (Ross, 1981) and can also be defined as state variables of the radiative transfer problem, in other words, the smallest set of variables which are needed to fully describe the physical state of the system at a given scale (Verstraete et al., 1996). However, mere description of the structure of the canopy through biophysical variables is not the only goal: the variables should be applicable for end-users in, for example, water and carbon cycle or climate modeling. It is obvious that measuring biophysical variables in situ is both laborious and time-consuming in forests – and impossible at global scale within a short time frame. Space-born monitoring is thus required, and algorithms for interpreting these variables from remotely sensed data need to be developed. Most common remotely sensed biophysical variables of forests include leaf area index (LAI), fraction of photosynthetically active radiation (400–700 nm) absorbed by vegetation (fPAR) and fraction of canopy cover (fCover). LAI and fCover are geometrical variables which are related to canopy gap fraction, i.e., the fraction of ground seen in a given direction (Nilson, 1977) – canopy gap fraction is in fact determined by LAI and its spatial distribution and leaf inclination distribution. fPAR, on the other hand, is an outcome of radiative transfer in vegetation, and the opposite of reflectance from vegetation. In addition, there are biophysical variables which are not geometric, but instead influence the spectral properties of scattering elements. An example of such variables is the chlorophyll content of green leaves (which also is closely connected to the nitrogen content of foliage). However, there is evidence that no clear distinction between foliar biochemistry and LAI can be made in practical remote sensing (Yoder and Pettigrew-Crosby, 1995). There are basically three different approaches: (1) statistical (empirical) and (2) physically based, and (3) various combinations of them (e.g., neural networks), to assess biophysical variables from spectral signals provided by optical satellite images. In the empirical approaches, commonly used in, for example, regional or national forest inventories, the vegetation characteristics of interest are estimated based on statistical relationships (regressions) obtained by collecting training data on the spectral signatures of a variety of objects. These methods are limited to a specified viewing and illumination configuration, and require large sets of reliable ground truth data. The physical approach, in contrast, builds upon an understanding of the physical laws governing the transfer of solar radiation in vegetative canopies and formulating it mathematically by reflectance models. Reflectance models, in turn, relate biophysical properties of the vegetation, or sets of canopy and stand parameters, to the spectral signal. Physically based methods are, at least in theory, more robust since they are not limited to a single configuration or vegetation biome type.

6 Modeling the Spectral Signature of Forests

149

Physically based methods have progressively become more and more attractive for the assessment or monitoring of biophysical characteristics of vegetation. They are the only methods which can fully take into use the versatile hyperspectral and multiangular information provided by the modern satellite sensors, and they are better suited for many current large scale applications than the empirical techniques. Several aspects must be considered when physically based models are developed for operational monitoring purposes. To begin with, the sensitivity of the model to the biophysical variable to be retrieved should be maximized. Simultaneously, undesired effects, for example from the atmosphere or terrain topography, should be eliminated as efficiently as possible using other available models and data. Finally, when considering the optimal model, the larger the area (e.g., the whole Earth), the more the model should rely on theory and the less on local data bases for input as the sets are usually not compatible from one region or country to another. In addition, we should take into account technical questions such as computation time required by the model. Canopy reflectance models were originally developed for agricultural crops, followed by broad-leaved forests, and even though presently there exist a number of reflectance models designed to be applicable also to conifers, some missing “difficult-to-model” properties and insufficient empirical data on some key parameters have still limited their optimal use. Several recent studies (Rautiainen and Stenberg, 2005a; Smolander and Stenberg, 2005) give quantitative support to the hypothesis put forward a long time ago (Norman and Jarvis, 1975; J. Ross, 1981– 2002, personal communication), that a major reason for the distinct spectral signature of coniferous forests lies in their hierarchical grouped structure, which governs the processes of multiple scattering within the canopy, as well as canopy absorption. Moreover, modeling tools are emerging by which grouping at different scales can be accounted for in canopy reflectance calculations. The effect of grouping on canopy PAR absorption and photosynthesis has been a well studied subject in forest production ecology (Oker-Blom, 1986; Leverenz and Hinckley, 1990; Wang and Jarvis, 1990; Oker-Blom et al., 1991; Nilson, 1992; Cescatti, 1997a) but only more recently, after becoming aware of the specific problems related to interpretation of satellite images over the boreal zone, has modeling the radiation regime of coniferous forests received increasing attention in the remote sensing community. The boreal zone spreads through Fennoscandia, Siberia, Alaska and Canada, and hosts a multitude of coniferous tree species adapted to the cold and drought climate conditions of the region. The forests are typically rather open with dense crowns (consisting of up to a few million needles per crown) and an abundant green understory and moss or lichen layer. The documented complex structure of the forests is further complicated by the fact that acquiring ground observations from many parts of the boreal zone is especially difficult due to the remoteness and climate of the region. For application of forest reflectance models in the boreal zone, remaining problems of high priority are thus how to model efficiently and sufficiently realistically (1) the hierarchic 3D canopy structure and (2) the contribution to the remotely sensed spectral signal from the background, typically composed of mixed green understory, and (3) how to separate the signals of the forest canopy layer and the understory from each other for correct interpretation of tree canopy biophysical variables from optical satellite images.

150

P. Stenberg et al.

In this chapter, we will first outline the basic principles and current status of existing models for simulating the spectral signature of forests. After this, we will focus on the specific issues related to applying these models in the boreal zone, including how to collect the required input data. Finally, more practical issues, such as application of ground truth data, model scaling and validation, will be addressed.

6.2 Modeling the Spectral Signature of Forests The basic premises for optical remote sensing of vegetation are that the solar radiation received by a remotely located sensor (e.g., on a satellite) upon interaction with the vegetation canopy carries in it the signature of the canopy, and that this spectral signature can be deciphered to obtain the information of interest (Goel, 1989). Satellite-borne sensors measure the mean intensities of radiation at different wavelengths emanating from a target on Earth. Correct interpretation of these measurements to yield the biophysical variables of interest requires an accurate specification of the relationships between these variables and the canopy leaving radiation field. This specification is quantified by canopy reflectance models. The models are parameterized using mathematical descriptions of canopy structure together with optical properties of the plant elements and the underlying surface to produce spectral signatures of canopy leaving radiation. In addition, the spectral and angular properties of the incoming radiation and the receiving sensor need to be specified (Table 6.1). Finally, the spectral signature depends on the resolution of the measurement (i.e., pixel size), which needs to be considered in defining the spatial scale of the model. Table 6.1 Variables affecting the spectral signature of a forest (excluding terrain topography and atmosphere) Sensor

Illumination

Forest tree layer canopy

Understory and soil

Zenith and azimuth viewing angles

Angles of incidence and azimuth

Wavelength band (i.e., spectral sensitivity)

Wavelength

Macroscale structure (distribution, size and shape of tree crowns) Microscale structure (distribution, size and shape of leaves, needles, shoots and branches) Other structural elements (e.g., distribution, size and shape of tree trunks and branches) Spectral properties of all the canopy elements

Geometrical structure (amount, distribution, size and shape of understory plants) Spectral properties of understory plants

Resolution

Soil optical properties (influenced by e.g., soil moisture and texture)

6 Modeling the Spectral Signature of Forests

151

The mathematical descriptions of the process of radiative transfer in vegetation canopies cover a range of different approaches, starting from the traditional turbid medium approach to geometric-optical and hybrid models, the main difference being the degree of detail at which canopy structure at different hierarchical levels is described (Table 6.1). The turbid medium approximation, introduced to modeling the radiation regime inside vegetation by Ross (1981), is based on the classical radiative transfer theory (Chandrasekhar, 1960). The radiative transfer theory has been successfully applied to predicting the distribution of shortwave (optical and infrared) photons in several media like atmospheres, stellar dust or water. Radiative transfer theory is, in principle, the law of conservation of energy written out for beams of radiation traveling in all possible directions in a volume filled with optically active, i.e., absorbing and scattering, material. It is thus a function of five coordinates: three spatial coordinates determining the elementary volume for which the energy conservation law is applied, and two variables defining the direction of the radiation beam. In the classical radiative transfer theory, the fate of a photon is determined by the probability of being absorbed or scattered while traveling a unit distance (these probabilities are called the absorption and scattering cross-sections, respectively). Additionally, interactions between different rays for each point are described by the probability distribution of directions the photon can be scattered in, measured relative to the photon’s traveling direction before scattering. This is called the photon scattering phase function. The equation of radiative transfer merely states that the number of photons exiting each elementary volume equals the number of photons entering the volume minus the number of absorbed photons, written out in a correct mathematical form. More generally, a source term has to be added for radiating media, e.g., inside stars or for the thermal infrared radiation in atmosphere. In the optical region under consideration in this chapter, no radiation is usually emitted by the medium. The difference between the classical radiative transfer theory and the turbid medium approach lies in the treatment of the directional properties of the medium: in the classical radiative transfer theory, the scattering phase function and the absorption and scattering cross-sections are assumed to be independent of the direction of photon travel. This assumption does not hold for the transfer of radiation inside a vegetation canopy. Thus, in the turbid medium approximation, plant cover is described as consisting of geometric elements which usually have a preferred orientation described, in turn, by a directional attribute. Now, the absorption and scattering cross-sections become functions of the direction of photon travel inside the canopy as well as of the actual shape of the elementary units (phytoelements) constituting the canopy. For example, in a canopy of preferably horizontal flat leaves, the probability of hitting a leaf while traveling a distance unit is much larger for a photon moving in the vertical direction than for one moving in the horizontal direction. Consequently, the interaction cross-section (sum of absorption and scattering crosssections) in such a canopy is larger in the vertical direction than in the horizontal. Orientation of plant elements is described by statistical distributions. For example, the orientation of leaves is described by the direction (inclination and azimuth) of their normal and that of stems, branches and needles by the direction of their axis.

152

P. Stenberg et al.

A number of different distributions suitable for describing the distribution of leaf normals of actual canopies are given in, e.g., Weiss et al. (2004). Cross-sections can be calculated from these distributions by integrating over all possible orientation angles. Other than implying that the scattering phase function and the cross-sections depend on direction within the canopy, the turbid media approach leaves the classical radiative transfer equations intact: canopy elements are still of infinitesimal size and do not cast shadows, i.e., the “far-field approximation” still holds. Although anisotropy (or dependence on photon travel direction) makes finding a solution to the radiative transfer equation more difficult and yields some traditional methods unpractical, a plethora of techniques is still available for finding an approximate solution. The basic equations for the radiation field can be found, for example, in the book by Ross (1981) or various other sources (e.g., Myneni et al., 1989). The crudest (and fastest) way of solving the radiative transfer problem for a plane-parallel turbid media (or in “slab geometry”) is by using the two-stream (Kubelka-Munk) equations. Only two directions, up and down, are used to describe the radiation field inside a horizontally homogeneous and infinite medium and the solutions can be found analytically. In its simplest form, the two stream model assumes isotropic distribution of both up- and downward traveling radiation, i.e., the intensity does not depend on view angle inside both hemispheres. This condition is clearly violated if direct solar radiation is present. To overcome the problem, the contribution of the direct beam is separated, thus creating an additional source for the remaining more isotropic diffuse radiation field. This technique is almost universally employed in all algorithms for solving the radiative transfer problem in plant canopies. This is because besides its extremely anisotropic character, direct illumination gives rise to the so-called hot-spot phenomenon (see discussion later in this chapter) that needs to be accounted for if measurements are taken near the backscattering direction. The solutions of the two-stream equation are analytical and thus extremely computer-efficient. It should also be noted that the two-stream equations cease to be approximations for a (theoretical) canopy of infinitesimally small horizontal ideally scattering leaves but, instead, give the exact analytical solution as the reflected and scattered radiation field inside such a canopy is isotropic. A classical example of an application of the two-stream approximation is the SAIL (Scattering and Arbitrarily Inclined Leaves) model (Verhoef, 1984, 1985). Besides the two fluxes, the model calculates intensities for two other directions: the solar direction (for reasons discussed above) and the view direction. This enables to take into account anisotropy in the distribution of reflected radiation. The absorption and scattering coefficients, and the scattering phase function are calculated from a distribution function of leaf inclination angles. Despite its age, the model is still used for remote sensing purposes, although sometimes modified to fit the particular needs or encapsulated in other models that include a higher level description of canopy structure (Kuusk, 1995; Kuusk, 2001; Andrieu et al., 1997). The SAIL model is reasonably accurate, and can take the competition from much more sophisticated 3D models, in the case of relatively homogeneous canopy types (crops, grasslands) that have a small number of structural parameters (Jacquemoud et al., 1995; Koetz et al., 2005; Meroni et al., 2004; Andrieu et al., 1997; Weiss and Baret, 1999; ZarcoTejada et al., 2003).

6 Modeling the Spectral Signature of Forests

153

If the media cannot be considered homogeneous in the horizontal directions, as is the case for many natural vegetation canopies, the radiative transfer equation can be solved in three dimensions (Myneni, 1990; Knyazikhin and Marshak, 1991). This approach is useful for highly structured scenes where the two-stream model cannot be adequately parameterized to include canopy heterogeneity. Once such a scene is generated based on measured data or a forest structure model, the canopy is divided into cells and the equation (or, after discretization, a set of coupled integrodifferential equations) is solved numerically using the discrete ordinates iterative approach on the grid (Myneni, 1990; Gastellu-Etchegorry et al., 1996; Knyazikhin et al., 1998a). This method is based on the calculation of radiation intensities in discretized directions. In each iteration, the contribution of radiation scattered to every discrete direction from all other directions is calculated for each canopy cell using a pre-calculated scattering matrix until the intensities converge to a solution. Naturally, this method, besides its mathematical complexity (which is further increased by using various convergence acceleration techniques), also requires a realistic description of the scene and reasonable cell-size to exclude shadowing effects. The method is computer-intensive but the results can be considered accurate as this method is based on the physical principles of radiative transfer. Another common method used for calculating radiation reflected by a canopy, for which a complete 3D description of the scene is given, is the Monte Carlo method. As the radiation field above plant canopy is composed of small contributions by individual photons, the angular and spatial variation in the intensity of reflected radiation can be viewed as a distribution function describing the possible exit directions and locations of light quanta. Monte Carlo methods (or ray tracing algorithms) are based on random sampling of this distribution. If a sufficient number of photons are traced, the counts of photons exiting in each direction are relatively close to the actual directional distribution of reflected radiation, although some numerical noise is inevitable. Although in studies of radiative transfer in canopies, the term “Monte Carlo” is commonly used to denote tracing the trajectories of photons in a realistic detailed 3D canopy scene (Gerstl et al., 1986; Ross and Marshak, 1988; North, 1996; Chelle and Andrieu, 1998; Govaerts and Verstraete, 1998; Thompson and Goel, 1998; Disney et al., 2006), Monte Carlo calculations can also be used to trace photons in a 3D turbid media with varying optical properties, like the geometricoptical approximation described below (Gerard and North, 1997; Garc´ıa-Haro and Sommer, 2002). An overview of the Monte Carlo modeling approach is given by Disney et al. (2000). If a 3D description of a scene is given, the reflected radiation field can also be calculated using the radiosity principle (Borel et al., 1991; Gerstl and Borel, 1992; Goel et al., 1991; Chelle and Andrieu, 1998; Garc´ıa-Haro et al., 1999; Qin et al., 2002; Soler et al., 2003). Radiation field inside the canopy is calculated using view factors for individual elements: if the illumination conditions of a particular canopy element are given, its brightness as viewed from the positions of all other elements can be calculated based on its geometry and scattering properties. The brightness of an object is a well-defined radiometric quantity that can be used to calculate the partial flux contributed by the object. Thus, if the magnitude and angular distribution

154

P. Stenberg et al.

of radiation field incident on a canopy element can be calculated from brightness of all other canopy elements and the intensity of the incident unscattered radiation, the brightness of the element as viewed from any point inside or above the canopy can be calculated. It is an iterative process complicated by the large number and mutual shading of elements present in a natural canopy. From the computing perspective, the most time-consuming task is calculating these view factors or solving the geometrical relations between each pair of foliage elements. Once these factors are known, calculations for different wavelengths for which only the scattering properties of elements are different, but not the geometry of the canopy as a whole are relatively fast. In the classical radiative transfer approach, the optical properties of the medium are assumed to change continuously (for the 3D case, this assumption is valid inside each cell). This is naturally not the case for actual vegetation canopies. The Monte Carlo and radiosity methods that describe the canopy as consisting of separate solid objects remove this restriction and take into account the mutual shading of elements (remove the far-field approximation). The assumption of discontinuities in canopy optical properties, or the existence of voids between scatterers, can also be inserted into the radiative transfer equation. The foliage are a density can be described using the statistical distribution of an indicator function, in other words, a function of the spatial coordinates that equals unity only if a scattering element is present at the point described by the coordinates. This leads to the stochastic radiative transfer equation in plant canopies (Menzhulin and Anisimov, 1991; Shabanov et al., 2000). Besides the mean reflecting properties of a canopy, this theory allows to calculate its statistical moments. Often the terms deterministic (“non-stochastic”) and stochastic are used as synonyms to 1D and 3D models, respectively. However, with the exception of Monte Carlo models, most of the existing canopy radiation/reflectance models, in fact use a non-stochastic approach in computing the spectral signatures. That is, even though the canopy structure may be described using statistical distributions, the canopy radiation field is solved for a mean realization (with average characteristics) of canopy structure. A truly stochastic approach, in contrast, would be to evaluate the 3D canopy radiation field for all possible statistical realizations of the canopy, and then average the corresponding radiation fields to obtain the ensemble average signature (Shabanov et al., 2000). The stochastic and non-stochastic approaches can result in different relationships between mean characteristics of canopy structure and canopy-leaving radiation. Sometimes, a more abstract description of the 3D structure of a natural forest canopy is required. The exact locations of all canopy elements are rarely known or reliably predicted for a forest canopy. Therefore, instead of describing the structure of a canopy by specifying the locations of the smallest canopy elements (e.g., leaves, needles or shoots), the canopy can be first divided into larger subunits. An obvious choice for the crudest first-level division would be the tree crown. The canopy can then be viewed as a congregation of geometrical tree crown envelopes with gaps between them. The locations of the tree crowns are described by a statistical distribution, thus accounting for mutual shadowing and the distribution of

6 Modeling the Spectral Signature of Forests

155

between-crown gaps. Depending on the accuracy by which the processes of radiative transfer are described in these models, the models are commonly categorized as geometric-optical or hybrid models, the term “hybrid” being a shortcut for “hybrid radiative transfer/geometric-optical model”. In the simplest form, a geometric-optical canopy reflectance model considers just the effect of the geometrical structure of a forest on the remotely sensed image (Li and Strahler, 1985, 1986, 1992; Welles and Norman, 1991; Chen and Leblanc, 1997). The image is supposed to consist of (sub-pixel-sized) regions of different brightness: directly illuminated tree crowns and ground, and shadowed tree crowns and ground. The fractions of these components are calculated using given illumination and view angles and mutual shadowing derived from geometrical considerations. These types of models can accurately account for only first-order scattered radiation, i.e., photons that reach the sensor having interacted only once with the geometric shapes comprising the canopy. Diffuse radiation is included using correction factors. This makes it difficult to use these models for different wavelengths where the optical properties of canopy elements are not similar. The share of first-order scattering can be calculated straightforwardly: this reflectance component, besides depending on canopy geometry, which is the same for all wavelengths, is a linear function of leaf single scattering albedo (leaf reflectance plus transmittance). Diffuse photons, on the other hand, interact with leaves (or other canopy elements) several times and the possible number of interaction is itself a function of leaf albedo at the specified wavelength. This makes canopy diffuse reflectance strongly nonlinear with leaf albedo, and results in highly wavelengthspecific diffuse radiation correction factors. Including the diffuse radiation field into the geometric-optical model leads to a hybrid radiative transfer/geometric-optical model (Nilson and Peterson, 1991; Li and Strahler, 1995; Ni et al., 1999; Atzberger, 2000; Chen and Leblanc, 2001; Huemmrich, 2001; Kuusk and Nilson, 2000; Peddle et al., 2004). The methods to include diffuse radiation can range from exact solutions (similar to those used in solving the 3D radiative transfer) to tracing the photons in the canopy comprised of crown envelopes seen as large chunks of turbid media (similar to the Monte Carlo approach). However, to maintain the high efficiency achieved by delineating the canopy into abstract crowns, a simple and fast approximation is often used. Now, the problem of finding reflected radiation intensity is divided into two sub-problems: (1) finding the first-order scattering component using a geometrical figure filled with absorbing and scattering foliage elements illuminated by a beam of direct radiation, taking into account mutual shadowing by other semi-transparent crowns, and (2) calculating the share of diffuse radiation using a simpler (e.g., twostream) solution of the radiative transfer equation. Naturally, in calculating both components one has to consider the effect of a partially reflecting ground surface or undergrowth. Joining the two sub-models is not an easy task: the multiple-scattering component has to be parameterized to include the effect of canopy structure so that conservation of energy and correct partitioning between first- and higher order scattering are maintained.

156

P. Stenberg et al.

Hybrid models are an efficient and flexible tool for describing the radiation regime of complex forests. The effect of different crown shapes can be included by using various geometrical shapes: ellipsoids and cones of different proportions are the most commonly used. Also, higher-order clumping can be introduced by further dividing the canopy into abstract objects and adding their contributions to total reflectance (e.g., Smolander and Stenberg, 2005). However, most models still assume that the crown envelopes are uniformly filled with scatterers, although from biological considerations and biometrical measurements it is known not to be true. Using more realistic foliage distributions would considerably affect the transparency of the crowns by introducing more clumping at both branch and whorl scale or by locating photosynthesizing material close to the crown edge. While accounting for higher-order structure usually leads to an increase in predicted transmittance if the higher-order objects are distributed randomly, a regular distribution of foliage clumps that has been suggested for natural canopies (Cescatti, 1997b) may diminish or even reverse the effect. The choice of the model to be used for solving a problem involving calculation of canopy reflectance or radiation balance clearly depends on many factors. The amount of required computer resources, manpower and time is clearly different for the different approaches. Also, the optimal choice depends on the object under investigation: for relatively homogeneous canopies (crops, grasslands), a two-stream model give very good results. Sometimes, if canopy reflectance calculations form only a small contribution to a larger problem under investigation that depends on many variables with possibly high uncertainties, the use of a simpler model is justified. This is clearly not the case for highly structured vegetation covers like shrublands or boreal coniferous forests where the canopy upper surface is not flat, and mutual shading and between-crown transmittance between tree crowns have to be taken into account. Also, the amount of a priori knowledge can be a limiting factor when constructing a detailed 3D description of a canopy. In this case, a less detailed approach with a smaller number of input parameters might be preferred, e.g., a hybrid geometric-optical/radiative transfer model. Another characteristic of the reflecting properties of a medium that contains finites objects filling a three-dimensional volume is the hot spot phenomenon. “Hotspot” is a bright area in a remotely sensed image opposite to the source of illumination caused by a lack of shadows in the exact backscattering direction. The width of this reflectance peak (or the size of the brighter area in an image) depends on the geometric properties of the reflecting medium. Based on their working principles, geometric-optical models take this phenomenon into account, but only partly. When looking from the direction of illumination, no shaded crowns can be seen. This results in an increase in the predicted reflectance. Yet, the hot-spot phenomenon is also produced at a finer level (e.g., leaves), which in turn leads to a distinct anisotropy of the brightness of tree crowns. In more sophisticated reflectance models, this leaflevel hot-spot is added to the wider hot spot created by crown structure. Also, a hot-spot correction can be added to models based on the radiative transfer equation (e.g., Gerstl et al., 1986; Marshak, 1989; Verstraete et al., 1990; Jupp and Strahler, 1991; Kuusk, 1991). The correction factor is semi-empirical due to the complex

6 Modeling the Spectral Signature of Forests

157

structure of vegetation canopies. The phenomenon itself is difficult to measure accurately since it tends to be in the shadow of the sensor. Due to common view angle configurations, it is also not registered by most remote sensing instruments. Apart from the choice of radiation transfer model and availability of adequate computing resources, quality of both structural and spectral input data has a large impact on model performance. Today, good-quality data of either sorts are scarce as they are difficult to measure and have a high natural variance. Moreover, as developing accurate radiative transfer models for vegetation canopies is a work in progress, a ready-made solution may very likely not yet exist. However, the models presented here (or, a quite up-to-date performance comparison can be found from RAMI website (http://rami-benchmark.jrc.it) (Pinty et al., 2004b)) cover a large number of approaches which, if adequate input is provided in terms of both measurement data and dedication, can be applied to a wide variety of problems. More information on the subject can be found, in addition to the works referred to in this section, from previously published reviews and textbooks (e.g., Myneni et al., 1991; Liang and Strahler, 2000). Canopy reflectance models can be constructed in many ways; some can be run in the forward mode, some both in the forward and inverse modes. In model inversion (discussed thoroughly in another chapter of this volume), radiation measurements are converted into variables of interest characterizing the target. In remote sensing of forests, invertible models can be used to infer biophysical variables from reflectances (or back-scattering) registered by the remote air-or satellite-borne sensors. Commonly used methods for the inverse estimation of vegetation characteristics from satellite images include (1) comparing the observed signal to a database of previously computed spectral signals for a wide selection of different canopies and choosing the closest matches (look-up tables, LUT) (Knyazikhin et al., 1998b), or (2) iteratively optimizing model input parameters to match the observed signal as closely as possible using different optimization routines (Kuusk and Nilson, 2000). Needless to say, the goodness of the estimates depends crucially on how realistic the model is. Another basic requirement for successful estimation is of course that the vegetation characteristic in question has a detectable influence on the spectral signal, and that measurement errors can be corrected for. Even so, a remaining limitation is the well known fact that the inversion problem is ill-posed: no unique solution exists but different combinations of input parameters produce the same spectral signal. To be able to solve the inverse problem, in other words, to reduce the array of possible solutions to one solution, we need to acquire information from outside the problem itself. Technical or mathematical advances do not remove the underlying ambiguity of the ill-posed nature of the inversion problem in remote sensing. Another central problem in inversion is scaling: the forest reflectance models assume that the given forest structure continues infinitely. However, often in the case coarse and medium resolution satellite images, the vegetation (forested area) may cover only part of the pixel. At best, thus, the “averaged solutions” to the inversion problem may be accurate at larger, often regional, scales but not at small scales.

158

P. Stenberg et al.

6.3 Structural and Spectral Characteristics of Coniferous Forests Northern hemisphere boreal forests, forming the largest unbroken, circumpolar forest zone in the world, are dominated by coniferous tree species. The radiation regime of coniferous forests is known to differ in many respects from that of other vegetation, e.g., broadleaved forests and agricultural crops (Norman and Jarvis 1975; Oker-Blom et al., 1991; Nilson, 1992). From the perspective of optical remote sensing, a specific problem encountered in the boreal forest zone is the poor performance of commonly used spectral vegetation indices (e.g., the NDVI) for the estimation of biophysical variables (such as LAI). One reason for the generally observed relative insensitivity of these indices to changes in LAI is apparently caused by an abundant presence of mixed green understory in boreal coniferous forests (Chen and Cihlar, 1996; Eklundh et al., 2001; Stenberg et al., 2004; Peltoniemi et al., 2005; Rautiainen et al., 2007). The influence of understory on the spectral signal naturally poses a problem also for the correct image interpretation using physically based models. In recent years, yet another phenomenon specific to coniferous forests has become a subject of increasing scientific interest by the optical remote sensing community. From empirical observations, it has long been known that coniferous forests have a lower reflectance, especially in the near-infrared (NIR), than broadleaved forests, but the physical background to this behavior is still partly unexplored. In the following, we will discuss how some conifer-specific characteristics affect the spectral signature of coniferous forests. There are several structural and spectral attributes which are specific to coniferous forests and require modifications to the radiative transfer models formulated for their broadleaved counterparts (Fig. 6.1). The non-flat, three-dimensional structure of conifer needles of varying, species specific shape, first of all, requires proper

Needles

Shoots Tree crowns, internal shoot distribution patttern

Fig. 6.1 Hierarchic structure of coniferous crowns: A three-layer scheme

6 Modeling the Spectral Signature of Forests

159

(re-)definition of some of the key variables and concepts of these models: e.g., the leaf orientation, the mean projection of unit foliage area (“G-function”), the leaf scattering phase function, and the leaf area index (Stenberg, 2006). These variables and distributions were all originally defined and derived under the assumption of flat leaves and, so, for example, equations given for the mean projection of planar leaf area (e.g., Warren Wilson 1967) are not directly applicable to needles but must be derived with consideration given to the needle shape (Oker-Blom and Kellom¨aki, 1982; Lang, 1991; Chen and Black, 1992). A more serious theoretical problem for modeling however arises from that in many coniferous species needles are closely grouped together as shoots. The resulting small-scale variation in needle area density cannot readily be represented in the formulation of the radiative transfer equation based on the concept of an “elementary volume”, which should be small enough that essentially no mutual shading between the elements exists but large enough for statistical laws to apply (Ross, 1981). It has been proposed, therefore, that the shoot should be treated as the basic structural element of coniferous canopies – an approach that has actually long been used in models of canopy light interception and photosynthesis (e.g., Oker-Blom and Kellom¨aki, 1983; Nilson and Ross, 1997; Cescatti, 1997a), but has not yet been fully implemented in forest reflectance models due to the lack of data and models describing the scattering properties of shoots (but see Smolander and Stenberg, 2003). More data are available on another key parameter entering the shoot based models, namely the shoot silhouette to total area ratio (STAR) (Oker-Blom and Smolander, 1988), which is conceptually analogous to the G-function defined for flat leaves (Nilson, 1971), but includes a clumping coefficient accounting for the mutual shading of needles in the shoot. The clumping or mutual shading of needles in shoots acts to decrease the interaction cross section area of a given amount of total needle area, i.e., the extinction coefficient, and this effect can in radiative transfer models be parameterized (quantified) using the STAR. The decrease in shoot single scattering albedo, as compared to that of a single needle, has also been shown to be closely related to STAR, which thus can be used to modify the scattering properties of an elementary volume containing shoots (Smolander and Stenberg, 2003). However, theoretical models and, above all, empirical data on the shoot scattering phase function for different species are still needed for correct parameterization of coniferous canopy reflectance models. In some current models (Knyazikhin et al., 1998b; Kuusk and Nilson, 2000), shoot level grouping is accounted for in quantifying (i.e., reducing) the extinction coefficient (interaction cross section area of the elementary foliated volume), but its effect on the volume scattering phase function has not yet been fully implemented. It should be noted that when shoots (instead of single needles, or needle surface area elements) are treated as the basic elements, the optical properties (transmittance and reflectance) of single needles no longer suffice to describe the scattering properties of the elementary volume in the radiative transfer equation. Also at higher levels of organization, coniferous forests display a distinct grouped pattern (Fig. 6.1). The canopies are typically formed of dense, narrow and deep tree crowns. Crown shape, volume, and density have a considerable effect on the total

160

P. Stenberg et al.

canopy reflectance, as well as on the bidirectional reflectance factor (BRF) of conifer stands (Rautiainen et al., 2004). The same properties, together with the stand density (number of trees per hectare), also determine how much of the background (forest floor) is visible to the satellite instruments. Because northern coniferous forests are often rather open, with an abundant green understory, the spectral signal is a complex mixture of (overstory) canopy and background reflectance.

6.4 Effect of Clumping on the Spectral Signal: p-Theory as an Example The high degree of grouping (or clumping) of foliage at different hierarchical scales in coniferous forest stands (Fig. 6.1) is considered to be the most important reason behind the different spectral signature of these forests, in particular the low reflectance in NIR, as compared to broadleaved forests (Nilson et al., 1999; Rautiainen and Stenberg, 2005a). Incorporating the effects of grouping in canopy reflectance models poses a true challenge as very detailed canopy descriptions cannot readily be integrated into models operating at large spatial scales – this is typically the case in remote sensing of vegetation. Thus, especially for large scale applications, it is important to search for some key parameters that could capture the most essential structural features of a forest stand. One such canopy structural parameter, proposed to govern canopy absorption and scattering, is the spectrally invariant (i.e., wavelength independent) “p-parameter” introduced by Knyazikhin et al. (1998b). The “p-theory”, in short, predicts that the amount of radiation scattered by a canopy (bounded underneath by a black surface) should depend only on the wavelength and the spectrally invariant parameter (p), which can be interpreted as the probability that a photon scattered from a leaf (or needle) in the canopy will interact within the canopy again – the “recollision probability” (Smolander and Stenberg, 2005; Rautiainen and Stenberg, 2005a). The usefulness of this parameter in practical applications depends on whether and how well p can be related to (or derived from) other commonly available forest stand data. Nonetheless, it is a powerful modeling tool because it links canopy absorption (α C ) and scattering (ωC ) at any wavelength (λ ) to the phytoelement (leaf or needle) scattering coefficient (ωL ) at the considered wavelength, while simultaneously preserving the law of energy conservation (Panferov et al., 2001; Wang et al., 2003). The relationship between canopy and leaf scattering coefficients is described by the simple equation:

ωC (λ ) =

ωL (λ ) − pωL (λ ) 1 − pωL (λ )

(6.1)

It follows from Eq. (6.1) that, at any given wavelength, the canopy spectral scattering coefficient (ωC ) decreases in a nonlinear fashion with increasing canopy aggregation or grouping (quantified by the p parameter). More importantly, in any given canopy (fixed p), the relationship between ωC and ωL is also nonlinear, implying that canopy

6 Modeling the Spectral Signature of Forests

161

structure (or grouping, specifically) does not only affect the magnitude of the canopy leaving radiation but also modifies its spectral distribution. Simulation studies and empirical measurements have provided support for the validity and usefulness of the p-theory. Equation (6.1) was shown to precisely predict the absorption and scattering in structurally homogeneous canopies simulated with a Monte Carlo model (Smolander and Stenberg, 2003; Smolander and Stenberg, 2005), and to hold true also in heterogeneous canopies (M˜ottus et al., 2007) simulated using the hybrid FRT model (Kuusk and Nilson, 2000). Another spectral invariant parameter (pt ) has been proposed to control canopy transmission, i.e., the part of the scattered radiation that exits the canopy downwards (Knyazikhin et al., 1998a, b; Panferov et al., 2001; Shabanov et al., 2003). This structural parameter so far lacks a clear physical interpretation, as has been given for the other spectral invariant – the recollision probability (p), but if it can be formulated and shown to be valid (both theoretically and empirically), the two parameters offer a simple and effective tool for parameterization of the canopy radiation budget. Namely, given the absorption (p value) and transmission (pt value), total reflectance (the upward scattered part of the incident radiation) is also known (because they all sum up to one). Thus, the spectral invariants p and pt would allow calculation of all components, i.e., spectral absorptance, transmittance and reflectance of the canopy shortwave radiation budget for any given wavelength knowing the leaf (or needle) scattering coefficient at the same wavelength. If the relationships between p and canopy structural parameters such as LAI are known, the spectral signature of the canopy can be predicted in terms of LAI or, conversely, inverse estimation LAI can be done based on measured canopy reflectance. The relationship described by Eq. (6.1), linking together canopy scattering coefficients at a specific wavelength to the leaf albedo at the same wavelength, can actually be applied at different hierarchical levels and provides a tool for scaling grouping effects. In the simulation studies by Smolander and Stenberg (2003, 2005), it was found that Eq. (6.1) could be used to scale from needle to shoot scattering by replacing p by the “recollision probability within a shoot” (psh ). Moreover, it was shown that the canopy level recollision probability (p) could be decomposed into (1) the probability (psh ) that the new interaction occurs within the same shoot where the first (former) interaction took place, and (2) the probability (pc ) that interaction occurs with another shoot in the canopy, according to the formula: p = psh + (1 − psh )pc

(6.2)

Equation (6.2), which in a similar manner could be further developed (decomposed) to account, e.g., for grouping at the crown level, shows how the whole canopy p value is affected by grouping at different spatial scales. For a given LAI, the p value increases with the degree of grouping present in the spatial dispersion of the leaf area. Thus, for example, considering a broadleaved and coniferous canopy with similar LAI and macro-scale structure (Table 6.1), the coniferous canopy would have higher p or, equivalently, a smaller escape probability for the radiation incident

162

P. Stenberg et al.

on the canopy. This explains, at least partly, the observed lower canopy reflectance (albedo) of coniferous forests as compared to broadleaved forests. Two shortcomings of the p-theory for practical applications should however be noted: (1) It describes only canopy scattering, i.e., the contribution from background reflectance must be modelled separately and (2) it is not able to describe the angular distribution of scattered radiation. To solve these problems, the p-theory should be combined with other physically-based reflectance modeling concepts. With increasing p, the scattered part of the radiation intercepted by the foliated canopy (the crowns), and thus also canopy reflectance, decreases. However, the effect of groping on whole stand reflectance is not as straightforward. To demonstrate the combined effect grouping has on canopy reflectance, on one hand, and on the contribution of understory reflectance, on the other hand, we use the simple parameterization model, PARAS (Rautiainen and Stenberg, 2005a). In this semi-physical model, forest BRF is calculated as a sum of the ground and canopy components: BRF = cg f (θ1 )cg f (θ2 )ρground + f (θ1 , θ2 )i0 (θ2 )

ωL − pωL 1 − pωL

(6.3)

The parameters of Eq. (6.3) are defined as follows: θ1 and θ2 are the viewing and illumination zenith angles, cgf denotes the canopy gap fraction in the directions of view and illumination (Sun), ρground is the BRF of the ground, f is the canopy scattering phase function, i0 (θ2 ) is canopy interceptance or the fraction of the incoming radiation interacting with the canopy, and ωL is needle (or leaf) scattering coefficient. With increasing degree of grouping (larger p), canopy interceptance (i0 ) simultaneously decreases while the canopy gap fractions (cgf) increase and, thus, the contribution from ground (understory) reflectance increases. Especially grouping at larger scale (between crowns) may more importantly influence the total stand reflectance through its effect on increasing the contribution from the background than through its effect on the canopy contribution. Incorporating the effect of shoot scale clumping in the PARAS canopy reflectance model was shown to produce realistic reflectance values in the nearinfrared (NIR) of coniferous forests. This can be seen as a major improvement since the low NIR reflectance observed in coniferous areas is one of the main anomalies that models have not been able to account for. The results give support to the hypothesis that in coniferous canopies large part of the clumping occurs at the shoot level and, thus, that the incorporation of shoot structural and spectral properties into current forest reflectance models will significantly improve their performance.

6.5 Scaling of Canopy Reflectance In the previous section, the term “scaling” was introduced to describe an application of recollision probability at different canopy grouping levels, i.e., the dimensionless scattering characteristics calculated at one level were applied at another, larger scale. This approach is similar to using downscaled models for predicting the behavior

6 Modeling the Spectral Signature of Forests

163

of the full-scale object. The procedure can be applied repeatedly as long as the character of the process (in our case, interaction of photons with canopy elements) remains unchanged. In optical remote sensing, the limit is determined by the ratio of photon wavelength to the size of the scatterers, i.e., radiative transfer is still determined by geometric optics, and refraction can be ignored. In the current section, “scaling” has a different meaning that is more familiar to the remote sensing community. Here, we use it to describe the spatial variation of canopies that has to be taken into account when comparing sensors with different view configurations (i.e., sensors that produce images with different pixel sizes). Generally, models can produce only point estimates of canopy reflectance, although most of them have a method to account for within-plot variability (statistical distribution of tree locations, etc.). This works well as long as the canopy is horizontally homogeneous (which is what these models actually presume) and its structure does not vary across a relatively large area, substantially exceeding the dimensions of a pixel in the image produced by the sensor. Natural forest canopies, however, almost never possess such a property: they include clusters of high tree concentrations and canopy openings, due to harvesting or windfall. As the reflectance process is strongly nonlinear, this heterogeneity is difficult to take into account in the spatially averaged pixels of remote sensing instruments. The scale at which model results can be related to the reflected signal in a straightforward way depends on the size of the structural units of the canopy. For example, models that utilize an assumption of a statistical distribution of tree locations predict an average signal produced by such a canopy and cannot be applied to model distinct patterns created by a particular configuration of individual trees or the variance of intensity in the image of a single crown. They were not designed for these purposes as overly detailed patterns are commonly not required for remote sensing of larger areas. Another reason is that biophysical variables (LAI, fCover) are defined for a larger canopy area and cannot be used to describe a single tree. When moving to larger resolutions, the models may also fail due to the nonlinearity of the radiative transfer problem and a highly varying or discontinuous distribution of tree locations. Canopy spectral properties are a function of spatial resolution (Tian et al., 2000; Pinty et al., 2004a), and especially in regions with a fragmented forest area, the signal will always be a mixture of reflectance from many different ground (soil/understory) and forest canopy compartments. This may lead to a situation where a model cannot be used as such (without alterations) to model signals by sensors with different spatial resolutions (Tian et al., 2002). The scaling problem affects the retrieval of different biophysical variables with varying severity. While some variables can be considered almost scale-independent (fCover, fPAR), predicting others (e.g., LAI) requires a careful consideration of spatial variation (Weiss et al., 2000). Algorithms for a correct treatment of the scaling problem of vegetation reflectance modeling are scarce although a physically-based theory for scaling was developed by Tian et al. (2002), for example. The incorporation of scaling algorithms directly in radiation transfer models is so far an almost unexploited subject (Widen, 2004).

164

P. Stenberg et al.

6.6 Use of Stand Inventory Data as Model Input Ground truth, or ground observations, is required for both developing and testing physically based forest reflectance models and their inversion. There are two general paths for acquiring this information: by conducting measurements or by using existing data bases (e.g., land cover maps, regional or national forest inventories, spectral data banks of common plant species). However, a general problem with all these data sources may be their age when compared to the more recently obtained remotely sensed image. Stand inventory data (which typically consists of information on tree species, tree height and diameter, and stand density or basal area) form a common input set required by most 3D forest reflectance models. However, stand inventories do not provide all the parameters typically included in physically based reflectance models, e.g., crown shape, leaf area density and orientation, nor do they include the biophysical variables of interest in this chapter (LAI, fPAR, fCover). Testing reflectance models should be done with a carefully measured, comprehensive input data set which includes all the input required by the model and which does not rely on other models (i.e., is not generated by other models). However, when operational inversion for obtaining biophysical variables (LAI, fPAR, or fCover in this case) is considered, such comprehensive data sets are not available, and other, let us say less appropriate, sets need to be utilized. A typical source would be regional or national forest inventory data bases which have recently become more suitable for the purpose in the boreal zone of North America and Europe. Traditionally forest mensuration science dealt with determining the volume of stands and logs and then studying stand growth and yield. Nowadays, however, the scope has widened and regional and national forest inventories have become more efficient environmental monitoring tools and are conducted to ensure the sustainable use of forests. Thus, they include more stand variables than tree height and diameter and form a better interface for forest reflectance models. Nevertheless, if forest reflectance models are run through a routine forest management data base (e.g., Rautiainen, 2005), there are several structural input parameters, such as crown shape, woody area index or shoot size, in the models which can hardly be obtained from these data bases. Thus, alternative solutions for obtaining the structural information need to be considered. A possibility worth exploring further would be to create realistic input data on these additional structural parameters from basic stand variables through regression models and allometric relationships. Such models exist for relating different biomass components of a tree or forest (e.g., Marklund, 1988) but parameters describing crown architecture have been of less interest in forest mensuration. In the context of radiative transfer modeling, one of the most central parameters not available from routine stand inventory data sets is crown shape which here serves as an example. Crown shape determines the limits of integration over the crown envelope and thus the scattering volume. Therefore, crown shape needs to be predicted from routine stand variables such as tree height and breast height diameter. This can be done by, for example, developing a crown shape model based on extensive measurements and then using the model to predict crown shape for a stand where it has not

6 Modeling the Spectral Signature of Forests

165

been measured (e.g., Baldwin and Peterson, 1997; Rautiainen and Stenberg, 2005b; M˜ottus et al., 2006). Alternatively, crown shape can be assumed to have a constant geometric shape (e.g., a cone or an ellipsoid) for a given tree species.

6.7 Concluding Remarks Remote sensing mapping techniques have traditionally been statistically oriented but physically based methods have progressively become more and more attractive because they are better suited for many current large scale applications related to global mapping of vegetation (e.g., Knyazikhin et al., 1998b). Implementation of physically based remote sensing methodology has become feasible thanks to the development of fast computer hardware and high resolution and multi-angular satellite instruments. The rapid technological and methodological development of satellite derived products offers a range of new possibilities for the mapping of terrestrial biophysical parameters such as vegetation type, forest cover (fCover) and leaf area index (LAI) from remotely sensed data (Myneni et al., 1997; Chen et al., 2002). Remote sensing based methodologies are especially important, and in fact the only feasible alternative, in regions where reliable field information is not available, is difficult to reach, or is too expensive. This, on the other hand, makes the importance of validation an even more crucial part of the development process aiming at producing trustworthy information righteously required by the end-users. Several research networks operate today on the refinement of remote sensing mapping methods and on validation of satellite derived biophysical products, such as LAI, fCover and fPAR, that are presently routinely generated by a range of sensors (Knyazikhin et al., 1998b; Morisette et al., 2006). For critical assessment of the accuracy of remotely sensed estimates of biophysical variables, there should be available a set of statistically representative (sufficient) and independent reference data with known accuracy. Also, special care should be taken that similar (standardized) measurement methodologies and designs are used to produce reference data of the variable of interest (Morisette et al., 2006). Global validation networks have a very important mission in providing these validation data needed for the further development of the remote sensing methodology. Physically based remote sensing methods still have room for improvement, and we have claimed in this chapter that there is especially a need to further develop reflectance models designed for coniferous forests, accounting for their complex 3D canopy structure and the small-scale variation in needle area density. The widely discussed global trends of greening and vegetation status require accurate algorithms for boreal forests. An efficient use of physically based forest reflectance models in the boreal coniferous forest zone would require that the model input data include variables such as crown shape and canopy cover, and spatial pattern of trees, to represent their clumped structure. In combination with standard forest inventory data, these key structural parameters would form a very valuable data base. It seems reasonable to believe that with better representation of the 3D canopy structure, the

166

P. Stenberg et al.

accuracy of remotely sensed estimates of canopy biophysical characteristics would improve considerably. Accordingly, future tasks of high priority include means to create realistic input data on crown shape, tree pattern and canopy cover from basic stand variables, for example through regression models and allometric relationships. In boreal forests, another limitation to successful application of canopy reflectance models is a typically large proportion of non-contrasting background reflectance in the visible and near-infrared (NIR) part of the spectra due to abundant mixed green understory. Stand inventory data do not include spectral properties of the forests, and there is a need for more data on the angular reflectance properties of common forest species, including not only understory vegetation but tree components, such as needles, shoots and bark. Use of the p-theory is for application in large scale remote sensing mapping methods is another question of large current interest. Can a simple parameterization model using the spectral invariants be built to mimic canopy reflectance with sufficient accuracy? It has already been shown, and put into practice in the MODIS LAI/FPAR algorithm, that the spectral invariants provide a powerful calculation tool in reflectance models and preserve the law of energy conservation. However, if one could derive specific relationships between the spectral invariants (p and pt ) and basic canopy structural parameters such as LAI and canopy cover, it would provide an effective tool by which the effect of the clumping of foliage at different hierarchical levels of the forest stand structure could be incorporated in forest reflectance models.

References Andrieu B, Baret F, Jacquemoud S, Malthus T, Steven M (1997) Evaluation of an improved version of SAIL model for simulating bidirectional reflectance of sugar beet canopies. Remote Sens. Environ. 60:247–257 Atzberger C (2000) Development of an invertible forest reflectance model: The INFOR-Model. In: Buchroithner (ed), A decade of Trans-European Remote Sensing Cooperation, Proc. 20th EARSeL Symposium, 14–16 June 2000, Dresden, pp 39–44 Baldwin VC, Peterson KD (1997) Predicting the crown shape of loblolly pine trees. Can. J. For. Res. 27:102–107 Borel CC, Gerstl SAW, Powers BJ (1991) The radiosity method in optical remote sensing of structured 3-D surfaces. Remote Sens. Environ. 36:13–44 Cescatti A (1997a) Modelling the radiative transfer in discontinuous canopies of assymmetric crowns. I. Model structure and algorithms. Ecol. Model. 101:263–274 Cescatti A (1997b) Modelling the radiative transfer in discontinuous canopies of assymmetric crowns. II. Model testing and application in a Norway spruce stand. Ecol. Model. 101:275– 284 Chandrasekhar S (1960) Radiative Transfer. Oxford University Press, London, UK Chelle M, Andrieu B (1998) The nested radiosity model for the distribution of light within plant canopies. Ecol. Model. 111:75–91 Chen J, Pavlic G, Brown L, Cihlar J, Leblanc SG, White HP, Hall RJ, Peddle DR, King DJ, Trofymow JA, Swift E, van der Sanden J, Pellikka PKE (2002) Derivation and validation of Canada-wide coarse-resolution leaf area index maps using high-resolution satellite imagery and ground measurements. Remote Sens. Environ. 80:165–184

6 Modeling the Spectral Signature of Forests

167

Chen JM, Black TA (1992) Defining leaf area index for non-flat leaves. Plant Cell Environ. 15: 421–429 Chen JM, Cihlar J (1996) Retrieving leaf area index of boreal conifer forests using Landsat TM images. Remote Sens. Environ. 55:153–162 Chen JM, Leblanc SG (1997) A four-scale bidirectional reflectance model based on canopy architecture. IEEE Trans. Geosci. Remote Sens. 35:1316–1337 Chen JM, Leblanc SG (2001) Multiple-scattering scheme useful for geometric optical modeling. IEEE Trans. Geosci. Remote Sens. 39:1061–1071 Disney MI, Lewis P, North PRJ (2000) Monte Carlo ray tracing in optical canopy reflectance modelling. Remote Sens. Rev. 18:163–196 Disney MI, Lewis P, Saich P (2006) 3D modelling of forest canopy structure for remote sensing simulations in the optical and microwave domains. Remote Sens. Environ. 100:114–132 Eklundh L, Harrie L, Kuusk A (2001) Investigating relationships between Landsat ETM + sensor data and leaf area index in boreal conifer forest. Remote Sens. Environ. 78:239–251 Garc´ıa-Haro FJ, Sommer S (2002) A fast canopy reflectance model to simulate realistic remote sensing scenarios. Remote Sens. Environ. 81:205–227 Garc´ıa-Haro FJ, Gilabert MA, Meli´a J (1999) A radiosity model for heterogeneous canopies in remote sensing. J. Geophys. Res. 104:12159–12175 Gastellu-Etchegorry J, Demarez V, Pinel, V, Zagolski F (1996) Modeling radiative transfer in heterogeneous 3-D vegetation canopies. Remote Sens. Environ. 58:131–156 Gerard FF, North PRJ (1997) Analyzing the effect of structural variability and canopy gaps on forest BRDF using a geometric-optical model. Remote Sens. Environ. 62:46–62 Gerstl SAW, Borel CC (1992) Principles of the radiosity method versus radiative transfer for canopy reflectance modeling. IEEE Trans. Geosci. Remote Sens. 30:271–275 Gerstl SAW, Simmer C, Powers BJ (1986) The canopy hot-spot as crop identifier. In: MSJ Damen, G Sicco Smit, HT Verstappen (eds), Remote Sensing for Resources Development and Environmental Management. Proc. 7th International Symposium, Enschede, 25–29 August 1986. AA Balkema, Rotterdam Goel NS (1989) Inversion of canopy reflectance models for estimation of biophysical parameters from reflectance data. In: G Asrar (ed), Theory and Applications of Optical Remote Sensing. Wiley, New York, pp 205–251 Goel NS, Rozehnal I, Thompson RL (1991) A computer-graphics based model for scattering from objects of arbitrary shapes in the optical-region. Remote Sens. Environ. 36:73–104 Govaerts YM, Verstraete MM (1998) Raytran: a Monte Carlo ray-tracing model to compute light scattering in three-dimensional heterogeneous media. IEEE Trans. Geosci. Remote Sens. 36:493–505 Huemmrich KF (2001) The GeoSail model: a simple addition to the SAIL model to describe discontinuous canopy reflectance. Remote Sens. Environ. 75:423–431 Jacquemoud S, Baret F, Andrieu B, Danson FM, Jaggard K (1995) Extraction of vegetation biophysical parameters by inversion of the PROSPECT + SAIL models on sugar beet canopy reflectance data. Application to TM and AVIRIS sensors. Remote Sens. Environ. 52:163–172 Jupp DLB, Strahler AH (1991) A hotspot model for leaf canopies. Remote Sens. Environ. 38: 193–210 Knyazikhin Y, Marshak A (1991) Fundamental equations of radiative transfer in leaf canopies, and iterative methods for their solution. In: RB Myneni, J Ross (eds), Photon–Vegetation Interactions, Springer, Berlin-Heidelberg, pp 9–43 Knyazikhin Y, Kranigk J, Myneni RB, Panfyorov O, Gravenhorst G (1998a) Influence of smallscale structure on radiative transfer and photosynthesis in vegetation cover. J. Geophys. Res. 103(DD6):6133–6144 Knyazikhin Y, Martonchik JV, Myneni RB, Diner DJ, Running SW (1998b) Synergistic algorithm for estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation from MODIS and MISR data. J. Geophys. Res. 103(D24):32257–32276 Koetz B, Baret F, Poilv´e H, Hill J (2005) Use of coupled canopy structure dynamic and radiative transfer models to estimate biophysical canopy characteristics. Remote Sens. Environ. 95: 115–124

168

P. Stenberg et al.

Kuusk A (1991) The hot spot effect in plant canopy reflectance. In: RB Myneni, J Ross (eds), Photon–Vegetation Interactions, Springer, Berlin-Heidelberg, pp 139–159 Kuusk A (1995) A fast, invertible canopy reflectance model. Remote Sens. Environ. 51:342–350 Kuusk A (2001) A two-layer canopy reflectance model. J. Quant. Spectrosc. Radiat. Transfer 71:1–9 Kuusk A, Nilson T (2000) A directional multispectral forest reflectance model. Remote Sens. Environ. 72:244–252 Lang ARG (1991) Application of some of Cauchy’s theorems to estimation of surface areas of leaves, needles and branches of plants, and light transmittance. Agric. For. Meteorol. 55: 191–212 Leverenz JW, Hinckley TM (1990) Shoot structure, leaf area index and productivity of evergreen conifer stands. Tree Physiology. 6:135–149 Li X, Strahler AH (1985) Geometric-optical modeling of a conifer forest canopy. IEEE Trans. Geosci. Remote Sens. GE-23:705–721 Li X, Strahler AH (1986) Geometric-optical bidirectional reflectance modeling of a conifer forest canopy. IEEE Trans. Geosci. Remote Sens. GE-24:906–919 Li X, Strahler AH (1992) Geometric-optical bidirectional reflectance modeling of the discrete crown vegetation canopy: effect of crown shape and mutual shadowing. IEEE Trans. Geosci. Remote Sens. 30:276–292 Li X, Strahler AH (1995) A hybrid geometric-optical radiative transfer approach for modeling albedo and directional reflectance of discontinuous canopies. IEEE Trans. Geosci. Remote Sens. 33:466–480 Liang S, Strahler AH (guest editors) 2000. Land surface bi-directional reflectance distribution function (BRDF): Recent advances and future prospects. Special issue of Remote Sensing Reviews. 18:83–511 Marklund L (1988) Biomassafunktioner f¨or tall, gran och bj¨ork i Sverige. Report 4 (in Swedish). Department of Forest Survey, Swedish University of Agricultural Sciences, Ume˚a, Sweden. 73 pp Marshak AL (1989) The effect of the hot spot on the transport equation in plant canopies. J. Quant. Spectrosc. Radiat. Transfer L.2:615–630 Menzhulin GV, Anisimov OA (1991) Principles of statistical phytoactinometry. In: RB Myneni, J Ross, (eds), Photon–Vegetation Interactions, Springer, Berlin-Heidelberg, pp 111–137 Meroni M, Colombo R, Panigada C (2004) Inversion of a radiative transfer model with hyperspectral observations for LAI mapping in poplar plantations. Remote Sens. Environ. 92:195–206 Morisette JT, Baret F, Privette JL, Myneni RB, Nickeson J, Garrigues S, Shabanov N, Weiss M, Fernandes R, Leblanc S, Kalacska M, S´anchez-Azofeifa GA, Chubey M, Rivard B, Stenberg P, Rautiainen M, Voipio P, Manninen T, Pilant A, Lewis T, Iiames J, Colombo R, Meroni M, Busetto L, Cohen W, Turner D, Warner ED, Petersen GW, Seufert G, Cook R (2006) Validation of global moderate resolution LAI Products: a framework proposed within the CEOS Land Product Validation subgroup. IEEE Trans. Geosci. Remote Sens. 44(7):1804–1817 M˜ottus M, Stenberg P, Rautiainen M (2007) Photon recollision probability in heterogeneous forest canopies: Compatibility with a hybrid GO model. J. Geophys. Res. 112:D03104, doi:10.1029/2006JD007445 M˜ottus M, Sulev M, Lang M (2006) Estimation of crown volume for a geometric radiation model from detailed measurements of tree structure. Ecological Modelling 198(3–4):506–514. Myneni RB (1990) Modeling radiative transfer and photosynthesis in three-dimensional vegetation canopies. Agric. For. Meteorol. 55:323–344 Myneni RB, Ross J, Asrar G (1989) A review on the theory of photon transport in leaf canopies. Agric. For. Meteorol. 45:1–153 Myneni RB, Ross J (eds) (1991) Photon–Vegetation Interactions, Springer, Berlin-Heidelberg, 565 pp Myneni RB, Keeling CD, Tucker CJ, Asrar G, Nemani RR (1997) Increased plant growth in the northern high latitudes from 1981 to 1991 Nature. 386 (6626):698–702

6 Modeling the Spectral Signature of Forests

169

Ni W, Li X, Woodcock C, Caetano M, Strahler A (1999) An analytical hybrid GORT model for bidirectional reflectance over discontinuous plant canopies. IEEE Trans. Geosci. Remote Sens. 37:987–999 Nilson T (1971) A theoretical analysis of the frequency of gaps in plant stands. Agric. Meteorol. 8:25–38 Nilson T (1977) A theory of radiation penetration into nonhomogeneous plant canopies. In: The Penetration of Solar Radiation into Plant Canopies, Acad. Sci. ESSR Report, Tartu, 5–70 (in Russian) Nilson T (1992) Radiative transfer in nonhomogeneous plant canopies. In: G Stanhill (ed), Advances in Bioclimatology 1. Springer, Berlin-Heidelberg, pp 59–88 Nilson T, Peterson U (1991) A forest canopy reflectance model and a test case. Remote Sens. Environ. 37:131–142 Nilson T, Ross J (1997) Modeling radiative transfer through forest canopies: implications for canopy photosynthesis and remote sensing. In: HL Gholz, K Nakane, H Shimoda (eds), The Use of Remote Sensing in the Modeling of Forest Productivity. Kluwer, Dordrecht, The Netherlands, pp 23–60 Nilson T, Anniste J, Lang M, Praks J (1999) Determination of needle area indices of coniferous forest canopies in the NOPEX region by ground-based optical measurements and satellite images. Agric. For. Meteorol. 98–99:449–462 Norman J, Jarvis P (1975) Photosynthesis in Sitka spruce (Picea sitchensis (Bong.) Carr.). V. Radiation penetration theory and a test case. J. Appl. Ecol. 12(3):839–877 North PRJ (1996) Three-dimensional forest light interaction model using a Monte Carlo method. IEEE Trans. Geosci. Remote Sens. 34:946–956 Oker-Blom P (1986) Photosynthetic radiation regime and canopy structure in modeled forest stands. Acta For. Fenn. 197:1–44 Oker-Blom P, Kellom¨aki S (1982) Effect of angular distribution of foliage on light absorption and photosynthesis in the plant canopy: theoretical computations. Agric. Meteorol. 26:105–116 Oker-Blom P, Kellom¨aki S (1983) Effect of grouping of foliage on the within-stand and withincrown light regime: comparison of random and grouping canopy models. Agric. Meteorol. 28:143–155 Oker-Blom P, Smolander H (1988) The ratio of shoot silhouette area to total needle area in Scots pine. For. Sci. 34:894–906 Oker-Blom P, Lappi J, Smolander H (1991) Radiation regime and photosynthesis of coniferous stands. In: RB Myneni, J Ross (eds), Photon-Vegetation Interactions: Applications in Optical Remote Sensing and Plant Ecology. Springer, New York, pp 469–499 Panferov O, Knyazikhin Y, Myneni RB, Szarzynski J, Engwald S, Schnitzler KG, Gravenhorst G (2001) The role of canopy structure in the spectral variation of transmission and absorption of solar radiation in vegetation canopies. IEEE Trans. Geosci. Remote Sens. 39:241–253 Peddle DR, Johnson RL, Cihlar J, Latifovic R (2004) Large area forest classification and biophysical parameter estimation using the 5-scale canopy reflectance model in multiple-forward-mode. Remote Sens. Environ. 89:252–263 Peltoniemi JI, Kaasalainen S, N¨ar¨anen J, Rautiainen M, Stenberg P, Smolander H, Smolander S, Voipio P (2005) BRDF measurement of understorey vegetation in pine forests: dwarf shrubs, lichen and moss. Remote Sens. Environ. 94:343–354 Pinty B, Gobron N, Widlowski J-L, Lavergne T, Verstraete MM (2004a) Synergy between 1-D and 3-D radiation transfer models to retrieve vegetation canopy properties from remote sensing data. J. Geophys. Res. 109, D21205, doi:10.1029/2004JD005214 Pinty B, Widlowski JL, Taberner M, Gobron N, Verstraete MM, Disney M, Gascon F, Gastellu JP, Jiang L, Kuusk A, Lewis P, Li X, Ni-Meister W, Nilson T, North P, Qin W, Su L, Tang S, Thompson R, Verhoef W, Wang H, Wang J, Yan G, Zang H (2004b) Radiation transfer model intercomparison (RAMI) exercise: results from the second phase. J. Geophys. Res. 109, D06210, doi:10.1029/2003JD004252 Qin W, Gerstl SAW, Deering DW, Goel NS (2002) Characterizing leaf geometry for grass and crop canopies from hotspot observations: a simulation study. Remote Sens. Environ. 880:100–113

170

P. Stenberg et al.

Rautiainen M (2005) Retrieval of leaf area index for a coniferous forest by inverting a forest reflectance model. Remote Sens. Environ. 99:295–303 Rautiainen M, Suomalainen J, M˜ottus M, Stenberg P, Voipio P, Peltoniemi J, Manninen, T (2007) Coupling forest canopy and understory reflectance in the Arctic latitudes of Finland. Remote Sens. Environ. 110:332–343 Rautiainen M, Stenberg P (2005a) Application of photon recollision probability in coniferous canopy reflectance simulations. Remote Sens. Environ. 96:98–107 Rautiainen M, Stenberg P (2005b) Simplified tree crown model using standard forest mensuration data for Scots pine. Agric. For. Meteorol. 128:123–129 Rautiainen M, Stenberg P, Nilson T, Kuusk A (2004) The effect of crown shape on the reflectance of coniferous stands. Remote Sens. Environ. 89:41–52 Ross J (1981) The Radiation Regime and Architecture of Plant Stands. Junk Publishers, The Hague, 391 pp Ross KJ, Marshak AL (1988) Calculation of canopy bidirectional reflectance using the Monte Carlo method. Remote Sens. Environ. 24:213–225 Shabanov NV, Knyazikhin Y, Baret F, Myneni RB (2000) Stochastic modeling of radiation regime in discontinuous vegetation canopies. Remote Sens. Environ. 74:125–144 Shabanov NV, Wang Y, Buermann W, Dong J, Hoffman S, Smith GR, Tian Y, Knyazikhin Y, Myneni RB (2003) Effect of foliage spatial heterogeneity in the MODIS LAI and FPAR algorithm over broadleaf forests. Remote Sens. Environ. 85:410–423 Smolander S, Stenberg P (2003) A method to account for shoot scale clumping in coniferous canopy reflectance models. Remote Sens. Environ. 88:363–373 Smolander S, Stenberg P (2005) Simple parameterizations for the radiation budget of uniform broadleaved and coniferous canopies. Remote Sens. Environ. 94:355–363 Soler C, Sillion FX, Blaise F, Dereffye P (2003) An efficient instantiation algorithm for simulating radiant energy transfer in plant models. ACM Trans. Graphics. 22:204–233 Stenberg P (2006) A note on the G-function for needle leaf canopies. Agric. For. Meteorol. Stenberg P, Rautiainen M, Manninen T, Voipio P, Smolander H (2004) Reduced simple ratio better than NDVI for estimating LAI in Finnish pine and spruce stands. Silva Fennica. 38(1):3–14 Thompson RL, Goel NS (1998) Two models for rapidly calculating bidirectional reflectance: Photon spread (ps) model and statistical photon spread (sps) model. Remote Sensing Reviews 16:157–207 Tian Y, Zhang Y, Knyazikhin Y, Myneni RB, Glassy JM, Dedieu G, Running SW (2000) Prototyping of MODIS LAI and FPAR algorithm with LASUR and LANDSAT data. IEEE Trans. Geosci. Remote Sens. 38:2387–2401 Tian Y, Wang Y, Zhang Y, Knyazikhin Y, Bogaert J, Myneni RB (2002) Radiative transfer based scaling of LAI retrievals from reflectance data of different resolutions. Remote Sens. Environ. 84:143–159 Verhoef W (1984) Light-scattering by leaf layers with application to canopy reflectance modeling – the SAIL model. Remote Sens. Environ. 16:125–141 Verhoef W (1985) Earth observation modeling based on layer scattering matrices. Remote Sens. Environ. 17:165–178 Verstraete MM, Pinty B, Dickinson RE (1990) A physical model of the bidirectional reflectance of vegetation canopies. Part 1: theory. J. Geophys. Res. 95:11765–11775 Verstraete M, Pinty B, Myneni R (1996) Potential and limitations of information extraction on the terrestrial biosphere from satellite remote sensing. Remote Sens. Environ. 58:201–214 Wang Y, Buermann W, Stenberg P, Smolander H, H¨ame T, Tian Y, Hu J, Knyazikhin Y, Myneni RB (2003) A new parameterization of canopy spectral response to incident solar radiation: case study with hyperspectral data from pine dominant forest. Remote Sens. Environ. 85:304–315 Wang YP, Jarvis PG (1990) Influence of crown structural properties on PAR absorption, photosynthesis, and transpiration in Sitka spruce: application of a model (MAESTRO). Tree Physiology. 7:297–316 Warren Wilson J (1967) Stand structure and light penetration. III. Sunlit foliage area. J. Appl. Ecol. 4(1):159–165

6 Modeling the Spectral Signature of Forests

171

Weiss M, Baret F (1999) Evaluation of canopy biophysical variable retrieval performances from the accumulation of large swath satellite data. Remote Sens. Environ. 70:293–306 Weiss M, Baret F, Myneni RB, Pragn`erea A, Knyazikhin Y (2000) Investigation of a model inversion technique to estimate canopy biophysical variables from spectral and directional reflectance data. Agronomie. 20:3–22 Weiss M, Baret F, Smith GJ, Jonckheere I, Coppin P (2004) Review of methods for in situ leaf area index (LAI) determination. Part II. Estimation of LAI, errors and sampling. Agric. For. Meteorol. 121:37–53 Welles JM, Norman JM (1991) Photon transport in discontinuous canopies: a weighted random approach. In: RB Myneni, J Ross (eds), Photon-Vegetation Interaction, Springer, BerlinHeidelberg, pp 389–414 Widen N (2004) Assessing the accuracy of land surface characteristics estimated from multiangular remotely sensed data. Int. J. Remote Sens. 25:1105–1117 Yoder B, Pettigrew-Crosby R (1995) Predicting nitrogen and chlorophyll content and concentrations from reflectance spectra (400–2500 nm) at leaf and canopy scales. Remote Sens. Environ. 53:199–211 Zarco-Tejada PJ, Rueda CA, Ustin SL (2003) Water content estimation in vegetation with MODIS reflectance data and model inversion methods. Remote Sens. Environ. 85:109–124

Chapter 7

Estimating Canopy Characteristics from Remote Sensing Observations: Review of Methods and Associated Problems Fr´ed´eric Baret and Samuel Buis

Abstract This article describes the methods and problems associated to the estimation of canopy characteristics from remote sensing observations. It is illustrated over the solar spectral domain, with emphasis on LAI estimation using currently available algorithms developed for moderate resolution sensors. The principles of algorithms are first presented, distinguishing between canopy biophysical and radiometric data driven approaches that may use either radiative transfer models or experimental observations. Advantages and drawback are discussed with due attention to the operational character of the algorithms. Then the under-determination and ill-posedness nature of the inverse problem is described and illustrated. Finally, ways to improve the retrieval performances are presented, including the use of prior information, the exploitation of spatial and temporal constraints, and the interest in using holistic approaches based on the coupling of radiative transfer processes at several scales or levels. A conclusion is eventually proposed, discussing the three main components of retrieval approaches: retrieval techniques, radiative transfer models, and the exploitation of observations and ancillary information.

7.1 Introduction Many applications require an exhaustive description of the spatial domain of interest that may cover a large range of scales: from the very local one corresponding to precision agriculture where cultural practices are adapted to the within field variability, through environmental management generally approached at the landscape scale, up to biogeochemical cycling and vegetation dynamics investigated at national, continental and global scales. Most of these applications are using our knowledge on the F. Baret and S. Buis UMR1114, INRA-CSE, 84 914 Avignon, France [email protected] S. Liang (ed.), Advances in Land Remote Sensing, 173–201. c Springer Science + Business Media B.V., 2008

173

174

F. Baret, S. Buis

main physical, chemical and biological processes involved such as energy balance, evapotranspiration, photosynthesis and respiration. This knowledge is encapsulated into a variety of surface process models. However, to account for the spatial heterogeneity observed at all scales, dedicated imaging systems are required to get a distributed description of surface characteristics within the domain of interest. By its capacity to cover exhaustively large space areas, remote sensing provides a very pertinent answer to those requirements. However, remote sensing observations sample the radiation field reflected or emitted by the surface, and thus do not provide directly the biophysical characteristics required by the models for describing some state variables of the surface. An intermediate step is therefore necessary to transform the remote sensing measurements into estimates of the surface biophysical characteristics. Many methods have been proposed to retrieve surface characteristics from remote sensing observations. They span from simple empirical ones with calibration over experimental data sets, up to more complex ones based on the use of radiative transfer models. Radiative transfer models summarize our knowledge on the physical processes involved in the photon transport within vegetation canopies or atmosphere, and simulate the radiation field reflected or emitted by the surface for given observational configuration, once the vegetation and the background as well as possibly the atmosphere are specified. Retrieving canopy characteristics from the radiation field as sampled by the sensor aboard satellite needs to “invert” the radiative transfer model. This article aims at presenting the state of the art in the estimation of surface characteristics from remote sensing observations. Although this is a very general problem in remote sensing, it will be illustrated by examples taken in the solar domain (400–2,500nm), with emphasis put on the current operational algorithms that are mainly used for medium resolution sensors such as MODIS, MERIS, AVHRR, VEGETATION, POLDER and SEAWIFS. Among the possible canopy characteristics accessible from remote sensing in the reflective solar domain, we will focus on leaf area index (LAI), defined as half the developed area of green elements per unit horizontal soil (Stenberg, 2006). As a matter of fact, LAI is one of the key canopy state biophysical variables required by many process models to describe energy and mass exchanges in the soil/plant/atmosphere system.

7.2 Principles of Biophysical Variable Retrieval Algorithms Remote sensing data result from radiative transfer processes within canopies that depend on canopy variables, and observational configuration (wavelength, view and illumination directions). Canopy variables include the variables of interest for the applications such as LAI, and the other variables that are not of direct use for the applications but that influence the radiative transfer, such as soil background properties. The causal relationship between the variables of interest and remote sensing data corresponds to the forward (or direct) problem (Fig. 7.1). They could be

7 Estimating Canopy Characteristics from Remote Sensing Observations

175

Other variables Radiative transfer Variables of interest

Inverse problem

Observation configuration

Forward problem

Remote sensing data

Retrieval algorithm Prior information

Fig. 7.1 Forward (solid lines) and inverse (dashed lines) problems in remote sensing

either described through empirical relationships calibrated over experiments or using radiative transfer models based on a more or less close approximation of the actual physical processes. Conversely, retrieving the variables of interest from remote sensing measurements corresponds to the inverse problem, i.e., developing algorithms to estimate the variables of interest from remote sensing data as observed in a given configuration. Prior information on the type of surface and on the distribution of the variables of interest can also be included in the retrieval process to improve the performances as we will see later. A panoply of retrieval techniques currently used have been reviewed in the early 1990s by several authors (Asrar et al., 1989; Goel, 1989; Pinty and Verstraete, 1991) and more recently by Kimes et al. (2000) and Liang (2004). They can be split into two main approaches (Fig 7.2) depending if the emphasis is put on remote sensing data (radiometric data) or on the variables of interest to be estimated (canopy biophysical variables).

7.2.1 Canopy Biophysical Variables-driven Approach The approach requires first to calibrate the inverse model: a parametric model representing the inverse model is adjusted over a learning data set (Fig. 7.2, left). It mainly consists in adjusting the parameters to fit a response surface between reflectance values and the corresponding canopy variables of interest (LAI in this example). Once calibrated, the parametric model is run to compute the variables of interest from the observed reflectance values. The learning data set can be generated either using simulations of radiative transfer models, or based on concurrent experimental measurements of the variables of interest and reflectance data.

176

F. Baret, S. Buis

Radiometric data driven approach

Biophysical variables driven approach Learning data set LAI

RT process

Reflectance

LAI

∆ LAI*

RT process

Reflectance

∆ Inverse Model

Reflectance Measurement

Parameters

Fig. 7.2 The two main approaches used to estimate canopy characteristics from remote sensing data for LAI estimation. On the left side the approach focusing on the biophysical variables showing the calibration of the inverse model. Once the inverse model is calibrated it can be applied using the measured reflectance as input. On the right side, the approach focusing on radiometric data showing the solution search process leading to the estimated LAI value, LAI ∗ . “∆” represents the cost function to be minimized over the biophysical variables (left) or over the radiometric data (right)

7.2.1.1 Calibration over Experimental Data Sets This was the first approach historically used, the reflectance in few bands being generally combined into vegetation indices (VI) designed to minimise the influence of confounding factors such as soil reflectance and atmospheric effects (Baret and Guyot, 1991). The relationships between VIs and canopy variables are calibrated over experimental observations (Asrar et al., 1984; Huete, 1988; Wiegand et al., 1990; Wiegand et al., 1992; Richardson et al., 1992). Recently Chen et al. (2002) used simple VIs to derive LAI estimates from AVHRR and VEGETATION across Canada. This was extended at the global scale by Deng et al. (2006). In agreement with several observations, these authors found that the relationships vary from one cover type to another as illustrated by Fig. 7.3. The development of such empirical transfer functions is limited by the difficulty to get a training data base that represents the whole range of possible conditions encountered over the targeted surfaces, i.e., combinations of geometrical configurations, type of vegetation and states including variability in development stage and stress level, and type of background and state (roughness, moisture). Measurement errors associated both to the variables of interest and to radiometric data may also propagate into uncertainties and biases in the algorithm and should be explicitly accounted for Fernandes and Leblanc (2005) and Huang et al. (2006). Further, since ground measurements having a footprint ranging from few meters to few decametres, specific sampling designs should be developed to represent the sensor pixel. This task is obviously more difficult for medium and coarse resolution sensors as outlined by Morisette et al. (2006). Higher spatial resolution observations could be used to extend the local ground measurements to the actual pixel size of medium or coarse resolution sensors.

7 Estimating Canopy Characteristics from Remote Sensing Observations

177

10 Coniferous Deciduous Mixed Others

8

LAI

6

4

2

0

0

5 10 RSR VEGETATION

Fig. 7.3 Empirical relationships between LAI values as a function of the simple ratio vegetation index (RSR) computed for VEGETATION data for four types of canopies (After Chen et al., 2002)

7.2.1.2 Calibration over Radiative Transfer Model Simulations To avoid limitations associated to the empirical nature of the training data base, radiative transfer models could be used alternatively to generate a training data base. Radiative transfer models can be used to create a data base covering a wide range of situations and configurations. Several authors have therefore proposed replacing actual observations by numerical experiments based on radiative transfer model simulations to calibrate empirical relationships (Sellers, 1985; Baret and Guyot, 1991; Rondeaux et al., 1996; Leprieur et al., 1994; Banari et al., 1996; Huete et al., 1997; Verstraete and Pinty, 1996). Based on these principles, operational algorithms developed for medium resolution sensors are currently used: MGVI for MERIS (Gobron et al., 2000) further extended to other sensors, MODIS back-up algorithm based on NDVI (Knyazikhin, 1999), POLDER algorithm based on DVI computed from bidirectional reflectance factor (BRF) measurements normalized to a standard geometrical configuration (Roujean and Lacaze, 2002). Nevertheless, although quite often effective, VIs are intrinsically limited by the empiricism of their design and the small number of bands concurrently used (generally 2–3). This might not be a major problem for fAPAR and fCover variables that are relatively simple to estimate, but would be more difficult for variables such as LAI or chlorophyll content (Cab ) showing higher level of non linearity with reflectance measurements (Weiss et al., 2000). The efficient interpolation capacity of neural network (NNT) can be exploited to adjust surface responses (Leshno et al., 1993). Several authors have proposed such an approach since the beginning of the 1990s (Smith, 1992; Smith, 1993; Atkinson and Tatnall, 1997; Kimes et al., 1998; Abuelgasim et al., 1998; Gong et al., 1999;

178

F. Baret, S. Buis

Danson et al., 2003). Neural networks were compared with a specific implementation of multiple regression, the projection pursuit regression, and were concluded to achieve very similar performances (Fang and Liang, 2005). Baret et al. (1995) demonstrated that NNT used with individual bands were performing better than classical approaches based on vegetation indices especially when calibrated with radiative transfer model simulations rather than with experimental observations. Weiss et al. (2002a) validated such techniques over a range of crops for estimating the main canopy biophysical parameters LAI and fCover from airborne POLDER instrument. Recently, several authors developed operational products for medium resolution sensors, starting from top of canopy level: Lacaze (2005) for POLDER, Bacour et al. (2006) for MERIS, and Baret et al. (2007) for VEGETATION instruments. Baret et al. (2006b) proposed an operational algorithm from the MERIS top of atmosphere data by coupling an atmospheric radiative transfer model to the surface one, exploiting explicitly 13 over the 15 bands of MERIS. Several ways may be used to build a data set for training empirical relationships depending on the performances targeted. Evaluation of the performances of an algorithm is generally achieved by computing the Root Mean Square Error (RMSE) value over a test data base made of representative cases. Best performances will therefore be obtained when the variables in the training data base are distributed similarly to those in the testing one, i.e., close to the actual distribution of the variables: the coefficients of the empirical transfer function will be optimized for these conditions, and uncertainties will be minimal for the most frequent cases. Although achieving poorer performances in term of RMSE, a more even distribution of the uncertainties may be alternatively obtained using uniform distributions of the variables. Note that, for a given number of cases simulated in the training data base, the density of cases that populate the space of canopy realization may rapidly decrease as a function of the number of required variables. Experimental plans may be used in this situation as proposed by Bacour et al. (2002b), in order to focus on the first order effects and interactions. Additionally, Baret et al. (2006b) proposed to steamline the data base in the reflectance space by retaining the cases that belong both to the simulated and actual remote sensing measurements spaces (Fig. 7.4). This allows discarding cases that were simulated but not actually observed. Conversely, it allows also identifying cases which are observed but not simulated. This is achieved by first compiling a large data base of reflectance measurements that should be representative of the possible situations available. Then the reflectance mismatch is

Simulations Cases not represented in the measured database

Actual measurements Cases not represented in the simulated database Training database: selection of cases in the intersection space

Fig. 7.4 Streamlining the simulated training data set by comparison to actual measurements. The intersection between the space of simulated radiometric data (in dark gray) with that of the actual measurements (in light gray) is used as the training data base

7 Estimating Canopy Characteristics from Remote Sensing Observations

179

computed for each case in the simulated data base: it is the minimum RMSE value computed between the reflectance in the simulated data base and the ensemble of actual measurements. A threshold corresponding to the uncertainties in the radiometric measurements is then used to decide whether a simulated case is rejected from the training data base. Additional criterions could be used to streamline the training data base, based on the expected consistency between several products such as LAI and fAPAR as proposed by Bacour et al. (2006). Although the use of radiative transfer models appears very appealing, this approach is however limited by several aspects. The first one is the capacity of the models to get a faithful description of the radiative transfer in canopies. Up to now, most radiative transfer models used are computer efficient ones allowing populating large training data base within few hours/days with a single regular computer. They generally correspond to simple description of canopy architecture which may not represent the actual one, particularly regarding the clumped nature of many vegetation types. This leads to model uncertainties that may dominate all other sources of uncertainties for some of the vegetation types. Recent advances in modeling more complex canopy architecture (e.g., Soler et al., 2001; Lewis et al., 2004) offer great potential for improvement. However, the second limitation will probably counterbalance these advancements: building a realistic training data set requires a fair description of the distribution and co-distribution of the corresponding architectural variables to define the actual space of canopy realization. For the simplest radiative transfer models (e.g., Verhoef, 1984; Kuusk, 1995; Gobron et al., 1997) at least three architectural variables are required (LAI, leaf angle distribution function and size of the leaves relative to canopy height), the distribution of which being very poorly known. This is even more difficult when using more complex and realistic architectural description that requires more variables. Note that in these approaches based on radiative transfer model simulations, radiometric measurements uncertainties have to be added to the simulations when building up the training data base. This allows more robustness within the training process and thus improved retrieval performances. Accounting for these uncertainties is also critical when large differences exist among bands used or when these uncertainties are strongly correlated. Canopy biophysical variables driven approaches present the advantage of being very flexible. For example, estimates of biophysical variables from one sensor could be used to constitute the training data base for another sensor. This could be applied over high spatial resolution products that are aggregated to coarser spatial resolution to generate an appropriate training data base. This could also apply to generate consistent products between sensors.

7.2.2 Radiometric Data-driven Approach While the previous approach was focusing on minimizing the distance between the variables retrieved from the inverse model and those from the training data set, the alternative approach is based on finding the best match between the measured

180

F. Baret, S. Buis

reflectance values and those either simulated by a radiative transfer model or stored within a database made of experimental observations. No proper calibration step is required in this approach. However, several ingredients of these techniques are difficult to evaluate (uncertainties, parameters of the search algorithm) and need generally some tuning over a “prototyping” data set. The performances of the approach will both depend on the minimization algorithm itself and on the level of ill-posedness of the inverse problem as a function of measurement configuration and model and measurement uncertainties. Several minimization techniques have been used: classical iterative optimization, simulated annealing (Bacour, 2001), genetic algorithms (Fang et al., 2003; Renders and Flasse, 1996), look up tables and Monte Carlo Markov Chains (Zhang et al., 2005). However, classical iterative optimization techniques (OPT) and look up tables (LUT) have been the most widely used and will be described with more details below.

7.2.2.1 Iterative Optimisation (OPT) This classical technique consists in updating the values of the unknown input biophysical canopy radiative transfer model variables until the simulated reflectance closely fit the corresponding measurements (Goel and Deering, 1985; Kuusk, 1991a and 1991b; Goel, 1984a and b; Pinty et al., 1990; Jacquemoud et al., 1995; Privette et al., 1996; Bicheron and Leroy, 1999; Combal et al., 2000; Bacour et al., 2002a; Combal et al., 2002). A good review on optimization methods used in remote sensing for land applications can be found in Bacour (2001). The goodness of fit between measured and simulated reflectance spectra is quantified by a cost function (J) that may account explicitly for measurements and model uncertainties. The cost function may be theoretically derived from the maximum likelihood (Tarentola, 1987). When no prior information is available and when uncertainties associated to each configuration used are assumed independent and gaussian, J is assessed using norm L2, i.e., sum over the N observational configurations of the square of the difference beˆ weighed by the tween the measured reflectance values (R) and those simulated (R), variance (σ2 ) associated to both reflectance measurements and model uncertainties: (Rn − Rˆ n )2 σn2 n=1 N

J=

∑

(7.1)

However, because of the difficulty to provide an estimate of σ2 , several approximations have been used as shown in Bacour (2001). It spans from the simple ones such as norm L1 to norm L2 with no weighing of the configurations, up to more complex based on some modeling of the variance term (Table 7.1). The main limitation of OPT techniques is twofold. (1) Firstly, the algorithm might converge to a local minimum of the cost function that could be far away from the global one expected to correspond to the actual solution. This can be partly avoided by using a range of initial solutions, coupled with constraints on the range of variation of the variables to be estimated. The use of a priori information in the

7 Estimating Canopy Characteristics from Remote Sensing Observations

181

Table 7.1 The cost functions (J) used in several studies dealing with radiative transfer model inversion for canopy biophysical variables retrieval. N is the number of configurations (bands and directions); Rˆ n and Rn being respectively the simulated and measured reflectance values for configuration n. θν and φ are the zenith and relative azimuth view angles Cost function N

J= ∑

References

Rn −Rˆ n Rn

Gao and Lesht, 1997)

n=1 N

ˆ J = ∑ RnR−nRn n=1 N

J= ∑

n=1 N

Rn −Rˆ n Rn +Rˆ n

(Qiu et al., 1998)

2 (Gobron et al., 1997)

J = ∑ (Rn − Rˆ n )2

(Goel and Thompson, 1984; Pinty et al., 1990; Privette et al., 1996; Braswell et al., 1996; Jacquemoud et al., 2000; Combal et al., 2002)

n=1 N

J= ∑

n=1 N

Rn −Rˆ n Rn

J = ∑ ωn n=1

2

Rn −Rˆ n Rn

2

(Nilson and Kuusk, 1989; Kuusk, 1991a and b; Bicheron and Leroy, 1999; Weiss et al., 2000) ; ωn =

cos(θν ·sin(φ ))+1 2

(Bacour et al., 2002a)

cost function generally improves the convexity of the error surface, which is critical as we will see later (Combal et al., 2002). The descent algorithm may also limit the trapping in a local minimum by reducing the rate of descent. However, a compromise has to be chosen between rapid convergence achieved with large descent rate, and limiting the probability of falling in a local minimum achieved with a slow descent rate. Further, the optimization algorithm may sometimes lack of robustness due to numerical problems occurring generally with very small values of J. The criterion used to stop the iterations is in addition not always easy to adjust, requiring some preliminary tests (Bonnans et al., 2006). (2) Secondly, the OPT algorithm requires large computer resources because of its iterative nature. However, there are ways to speed up the process by limiting the number of model runs for each iteration using the adjoint model that provides an analytical expression of the gradient of the cost function (Lauvernet et al., 2007). Nevertheless, OPT techniques are still difficult to use routinely and exhaustively over large images, although image segmentation may help reducing significantly the number of pixels to process, the optimization process being restricted to a limited set of representative pixels. Note that these techniques allow getting some estimates of the uncertainties associated to the solution under some assumptions. However, the distribution of the solution will be here always unimodal, conversely to what could be achieved with the other radiometric driven approaches. The main advantage of iterative optimization methods is their flexibility, allowing retrieving canopy characteristics from several observational configurations. It is even possible to invert radiative transfer models concurrently over several pixels. This opens great potentials for exploiting additional temporal or spatial constraints as we will see later.

182

F. Baret, S. Buis

7.2.2.2 Look Up Tables This is conceptually the simplest technique, although its implementation is not trivial (Weiss et al., 2000). It is the basis of the MODIS and MISR LAI and fAPAR products (Knyazikhin et al., 1999). Firstly a large data base (the Look Up Table, LUT) is generated, consisting of sets of input variables of the canopy radiative transfer model used. Then, the corresponding reflectance values are simulated. The LUT can alternatively be based on experimental observations, although this requires a very good sampling of the space of canopy realization. Once the LUT has been generated, finding the solution for a given set of reflectance measurements consists in selecting the closest cases in the reflectance table according to a cost function, and then extracting the corresponding set of canopy biophysical variables. Note that the distribution of the solution could be obtained by accounting for the uncertainties associated to the reflectance values as discussed by Knyazikhin et al. (1998a and b). This technique overcomes some of the limitations of iterative optimization techniques. As a matter of fact, the search for the solution is global here, leading to the true minimum if the space of canopy realisation is sufficiently well sampled. Note that for generating the LUT, the space of canopy realization has to be sampled to represent the surface response, i.e., with better sampling where the sensitivity of reflectance to canopy characteristics is the higher (Weiss et al., 2000; Combal et al., 2002). This is different from the sampling of the training data base required in canopy biophysical variables driven approaches. The implementation of a LUT technique in algorithmic operational chains is very efficient because the radiative transfer model is run off-line. However, LUT techniques require a fixed number of inputs unless having very large tables that could be more difficult to manipulate. In addition, the way the solution is defined is not always based on solid theoretical background. The cases selected as possible solutions are either defined as a fraction of the initial population of cases (after tests and trials) such as in Weiss et al. (2000) or Combal et al. (2002). It can be also defined by a threshold corresponding to measurement and model uncertainties as in Knyazikhin et al. (1998a and b).

7.2.2.3 Bayesian Methods: Importance Sampling and MCMC Alternative methods are available which are based on statistical backgrounds: MonteCarlo Markov Chains (MCMC) and Importance Sampling (IS) (Makowski D., J. Hiller, et al., 2006). These two Bayesian methods approximate the posterior distribution, i.e., the distribution of the variables when the reflectance measurement is known. Although very little attention has been paid to these techniques at the exception of Zhang et al. (2005) who used with success the MCMC MetropolisHastings algorithm with MODIS data. However, Metropolis-Hastings algorithm is an iterative process that might not be well suited for operational applications at

7 Estimating Canopy Characteristics from Remote Sensing Observations

183

large scale, similarly to OPT methods. Conversely, IS methods that do not require multiple iterations might be efficient for this purpose and need to be properly evaluated for remote sensing applications.

7.3 The Under-determined and Ill-posed Nature of the Inverse Problem in Remote Sensing 7.3.1 Under-Determination of the Inverse Problem Estimating biophysical variables from remote sensing measurements is often an under-determined problem: the number of unknowns is generally larger than the number of independent radiometric information remotely sampled by sensors. In the case of a simple canopy radiative transfer model such as SAIL (Verhoef, 2002), canopy reflectance at the top of canopy (ρ toc ) for a given illumination and view geometry (θ s, θ v, ϕ ) is simulated (Eq. (7.2)) using three variables describing canopy structure that do not depend on wavelength (LAI, average leaf angle (ALA) and hot spot parameter (hot) as modelled by Kuusk (1995)), and leaf reflectance (refl) and transmittance (tran) as well as soil reflectance (Rs) that obviously depend on wavelength (λ ). ρtoc (λ, θs, θv, ϕ) = CAN(LAI, ALA, hot, re f l(λ ), tran(λ ), Rs(λ, θs, θv, ϕ), θs, θv, ϕ) (7.2) Several studies report that canopy (and soil) bidirectional reflectance distribution function (BRDF) could be decomposed using empirical or semi-empirical orthogonal functions with generally 2–4 kernels (Lucht, 1998; Br´eon et al., 2002; Weiss et al., 2002). Therefore, 7–9 characteristics (3 canopy structure, 2 leaf properties [refl, tran] input variables and the 2–4 terms describing soil BRDF, Rs(λ, θs, θv, ϕ) have to be estimated out of a maximum of 4 independent information derived from BRDF measurements in a single band. Retrieval of canopy characteristics from BRDF measurements in a single band is therefore not possible without introducing other information in the system, particularly when soil background plays a significant role, i.e., for low to medium LAI values. Similar observations are made when considering the reflectance spectral variation: leaf spectral properties may be described by a dedicated model such as PROSPECT (Jacquemoud and Baret, 1990) requiring at least 5 input variables: mesophyll structure parameter (N), chlorophyll (Cab ), dry matter (Cdm ), brown pigment (Cbp ) and water (Cw ) contents: [re f l(λ ), tran(λ )] = LEAF(N, Cab , Cdm , Cbp , Cw , λ )

(7.3)

184

F. Baret, S. Buis

Soil reflectance Rs(λ , θ s, θ v, ϕ ) may be described by a model such as that proposed by Jacquemoud et al. (1992) and derived from that of Hapke (1981). It requires a single scattering albedo ω (λ ) that varies with wavelength and soil composition, between 1 to 4 phase function coefficients (αi ), and a roughness parameter (r). According to Price (1990), soil spectral variation, may be approximated as a linear combination of 2–10 end-members. This is assumed to apply similarly to the spectral variation of the single scattering albedo with weigh w j and end members ω j (λ ): ω (λ ) = ∑ w j · ω j (λ ) (7.4) j

The whole soil spectral and directional reflectance field could subsequently be simulated with at least five parameters: Rs(λ , θ s, θ v, ϕ ) = SOIL([w j ], [αi ], r, λ , θ s, θ v, ϕ )

(7.5)

Consequently, the whole spectral and directional top of canopy reflectance field could therefore be modelled by coupling together the soil, leaf and canopy reflectance models, which leads to at least 13 input variables. These 13 unknowns have to be estimated from the information content in remote sensing measurements. Most of currently available sensors for which operational biophysical products are available have a relatively small number of configurations: from two for AVHRR (red and near infrared bands), to 15 bands for MERIS (VIS and NIR) and MODIS (VIS, NIR, SWIR) with several bands dedicated to particular atmosphere, cloud, snow/ice, or ocean characteristics. In the case of multidirectional sensors, the number of configurations may be larger as in the case of MISR (36 configurations = 9 cameras ×4 bands), or POLDER (84 configurations = 14 directions ×6 bands). However, the actual dimensionality of remote sensing measurements is much smaller than the number of available configurations considering the relatively high level redundancy between bands (Price, 1994; Price, 1990; Liu et al., 2002; Green and Boardman, 2001) and directions (Zhang et al., 2002a and b; Weiss et al., 2002b). Although further investigation is required to better quantify the actual dimensionality of remote sensing observations, it is clear that retrieval of surface characteristics from reflectance measurements is an under-determined problem in many cases. Improving retrieval performances will require introducing ancillary information and constraints in the system.

7.3.2 Evidence of the Ill-posed Problem A problem is well posed if and only if its solution exists, is unique, and depends continuously on the data (Garabedian, 1964). Several authors have reported that the inverse problem in remote sensing is ill-posed (Knyazikhin et al., 1999; Combal et al., 2001; Baret et al., 2000) because of its under-determination and uncertainties attached to models and measurements. In addition, models may incorporate sets of

7 Estimating Canopy Characteristics from Remote Sensing Observations 80

Cab (µg.cm-2)

Reflectance

0.3

0.2

0.1

0 400

185

600 800 Wavelength (nm)

60

40

20

0

0.5

1 LAI

1.5

2

Fig. 7.5 Actual reflectance measurements (left plot, solid lines representing the mean and standard deviations) and the corresponding closer simulations achieved with a simple turbid medium radiative transfer model (the series of dots). On the right, the input LAI and Cab (the “+” symbols) variables used to simulate the reflectance spectra shown on the left plot. The actual LAI and Cab measurements are displayed with their associated confidence interval (bold line corresponding to 1 standard deviation). Data acquired over a sugar beet experiment conducted in 1990

variables that appear always in combinations such as products between variables. In these conditions, very similar reflectance spectra simulated by a radiative transfer model (Fig. 7.5, left) may correspond to a wide range of solutions (Fig. 7.5, right). In the case illustrated by Fig. 7.5, high correlation is found between LAI and those leaf chlorophyll content estimated values. This compensation between variables was sometimes termed “ambiguity” (Baret et al., 1999) or “equi-finality” (Shoshany, 1991; Teillet et al., 1997). This may also indicate that the product LAI · Cab should be used in place of individual estimates of LAI and Cab . Although not appearing formally in the radiative transfer model, this product is physically meaningful from the radiative transfer processes perspective and corresponds to the actual optical thickness of the medium (Weiss et al., 2000). Measurement and model uncertainties may also induce instability in the solution of the inverse problem. This is particularly true for well developed canopies, where a small variation in the measured reflectance can translate into large variation of variables such as LAI, for which reflectance “saturates”, i.e., is very little sensitive to LAI variation. A proper sensitivity analysis should help quantifying interactions between input variables. A complementary sensitivity analysis conducted over the cost function could also help evaluating the identifiability of the solution, i.e., if output variables could be accurately retrieved from a given set of observations (Salteli, 2004). Regularization techniques are thus necessary to obtain a stable and reliable solution of the ill-posed inverse problem. This could be achieved both by using prior information on the distribution of the variables, and by exploiting some constraints on the variables. These two issues will be investigated separately in the following.

186

F. Baret, S. Buis

7.4 Improving the Retrieval Performances 7.4.1 Using Prior Information If no remote sensing measurement is available, the best estimates of the variables would come from the prior information on their distribution (Fig. 7.6d), capitalizing, all the knowledge coming from bibliography, past experiments or experts. Conversely, when a radiative transfer model is available along with remote sensing measurements, the variables can be estimated by inverting the RT model without using any prior information. This will be illustrated using a simple example: estimating LAI from NDVI vegetation index. In this case the RT model consists in an analytical relationship as proposed by Baret and Guyot (1991): NDV I = NDV I∞ + (NDV Is -NDV I∞ ) · e−K·LAI 1

b

c

0.8

0.8

0.6

0.6

0.6

0.4 0.2 0

NDVI

0.8

NDVI

NDVI

1

1

a

0.4 0.2

Remote Sensing measurements (NDVI)

0

0.05

0.02 PDF

d

2

4 LAI

6

0

8

Estimates from RS measurements and RT model without prior information

0

2

4 LAI

6

8

0.05

1

Prior information on LAI

0.4 0.2

Radiative Transfer model: NDVI=f(LAI)+e

0 0

0.04

(7.6)

e

NDVI estimates with PI

f

0.04

0.8

0.04

0.03

0.6

0.03

0.02

0

0.4

2

4 LAI

6

8

0

NDVI estimates without PI

0.02 Estimates from RS measurements and RT model with prior information

0.2

0.01 0

PDF

NDVI

PDF

Prior Information

0

2

4 LAI

6

0.01

8

0

0

2

4 LAI

6

8

Fig. 7.6 Estimation of canopy variables by combining remote sensing measurements, radiative transfer model and prior information. All these pieces of information are represented by their probability distribution function (PDF): (a) PDF of remote sensing measurements in the simple case of NDVI; (b) PDF of RT model simulations (NDV I = f(LAI)) accounting for model uncertainties; (c) PDF of LAI as retrieved from RT model and NDVI measurement and their associated uncertainties, without using prior information; (d) PDF of LAI used as prior information; (e) Computation of LAI PDF as estimated from NDVI measurements and RT model, using prior information on LAI; (f) PDF of the solution (posterior distribution) when using only prior information (idem as plot d), using RT model and NDVI measurements and their associated uncertainties only, and using all the information available (RT model and NDVI measurements and their associated uncertainties and prior information). The three contour plots (b, c, e) are coded from white to black for zero to max PDF values with the same gray scale. Very simple assumptions on uncertainties models and values are used here just for illustration

7 Estimating Canopy Characteristics from Remote Sensing Observations

187

with NDV Is and NDV I∞ being respectively the bare soil and asymptotic values of NDVI, and K an extinction coefficient (K = 0.8). However, uncertainties are associated both with remote sensing measurements (Fig. 7.6a NDV Ir = ℵ(0.8, 0.1) where ℵ(x, σ2 ) means a Gaussian distribution with mean x and variance σ2 ) and the RT model (Fig. 7.6b RT model represented by Eq. (7.2) with a Gaussian noise ℵ(0, 0.1)). Accounting for these uncertainties in the form of the corresponding probability distribution function (PDF) allows deriving the PDF of the estimated variable (Fig. 7.6c). The small sensitivity of NDVI to LAI as compared to measurement and model uncertainties induce a relatively broad PDF for the larger LAI values (Figs. 7.6c, f). This corresponds to an ill-posed problem, where a wide range of possible solutions match very similar measurements. The combination of RT model, remote sensing measurements and prior information on the variables (here LAI = ℵ(2, 1.5) allows getting more reliable solutions accounting for all the sources of information available in an optimal way (Figs. 7.6e, f). The example provided above for a measurement value of NDV I = 0.8 could be extended to the whole range of NDVI values. It shows that the mode of the distribution of the solution corresponding to the maximum likelihood (maximum of the PDF) strongly depends on the type of input information used (Fig. 7.7, left). When only prior information is used, the mode stays constant and obviously independent from measurements. When RT model and measurements are used with their uncertainties, the LAI mode is generally close to the values obtained without considering uncertainties, assuming perfect model and measurements. However, over the saturation domain corresponding to NDVI values higher than 0.85, accounting for the uncertainties provides lower modal values because of the non linearity of the

5

LAI

4 3

3

Mode

Standard deviation

RT model

2.5

RT model and measurement uncertainties

RT model and measurement uncertainties

2

RT model and measurement uncertainties and prior information

LAI

6

1.5 Prior information

1

2 Prior information

1 0 0

RT model and measurement uncertainties and prior information

0.5 0.5 NDVI

1

0

0

0.5 NDVI

1

Fig. 7.7 Mode (plot on the left) of the distribution of the solution (LAI) of the inverse problem as a function of the measured value (NDVI). The mode corresponds to the maximum PDF value, i.e., the maximum likelihood. Four estimates are displayed: using only prior information; using RT model (LAI = RT −1 (NDV I)) assumed to be perfect with perfect measurements (no uncertainities accounted for); using RT model and measurements with their associated uncertainities; using RT model and measurements with their associated uncertainities and prior information. On the right, the standard deviation of the distribution of the solution is also displayed for the several cases. The case with perfect RT model and measurements is not displayed here because its standard deviation is null by definition

188

F. Baret, S. Buis

model. When prior information is used in addition to RT model and remote sensing measurements, differences of LAI mode are marginal over the domain where NDVI is sensitive enough to LAI. Conversely, over the saturation domain, LAI modal values are always lower (closer to prior information value) than those observed when not using prior information which would lead to a bias. However, the interest of using the prior information is clearly demonstrated when considering the standard deviation of the distribution of the solutions (Fig. 7.7, right). Introducing prior information in the inversion process provides a very significant reduction of the variability of the posterior distribution. This is obviously more important for the larger NDVI values corresponding to the saturation domain: in this case, very large scattering of the retrieved LAI values is expected when no prior information is used. Although the maximum likelihood is often used as “the solution”, the variability within the posterior distribution as represented by its standard deviation appears to be very informative and useful. The theory behind this Bayesian approach has been extensively described by Tarantola (2005). When restricting the solution as that maximizing the likelihood, i.e., corresponding to the maximum of the PDF, a general formulation of the cost function may be derived under Gaussian distribution assumption: ˆ + (Vˆ −Vp )t ·C−1 · (Vˆ −Vp ) ˆ t ·W −1 · (R − R) J = (R − R) Radiometricinformation

(7.7)

Prior information

where Vˆ is the vector of the input biophysical variables estimates, R corresponds to the vector of remote sensing measurements of dimension N (the number of bands and directions used), Rˆ is the vector of the simulated reflectance corresponding to the solution Vˆ (the vector of canopy biophysical variables) and Vp the vector of prior values of biophysical variables. Matrices W and C are the covariance matrices characterizing respectively the radiometric and model uncertainties, and that of the prior information. Note that the first part of this equation corresponds to the distance between the measured and the simulated radiometric data. It simplifies into Eq. (7.1) if the covariance terms of matrix W are assumed to be zero, i.e., measurement and model uncertainties are independent between configurations. The second part of Eq. (7.7) corresponds to the distance between the values of the estimated variables and those of the prior information. Very few studies are currently based on this formulation of the cost function where prior information is explicitly used (Combal et al., 2002). Implementing the cost function as expressed by Eq. (7.7) requires some reasonable estimates of covariance matrices W and C as well as prior values Vp . The terms of W should reflect both measurement and model uncertainties. While some rough estimates of measurement uncertainties could be derived from the sensor specification, model uncertainties are far more difficult to estimate. Further, they may depend significantly on the type of situation considered, such as low or high vegetation amount. Even more difficult to estimate, are the covariance terms in W : measurement and model uncertainties may have important structure that translates into high covariance terms which are however very poorly known. When using simul-

7 Estimating Canopy Characteristics from Remote Sensing Observations

189

taneously a large number of configurations as in the case of hyperspectral observations, these covariance terms will be very important to account for: they will allow weighing properly the several configurations used. The difficulty to estimate the covariance terms explains why a small number of configurations is often selected when a larger number is available as in the case of hyperspectral and/or directional observations. Retrieval approaches should be used within well defined and if possible restricted domains. Larger domains will generally degrade retrieval performances since the prior information will be looser defined, similarly to the covariance matrices characterizing uncertainties. However, splitting the whole domain into a set of subdomains may introduce problems due to misclassification and attribution errors as observed by Lotsch et al. (2003), and artefacts at the limit between classes translating into more chaotic spatial or temporal variation of the solution. The way prior information is introduced in the inversion process depends on the inversion technique used. The cost function represented by Eq. (7.7) is used within iterative optimization and LUTs. Bayesian methods include the a priori distribution through the use of the Bayes theorem to estimate the a posteriori distribution. For biophysical variables driven approaches the training data base should reflect the actual knowledge on the distribution of the variables. Note that the difficulty in defining explicitly the covariance terms in the uncertainties on remote sensing inputs (RT model and measurements) for the radiometric data driven approaches remains in the biophysical variables driven approaches for the generation of the training data base. However, implicit introduction of these terms may be achieved when using a training data base made from actual satellite measurements as suggested by Bacour et al. (2006).

7.4.2 Using Additional Constraints 7.4.2.1 Coupling Models The radiative transfer in each element of the soil/leaf/canopy/atmosphere system is strongly coupled to the radiative transfer in the whole system. The simple example given previously to demonstrate the under-determined nature of the inverse problem in remote sensing shows that top of canopy reflectance could be written as:

ρ toc (λ , θ s, θ v, ϕ ) = CAN(LAI, ALA, hot, LEAF(N, Cab , Cdm , Cbp , Cw ), SOIL ([w j ], [αi ], r, λ , θ s, θ v, ϕ ) , θ s, θ v, ϕ ) (7.8) The same applies when retrieving some characteristics of the system from top of atmosphere reflectance (ρ toa ) measurements as usually achieved by sensors aboard satellite: ¨ Patm , Cwv , C03 , λ , θ s, θ v, ϕ ) ρ toa (λ , θ s, θ v, ϕ ) = ATM(ρ toc (λ , θ s, θ v, ϕ ), τ550 , A, (7.9)

190

F. Baret, S. Buis

where ATM represents an atmospheric RT model such as 6S (Vermote et al., 1997) ¨ being respectively the aerosol optical or MODTRAN (Berk et al., 1998), τ550 , A, thickness at 550 nm and the Angstr¨om coefficient, Patm is the atmospheric pressure, Cwv is the water vapor content and C03 the ozone content. Retrieval of characteristics of some element of the system without solving (implicitly or explicitly) the whole system will therefore be sub-optimal as demonstrated below. Let consider retrieving leaf biophysical properties [N, Cab , Cdm , Cbp , Cw ] from top of canopy remote sensing observations in B wavebands using a decoupled system and an iterative optimization technique. For sake of simplicity, soil reflectance will be assumed to be known. Estimates of leaf properties could be achieved in two steps. First, estimate the variables [LAI, ALA, hot, re f l(λ ), tran(λ )] from the reflectance in each of the B bands. A cost function accounting for the reflectance in the B bands should be minimized with the constraint that [LAI, ALA, hot] does not vary with wavelength. The number of unknowns in the system will therefore be (3 + 2 · B) corresponding to the 3 canopy structure variables and the 2 (reflectance and transmittance) leaf optical properties time the B bands. The second step of the process consists in estimating leaf biophysical properties [N, Cab , Cdm , Cbp , Cw ] from the retrieved leaf reflectance and transmittance in the B bands. The variables [N, Cab , Cdm , Cbp , Cw ] are tuned by minimizing a cost function accounting for leaf reflectance and transmittance in the B bands. Obviously, increasing the number of bands will not improve the underdetermined nature of the problem because the number of unknowns in the first step of the process will grow twice faster. In addition, since no biophysical constraints are set on the spectral variation of leaf optical properties, canopy structure variables derived from the first step may express larger and unrealistic range of variation. The proper way to solve this type of problem is to minimize a cost function accounting for canopy reflectance over the B wavebands based on the coupled leaf and canopy models. In this case, the number of unknowns will be eight (the three canopy structure variables and the five leaf characteristics) which is independent from the number of wavebands used. This allows limiting the under-determined nature of the problem by increasing the spectral sampling. Most of the retrieval approaches from top of canopy radiometric observations are now using implicitly or explicitly coupled models as shown in Table 7.2. However, although offering great potentials as demonstrated recently (Baret, 2006b), the use of coupled atmosphere/surface models is still not very well developed because each sub-problem was handled by different communities.

7.4.2.2 Spatial Constraints Up to now, most retrieval algorithms are applied to independent pixels, neglecting the possible spatial structure as observed on most images. However, some authors attempted to exploit these very obvious patterns at high spatial resolution. The “object retrieval” approach proposed by Atzberger (2004) is based on the use of covariance between variables as observed over a limited cluster of pixels representing the same

7 Estimating Canopy Characteristics from Remote Sensing Observations

191

Table 7.2 Synthesis of the several algorithms currently used operationally to retrieve canopy biophysical variables. 1: (Lacaze, 2004); 2: (Knyazikhin et al., 1999); 3: (Gobron et al., 1999); 4: (Weiss et al., 2002; Baret et al., 2007); 5: (Chen et al., 2002; Deng et al., 2006); 6: (Bacour et al., 2006) #

Algorithm

1

RT models Canopy

Atmosphere

Inversion technique

uncertainties

prior information

Kuusk LAI, ALA, hot

TOC

NNT

measurements

Some variables fixed Range of variation

Hapke 3 typical + understorey

DISORD 6 biomes

TOC

LUT

measurements prescribed at 20%

specific values for 6 biomes

PROSPECT N, Cab, (Cw,Cdm, Cs)

5 typical soil unique BRDF

Gobron LAI, ALA, hot

TOA (MODTRAN)

Parametric

not specified

Range of variation (uniform distribution)

PROSPECT N, Cab, Cw,Cdm, Cbp

brightness parameter &reference spectra

SAIL LAI, ALA, hot, vCover

TOC

NNT

TOC

Parametric

2 versions: - TOC version - TOA version (SMAC)

NNT

leaf

soil

POLDER LAI, fAPAR

PROSPECT N, Cab, (Cw,Cdm, Cs)

PRICE 2 abundances

2

MODIS/MISR LAI, fAPAR

prescribed for each biome

3

MERIS MGVI fAPAR

4

VEGETATION CYCLOPES LAI, fAPAR, fCover

5

VEGETATIONCanada-Global LAI

Empirical relations for specific biomes using TM sensor and the corresponding ground measurements over some sites Prescribed BRDF model

6

MERIS LAI, fAPAR, fCover, LAIxCab

PROSPECT green/brown separated N, Cab, Cdm, Cw, Cbp

(1)

(2)

brightness parameter &reference spectra

SAIL, LAI, ALA, hot, vCover

model and measurements approximation of actual prescribed at 4% distribution (relative)

not specified

Specific relations for each biome

model and measurements approximation of actual prescribed at 4% distribution (relative)

class of object such as an agricultural field. Results show quite significant improvement of the retrieval performances for LAI, Cab and Cw , presumably because of a better handling of the possible compensation between LAI and ALA in the retrieval process as suggested by Atzberger (2004) and outlined by Jacquemoud (1993). Other approaches based on models with random effects (Faivre and Fischer, 1997) may be also very attractive, although rarely used within the land remote sensing community. They allow characterizing a population by their two first statistical moments (mean and variance). In the case of remote sensing applications, this could be applied over a cluster of P pixels belonging to the same class of surface as in the “object retrieval” approach of Atzberger (2004). The inversion process could be achieved by tuning both the mean and variance values of each input variable over the P pixels using iterative optimization techniques. The individual values of each pixel could be derived from the estimated mean and variance values of the variables and the departure between the actual radiometric measurements of the pixels and the mean values over the object. The under-determination of the problem could significantly decrease with this approach: the number of unknowns to estimate is independent on the number of pixels considered in the cluster and is just twice the number of variables to estimate (mean and variance). Although quite promising, these methods need further evaluation, and probably adaptation before being accepted and used by the remote sensing community. Note that only statistical distributions are used for both methods presented, although additional geo-statistical constraints could be exploited particularly for the higher spatial resolutions, based on variograms (Garrigues et al., 2006).

192

F. Baret, S. Buis

7.4.2.3 Temporal Constraints The dynamics of canopies results from elementary processes under the control of climate, soil and the genetic characteristics of the plants that incrementally change canopy structure and optical properties of the elements. Very brutal and chaotic time course are therefore not expected, at the exception of accidents such as fire, flooding, harvesting, or lodging. The smooth character of the dynamics of canopy variables may be exploited as additional constraint in the retrieval process. K¨otz et al. (2005) proposed using a semi-empirical model of canopy structure dynamics to improve remote sensing estimates of LAI over maize crops. Results show a significant improvement of estimates, particularly for the larger LAI values where saturation of reflectance is known to be a problem. This approach requires a semi-empirical model of canopy structure dynamics (here LAI) describing the whole growth cycle with few parameters. In the case of the model used by K¨otz et al. (2005) five parameters are needed. In this case, the under-determined nature of the inverse problem will decrease only if more than five dates of remote sensing observations are available and well distributed over the growth cycle. However, because the parameters of the model of LAI dynamics have some biological meaning, prior information on them could be accumulated and efficiently exploited. More recently, Lauvernet et al. (2007) proposed a “multitemporal patch” inversion scheme to account for both spatial and temporal constraints. Reflectance data are here considered observed from top of atmosphere. Atmosphere/canopy/ leaf/soil RT models are thus coupled to simulate top of atmosphere reflectance from the set of input variables as stated by Eqs. (7.8) and (7.9). Spatial and temporal constraints are based on the assumption that the atmosphere is considered stable over a limited area (typically few kilometres) but varies from date to date, and that surface characteristics vary only marginally over a limited temporal window (typically ±7 days) but may strongly change from pixel to pixel. This has obviously important consequences on the under-determined nature of the inverse problem as demonstrated hereafter. The atmosphere characteristics [Patm , Cwv , C03 ], except the ¨ are assumed to be known from independent observations aerosol ones [τ550 , A], such as meteorological estimation or dedicated sensors or algorithms. The observational configuration [λ, θ s, θ v, ϕ ] is also known at the time of image acquisition. Soil reflectance was simply approximated as lambertian, with reflectance proportional to a reference soil spectra according to a brightness parameter Bs (Bacour et al., 2006). The brightness parameter is assumed to vary both from date to date and pixels to pixels, without any constraints. The forward model resulting from nesting the RT models presented previously could be written as a function of the ¨ with nA = 2 ten unknowns [N, Cab , Cdm , Cbp , Cw , LAI, ALA, hot, Bs, τ550 , A] atmosphere variables, nC = 8 canopy and leaf variables and ns = 1 soil variables. Let consider d dates of observation available over a limited temporal window during which the canopy variables are about constant, and a spatial window of p pixels for which the atmosphere is considered homogeneous. The number of unknowns, N(p, d) in the case of concurrent inversion of an ensemble of d dates and p pixels using the spatial and temporal constraints described above is therefore:

7 Estimating Canopy Characteristics from Remote Sensing Observations

193

N(p, d) = d · nA + p · nC + d · p · nS

(7.10)

Inverting the nested radiative transfer models concurrently over an ensemble of d dates and p pixels will significantly reduce the total number of unknowns. (N(p, d)) as compared to p times d independent instantaneous pixel inversions (p.d.N(1, 1)). Figure 7.8 shows that the number of unknowns to be estimated within the same inversion process for p pixels and d dates as compared to p.d single pixel and single date inversions (N(p, d)/(p.d.N(1, 1))) decreases significantly up to about 10 pixels. However, the main advantage over “ensemble” inversion is reached when applying concurrently the inversion process to several dates. Using two dates and more than 10 pixels allows dividing by almost 2 the number of unknowns. Note that these results concern only the number of unknowns, and is therefore applicable to any observational configuration characterized by a set of bands and directions. Results on the performances achieved demonstrate the interest of the approach for the estimation of most of the variables, particularly for the aerosol characteristics and for LAI, LAI × Cab and ALA canopy characteristics. However, again, this new approach was only demonstrated over RT simulations, and its interest should be verified over experiments with actual remote sensing data and the corresponding ground truth.

1 0.9

d=1

N(p,d) / (p.d.N(1,1 ))

0.8 0.7 0.6 0.5

d=2

0.4 d=3 0.3

d=4 d=5 d=6

0.2 0.1

0

0

5

10

15

20

25

30

35

40

45

50

Number of pixels (p)

Fig. 7.8 Ratio, N(p, d)/(p.d.N(1, 1)), between the number of unknowns when applying the inversion process concurrently to p pixels and d dates with that of single pixel and date inversion as a function of the number of dates and pixels considered

194

F. Baret, S. Buis

7.5 Conclusion This overview of retrieval approaches is based on methods currently used, while alternative ways to solve the problem and hopefully improve the accuracy and robustness of estimates were briefly introduced. Several ingredients of the algorithms were identified apart from the retrieval techniques themselves: radiative transfer models, observations, additional information and constraints. We will briefly summarize the conclusions for each of these ingredients in the following.

7.5.1 Retrieval Techniques The several techniques investigated have been classified as radiometric variables or biophysical variables driven approaches. However, both types of methods could be either derived from actual measurements or based on radiative transfer model simulations. The best approaches are obviously the ones that will be trained over data sets that are as close as possible to the evaluation data set. For this very reason, canopy biophysical variables trained over empirical data sets would be ideal. In addition, canopy biophysical variables driven approaches present the advantage of being very computer efficient once trained, allowing easy implementation within operational processing chains. However, because of the difficulty of getting a large enough training data set representing the actual distribution of cases (observational configuration, type of canopies and state, background properties, eventually atmosphere characteristics), training data base made of radiative transfer model simulations is preferred. These hybrid techniques as termed by Liang (2004) require however the radiative transfer models to be well adapted to the type of canopy they target, and their adequacy to be quantified to properly input model uncertainties. In addition, the structure of uncertainties on the radiometric variables and distribution and codistribution of the input biophysical variables should be also known. An alternative approach currently not yet explored would consist in bridging the two retrieval approaches: actual sensor measurements are used to build the training data base allowing to keep all the structure of measurement uncertainties. This data base should be representative of the cases investigated, which might be possible by specific spatial and temporal sampling schemes as proposed by Baret et al. (2006a) in the case of global observations. The corresponding best estimates of canopy biophysical variables could be derived from inversion methods such as iterative optimization techniques for which all the information available should be exploited: fusion of all currently available sensors observations, prior information and spatial and temporal constraints. As a matter of fact, most radiometric variables driven approaches are very flexible and could easily ingest data from several sensors, bands and directions, at the expense of computer requirements which make these methods more difficult for an operational use. Conversely, canopy biophysical driven approaches are not as flexible as radiometric driven approaches: they are generally tuned for a limited set of

7 Estimating Canopy Characteristics from Remote Sensing Observations

195

observational conditions: using other configurations would require a specific training or a dramatic enlargement of the training data base. Retrieval methods will be more efficient when applied to a limited set of surface types as compared to a very generic (global) solution. Approaches based on a classification would thus allow closer adaptation to each class of both the radiative transfer model and prior information. However, attribution errors may significantly alter the performances. Using a continuous classification (Hansen et al., 2002; Hansen and DeFries, 2004; Schwartz and Zimmermann, 2005) will probably limit this source of uncertainties and avoid getting artefacts when two consecutive pixels will jump from one class to another. Biophysical variables estimates are generally integrated within other process models such as hydrology or biogeochemical cycling along with other ground observations. Quantification of the associated uncertainties is therefore required to properly merge these several sources of information. Current available products did not provide quantitative evaluation of the confidence interval around the solution, but are limited to qualitative indices. Bayesian approaches provide a direct access to the distribution of the solution of the inverse problem and may be very useful for estimating the uncertainties. Current operational algorithms need further developments to fully satisfy this important user requirement.

7.5.2 Radiative Transfer Models Performances of methods based on radiative transfer models are largely depending on the realism of the simulations. Radiative transfer models are based on a set of assumptions, particularly regarding the description of canopy architecture. A more realistic description of canopy architecture will require additional input variables and will be probably more demanding in computer resources. Knowledge of prior distribution and co-distribution of these additional canopy structure variables will constitute a limitation. Further, using such more realistic radiative transfer model requiring a larger number of unknowns will not necessarily improve the retrieval performances because the under-determination of the problem will be even more limiting. A compromise should therefore be found between the realism of the description of canopy structure, and its complexity. Particular attention should be paid on the definition of the variables used in the radiative transfer model that should match the one required for the application. For example, the original LAI definition (Stenberg, 2006) may be altered depending on the way and scale at which leaf clumping is accounted for (Chen and Leblanc, 1997). Great caution should be also paid when comparing retrievals with ground measurements or inter-comparing several products. As demonstrated here, holistic approaches based on the coupling of canopy, leaf and soil models are optimal for best performances. Eventually, coupled surface and atmosphere models would certainly help solving in an elegant way the retrieval of surface variables from top of the atmosphere observations.

196

F. Baret, S. Buis

7.5.3 Observations and Ancillary Information The observational configuration is an important element that drives the accuracy of canopy biophysical variables estimation. It depends obviously on the variables targeted. For the time being, sufficient maturity is achieved for the estimation of LAI, fAPAR, the cover fraction, chlorophyll and water contents variables to implement operational algorithms for delivering the corresponding products to the user community. The interest of multidirectional and hyperspectral observations is still to be rigorously demonstrated for these variables by comparison over actual ground measurements. Frequent observations are required to monitor the dynamics of the vegetation that conveys a large amount of information on the functioning of the surface. With the hopefully venue of systems capable of high revisit frequency with high spatial resolution, new retrieval methods should be developed to exploit the temporal and spatial dimensions in addition to the more classical spectral and in a lesser way directional ones. This would allow benefiting from the spatial and temporal constraints and consequently reduce the number of unknowns to be retrieved. Ultimately, this approach will converge towards direct assimilation of top of atmosphere radiances into surface process models. However, the research community is not mature enough on the coupling between radiative transfer models and canopy process models. Radiative transfer model inversion had still to mature and improve the accuracy of surface variables estimation before jumping towards radiance data assimilation. Knowledge and management of uncertainties is one of the critical issues for the retrieval algorithms. If measurement uncertainties coming from the sensor are relatively well known, their structure (covariance between bands and directions for example) is poorly documented. This is even worse when considering model uncertainties that may change dramatically from place (and time) to place (and time) with presumably specific features (covariance between configurations). The other critical issue is the lack of prior information on the distribution of most land surface attributes. However, this could be accumulated from the numerous experiments organized in support of satellite images. A mechanism should thus be developed to capitalize on the information gathered within the remote sensing research community as well as other communities working with ecosystems. Note that getting high spatial resolution data will considerably ease the characterization of prior distribution of the variables, provided that each pixel could be properly classified. Any retrieval algorithm should be properly validated before delivering its products to the user community according to consensus protocols (Morisette et al., 2006). This process will not only provide a way to characterize the associated uncertainties, it will be also critical for improving the algorithms. A short feedback loop should therefore be set-up between algorithm prototyping and validation. When retrieval algorithms are based on radiative transfer modeling, this will implicitly merge observations and model to improve robustness and accuracy of the products at the expense of a decrease in the desired independency between the validation and calibration processes.

7 Estimating Canopy Characteristics from Remote Sensing Observations

197

References Abuelgasim AA, Gopal S, Strahler AH (1998) Forward and inverse modelling of canopy directional reflectance using a neural network. Int. J. Remote Sens. 19(3):453–471 Asrar G, Fuchs M, Kanemasu ET, Hatfield JL (1984) Estimating absorbed photosynthetic radiation and leaf area index from spectral reflectance in wheat. AJ, 76:300, 306 Asrar G, Myeni RB, Kanemasu ET (1989) Estimation of plant canopy attributes from spectral reflectance measurements. Theory and Application of Optical Remote Sensing. J. Wiley, Indianapolis: 252–296 Atkinson PM, Tatnall ARL (1997). Neural network in remote sensing. Int. J. Remote Sens. 18(4):699–709 Atzberger C (2004) Object-based retrieval of biophysical canopy variables using artificial neural nets and radiative transfer models. Remote Sens. Environ. 93:53–67 Bacour C (2001) Contribtion a` la d´etermination des param`etres biophysiques des couverts v´eg´etaux par inversion de mod`eles de r´eflectance: analyse de sensibilit´e et configurations optimales. PhD Thesis, Universit´e Paris 7 – Denis Diderot, Paris (France), 206 pp Bacour C, et al. (2002a) Reliability of the estimation of vegetation characteristics by inversion of three canopy reflectance models on airborne POLDER data. Agronomie, 22:555–566 Bacour C, Jacquemoud S, Tourbier Y, Dechambre M, Frangi JP (2002b) Design and analysis of numerical experiments to compare four canopy reflectance models. Remote Sens. Environ. 79:72–83 Bacour C, Baret F, B´eal D, Weiss M, Pavageau K (2006) Neural network estimation of LAI, fAPAR, fCover and LAIxCab, from top of canopy MERIS reflectance data: principles and validation. Remote Sens. Environ. 105:313–325 Banari A, Huete AR, Morin D, Zagolski F (1996) Effets de la couleur et de la brillance du sol sur les indices de v´eg´etation. Int. J. Remote Sens. 17(10):1885–1906 Baret F, Guyot G (1991) Potentials and limits of vegetation indices for LAI and APAR assessment. Remote Sens. Environ. 35:161–173 Baret F, Clevers JGPW, Steven MD (1995) The robustness of canopy gap fraction estimates from red and near infrared reflectances: a comparison of approaches. Remote Sens. Environ. 54:141–151 Baret F, Knyazikhin J, Weiss M, Myneni R, Pragn`ere A (1999) Overview of Canopy Biophysical Variables Retrieval Techniques. ALPS’99, Meribel, France, CNES, Paris Baret F, Weiss M, Troufleau D, Pr´evot L, Combal B (2000) Maximum information exploitation for canopy characterisation by remote sensing. Remote sensing in agriculture. Aspects of Applied Biology. Association of Appied Biologists, pp. 71–82 Baret F, et al. (2006a) Evaluation of the representativeness of networks of sites for the global validation and inter-comparison of land biophysical products. Proposition of the CEOS-BELMANIP. IEEE Trans.Geosci. Remote Sens. 44(7:special issue on global land product validation):1794– 1803 Baret F, et al. (2006b) Algorithm Theoretical Basis Document for MERIS Top of Atmosphere Land Products (TOA VEG), INRA-CSE, Avignon Baret F, et al. (2007) LAI, fAPAR and fCover CYCLOPES global products derived from VEGETATION. Part 1: principles of the algorithm. Remote Sens. Environ. 110:275–286 Berk A, et al. (1998) MODTRAN cloud and multiple scattering upgrades with application to AVIRIS. Remote Sens. Environ. 65:367–375 Bicheron P, Leroy M (1999) A method of biophysical parameter retrieval at global scale by inversion of a vegetation reflectance model. Remote Sens. Environ. 67:251–266 Bonnans, JF, Gilbert, JC, Lemar´echal, C, and Sagastiz´abal, CA (2006) Numerical Optimization. Theoretical and Practical Aspects (Second Edition). Springer, New-York. 494pp. Braswell BH, et al. (1996) Extracting ecological and biophysical information from AVHRR optical data: an integrated algorithm based on inverse modeling. J. Geophys. Res. 101(D18):23335– 23348

198

F. Baret, S. Buis

Br´eon FM, Maignan F, Leroy M, Grant I (2002) Analysis of hot spot directional signatures measured from space. J. Geophys. Res. 107(D16):AAC 1 1–15 Chen JM, Leblanc SG (1997) A four-scale bidirectional reflectance model based on canopy architecture. IEEE Trans. Geosci. Remote Sens. 35:1316–1337 Chen JM, et al. (2002) Derivation and validation of Canada wide coarse resolution leaf area index maps using high resolution satellite imagery and ground masurements. Remote Sens. Environ. 80:165–184 Combal B, Ochshepkov SL, Sinyuk A, Isaka H (2000) Statistical framework of the inverse problem in the retrieval of vegetation parameters. Agronomie 20:65–77 Combal B, Baret F, Weiss M (2001) Improving canopy variables estimation from remote sensing data by exploiting ancillary information. Case study on sugar beet canopies. Agronomie 22:2–15 Combal B, et al. (2002) Retrieval of canopy biophysical variables from bi-directional reflectance data. Using prior information to solve the ill-posed inverse problem. Remote Sens. Environ. 84:1–15 Danson FM, Rowland CS, Baret F (2003) Training a neural network with a canopy reflectance model to estimate crop leaf area index. Int. J. Remote Sens. 24(23):4891–4905 Deng F, Chen JM, Chen M, Pisek J (2006) Algorithm for global leaf area index retrieval using satellite imagery. IEEE Trans. Geosci. Remote Sens. 44(8):2219–2229 Faivre R, Fischer A (1997) Predicting crop reflectances using satellite data observing mixed pixels. J. Agric., Biol. Environ. Stat. 2:87–107 Fang H, Liang S (2005) A hybrid inversion method for mapping leaf area index from MODIS data: experiments and application to broadleaf and needleleaf canopies. Remote Sens. Environ. 94:405–424 Fang H, Liang S, Kuusk A (2003) Retrieving leaf area index using a genetic algorithm with a canopy radiative transfer model. Remote Sens. Environ. 85:257–270 Fernandes R, Leblanc SG (2005) Parametric (modified least squares) and non parametric (TheilSen) linear regressions for predicting biophysical parameters in the presence of measurements errors. Remote Sens. Environ. 95:303–316 Gao W, Lesht BM (1997) Model inversion of satellite-measured reflectance obtaining surface biophysical and bidirectional characteristics of grassland. Remote Sens. Environ. 59:461–471 Garabedian P (1964) Partial differential equations. Wiley, New York, 672pp. Garrigues S, Allard D, Baret F, Weiss M (2006) Quantifying spatial heterogeneity at the landscape scale using variogram models. Remote Sens. Environ. 103:81–96 Gobron N, Pinty B, Verstraete MM, Govaerts Y (1997) A semi-discrete model for the scattering of light by vegetation. J. Geophys. Res. Gobron N, Pinty B, Verstraete M, Govaerts Y (1999) Level 2 vegetation index algorithm theoretical Basis. Joint Research Center, Ispra, Italy Gobron N, Pinty B, Verstraete M, Widlowski JL (2000) Advanced vegetation indices optimized for upcoming sensors: design, performances and applications. IEEE Trans. Geosci. Remote Sens. 38(6):2489–2505 Goel NS (1989) Inversion of canopy reflectance models for estimation of biophysical parameters from reflectance data. In: G. Asrar (Editor), Theory and Applications of Optical Remote Sensing. Wiley, New York, pp. 205–251 Goel NS, Deering DW (1985) Evaluation of a canopy reflectance model for LAI estimation through its inversion. IEEE Trans. Geosci. Remote Sens. GE-23(5):674–684 Goel NS, Thompson RL (1984) Inversion of vegetation canopy reflectance models forestimating agronomic variables. III. Estimation using only canopy reflectance data as illustrated by the SUIT model. Remote Sens. Environ. 15:223–236 Goel NS, Strebel DE, Thompson RL (1984) Inversion of vegetation canopy reflectance models for estimating agronomic variables. II. Use of angle transforms and error analysis as illustrated by the SUIT model. Remote Sens. Environ. 14:77–111 Gong P, Wang DX, Liang S (1999) Inverting a canopy reflectance model using a neural network. Int. J. Remote Sens. 20(1):111–122

7 Estimating Canopy Characteristics from Remote Sensing Observations

199

Green RO, Boardman J (2001) Exploration of the relationship between information content and signal to noise ratio and spatial resolution in AVIRIS spectral data. In: RO Green (ed), AVIRIS Workshop, Pasadena, CA Hansen MC, et al. (2002) Towards an operational MODIS continuous firld percent tree coveralgorithm: examples using AVHRR and MODIS data. Remote Sens. Environ. 83:309–319 Hansen MC, DeFries RS (2004) Detecting long-term global forest change using continuous fields of tree-cover maps from 8-km Advanced Very High Resolution Radiometer (AVHRR) data for the years 1982–99. Ecosystems 7:695–716 Hapke B (1981) Bidirectional reflectance spectroscopy. 1. Theory. J. Geophys. Res. 86:3039–3054 Huang D, et al. (2006) The importance of measurement error for deriving accurate reference leaf area index maps and validation of the MODIS LAI products. IEEE Trans. Geosci. Remote Sens. 44(7):1866–1871 Huete AR (1988) A soil adjusted vegetation index (SAVI). Remote Sens. Environ. 25:295–309 Huete AR, Liu HQ, Batchily K, van Leeuwen W (1997) A comparison of vegetation indices over a global set of TM images for EOS-MODIS. Remote Sens. Environ. 59:440–451 Jacquemoud S (1993) Inversion of the PROSPECT+SAIL canopy reflectance model from AVIRIS equivalent spectra: theoretical study. Remote Sens. Environ. 44:1–25 Jacquemoud S, Baret F (1990) PROSPECT: a model of leaf optical properties spectra. Remote Sens. Environ. 34:75–91 Jacquemoud S, Baret F, Hanocq JF (1992) Modeling spectral and directional soil reflectance. Remote Sens. Environ. 41:123–132 Jacquemoud S, Baret F, Andrieu B, Danson M, Jaggard K (1995) Extraction of vegetation biophysical parameters by inversion of the PROSPECT+SAIL model on sugar beet canopy reflectance data. Application to TM and AVIRIS sensors. Remote Sens. Environ. 52:163–172 Jacquemoud S, Bacour C, Poilv´e H, Frangi JP (2000) Comparison of four radiative transfer models to simulate plant canopies reflectance: direct and inverse mode. Remote Sens. Environ. 74:471–481 Kimes DS, Nelson RF, Manry MT, Fung AK (1998) Attributes of neural networks for extracting continuous vegetation variables from optical and radar measurements. Int. J. Remote Sens. 19(14):2639–2663 Kimes DS, Knyazikhin Y, Privette JL, Abuelgasim AA, Gao F (2000) Inversion methods for physically-based models. Remote Sens. Rev. 18:381–439 Knyazikhin Y, et al. (1998a) Estimation of vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation from atmosphere-corrected MISR data. J. Geophys. Res. 103(D24):32239–32256 Knyazikhin Y, Martonchik JV, Myneni RB, Diner DJ, Running SW (1998b) Synergistic algorithm for estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation from MODIS and MISR data. J. Geophys. Res.103(D24):32257–32275 Knyazikhin Y, et al. (1999) MODIS leaf area index (LAi) and fraction of photosynthetically active radiation absorbed by vegetation (FPAR) product (MOD15) algorithm theoretical basis document. http://eospso.gsfc.nasa.gov/atbd/modistables.html K¨otz B, Baret F, Poilv´e H, Hill J (2005) Use of coupled canopy structure dynamic and radiative transfer models to estimate biophysical canopy characteristics. Remote Sens. Environ. 95:115–124 Kuusk A (1991a) Determination of vegetation canopy parameters from optical measurements. Remote Sens. Environ. 37:207–218 Kuusk A (1991b) The inversion of the Nilson-Kuusk canopy reflectance model, a test case. Int. Geosci. Remote Sens. Symp. (IGARSS’91), Helsinki, Finland, pp 1547–1550 Kuusk A (1995) A fast invertible canopy reflectance model. Remote Sens. Environ. 51:342–350 Lacaze R (2004) Land Surface Level3 Algorithm: The inversion of a vegetation model by a neural network. short report. http://smsc.cnes.fr/POLDER/SCIEPROD/lsp2algol3.htm Lacaze R (2005) POLDER-2 Land surface level 3 products. User manual and product validation, Medias/CNES, Toulouse

200

F. Baret, S. Buis

Lauvernet C, Baret F, Hasco¨et L, Buis S, Le Dimet FX (2008) Multitemporal-patch ensemble inversion of coupled surface-atmosphere radiative transfer models for land surface characterization. Remote Sens. Environ. (in press) Leprieur C, Verstraete MM, Pinty B (1994) Evaluation of the performance of various vegetation indices to retrieve vegetation cover from AVHRR data. Remote Sens. Rev. 10:265–284 Leshno M, Ya Lin V, Pinkus A, Shocken S (1993) Multilayer feedforward networks with non polynomial activation function can approximate any function. Neural Networks 6:861–867 Lewis P, et al. (2004) Calibration of and L-system model of winter wheat for remote sensing modelling and inversion. In: C. Godin (ed), 4th Int. Workshop on Functional-Structural Plant Models, Montpellier, France, pp 257–261 Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, Hoboken, NJ Liu W, et al. (2002) Relating soil surface moisture to reflectance. Remote Sens. Environ. 81:238–246 Lotsch A, Tian Y, Friedl MA, Myneni RB (2003) Land cover mapping in support of LAI and fPAR retrievals from EOS-MODIS and MISR: classification methods and sensitivities to errors. Int.J. Remote Sens. 24(10):1997–2016 Lucht W (1998) Expected retrieval accuracies of bidirectional reflectance and albedo from EOSMODIS and MISR angular sampling. J. Geophys. Res. 103(D8):8763–8778 Makowski, D, Hillier, J, Wallach, D, Andrieu, B Jeuffroy, MH (2006) Parameter estimation for crop models. In D. Wallach, D. Makowski and J.W Jones (eds.), Working with dynamic crop models: evaluation, analysis, parameterization, and applications, pp. 101–149. Elsevier, Amsterdam Morisette J, et al. (2006) Validation of global moderate resolution LAI Products: a framework proposed within the CEOS Land Product Validation subgroup. IEEE Trans. Geosci. Remote Sens. 44(7):1804–1817 Nilson T, Kuusk A (1989) A reflectance model for homogeneous plant canopy and its inversion. Remote Sens. Environ. 27:157–167 Pinty B, Verstraete MM (1991) Extracting information on surface properties from bidirectional reflectance measurements. J. Geophys. Res. 96:2865–2874 Pinty B, Verstraete MM, Dickinson RE (1990) A physical model of the bidirectional reflectance of vegetation canopies. 2. Inversion and validation. J. Geophys. Res. 95(D8):11767–11775 Price J (1994) How unique are spectral signatures? Remote Sens. Environ. 49:181–186 Price JC (1990) On the information content of soil reflectance spectra. Remote Sens. Environ. 33:113–121 Privette JL, Emery WJ, Schimel DS (1996) Inversion of a vegetation reflectance model with NOAA AVHRR data. Remote Sens. Environ. 58:187–200 Qiu J, Gao W, Lesht BM (1998) Inverting optical reflectance to estimate surface properties of vegetation canopies. Int. J. Remote Sens. 19(4):641–656 Renders J-M, Flasse SP (1996) Hybrid methods using genetic algorithms for global optimization. IEEE Trans. Systems, Man, and Cybernetics 26:243–258 Richardson AJ, Wiegand CL, Wanjura DF, Dusek D, Steiner JL (1992) Multisite analyses of spectral-biophysical data for sorghum. RSE, 41:71:82 Rondeaux G, Steven MD, Baret F (1996) Optimization of soil adjusted vegetation indices. Remote Sens. Environ. 55:95–107 Roujean JL, Lacaze R (2002) Global mapping of vegetation parameters from POLDER multiangular measurements for studies of surface-atmosphere interactions: a pragmatic method and validtion. J. Geophys. Res. 107(D12):ACL 6 1–14 Saltelli, A, Tarantola, S, Campolongo, F, Ratto, M (2004) Sensitivity Analysis in practice. A guide to assessing scientific models. Wiley & Sons, Chichester. 217pp. Schwartz MD, Zimmermann G (2005) A new GLM-based method for mapping tree cover continuous fields using regional MODIS reflectance data. Remote Sens. Environ. 95:428–443 Sellers PJ (1985) Canopy reflectance photosynthesis and transpiration. Int. J. Remote Sens. 3:1335–1372 Shoshany M (1991) The equifinality of bidirectional distribution functions of various microstructures. Int. J. Remote Sens. 12(11):2267–2281

7 Estimating Canopy Characteristics from Remote Sensing Observations

201

Smith JA (1992) LAI inversion from optical reflectance using neural network trained with multiple scattering model. In: WR (ed), IGARS’92. IEEE, South Shore Harbour Resort and Conference Center, Houston, TX, pp 757–759 Smith JA (1993) LAI inversion using backpropagation neural network trained with multiple scattering model. IEEE Trans. Geosci. Remote Sens. 31(5):1102–1106 Soler CFS, Blaise F, de Reffye P (2001) A physiological plant growth simulation engine based on accurate radiant energy transfer. 4116, INRIA, Montbonnot-Saint-Martin, France Stenberg P (2006) A note on the G-function for needle leaf canopies. Agric. For. Meteorol. 136 (1–2):76 Tarantola A (2005) Inverse problem theory problem and methods for model parameter estimation. Soc. Ind. Appl. Math., 358 pp Tarentola A (1987) Inverse problem theory. Methods for data fitting and model parameter estimation. Elsevier, New York Teillet PM, Gauthier RP, Staenz K, Fournier RA (1997) BRDF equifinality studies in the context of forest canopies. In: G Guyot, T Phulpin (eds), Physical Measurements and Signatures in Remote Sensing, Balkema, Rotterdam, Courchevel, France, pp 163–170 Verhoef W (1984) Light scattering by leaf layers with application to canopy reflectance modeling: the SAIL model. Remote Sens. Environ. 16:125–141 Verhoef, W (2002) Improved modeling of multiple scattering in leaf canopies: The model SAIL++. In J. A. Sobrino (ed.), 1st Int. Symp. on Recent Advances in Quantitative Remote Sensing, pp. 11–20, Universitat de Valencia, Spain Vermote EF, Tanr´e D, Deuz´e JL, Herman M, Mocrette JJ (1997) Second simulation of the satellite signal in the solar spectrum, 6S: an overview. IEEE Trans. Geosci. Remote Sens. 35(3):675– 686 Verstraete M, Pinty B (1996) Designing optimal spectral indexes for remote sensing applications. IEEE Trans. Geosci. Remote Sens. 34(5):1254–1265 Weiss M, Baret F, Myneni R, Pragn`ere A, Knyazikhin Y (2000) Investigation of a model inversion technique for the estimation of crop charcteristics from spectral and directional reflectance data. Agronomie 20:3–22 Weiss M, et al. (2002a) Validation of neural net techniques to estimate canopy biophysical variables from remote sensing data. Agronomie 22:547–554 Weiss M, et al. (2002b) Evaluation of kernel-driven BRDF models for the normalization of Alpilles/ReSeDA POLDER data. Agronomie 22:531–536 Wiegand CL, Gerberman AH, Gallo KP, Blad BL, Dusek D (1990) Multisite analyses of spectralbiophysical data for corn. Remote Sens. Environ. 33:1–16 Wiegand CL, et al. (1992) Multisite analysis of spectral biophysical data for wheat. Remote Sens. Environ. 42(1):1–22 Zhang Y, Shabanov N, Knyazikhin Y, Myneni RB (2002a) Assessing the information content of multiangle satellite data for mapping biomes I. Theory. Remote Sens. Environ. 80:435–446 Zhang Y, Tian Y, Myneni RB, Knyazikhin Y, Woodcock CE (2002b) Assessing the information content of multiangle satellite data for mapping biomes I. Statistical analysis. Remote Sens. Environ. 80:418–434 Zhang Q, et al. (2005) Estimating light absorption by chlorophyll, leaf and canopy in a deciduous broadleaf forest using MODIS data and a radiative transfer model. Remote Sens. Environ. 99(3):357

Chapter 8

Knowledge Database and Inversion Jindi Wang and Xiaowen Li

Abstract Physical remote sensing models usually need dozens of parameters to have reasonable degree of precision. With limited information provided by remotely sensed observations, it is an ill-posed problem to estimate parameters through model inversion. Many researches have developed inversion models and algorithms. We developed a priori knowledge based inversion strategy and algorithm. The spectrum knowledge database of typical land surface objects has been established to provide the prior knowledge of model parameters. Some approaches are presented in this chapter, which include the uncertainty and sensitivity matrix for analysis of observation data and parameters, the model inversion method supported by the knowledge database, the scaling correction on estimated parameters. Some study directions in model inversion, such as how to accumulate and use spatial and temporal change knowledge, how to validate the parameter inversion results, are also discussed.

8.1 Questions on and Possible Answers to Physical Remote Sensing Model Inversion One of the primary problems of remote sensing science is the retrieval of information on land surface parameters from remotely sensed data. Since satellite remote sensing deals with a complex system coupling atmosphere and land surface, physical remote sensing models usually need several to tens parameters to describe the relation between land surface parameters and remote sensing signals for having reasonable degree of model precision. But, with limited information provided by remotely sensed observations, it remains a challenge to estimate parameters through remote sensing model inversion. In mathematics, it is an ill-posed problem Jindi Wang and Xiaowen Li Research Center for Remote Sensing & GIS Beijing Normal University, Beijing, China [email protected] S. Liang (ed.), Advances in Land Remote Sensing, 203–217. c Springer Science + Business Media B.V., 2008

203

204

J. Wang, X. Li

to estimate more model parameters from less observations. In past years, many researchers have developed new inversion models and algorithms (Verstraete et al., 1996; Myneni et al., 1995). Some inversion research has developed better cost function for physical model inversion. In addition, different kinds of mathematical optimization algorithms have been developed (Kimes et al., 2000; Kunsk, 1991). Verstraete et al. (1996) pointed out the main limitation and the necessary observations for obtaining successful model inversion. Kimes et al. (1991) presented the knowledge-based expert system for inferring vegetation characteristics. Fang et al. (2005) used a genetic algorithm to estimate leaf area index (LAI) of vegetation canopy with radiative transfer model inversion. Neural network methods were also used for model inversion (Smith, 1993). Liang (2004) summarized the main research achievements in estimating land surface biophysical variables and surface radiation budgets. As advances in the field of multi-angular remote sensing progress, bidirectional reflectance distribution function (BRDF) models can be inverted to estimate the structural parameters and spectral component signatures of land surface cover types, such as the MODIS albedo and LAI products. Some operational algorithms have been implemented to generate products of land surface parameters, such as albedo, LAI, fraction of photosynthetically active radiation (FPAR), net primary production (NPP) from MODIS and MISR observations at the spatial resolution of 250 m or 1 km. The surface BRDF and albedo product from POLDER has been developed at a pixel resolution of approximately 6 km. Some validation work on these data products has been carried out and it has been found that there are still some uncertainties on model, observation and reference data which influence the accuracy of these products (Morisette et al., 2006; Tan et al., 2005). However, when we want to improve the estimation accuracy of retrieved surface parameters to meet requirements of applications at different spatial scales, the inversion of physical remote sensing models is a very difficult problem that still requires further studies from the viewpoint of both information theory (Li et al., 1998) and the comprehensive practice of model inversion (Privette et al., 1997; Wanner et al., 1997; Li et al., 2000b; Liang, 2004). The real physical system that couples the atmosphere and land surface is extremely complex and it requires many parameters to describe it faithfully. Any physical model can only be an approximation of this real system, and a good model will have many important parameters to capture the major variations of the real system. However, remotely sensed observations are usually more or less correlated. The remotely sensed signal, no matter how fine its spectral and angular resolution, contains only limited information. Therefore, BRDF model inversion problems, such as those in geoscience generally, are usually underdetermined, making the use of a priori knowledge necessary. As the ancient philosopher Confucius pointed out, “Our knowledge consists of two parts – what we know, and what we know we don’t know”. In the case that remote sensing signals contain limited but valuable information, it is important to extract information about what we don’t know or what is uncertain, rather than to invert all model parameters at the same time, pretending that we know nothing. Using this principle in earlier work, we expressed a priori knowledge of model

8 Knowledge Database and Inversion

205

parameters as the best guesses for the associated uncertainties, and developed the a priori knowledge based inversion strategy and algorithm (Li et al., 1997). The results were encouraging, and thus we tried to formalize the approach, and to establish the a priori knowledge database of typical land surface parameters (Li et al., 2002).

8.1.1 A Priori Knowledge-based Inversion Strategy We developed a priori knowledge based inversion strategy and algorithm (Li et al., 1998, 2001) called, Multi-stage, Sample-direction dependent, Target-decision (MSDT) (Li et al., 1997). There are three questions should be answered in order to run the strategy. The first question is how to express the a priori knowledge and to make it available in model inversion algorithm. The second question is how to divide the model parameters set into subsets, in order to invert the parameters by using the most sensitive data. The third question is how to accumulate the knowledge during the inversion procedure, when we have one scene observing data or a data set containing more scenes of continuous observations. In our previous study (Li et al., 1998), the a priori knowledge of model parameters is expressed as a joint probability density, while a priori knowledge of the model accuracy and measurement noise is expressed as a conditional joint probability density. In inversion model, the a priori probability density function (PDF) of the observations can be defined, and can be used in the cost function based Bayesian inference theory. An important feature of Bayesian inversion is that there is no prerequisite number of independent observations for a successful inversion. So long as new observations are acquired, a priori probability density in parameter space can be modified to obtain posterior density, allowing knowledge to be accumulated. Taking the land surface spectral albedo inversion as an example, Li et al. (2001) showed how the a priori knowledge significantly improves the retrieval of surface bidirectional reflectance and spectral albedo from satellite observations. In the paper, the a priori knowledge are extracted from field measurements of bidirectional reflectance factors for various surface cover types in red and near-infrared bands. Bidirectional reflectance and albedo are retrieved by the kernel-driven BRDF model inversion that uses surface reflectance observations derived from orbiting satellites. A priori knowledge is applied when noise and poor angular sampling reduce the accuracy of model inversion. In such cases, a priori knowledge can indicate when retrieved kernel weights or albedos are outside expected bounds, leading to a closer examination of the data. If data are noisy, a priori knowledge can be used to smooth the data. If the data exhibit poor angular sampling, a priori knowledge can be used according to Bayesian inference theory to yield a posteriori estimates of the unknown kernel weights, where Bayesian theory is applied in the data space rather than in the parameter space. Extensive studies and simulations using 73 sets of field observations and 395 space-borne observation sets from the POLDER instrument demonstrates the importance of a priori information in improving inversions and BRDF retrievals.

206

J. Wang, X. Li

The further studies introduce Tarantola’s inversion cost function in to BRDF model inversion algorithm. Tarantola’s inversion theory has also been widely applied in geophysical inversion (Tarantola, 1987) and atmospheric remote sensing (Rodgers, 1976), because the inverse problems in those fields are even more illposed than in land surface remote sensing. Based on Tarantola’s theory, the cost function used in land surface parameters inversion is defined as the parameters’ posterior probability density, and can be expressed as

1 pM (X) = const · exp − ( f (X) −Yobs )T CD−1 ( f (X) −Yobs ) 2 −1 (X − Xprior ) (8.1) + (X − X prior )T CM where Yobs is for observation data and Xprior is for a priori knowledge of parameters. The principle of the cost function is very similar to Bayesian inversion. One of its advantages is its ability to clearly express the errors of the model, the observed data, and the parameters’ initial value, with co-variances matrixes CD and CM respectively. This creates a new challenge on forming the error co-variances of model, data and parameters, this requires a great deal of a priori knowledge. Previous studies have demonstrated the effectiveness of the inversion model, using simulated data and field measurements (Yan et al., 2001; Wang et al., 2000). In these studies, the parameters’ initial values can be estimated by our field measurements, while the co-variance was set as the parameters’ standard deviation under the assumption that their estimated values have a normal distribution.

8.1.2 Uncertainty and Sensitivity Matrix of Parameters for Model Inversion It is important to examine the uncertainty and sensitivities of the parameters in model inversion, since we expect to estimate the values of parameters that are sensitive to observations and model, while giving certain values for other parameters that have relatively fewer uncertainties. To make the analysis of parameter uncertainty and sensitivity more effective, we defined the uncertainty and sensitivity matrix (USM), which is an objective expression of the prior knowledge. In order to construct the USM, we assume that a BRDF model has N spectral bands and K structural parameters and L spectral parameters of component materials, making K + N × L in total. Since the multiangular observations have M samples, the USM will have M × N rows and K + N × L columns. This matrix is too large to effectively use for calculations. Because the structural parameters are independent on the spectral band, we decomposed the matrix into a structural matrix that has M × N rows, K columns, where the N matrix of M × L spectral parameters corresponds to N bands, making a total N + 1 matrices. An element of the USM can be expressed as (Li et al., 1997),

8 Knowledge Database and Inversion

USM[i][ j] =

207

BRDF( j|i) BRDF(i)

(8.2)

where BRDF( j|i) is the maximum difference of BRDF as a function of the jth parameter within its uncertainty, given the ith geometry of illumination and viewing; BRDF(i) is BRDF as predicted by the model at the ith geometry, when all parameters at their best guess values. These best guess values of parameters are also from the accumulation of prior knowledge, and they can be updated when further parameter estimates are obtained during the MSDT inversion procedure. This USM definition has three advantages: (1) the uncertainty of the initial guess for being inverted parameter is taken into account; (2) the USM is less dependent on the initial guess; (3) all elements of the USM have the same units, and are therefore quantitatively comparable. This USM has thus been used widely in our recent model inversion studies. It can also be applied for parameters sensitivity analysis for models and data in relative studies (Li et al., 1997; Gao and Zhu, 1997; Yan et al., 2001).

8.1.3 Getting the Prior Knowledge on Typical Land Surface Parameters As stated above, after developing the model inversion strategy, we recognized that the most important issue in remote sensing model inversion is accumulation of prior knowledge regarding the model parameters. Given that we are not only concerned with the validation of the model, but also with how to apply the inversion algorithm for land surface parameter estimation at the scale of remote sensing images, we have to consider establishing a very powerful prior knowledge database that assembles the initial estimates of the parameters and also their co-variance values during the inversion procedure. These initial parameter values and their co-variance matrix are also necessary for supporting data assimilation algorithms. In practice, we firstly classify the a priori knowledge into different levels: the general knowledge about the land surface, or “global knowledge”, knowledge related to land cover type, and target-specific knowledge. The means to accumulate knowledge at these levels may be different, but should include the following: 1. 2. 3. 4. 5.

Applicable forward model(s) Physical limits and probability density in model parameter space Statistics of model accuracy and noise in remote sensing signals Seasonal change associated with land cover types or targets Confidence of the above knowledge

Note that even a single observation can change the a priori PDF of more than one parameter significantly. Li et al. (1998) provided an example on how the accumulation of knowledge is achieved in parameter space. However, the required numerical integration in parameter space is time-consuming whenever parameters set is large, as is the case with the retrieval of surface BRDF and albedo from satellite data.

208

J. Wang, X. Li

Our recent approaches are to establish an a priori knowledge database of the parameters of typical land surface targets, including the spatial and temporal distribution knowledge of parameters. This knowledge database also includes the forward models, remote sensing data and a priori knowledge at both global and land cover type related levels. Based on the knowledge base, it is expected that the accuracy of land surface parameters estimation at the scale of remote sensing observations will improve (Wang et al., 2003).

8.2 The Spectrum Knowledge Database of Typical Land Surface Objects Establishing the spectrum knowledge database of typical land surface objects is significant for the development of quantitative remote sensing. The research on spectral features of objects is the foundation of remote sensing applications. Land cover and land use classification and image interpretation are usually based on the spectrum recognizing method. Many kinds of spectrum matching techniques have been developed. However, users cannot always obtain the desired accuracy for classification and identification. One of the problems is that the spectral data of objects measured at different scales are not comparable. It is rare to have the correct relation between the spectrum of objects measured indoors or in the field, and the spectral image data acquired through remote sensing observations. The available spectrum data measured indoors and in the field do not have enough corresponding descriptors to clearly determine the related spectral environmental variables. This can result in some confusion in spectrum applications using spectra matching and other image processing algorithms. The second problem is that we need to integrate the physical models and measurement data into a close linked system to allow models and data to be effectively applied for land surface parameters estimation. When the linked system is developed, the prior knowledge of the model parameters can be extracted from the database and the model prediction and parameters inversion algorithm can also be included in the system. That is our approach to convert the individual measurement data sets to be the knowledge database with the stated goal of having effective physical model explanations and predictions. The establishment of the spectrum knowledge database of typical land surface objects is supported by China’s National High Technology Research and Development Program. The spectrum knowledge database consists of the measuring spectral data set and related environmental variables of objects at different observing scales, the physical models set and a priori knowledge set which provide the main geographical background data for models.

8 Knowledge Database and Inversion

209

8.2.1 The Definition of Spectrum at Three Measured Scales To avoid confusion in spectral matching for land cover type recognition, we clearly defined and measured three kinds of spectrum for different observed objects, while considering the application requirements. The three spectrums are the spectrums of materials, endmembers and remote sensed pixels. The spectrum of materials is usually measured under well controlled measurement conditions in the laboratory. The measured objects can be crop leaves, mineral samples, and water samples. They are essential spectrum data of the spectrum database, and are usually taken as given variables in physical models. The so-called spectrum of endmembers is the spectrum of components of remotely sensed pixel, it is the basic parameter in general remote sensing models. The spectrum of endmembers is usually measured in the field, where the surface of the measured object is relatively uniform, and where the measuring FOV of sensor is less than pixel size. For example, in the scene integrated geometric optical BRDF model on canopy reflectance, the sunlit crown, sunlit background, shaded crown and shaded background are endmembers of the modeled forest pixel. Their spectrums are usually the main parameters of scene integrated models, which may differ from the spectrum of leaves and soils. The spectrum of remotely sensed pixels comes from remotely sensed observations, and is usually the main concern of users. Note that when we consider different modeling scales, the measured endmember spectrum and the spectrum of remotely sensed pixels should be used according to the scale of the application (Li et al., 2002; Wang et al., 2003).

8.2.2 Typical Land Surface Objects Spectrum Database In the spectral data set, for every typical object the measured data includes not only its surface spectra, but also the environmental variables of each observed object. The variables are all physical model parameters, which can be used to predict the spectra of a given surface object by a physical model. In the first step of the database establishment, the main typical land surface objects include three land cover types: the main crops growing in China, which are winter wheat, rice, maize, cotton and cole; rocks and minerals; and bodies of water. The collected data consist of two parts. One part is the individual spectrum data sets measured during the last 20 years. In order to keep this data available, we compiled the data following our own data collection standards. For instance, some of the data were measured in the laboratory and in standard well controlled measuring conditions, such as for minerals. The other part consists of new measurements. We first determined the environmental variables to be measured, which are dependent on the mature physical models, and widely required parameters for remote sensing applications. We established the technical criteria for remote sensing experiment instruments, laboratories and

210

J. Wang, X. Li

field sites, and also the technical criteria and requirements for the measurement of object’s spectrum and environmental variables. All new data measurements follow these criteria and regulations, ensuring that the measurements of surface spectra and environmental variables are made together. For those objects which surface features may change with time, such as crop vegetation canopy, we made measurements during each of their main growth stages across their entire growing season. These data sets with temporal changes can be used to extract the temporal change information of the canopy surface reflectance and canopy structural parameters, which is the most important prior knowledge for canopy reflectance model inversion and data assimilation.

8.2.3 From Spectrum Database to Knowledge Base To make the spectrum database applicable for quantitative remote sensing, we further describe how to develop the spectrum database to the spectrum knowledge base. We do this by integrating the spectrum database, the remote sensing image base, the remote sensing model base, and the geographic background database to create the spectrum knowledge database. The remote sensing model base is the key part of the knowledge base. The models describe the relationship between the spectrum reflectance measurements and related environmental variables. The spectra of materials, of endmembers, of remote sensing pixels and the measuring environment variables can be linked by models when the scale of observation is known. The geographic background knowledge data sets provide the prior knowledge of the model parameters for running the models. A significant function of the knowledge base is to make surface spectra simulation or prediction. Because there are too many types of land surface objects, the amount of objects we can measure is always limited. The potential change of the objects are infinite, we can not include all the cases for even one given object. For example, for winter wheat, we almost cannot obtain all the spectrum for its entire growing seasons from the different regions where it is planted. Considering the application of spectrum database, users may occasionally require the spectra of winter wheat for a given date from a specific place, which may not be stored in the database. The simulation or prediction of spectrum is achieved by using physical-based remote sensing models, and applicable empirical models on objects, such as crop simulation models, as well as the knowledge base. Therefore, our knowledge base developed the ability to simulate spectra. When the spectra of an object requested by the user is not in the database, the remote sensing models can be used to simulate the spectra, based on measured spectrum and environmental variables of similar land surface objects saved in the knowledge base. For example, our database stored some standard spectrum of winter wheat at its different growing stages, while the user requires the spectrum at a specific stage in order to predict its growing state from his remotely sensed image. The spectrum from existing spectrum and from related structural parameters of several growing seasons will be calculated with a

8 Knowledge Database and Inversion

211

spectral prediction modular. The crop simulation model will provide information about the growing tend, the geographic knowledge base will provide the information about phenological and regional characteristics. The spectral database will provide the spectrum of material and components, and also the structural parameters of a typical related winter wheat. In this way, the modular can interpolate and extend the spectrum in temporal and spatial scales, and also predict the spectrum at a given field of view of sensor, a given sunlight, and under the specific atmospheric conditions. The simulated spectra will provide the user with a reference. Spectra data extension is an important feature of our knowledge base. It can also be available as a research platform for studies on object surface spectra prediction and model validation. At present, in our knowledge base, the model base includes general models, physical models, and some special application models. The geographic background data includes the following: the DEM of 1:100,000–1:250,000 of the demonstration region, and land use map of the demonstration region, the base data of 1:4,000,000 covering the whole of China (DEM, physiognomy, vegetation, soil, geosciences, rivers and lakes), phenological phase of typical crops, Chinese geological map (1:5,000,000), such as Nonmetals Metals Mineral Resource Map, Mineral Resource Map and China water resource map (1:4,000,000). From the measured database, we can construct the prior knowledge of land surface parameters by the statistical data analysis. The a priori knowledge of parameters can be expressed as their mean and variance at spatial and temporal scales of the accumulated data. The establishment of the spectrum knowledge base allows knowledge based inversion strategy and algorithm to be used to process remotely sensed image (Wang and Li, 2004).

8.2.4 The Internet-based Spectrum Knowledge Database Service System The internet-based spectrum knowledge database service system includes the software system and its corresponding hardware environment. This system consists of several modules, such as spectrum querying, model calculation, knowledge querying, image management, application demonstrations, and system management. The spectrum and models in our knowledge database were all imported according to the data quality level regulations. The system also has several demonstrations of typical field remote sensing applications, which include the following: production estimation and growth monitoring of crops in North China; precise agriculture demonstration of winter wheat, demonstrations of cotton growth monitoring; rock and mineral mapping using airborne hyperspectral remotely sensed data; water quality evaluation of the Huangpu River in Shanghai with airborne hyperspectral remotely sensed data. By employing the explanations in the demonstration system, users can learn how to apply the spectrum and other data to extract required land surface parameters from remotely sensed images.

212

J. Wang, X. Li

8.3 Land Surface Parameters Inversion Supported by the Knowledge Database Based on the spectrum database, we propose our methods to estimate land surface parameters, which make effective use of the spectrum knowledge. There are two kinds of questions that need to be considered. One is how to abstract the spatial and temporal distribution of the prior knowledge of parameters from database. When we estimate vegetation parameters from remotely sensed data, the data we used can be from a scene covering large region, so we need to have the spatial distribution prior knowledge of parameters around the same region to support the prior knowledgebased inversion algorithm. When we want to understand the temporal changes of crop growth parameters during its growing seasons, we need to have the knowledge with temporal distribution of estimated parameters, such as when we perform data assimilation using a crop growth model. The second kind of questions we should consider is how to introduce the information on the spatial and temporal distribution of parameters into the inversion algorithm. The model we use for the inversion should be not only suitable for processing one special scene remotely sensed data, but also for the data that changes with time. A dynamic model should then be developed, and a data assimilation algorithm also should be developed. The new method should be able to perform time sequential matching for better estimation of parameters at given time and date.

8.3.1 Extract Spatial and Temporal Distribution of Land Surface Parameters From the spectrum database, we can extract the spatial and temporal distributed a priori knowledge of land surface parameters by means of statistical data analysis, as well as model predictions. To do this, we should first understand the kinds of spatial and temporal data that are available, and then how to make use of them. The prior knowledge data of parameters, which are applied in inversion and data assimilation algorithms, can be at several spatial and temporal scales. The prior knowledge data at different spatial scales can be categorized in two types. One type of data relate to the observed objects, but at different resolutions at which the data is acquired. The spatial resolution of acquired data could be at tens of centimeters, at tens of meters, or even at several kilometers. Another type of data relate to the spatial distribution variability from the same or different types of land surface objects. The spectrum database provides knowledge data by mainly focusing on the former kinds of data. For the latter, some geographic background data with several certain scales can be available. Looking at the example of inversion of vegetation canopy reflectance model, the function of prior knowledge provided by spectrum database can be the following: first, spectrum database can provide the spectrum of materials or endmembers, such as those of leaves and soil from the

8 Knowledge Database and Inversion

213

background of crop canopy, which can be used to fix some parameter values, if a model has too many parameters and there are only limited observations to be available for estimate the all model parameters. Secondly, the spectrum database can provide some statistical information of parameters, such as the parameters’ mean value and standard deviation, which can be used as initial guesses and the uncertainty index of parameters to be retrieved. For prior knowledge data at temporal scale, the spectrum database can provide some referencable parameters for dynamic models, while retrieving the variable information of land surface parameters. Looking again at the example of crop canopy reflectance model inversion, due to crop growth over time, the observed spectrum will change. The dynamic change spectral observations can provide an opportunity to retrieve biophysical and biochemical parameters of the crop surface. Thus, the dynamic spectral data of continuing growing vegetation can be used to extract some specific knowledge from the temporal sequence. This type of knowledge may be used to compensate for the limited information from remotely sensed data. In addition, the knowledge is useful for data assimilation. The continuous temporal knowledge of the parameters can be applied to compare with the dynamic model prediction, and then to provide reference feedback to adjust the model’s parameter estimation for the data assimilation procedure. This should result in a better estimation for parameters when their values change with time and with natural features, such as the leaf area index in the crop growth model.

8.3.2 Inversion Model Based on Bayesian Network to Integrate Physical Model and Prior Knowledge Recent research has shown that combining the empirical formula method and the physical model inversion into a new hybrid inversion scheme for estimating surface parameters will be a promising trend (Fang and Liang, 2005; Liang, 2004). Following previous work on hybrid inversion, we build a new hybrid inversion scheme which uses a Bayesian Network (BN) to determine the mapping relationship between simulated reflectances and their corresponding biophysical parameters (Qu et al., 2005). As a hierarchical probability model, BN can not only be used as a non-parameters regression model, but can also be used to deduce information from multi-layered parameters (Marcot et al., 2001). This differs from other nonparameters regression methods, such as Neural Network. In our approach, we focus on the incorporation of prior knowledge derived from spectrum database and physical model into a unified framework. A simple Bayesian Network is illustrated as Fig. 8.1. Using the Bayesian Network to retrieve parameters, the posterior probability density distribution of A can be calculated using the observed data and their ancillary parameters, the following Eq. (8.3) can be derived using Bayesian theorems.

214

J. Wang, X. Li

Fig. 8.1 A simple Bayesian Network

C

A

Ancillary

Parameter space

B

p(A|B = bi ,C = ck ) ∝ p(C = ck )p(A|C = ck )p(B = bi |A)

Data

(8.3)

Where p(A|C = ck ) presents the prior knowledge of interested parameters, p(C = ck ) represent the probability distribution of other variables, such as time, location, elevation, land use, and other assistant information, which can affect the ancillary parameters, p(A|C = ck ) is the probability density distribution of the parameters to be derived after obtaining the above information. The two probability distributions can be obtained from the spectrum database, and the quantitative influence between them, i.e., p(A|C) can be obtained statistically by using data simulated by physical models. By extending the Bayesian theorem into the Bayesian Network, which uses a multifactor deducing method, the prior knowledge about the land surface parameters can be extracted from the spectrum database, and then can be combined with the physical model to retrieve the information on the desired parameters. By integrating the observed data and new information into a unified framework to infer knowledge, this new hybrid inversion scheme has shown that it can incorporate more information besides model parameters into the process of remote sensing model inversion to retrieve biophysical and biochemical parameters. The process of extracting knowledge from the historical database, and using it in the retrieving information from remote sensing models and remotely sensed data is the summingup of prior knowledge and new information. In this process, the information from newly obtained data can be updated and added to the existing knowledge on parameters. In our preliminary research, our proposed inversion approach was used and validated through retrieving the biophysical and biochemical parameters of winter wheat. We abstracted the prior knowledge of canopy reflectance model parameters from the spectrum database to obtain crop growth parameter distribution at different growth stages. We developed the inversion model based on the Bayesian Network to integrate the physical model and prior knowledge to estimated LAI and chlorophyll (a + b) with simulated BRDF spectral data, and to estimated LAI with ETM image (Qu et al., 2005).

8 Knowledge Database and Inversion

215

8.4 Parameters Scaling and Validation Scaling effects are the basic issues in remote sensing research. We studied the scaling effect in thermal infrared remote sensing models and the reciprocity principle when these basic physical principles are applied in processing special problems in remote sensing (Li et al., 2000a). When we estimate model parameters by remote sensing model inversion, one important question is whether the parameters require a uniform estimation value when observations are at different spatial scales. In our experiment of model inversion, one parameter may have different estimated values if the available remote sensing data are at different scales of observation. In such case, it is still a problem on how to improve the precision of model parameters estimation. Taking the estimation of crop planting area as example, we usually obtain different estimated area value from AVHRR data or from TM data. And the difference may not follow a certain formula when land cover types from the observed regions are different. This is mainly due to a greater amount of mixed pixels in AVHRR data, and the heterogeneity of land surface. It is very similar to the fractal problems, such as the coastline measurement using the data from different spatial resolution. In order to describe the difference resulting from different scales of observation, we proposed the concept of histo-variogram, defined as the “total fractal dimension”, and applied it to derive a formula for scaling correction on land surface parameter estimates (Zhang et al., 2003). We use LULC data as an example to study land use area estimation method using down-scaling. For the crop area estimates, we defined the initial area S0 which means the observed area at initial measured scale, stated its relation to the total fractal dimensions (d) and coefficients ( f (S0 )), and then applied it to up-scaling and down-scaling in crop area estimation. The estimated crop area of standard pixel (Sδ ) at different scale (δ) is expressed as Sδ = f (s0 )δ d−2

(8.4)

where ( f (S0 )) is the function of the initial area (S0 ) and with the same unit as S0 . δ is for measured scale, can be expressed with fractal of standard pixel. The relation between fractal (δ), coefficient (I) and standardized area (Sδ 0 ) is: (8.5) δi − δ0 (Iδ i − Sδ 0 )/Iδ i = 1.62Dδ i + 0.03 δ0 where I is the f (S0 ) expressed in Eq. (8.4), can be calculated by Eq. (8.4) with the initial measured scale δi in the process of scaling-up, δ0 is for finer measured scale than δi , Sδ 0 is the crop area we want to estimate by down-scaling method at the measured scale δ0 . We can obtain the estimated area when the scale of the data is changing based on the down-scaling method. In this way, using LULC data, we obtain the result of crop plant area estimation result by the down-scaling method, where the relative error is less than 5% when the estimated pixel size is 16 times that of the original data.

216

J. Wang, X. Li

8.5 Summary and Discussion The previous research we described are our main ideas on remote sensing model inversion strategy and algorithm, base on the analysis of the special ill-posed inversion problem in land surface parameters inversion. The available land surface spectrum knowledge database provides useful models and a priori knowledge of model parameters. We are presently researching the following: 1. How to use spatial and temporal change knowledge in remote sensing model inversion? Based on our previous research, we introduce a dynamic model to obtain a time consequence of model parameters. The idea on data assimilation will then be applied to improve the precision and reliability of parameters estimation. 2. How to validate and evaluate the parameters’ inversion result when the raw data are at different scales? This is still a major problem filled with many uncertainties. The problem regards the scaling effect of models, parameters and observations. 3. How to meet users’ requirements for land surface parameters estimated by remote sensing data, so as to provide input for common land model (CLM)? In this case, the a priori knowledge of the parameters’ spatial distribution and temporal signature will be obtained from the knowledge base, and then used to invert these dynamic changing parameters by introducing the main idea of data assimilation.

References Fang H, Liang S (2005) A hybrid inversion method for mapping leaf area index from MODIS data: experiments and application to broadleaf and needleleaf canopies. Remote Sens. Environ. 94(3):405–424 Gao F, Zhu Q (1997) Process system of measured bidirectional reflectance in Changchun laboratory. J. Remote Sens. 1(Suppl):123–130 Kimes DS, Harrison PR, Ratcliffe PA (1991) A knowledge-based expert system for inferring vegetation characteristics. Int. J. Remote Sens. 12:1987–2020 Kimes DS, Knyazikhin Y, Privette J, Abuelgasim A, Gao F (2000) Inversion methods for physically-based models. Remote Sens. Rev. 18:381–439 Kuusk A (1991) Determination of vegetation canopy parameters from optical measurements. Remote Sens. Environ. 47:194–202 Li X, Yan G, Liu Y, Wang J, Zhu C (1997) Uncertainty and sensitivity matrix of parameters in inversion of physical BRDF model. J. Remote Sens. 1(Suppl):113–122 Li X, Wang J, Hu B, Strahler AH (1998) On utilization of prior knowledge in inversion of remote sensing models. Sci. China (Series D) 41(6):580–586 Li X, Wang J, Strahler AH (2000a) Scale effects and scaling-up by geometric-optical model. Sci. China (Series E) 43(Suppl):17–22 Li X, Gao F, Wang J, Strahler AH (2000b) Estimation of the parameter error propagation in inversion based BRDF observations at single sun position. Sci. China (Series E) 43(Suppl):9–16 Li X, Gao F, Wang J, Strahler AH (2001) A priori knowledge accumulation and its application to linear BRDF model inversion. J. Geophys. Res. 106(D11):11925–11935

8 Knowledge Database and Inversion

217

Li X, Wang J, Gao F, Strahler A, Su L (2002) Accumulation of spectral BRDF knowledge for quantitative remote sensing of land surfaces. Proc. The First Int. Symposium on Recent Advances in Quantitative Remote Sensing, Torrent, Valencia, Spain, 16 September 2002 Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, New York Marcot BG, Holthausen RS, Raphael MG, Rowland M, Wisdom M (2001) Using Bayesian belief networks to evaluate fish and wildlife population viability under land management alternatives from an environmental impact statement. Forest Ecol. Manage. 153(1–3):29–42 Morisette JT, Baret F, Privette JL, Myneni RB, et al. (2006) Validation of global moderateresolution LAI products: a framework proposed within the CEOS land product validation subgroup. IEEE Trans. Geosci. Remote Sens. 44(2):1804–1817 Myneni RB, Maggion S, Iaquinta J, et al. (1995) Optical remote sensing of vegetation: modeling, caveats and algorithms. Remote Sens. Environ. 51:169–188 Privette JL, Eck TF, Deering DW (1997) Estimating spectral albedo and nadir reflectance through inversion of simple BRDF models with AVHRR/MODIS-like data. J. Geophys. Res. 102(D24):29529–29542 Qu Y, Wang J, Song J, Lin H (2005) A hybrid inversion scheme to estimate biophysical parameters of winter wheat: simulation and validation. Proc. ISPMSRS’05, Beijing,, 17–19 October 2005, pp 743–746 Rodgers CD (1976) Retrieval of atmospheric temperature and composition from remote measurement of thermal radiation. Rev. Geophys. Space Phys. 14(4):609–624 Smith JA (1993) LAI inversion using a back propagation neural network trained with a multiple scattering model. IEEE Trans. Geosci. Remote Sens. 31:1102–1106 Tan B, Hu J, Huang D, et al. (2005) Assessment of the broadleaf crop leaf area index product from the Terra MODIS instrument. Agric. Forest Meteorol. 135:124–134 Tarantola A (1987) Inverse Problem Theory – Methods for Data Fitting and Model Parameter Estimation. Elsevier, Amsterdam, The Netherlands, 613pp Verstraete MM, Pinty B, Myneni RB (1996) Potential and limitation of information extraction on the biosphere from satellite remote sensing. Remote Sens. Environ. 52:201–214 Wanner W, Strahler AH, Hu B, Lewis P, Muller J-P, Li X, Barker Schaaf CL, Barnsley MJ (1997) Global retrieval of bidirectional reflectance and albedo over land from EOS MODIS and MISR data: theory and algorithm. J. Geophys. Res. 102(D24):17143–17162 Wang J, Li X (2004) The spectrum knowledge base of typical objects and remote sensing inversion of land surface parameters. J. Remote Sens. 8(Suppl):4–7 (in Chinese) Wang J, Li X, Sun X, Liu Q (2000) Component temperatures inversion for remote sensing pixel based on directional thermal radiation model. Science in China (Series E) 43(Suppl):41–47 Wang J, Zhang L, Zhang H, Li X, Liu S (2003) Current progress of constructing the spectrum knowledge base of typical objects of land surfaces, ESA SPECTRA workshop in France (Italy), November 2003 Yan G, Wang J, Li X (2001) Making use of a priori knowledge of vegetation spectrum in inversion for canopy structure parameters. Proc. IGARSS’01, Sydney, Australia, 9–13 July 2001 Zhang H, Jiao Z, Yang H, Li X, et al. (2003) Research on scale effect of histogram. Sci. China (Series D) 45(10):1–12

Chapter 9

Retrieval of Surface Albedo from Satellite Sensors Crystal Schaaf, John Martonchik, Bernard Pinty, Yves Govaerts, Feng Gao, Alessio Lattanzio, Jicheng Liu, Alan Strahler, and Malcolm Taberner

Abstract Observations from a number of polar-orbiting and geostationary satellite sensors are now being used to produce operational land surface albedo products for range of modeling applications. The MODIS, MISR and Meteosat algorithms are presented as examples of the current strategies being employed to best exploit multi-day sequential, multi-angular instantaneous, and multi-temporal observations and accurately specify the reflective qualities of the underlying surface. While these retrievals represent a major advance in the remote sensing of the spatial and temporal heterogeneity of the surface, issues such as atmospheric correction, directional-tohemispherical conversion, and spectral interpolation remain to confound the satellite signal and introduce uncertainties and variability within and between products. Nevertheless, the potential of using multiple products and fusing recent observations with remotely sensed historical data must be explored as a realistic way to meet the needs of the modeling community. Keywords: MODIS · MISR · Meteosat · albedo · reflectance · anisotropy

Crystal Schaaf, Jicheng Liu and Alan Strahler Department of Geography and Environment, Boston University, Boston, USA [email protected] Feng Gao Earth Resources Technology, Jessup, USA John Martonchik Jet Propulsion Laboratory, Pasadena, USA Bernard Pinty and Malcolm Taberner Global Environment Monitoring Unit, IES, EC Joint Research Centre, Ispra (VA), Italy Yves Govaerts EUMETSAT, Darmstadt, Germany Alessio Lattanzio Makalumedia GMBH, Darmstadt, Germany S. Liang (ed.), Advances in Land Remote Sensing, 219–243. c Springer Science + Business Media B.V., 2008

219

220

C. Schaaf et al.

9.1 Introduction The availability of a large number of directional observations sampling the viewing hemisphere over a particular land surface can effectively capture its surface anisotropy and thus be used to accurately compute the surface albedo of that surface. While numerous samples may be possible in the field or laboratory, remotely sensed retrieval methods based on data from individual satellites usually must suffice with a limited number of directional reflectances of the surface, and the producers of such data sets must acknowledge that these observations may not necessarily represent a well-distributed sampling (Privette et al., 1997). Therefore a model is usually adopted to characterize the surface anisotropy – a model which can be inverted with a finite set of angular samples and then be used to predict surface reflectance in any sun-view geometry and derive surface albedo (Roujean et al., 1992; Walthall et al., 1985; Rahman et al., 1993; Engelsen et al., 1996; Wanner et al., 1997; Pinty et al., 2000a; Br´eon et al., 2002; Maignan et al., 2004). The acquisition of directional measurements from an individual sensor is determined by its scanning configuration and the platform’s orbital characteristics (Barnsley et al., 1994). However, cloud obscuration always reduces the number of clear-sky observations possible. Therefore, in the case of a single field of view sensor such as the MODerate Resolution Imaging Spectroradiaometer (MODIS), on board the polar orbiting Terra and Aqua platforms, an adequate directional sampling of surface reflectances can only be obtained by the accumulation of sequential observations over a specified time period. Multi-angular instruments such as the Multiangle Imaging SpectroRadiometer (MISR) instrument (also on board the Terra platform) obtain sufficient simultaneous directional observations to specify the surface anisotropy whenever a cloud-free acquisition is possible. Geostationary sensors (such as Meteosat) must trade numerous acquisitions under different illumination conditions during a day for directional observations to obtain the angular information necessary to sample the surface’s directional characteristics. Since 2000, all of these approaches have been implemented operationally to produce robust surface albedo fields for use in climate, hydrological, biogeochemical, and weather prediction models.

9.2 Background As a key land physical parameter controlling the surface radiation energy budget (Dickinson, 1983, 1995), global surface albedo with an absolute accuracy of 0.02–0.05 is required by climate models at a range of spatial and temporal scales (Henderson-Sellers and Wilson, 1983). Land cover-based schemes have historically been adopted in most of the land surface models and climate models for the parameterization and specification of surface albedo (Bonan et al., 2002; Sellers et al., 1996). Natural landscapes, however, are a collection of nested objects in a hierarchy and various processes control the biophysical characteristics at different spatial scales (Woodcock and Harward, 1992; Collins and Woodcock, 2000).

9 Retrieval of Surface Albedo from Satellite Sensors

221

Therefore, land surface models usually allow a sub-grid specification of land cover proportions to account for the heterogeneity of surface properties within a grid (Dickinson et al., 1995; Bonan et al., 2002), while the climate models are generally implemented at coarser spatial resolutions. However, the increasing spatial resolution of modern climate models makes it necessary to examine the spatial features of global surface albedo and the effect of spatial scales on the albedo specification. Therefore, a consistent and accurate global albedo data set is essential to the investigation of the sensitivity of climate to various types of forcing and to the identification of the effects of human activities. Satellite remote sensing represents the only efficient way to compile such consistent global albedo characterizations. Historically, global albedo data sets have been derived from the Advanced Very High Resolution Radiometer (AVHRR) (Csiszar and Gutman, 1999) and the Earth Radiation Budget Experiment (ERBE) radiometer (Li and Garand, 1994). With the advent of routine albedo products derived from MODIS (Gao et al., 2005; Schaaf et al., 2002; Lucht et al., 2000), MISR (Martonchik, 1997; Martonchik et al., 1998b), CERES (Clouds and the Earth’s Radiant Energy System), POLDER (Polarization and Directionality of the Earth’s Reflectances) which is currently on board PARASOL (Polarization & Anisotropy of Reflectances for Atmospheric Sciences) coupled with Observations from a Lidar (PARASOL) (Leroy et al., 1997; Hautecoeur and Leroy, 1998; Bicheron and Leroy, 2000; Maignan et al., 2004), and Meteosat (Pinty et al., 2000a,b), albedo data sets with spatial resolutions of 500 m to 20 km and temporal frequencies of daily to monthly are now available. Although the retrieval of albedo from these instruments represents a major advance in sensing the spatial and temporal surface heterogeneity, issues such as atmospheric correction, directional-to-hemispherical conversion, and spectral interpolation can still confound the satellite signal and introduce uncertainties. Most of these satellite products rely on sophisticated radiative transfer methods (Vermote et al., 1997; Berk et al., 1998; Liang et al., 1999; Liang, 2000) and bidirectional modeling (Roujean et al., 1992; Walthall et al., 1985; Rahman et al., 1993; Engelsen et al., 1996; Wanner et al., 1995; Wanner et al., 1997; Martonchik et al., 1998b; Pinty et al., 2000a, b) to obtain accurate surface quantities. The modeling community has enthusiastically begun to utilize these global and regional satellite albedo products as they have become available (Oleson et al., 2003; Zhou et al., 2003; Tian et al., 2004; Roesch et al., 2004; Knorr et al., 2001; Myhre et al., 2005a, b). With 5 or more years of data now available, interannual variations can be explored and short-term climatologies prepared which compensate for transient cloudiness or snowcover (Moody et al., 2005; Gao et al., 2005; Barlage et al., 2005). However, there remains the need to generate analogous surface albedo products prior to year 2000 and in particular over the last 25 years or so where Earth observing systems from space (e.g., the series of weather satellites) have been acquiring relevant data. Unfortunately, the design of the large majority of these global observation systems for environmental applications has been driven solely by demands in the domain of meteorology and weather forecasting and, as a consequence, these sensors do not fulfill some basic requirements for quantitative remote sensing applications over land, such as those related to the accurate sensor characterization, geolocation, and calibration.

222

C. Schaaf et al.

However, despite a number of technological limitations, these historical weather sensors constitute the only possible solution left to remote sensing scientists to assess such quantities as surface albedos at a global scale over the past decades. The required sequential accumulation of data over multiple days, i.e., for different view conditions, as adopted for MODIS sensors for inferring flux quantities from a number of instantaneous radiance measurements, can be extended to exploit the AVHRR series data archive (see for instance d’Entremont et al., 1999). An analogous strategy of sequential accumulation (but over every single day, i.e., for different solar illumination conditions), can be envisaged in the case of the archived measurements collected by sensors placed on a geostationary orbit. The exploitation of the reciprocity principle then allows the production for every sample area, of daily accumulated datasets of radiances measured at different viewing angles (see for instance, Lattanzio et al., 2006). Assuming thus that the geophysical system under investigation does not suffer from drastic changes during the period of data accumulation, e.g., multiple hours (days) for geostationary (polar) orbiting sensors, the temporal sampling of the radiance field for a given location can be interpreted as an angular sampling.

9.3 MODIS Albedo and Anisotropy Algorithm The operational MODIS albedo and anisotropy algorithm makes use of a kerneldriven, linear model of the Bidirectional Reflectance Factors (BRFs), which relies on the weighted sum of an isotropic parameter and two functions (or kernels) of viewing and illumination geometry (Roujean et al., 1992) to estimate the Bidirectional Reflectance Distribution Function (BRDF). One kernel is derived from radiative transfer models (Ross, 1981) and the other is based on surface scattering and geometric shadow casting theory (Li and Strahler, 1992). The kernel weights selected are those that best fit the cloud-cleared, atmospherically corrected surface reflectances available for each location over a 16-day period (Lucht et al., 2000; Schaaf et al., 2002). This model combination (Ross-Thick/Li-Sparse-Reciprocal or RTLSR) has been shown to be well suited to describing the surface anisotropy of the variety of land covers that are distributed world-wide (Privette et al., 1997; Lucht et al., 2000) and is similar to the kernel-driven schemes used to obtain anisotropy and albedo information by the POLDER (Leroy and Hautecoeur, 1998; Bicheron and Leroy, 2000; Maignan et al., 2004) satellite sensor. Once an appropriate anisotropy model has been retrieved, integration over all view angles results in a Directional Hemispherical Reflectance (DHR) or a black-sky albedo at any desired solar angle and a further integration over all illumination angles results in a BiMemispherical Reflectance (BHR) under isotropic illumination or a white-sky albedo. These albedo quantities are intrinsic to a specific location and are governed by the character and structure of its land cover. They can be combined with appropriate optical depth information to produce an actual (blue-sky) albedo for a specific time such as would be measured at the surface by field sensors under ambient

9 Retrieval of Surface Albedo from Satellite Sensors

223

illumination. The anisotropy models can also be used to compute surface reflectances at any other view or solar zenith angle desired. The spectral acquisitions can also be combined via narrow to broadband conversion coefficients (Liang et al., 1999; Liang, 2000) to provide broadband anisotropy information and thus broadband albedos similar to those routinely collected in the field with pyranometers and commonly used in large-scale models. The MODIS instruments on both Aqua and Terra have a 16-day repeat cycle and provide measurements on a global basis every 1–2 days. The 16-day period has also been chosen as an appropriate tradeoff between the availability of sufficient angular samples and the temporal stability of surface (Wanner et al., 1997; Gao et al., 2001). This assumption becomes tenuous during periods of strong phenological change such as vegetation greenup, senescence, or harvesting. By overlapping processing of the data such that retrievals are attempted every 8 days (based on all clear observations over the past 16 days), some of the phenological variability can be more accurately captured. Other periods of rapid change at the surface such as ephemeral snowfall also provide challenges in retrieving appropriate surface albedos. The MODIS algorithm addresses this by determining whether the majority of the clear observations available over a 16-day period represent snowcovered or snow-free situations and then retrieving the albedo of the majority condition accordingly. For those locations where the full anisotropic model described above can not be confidently retrieved due to poor or insufficient input observations, a backup algorithm is employed. This method (Strugnell and Lucht, 2001; Strugnell et al., 2001) relies on a global database of archetypal anisotropic model based on a land cover classification and historical high quality full model retrievals. This a priori data base is then used as a first guess of the underlying anisotropy and any available observations are used to constrain the model. While considered a lower quality result, Jin et al. (2003a, b) and Salomon et al. (2006) have found that this backup method often performs quite well under normal situations (e.g., Fig. 9.1). In view of the often insufficient angular sampling available, a synergistic use of multi-sensor observations has offered the best opportunity to improve both the coverage and the quality of global anisotropy and albedo retrievals. Terra has a descending equatorial crossing time of 10:30 a.m., while Aqua is flying in an ascending orbit with a 1:30 p.m. equatorial crossing time. By combining MODIS observations from both Terra and Aqua, more high-quality, cloud-free observations (under varying solar zenith angle) are available to generate better constrained model retrievals (see Fig. 9.1). Since the MODIS-Terra and MODIS-Aqua have similar instrument characterizations and utilize the same atmospheric correction algorithm, the combination of these data is fairly straightforward. However, the calibration and geolocation of both instruments must be continually monitored for compatibility and the quality of the aerosol retrieval from each sensor and its effect on the respective atmospherically-corrected surface reflectances must also be accounted for. In general, the combined Terra and Aqua MODIS product processing stream begins with a detailed quality check of each atmospherically corrected surface reflectance and then assigns various penalty weights to the individual observations according to

224

C. Schaaf et al. MODIS Surface Albedo vs. Boulder, Colorado, 2003 0.90 0.81

Ground Site (Daily at LSN) Ground Site (16-Day Average)

0.72

TERRA (Full Inv.) TERRA (Mag. Inv.)

0.63

Surface Albedo

AQUA + TERRA (Full Inv.) AQUA + TERRA (Mag. Inv.)

0.54 0.45 0.36 0.27 0.18 0.09 0.00

1

17 33 49 65 81 97 113 129 145 161 177 193 209 225 241 257 273 283 305 321 337 353 369

Julian Day

Fig. 9.1 MODIS 16-day broadband albedo as compared to field data from the Boulder BSRN tower site

the quality flag contained in each surface reflectance product (Schaaf et al., 2002). Thus, the quantified uncertainty of the sensor-specific surface reflectances is directly integrated into the retrieval. Results from the combined Terra-MODIS and Aqua-MODIS algorithm (Fig. 9.2) indicate that the increase in the number of observations does result in more higher quality retrievals and can decrease the use of backup retrievals by as much as 50% (Salomon et al., 2006). The MODIS BRDF/Albedo standard operational products (Lucht et al., 2000; Schaaf et al., 2002; Gao et al., 2005) provide the best fit RTLSR model parameters describing the surface anisotropy, black sky and white sky albedo quantities, the nadir (view-angle-corrected) surface reflectance of each location, and extensive quality information. The best fit RTLSR model parameters are retrieved for the first seven spectral bands of MODIS and three additional broadbands (0.3–0.7 µm, 0.7–5 µm, 0.3–5 µm). These anisotropy models are then used to compute white sky albedo and black sky albedo at local solar noon for the same seven spectral bands and three broadbands. The anisotropy models are also used to correct surface reflectances for view angle effects and provide BRFs at a common nadir view angle (Fig. 9.3). These Nadir BRDF-Adjusted Reflectances (NBAR) are computed for the seven spectral bands and are used as the primary input for the MODIS Land Cover and Land Cover Dynamics Products due to their stability and temporal consistency (Friedl et al., 2002; Zhang et al., 2003). In addition to the standard 500 m and 1 km tiled products in a sinusoidal projection, these same science data sets are

9 Retrieval of Surface Albedo from Satellite Sensors

225

Fig. 9.2 Quality improvement possible by combining MODIS observations from both the Terra and Aqua platforms. Top panel shows Terra alone (green high quality, red lower quality) while bottom panel shows Terra and Aqua (March 2006) MODIS Reflectance (MODO9GHK) 2004-126

Nadir BRDF-Adjusted Reflectance (NBAR) 2004-126

NIR(0.1-0.4) Red (0.0-0.2) Green (0.0-0.18)

NIR(0.1-0.4) Red (0.0-0.2) Green (0.0-0.18)

MODO9GHK. View Angular effects

Nadir BRDF-Adjusted Reflectance (NBAR).

between two swaths

Angular effect is removed

(North China Plain)

(North China Plain)

Fig. 9.3 Application of the MODIS anisotropy model parameter product to correct adjoining surface reflectance swaths for view angle effects by generating Nadir BRDF-Adjusted Reflectances (NBAR) 5 May 2004

226

C. Schaaf et al.

Fig. 9.4 MODIS global white sky albedo (from Terra and Aqua) March 2006

also routinely produced at a 0.05◦ spatial resolution in a global geographic (latitude/longitude) projection specifically for use by global modelers (Gao et al., 2005). In Fig. 9.4, the global false color field of spectral white sky albedo (March 2006) captures the seasonal variation due to vegetation phenology and snow cover extent.

9.4 MISR Albedo and Anisotropy Algorithm The Multi-angle Imaging SpectroRadiometer (MISR) on the EOS Terra platform consists of nine pushbroom cameras, viewing symmetrically about nadir in forward to aftward directions along the spacecraft track. Image data are acquired with nominal view zenith angles relative to the surface reference ellipsoid of 0.0◦ , 26.1◦ , 45.6◦ , 60.0◦ , and 70.5◦ in four spectral bands (446, 558, 672, and 866 nm) and with a crosstrack ground spatial resolution of 275 m to 1.1 km and a swath width of about 400 km (Diner et al., 1998). After these data are radiometrically calibrated, georectified, and averaged to a uniform resolution of 1.1 km, the land data undergo a series of processing steps, resulting in a myriad of surface parameters. The basic land surface products currently being generated include the spectral hemispherical-directional reflectance factor (HDRF) at the nine MISR view angles and the associated BHR. The HDRF is a measure of the directional reflectance of the surface under ambient atmospheric illumination (i.e., direct plus diffuse radiation). It is the ratio of the directionally reflected radiance from the surface to the reflected radiance from an ideal lambertian target under identical illumination conditions as the surface. The BHR is the HDRF integrated over all reflection angles in the upward hemisphere, i.e., it is the surface albedo under ambient atmospheric illumination. Related MISR surface parameters are the spectral BRFs at the nine MISR view

9 Retrieval of Surface Albedo from Satellite Sensors

227

angles and the DHR. The BRF and the DHR characterize the surface in the same way as the HDRF and BHR, respectively, but are defined for the condition of direct (i.e., collimated beam) illumination only. Thus, the top-of-atmosphere (TOA) MISR radiances are first atmospherically corrected to produce the HDRF and the BHR, surface reflectance properties as would be measured at ground level but at the MISR spatial resolution. The HDRF and BHR then are further atmospherically corrected to remove all diffuse illumination effects, resulting in the BRF and DHR. In addition to these spectral surface reflectance products, the BHR and DHR, integrated over the wavelength region of Photosynthetically Active Radiation (PAR) (400–700 nm), are also computed. The determination of these surface products requires that the atmosphere be sufficiently characterized in order for the correction process to occur. This characterization is accomplished by means of aerosol retrieval, a process performed on a region 17.6 × 17.6 km in size, containing the 1.1 × 1.1 km size subregions (Martonchik et al., 1998a, 2002a). After a surface BRF is determined at the subregion scale it is fitted to the three parameter modified Rahman–Pinty–Verstraete (MRPV) empirical model (Rahman et al., 1993; Engelsen et al., 1996), which provides a convenient representation of the surface scattering characteristics The details of the retrieval methodologies used to derive these various surface products have been described by Martonchik et al. (1998b). The unique capabilities of MISR’s multiple cameras allow for a simultaneous sampling of the surface anisotropy. By coupling the angular information with the spectral information, the MISR observations can be exploited to capture ephemeral effects such as springtime snow cover. On 17 April 2001 MISR observed a rural part of Manitoba and Saskatchewan about 110 km north of the US border (Path 34, Orbit 7083). Most of MISR’s imaging data have a resolution of 1.1 km, but all nine cameras in the red band (672 nm) and all four bands in the nadir camera take global data at the higher resolution of 275 m. Figure 9.5 shows two false color images of the Canadian scene at 275 m resolution, one emphasizing spectral information and the other, angular information. The image on the left is a multispectral color composite in which the MISR green band (558 nm), red band, and near IR band (866 nm) nadir view imagery are colored blue, green, and red. Here, vegetation appears red due to its high reflectivity in the near IR band and low reflectivity in the green and red bands. The image on the right is a multiangular color composite in which the 60◦ forward view, the nadir view, and 60◦ aftward view images are colored blue, green and red, respectively, essentially color coding the angular signature of the scene. Thus, for example, a region with a reflectance predominately in the nadir direction will appear green. Prominent features in both images are the Assiniboine and Qu’Appelle rivers, running southward and eastward, respectively. The bidirectional reflectance factors for three sites marked by yellow arrows in Fig. 9.5 are displayed in Fig. 9.6. The solar zenith angle is 42◦ and the azimuth angles of the BRFs are about 32◦ from the principal plane. The northern-most site is reddish in color in both composite images in Fig. 9.5, indicating vegetation with substantial backscatter. This backscatter signature is more clearly shown in the BRF plot in Fig. 9.6 The BHR (i.e., actual albedo), is 0.08 in the red band and the NDVI is 0.49, implying moderately dense vegetation. The eastern-most site also has a

228

C. Schaaf et al.

Fig. 9.5 Two false color images (275 m resolution) of an area (240 × 175 km) in central Canada on 17 April 2001 centered on the Saskatchewan–Manitoba border. The left image is a multispectral composite in which red (more like purple/blue in the pdf/doc versions of the images) indicates vegetation. The image on the right is a multiangular composite in which green indicates predominate scattering in the nadir direction Moderately dense vegetation (albedo = 0.08, NDVI = 0.49) Snowy forest (albedo = 0.18, NDVI = 0.24) Agricultural field with light snow (albedo = 0.18, NDVI = 0.13)

0.25

BRF (red band)

0.20

0.15

0.10

0.05 −80

−60

−40

−20

0

20

40

60

80

View Zenith angle (deg)

Fig. 9.6 Bidirectional reflectance factors for three sites marked by yellow arrows in Fig. 9.5

vegetative character, colored purple and red in the multiangular and multispectral composites, respectively. However, the BRFs for this site are higher and with more forward scattering than for the previous site, and with an increased red band BHR of 0.18 and a lower NDVI of only 0.13. The brightening of the BRF, the increase

9 Retrieval of Surface Albedo from Satellite Sensors

229

in forward scattering, and the decrease in the NDVI is indicative of the presence of some snow in and among small scale vegetation, probably agricultural. The last site in Fig. 9.5 is located in Duck Mountain Provincial Park and is colored green in the multiangular composite and red in the multispectral composite. This color combination implies vegetation but with strong scattering in the nadir direction. The BRF plot in Fig. 9.6 shows this signature in more detail, and it is quite different than those from the other two sites. Here, the vegetation is a forest with snow on the ground between the trees. When viewing in or near the nadir direction, the snow is highly visible and the BRF is at its highest. As the view angle progressively increases, the ratio of snow cover to tree structure decreases, lowering the reflectance until, at the extreme off-nadir view angles, the BRF is virtually all canopy with its characteristic pronounced backscatter. During the 7 weeks from 14 August to 29 September 2000, numerous orbits of MISR data were analyzed and compiled for southern Africa, as part of the dry season campaign of the Southern Africa Regional Science Initiative (SAFARI-2000), an international effort to study linkages between land and atmospheric processes. During this period a number of AERONET sunphotometer sites (Holben et al., 1998) were operational over the region, providing independent determinations of aerosol optical depth which were compared to those retrieved using MISR data (Diner et al., 2001). This validation study produced very favorable results, allowing considerable confidence to be placed in the subsequent atmospheric correction procedures and in the quality of the retrieved surface products. Figure 9.7 is a true color, 1.1 km resolution mosaic of the surface DHR for southern Africa, derived from 27 orbital swaths accumulated during this time period. The bright feature in the center is the Makgadikgadi Pans, an extensive salt bed in Botswana. In the interior part of southern Africa, much of the land can be classified as savanna and grassland. Figure 9.8 shows the HDRFs in all four MISR bands for grassland not far from Johannesburg. The grass is dried out, as can be discerned from the monotonically increasing HDRF with wavelength. Using data from 15 August (Path 168, Orbit 3509), this particular site was positioned on the extreme western edge of MISR’s orbital swath, providing multispectral measurements within 1◦ of the retro-solar direction (direct backscatter). The resulting hotspot, due to an almost complete lack of shadowing within the structured surface, can be seen very clearly in Fig. 9.8 as an enhancement of the HDRF in all bands at 49◦ view zenith angle, which is also the solar zenith angle. The hotspot, while not common in MISR surface retrievals, does occur for a wide range of latitudes, appearing at different camera view angles, depending on the season. In addition to the standard MISR products available at 275 m or 1.1 km, many are also available in a format of monthly global maps at a spatial resolution of 0.5◦ in both latitude and longitude. An example of this type of map is displayed in Fig. 9.9, showing surface DHR in natural color for the month of September 2005. The individual 0.5◦ pixels are created by averaging all 1.1 km DHR values accumulated within that pixel for that month. The white specks evident in some areas are fill pixels where no 1.1 km DHR values were available for the entire month, due mainly to cloud activity.

230

C. Schaaf et al.

Fig. 9.7 MISR true color, 1.1 km resolution mosaic of the surface directional-hemispherical reflectance (DHR) for southern Africa (14 August–29 September 2000) DRY GRASSLAND, SOUTH AFRICA Blue

Green

Red

Near-IR

0.45 0.40 0.35 HDRF

0.30 0.25 0.20 0.15 0.10 0.05 0.00 −80

−60

−40

−20

0

20

40

60

80

View Zenith angle (deg)

Fig. 9.8 Spectral and angular variation of HDRF for dried grassland on 15 August 2000 for a site near Johannesburg. Note the hotspot at the view zenith backscatter angle of 49◦ , which is also the solar zenith angle

9 Retrieval of Surface Albedo from Satellite Sensors

231

Fig. 9.9 Surface DHR in natural color for the month of September 2005

9.5 Meteosat Albedo and Anisotropy Algorithm The cornerstone of the retrieval algorithm for geostationary satellites developed by Pinty et al. (2000a) relies on the temporal sampling of geostationary satellites (data acquired every 15 or 30 min from sunrise to sunset) as if it were an instantaneous angular sampling of the radiance field emerging at the top of the atmosphere. The frequency of measurements of the same Earth location is indeed a unique capability offered by geostationary satellites that thus translates into an increasing number of conditions or positive constraints to be satisfied by the retrieval algorithm. The physics of the Meteosat retrieval aims at solving an inverse radiation transfer problem simultaneously with respect to the lower boundary condition, i.e., the surface bidirectional reflectance factor (BRF), and the aerosol optical thickness (Martonchik et al., 1998b, 2002b; Pinty et al., 2000a). All other effects due, for instance, to water vapour and ozone in the atmosphere are accounted for via a prescription of gas concentrations taken from either climatology and/or weather forecast models (re) analysis. In order to simplify the problem further, the gaseous absorption processes are treated separately from the aerosol-scattering-absorbing effects by the specification of two distinct atmospheric layers, one for the representation of the molecular absorption only and the other for the modeling of the coupled surfaceaerosol radiation transfer processes. The inverse algorithm is basically focusing on the estimates of key variables, namely the aerosol load and surface scattering properties, for which the a priori knowledge is quite limited or somewhat uncertain and the level of variability is quite high. The mathematics of the retrievals is established in such a way that the amplitude of the surface BRF is propagated to the top of the atmospheric scattering layer while its shape is modulated by the atmospheric scattering and absorbing properties

232

C. Schaaf et al.

(Pinty et al., 2000a). This is made possible thanks to (1) the mathematical formulation of the surface BRF model, namely the RPV model (Engelsen et al., 1996; Rahman et al., 1993) which separates the amplitude from the shape of the surface BRF and (2) the decomposition of some atmospheric functions like the upward and downward diffuse transmission with a Fourier expansion limited to the first two components. This approach proved to be computer efficient and accurate for modeling the radiance field at the top of a scattering-only atmosphere (Martonchik et al., 2002b). In this way, the angular field of the BRF at the top of the scattering atmosphere can be simply expressed as a sum of contributions invoking the coupling between the surface BRF shape and atmospheric scattering functions that can all be pre-computed and called during the retrieval procedure. For a given set of measured BRF values at the top of the atmosphere the latter procedure itself solves (1) a linear equation to calculate first the amplitude values of the surface BRF for the given pre-computed scattering functions and (2) a second order cost function estimating the closeness between the measured and the modeled BRF values at the top of the atmosphere. The retrieval procedure then ends up with an identification of probability distribution functions of the acceptable solutions, i.e., those satisfying one or multiple criteria depending on the number of degrees of freedom and the distributions of uncertainties in both the observations and in the forward model. The selection of the “most probable” solution for any given set of measured BRF amongst the set of acceptable solutions can be performed using various criteria including the identification of the solution corresponding to the arithmetic mean of the distribution of the amplitude of the surface BRF values. The solution retained is thus a set of model variables and parameters describing the surface scattering problem, such as the parameters characterizing the shape of the surface BRF and the effective aerosol loads, associated with the selected value for the amplitude of the surface BRF. The aerosol loads together with the surface scattering properties are given via effective optical thickness values for a prescribed aerosol type corresponding to average standard aerosol conditions regarding their detailed properties and vertical distribution as well. Since the retrieval strategy delivers the optimized set of the RPV model parameters characterizing the surface BRDF, one can generate DHR or black-sky albedo for any solar angle and/or BHR or white-sky albedo products (Pinty et al., 2000a). The abundance of cloudy conditions occurring during a day over the Earth disk sections sensed by geostationary satellites motivates the implementation of a procedure screening conditions that do not correspond to clear-sky cases. Pinty et al. (2000b) suggested adoption of an angular consistency check by which the daily top of the atmosphere radiance series for each individual pixel is used in an attempt to fit the MRPV model. This is based on a recursive filtering technique which identifies sequentially during the day, the outliers deviating significantly from the fitted MRPV model solutions. In the vast majority of the cases, larger (lower) radiance values than the MRPV fitted solutions are associated with cloudy (shadowed) conditions and the spatio-temporal fields of these outliers were shown to, indeed, display very consistent cloud fields and associated shadows along the course of the day

9 Retrieval of Surface Albedo from Satellite Sensors

233

(Pinty et al., 2000a). Measurements acquired at solar and view zenith angles larger than 70◦ or corresponding to cloud conditions are rejected. A minimum of six valid clear sky observations are necessary for the activation of the retrieval procedure. These cloudy as well as additional undesired conditions translate into incomplete geographic surface albedo map products. This caveat can be overcome by applying a time composite algorithm selecting, over a given composite time period, the particular day delivering the albedo value which is the closest to the average of the ensemble of values retrieved during that same time period. In this way, the geophysical values for each pixel can be delivered with all the relevant information used to generate them in the retrieval algorithm such as, for example, the number of observations used to perform the retrieval and estimation of the retrieval uncertainties among others. Sensors on board geostationary satellites sample the scattered solar radiance fields in a usually single, large (according to today’s standards) spectral band (see Fig. 9.10), e.g., ≈ 0.4–1.1 µm for Meteosat and GMS, 0.05–0.8 µm

Fig. 9.10 Examples of sensor spectral responses on board geostationary satellites in the solar domain (red line). The green solid line illustrates typical reflectance of green vegetation

234

C. Schaaf et al.

for GOES, which prevents us benefiting from advanced atmospheric correction algorithms based on multi-spectral information, as is possible for the MODIS and MISR sensors. Due to the fast spatio-temporal variability of the cloud, water vapor and aerosol fields, the crucial problem of accurately assessing/removing the undesired effects induced by these atmospheric components on the measured radiances remains to be solved on the sole basis of the available information gathered by the sensor or other sources of information. For those components, such as ozone, that have a non-negligible but still limited impact on the surface retrievals and whose variability remains somewhat small, the use of climatological values is generally acceptable. In most if not all cases to be addressed, the availability of a spectrally large single band only renders the partitioning between the surface and atmospheric contributions quite difficult since the scattering and absorption effects in the atmosphere are wavelength-dependent and coupled with spectrally variant surface properties. To date, the constraints imposed by operational exploitation infrastructures have not favored the processing of multi-sensor measurements assembled via data fusion procedures. This state of affairs encourages the development of albedo retrieval algorithms relying on the analysis of data (and data strings) collected by each geostationary sensor in stand alone mode. In turn, this places stringent requirements on the reliability and overall performance of the retrieval algorithm which thus, on the basis of a single spectrally large band, must be able to identify cloud occurrence and then solve, as well as possible, the coupled surface-atmosphere radiation transfer problem. In that context, multiple sensitivity test have to be conducted in order to optimize the crucial choices to be made such as, for instance, between the length of the period to perform sequential data accumulation, e.g., between a few hours and a few days, and the impacts of the assumption hindering this multi-angular data emulation procedure, e.g., no drastic changes in the geophysical system under investigation. Meteosat data processed with this algorithm have already been used in a variety of applications (see for instance, Pinty et al., 2000c; Knorr et al., 2001; Myhre et al., 2005b).

9.6 Fusion of Modern and Historical Surface Albedo Products Documenting the Earth climate and its variability requires access to reliable and accurate long time series of environmental products. Hence, archived meteorological satellite observations could contribute to the generation of climatic data records providing that (1) significant efforts are devoted to the improvement of the spectral characterization and calibration of these radiometers and (2) developing state of the art retrieval algorithms. The first point aims at reducing and controlling as much as possible the impact of measurements uncertainties on the accuracy of the retrievals while the second point similarly at model uncertainties. Previous studies have, indeed, already demonstrated the possibility to perform post-launch improvements of the radiometer characteristics (e.g., Govaerts, 1999). In the specific case of surface albedo retrieval, the science context set up by the requirements in the exploitation

9 Retrieval of Surface Albedo from Satellite Sensors

235

of data acquired by modern sensors such as those on board Terra and Meteosat Second Generation for instance has motivated a large number of studies addressing these points. The associated improved knowledge translates, in turn, into the development and use of better approaches for the optimal exploitation of measurements taken by “old generation” sensors. While active efforts are underway within the NASA community to couple MODIS observations to the historical AVHRR archives (Pedelty et al., 2007) considerable progress has already been made in exploiting historical geostationary satellite data to establish a historical data set of global surface albedo values. This requires the back processing and analysis of measurements assembled by the fleet of geostationary satellites over the past 25 years or so. The quality of these retrievals can be assessed by various means including first, the comparison of these retrieved albedo values against those operationally generated since year 2000 from modern and technologically advanced instruments such as MODIS and MISR (Pinty et al., 2004) and second, the intercomparison of surface albedo products generated over geographical regions of overlap that are, therefore, sampled simultaneously by two adjacent sensors together placed on geostationary orbit but located at different longitudes (Govaerts et al., 2004). In order to conduct comparison exercises of relevance for climate model applications, surface albedo products have to be made available over large spectral regions covering the energetically relevant solar domain [0.3–3.0 µm] split, whenever possible, into its broad visible [0.3–0.7 µm] and near-infrared [0.7–3.0 µm] parts. Achieving this step, usually called spectral conversion, requires developing appropriate tools to transform albedo product values, estimated over and weighted by the spectral response of the geostationary sensor, into values representative of the desired broad spectral range of interest (see for instance Liang (2000) and Govaerts et al. (2006)). One possible solution consists in approximating a parametric expression relating the measurements from the sensor to those that would be provided by an ideal rectangular shape sensor covering the solar domain of interest. Such an expression can be established on the basis of (1) a large number of simulations of top of atmosphere radiance fields of various geophysical situations that can be expected for the region of interest (e.g., vegetation with varying density, bare soils with different brightness, snow surfaces, coupled with a diversity of atmospheric conditions) and (2) a multi-regression analysis fitting at best the sensor-like values against those representative of the desired solar domain. This spectral conversion constitutes quite a delicate step and, as a matter of fact, its reliability relies strongly on the degree of coincidence between the distributions of the simulated and actual conditions. Its impact on the uncertainty of the final albedo products also depends crucially on the sensor spectral response function since the spectral conversion basically assumes that strong correlations exist between radiances taken across various wavelengths of the solar spectrum. Preliminary attempts to compare surface albedo products from modern sensors, such as MODIS and MISR, against those generated by the retrieval algorithm outlined here for geostationary satellites result into quite positive conclusions. Figure 9.11 illustrates an example of results to be expected when comparing

236

C. Schaaf et al.

Fig. 9.11 Scatterplot (density plot) between the MODIS white sky albedo and Meteosat BHR values retrieved between 20–65 ◦ W longitude and 40 ◦ S − 40 ◦ N latitude during the first 2 weeks of year 2001. This plot includes the high quality flag MODIS products only and the outliers from the two distributions have been removed (less than 5% of the total number of valid retrievals). The full, dashed, and dotted lines feature the fit obtained using the slope of the means, the linear regression, and the primary eigenvector, respectively

Meteosat and MODIS (considering high quality flags only) surface albedo products, in units of white sky albedo (BHR), over a large geographical region extending from Southern Europe and covering the entire African continent (between 20◦ W–65◦ N longitude and 40◦ S–40◦ N latitude. This figure is built from the analysis of products available during the first 2-week period of year 2001 after removal of outliers detected in large majority along the coastlines. The 10-day composite Meteosat products have been re-mapped into the MODIS Climate Modeling Grid at a spatial resolution of 0.05◦ . The MODIS and Meteosat spectral albedo products were both converted into an ideal rectangular shape (0.4–1.1 µm) in order to (1) make the best possible use of the available spectral information for both sensors, i.e., one large band in the Meteosat case and four narrow bands well distributed over this spectral interval in the MODIS case, and (2) minimize the uncertainties associated with the required spectral conversion. All three indicators used to characterize the statistical differences between the MODIS and Meteosat products, i.e., the slope of

9 Retrieval of Surface Albedo from Satellite Sensors

237

the means (full line), the regression line (dashed line) and the primary eigenvector (dotted line), show limited variability around the one-to-one line. The observed statistical differences are largely within the range of the systematic error/uncertainty due to the calibration knowledge (about 6% for Meteosat-7) (Govaerts et al., 2004) and/or model approximations (for instance, the decoupling of the gaseous absorption from the aerosol scattering effects). Figure 9.12 illustrates results from a comparison between the broadband black sky (DHR) albedo products generated, at 30◦ solar angle, by the analysis of the daily radiances collected during the first 10-day period of May 2001 by both GOES West (GOES-10) and GOES East (GOES-8) over the common land regions of the Earth disk that they jointly sample (see top panel in Fig. 9.12). These results illustrate the robustness of the albedo retrieval algorithm since its application yields differences between data sets from adjacent sensors that are well within the range of their estimated uncertainty level (about 10–15% for the GOES sensors) of their broadband albedo products. Figure 9.12 confirms earlier comparison results obtained by analyzing albedo products generated by two adjacent Meteosat sensors (Govaerts et al., 2004). The conclusions drawn from Figs. 9.11 and 9.12 suggest the generation of historical series of surface albedo products based on the exploitation of the fleet of geostationary satellites. Figure 9.13 is a demonstration example of an output from such an initiative which assembles broadband products (DHR at 30◦ Sun zenith angle) retrieved for the first 10-day period of May 2001 from five different satellites, namely GOES West and East, Meteosat-7 and Meteosat-5, and GMS-5. Table 9.1 provides statistics about these retrievals and the estimated uncertainties associated with each satellite retrievals. The measurement error includes both the radiometric uncertainties and approximations in the forward model. The estimated error on the DHR values is then derived from the uncertainty on the retrieved surface parameters. It thus looks feasible to build global albedo products for the last 25 years or so, for those places covered by archived data. These preliminary results open new avenues for the exploitation of geostationary archive data and prototype the fusion of such the generated products.

9.7 Summary The MODIS, MISR and Meteosat algorithms represent three complementary strategies for characterizing land surface reflectance anisotropy and obtaining measures of land surface albedo. Each algorithm makes use of the unique capabilities of its sensor to capture the spectral, spatial, temporal, and angular information necessary to accurately specify the reflective qualities of the underlying surface cover. With more than 6 years of MODIS and MISR observations now available, as well as the opportunity to utilize the historical geosynchronous satellite record, the modeling and data analysis communities enjoy unprecedented access to consistent, high-quality albedo and anisotropy data of the Earth’s land surface.

238

C. Schaaf et al.

Fig. 9.12 Comparison between broadband DHR (30◦ ) values retrieved over the common geographical region covered by both GOES West (GOES-10) and GOES East (GOES-8). The top panel corresponds to the density plot of the two DHR distributions for the period 1–10 May 2001. The bottom panel documents the histogram of the relative differences between the two distributions. The vertical lines colored in blue feature the mean value of these differences (dash-dotted) and one standard deviation from the mean (dashed)

9 Retrieval of Surface Albedo from Satellite Sensors

239

Fig. 9.13 Top panel: Location of operational geostationary satellites which archive data and are used to derived products shown in the bottom panel. The circles show the 60◦ viewing angle limit. Bottom panel: Illustration of the broadband surface albedo map derived from the application of the geostationary satellite retrieval algorithm on measurements taken simultaneously by GOES-8/10, Meteosat-5/-7 and GMS-5 over the period 1–10 May 2001 Table 9.1 Number of days processed during the 1–10 May 2001 period for each satellite. < Img/day > is the mean number of measurements available per day (note that some of the GOES images did not provide the nominal geographical coverage). <Meas. R. Err.> is the average measurement relative error, i.e., including both the radiometric error and forward model uncertainty. is the mean estimated DHR relative error Satellite GOES-10 GOES-8 MET-7 MET-5 GMS-5

Nbr days

<Meas. R. Err.>

10 10 10 10 10

22.9 13.7 17.3 16.3 9.9

5.2% 6.8% 7.4% 10.0% 8.4%

12.5% 14.4% 8.7% 10.1% 10.5%

240

C. Schaaf et al.

Acknowledgements Authors Schaaf, Strahler, and Liu are supported by NASA (under grant NNG04HZ14) and by their colleagues on the MODIS Science Team while Martonchik is supported by the MISR Science Team. Authors Pinty, Govaerts, Lattanzio, and Taberner are grateful to the Japan Meteorological Agency (JMA) and the Satellite Services Group of the National Oceanic and Atmospheric Administration (NOAA) for providing the GMS-5 and GOES-8/-10 data, respectively. Their contributions would not have been possible without the support of the Global Environment Monitoring unit of the Institute for Environment and Sustainability at the Joint Research Centre, and EUMETSAT.

References Barlage M, Zeng X, Mitchell K, Wei H (2005) A global 0.05-deg maximum albedo dataset of snow-covered land based on MODIS observations. Geophys. Res. Lett. 32, doi:10.1029/2005GL022881 Barnsley MJ, Strahler AH, Morris KP, Muller J-P (1994) Sampling the surface bidirectional reflectance distribution function (BRDF), 1, evaluation of current and future satellite sensors. Remote Sens. Rev. 8:271–311 Berk A, Bernstein LS, Anderson GP, Acharya PK, Robertson DC, Chetwynd JH, Adler-Golden SM (1998) MODTRAN cloud and multiple scattering upgrades with application to AVIRIS. Remote Sens. Environ. 65:367–375 Bicheron P, Leroy M (2000) Bidirectional reflectance distribution function signatures of major biomes observed from space. J. Geophys. Res. 105:26669–26681 Bonan GB, et al. (2002) The land surface climatology of the NCAR community land model coupled to the NCAR community ClimateModel. J. Clim. 15:3123–3149 Br´eon FM, Maignan F, Leroy M, Grant I (2002) Analysis of hot spot directional signatures measured from space. J. Geophys. Res. 107(16):4282–4296 Collins JB, Woodcock CE (2000) Geostatistical methods for analysis of multiscale variation in spatial data. Geographical Analysis 32(1):50–63 Csiszar, I, Gutman, G (1999) Mapping global land surface albedo from NOAA/AVHRR. J. Geophys. Res. 104:6215–6228 d’Entremont R, Schaaf CB, Lucht W, Strahler A (1999) Retrieval of red spectral albedo and bidirectional reflectance using AVHRR HRPT and GOES satellite observations of New England. J. Geophys. Res. 104:6229–6239 Dickinson RE (1983) Land surface processes and climate-surface albedos and energy balance. Adv. Geophys. 25:305–353 Dickinson RE (1995) Land processes in climate models. Remote Sens. Environ. 51:27–38 Diner DJ, Beckert JC, Reilly TH, Bruegge CJ, Conel JE, Kahn RA, Martonchik JV, Ackerman TP, Davies R, Gerstl SAW, Gordon HR, Muller J-P, Myneni RB, Sellers PJ, Pinty B, Verstraete MM (1998) Multi-angle Imaging SpectroRadiometer (MISR) instrument description and experiment overview. IEEE Trans. Geosci. Remote Sen. 36:1072–1087 Diner DJ, Abdou WA, Bruegge CJ, Conel JE, Crean KA, Gaitley BJ, Helmlinger MC, Kahn RA, Martonchik JV, Pilorz SH, Holben BN (2001) MISR aerosol optical depth retrievals over southern Africa during the SAFARI-2000 dry season campaign. Geophys. Res. Lett. 28:3127–3130 Engelsen O, Pinty B, Verstraete M M, Martonchik JV (1996) Parameteric bidirectional reflectance fractor models: evaluation, improvements, and applications. Technical Report EUR 16426 EN. EC Joint Research Centre Friedl MA, McIver DK, Hodges JCF, Zhang X, Muchoney D, Strahler AH, Woodcock CE, Gopal S, Schnieder A, Cooper A, Baccini A, Gao F, Schaaf C (2002) Global land cover from MODIS: algorithms and early results. Remote Sens. Environ. 83:287–302 Gao F, Schaaf C, Strahler AH, Lucht W (2001) Using a multi-kernel least variance approach to retrieve and evaluate albedo from limited BRDF observations. Remote Sens. Environ. 76:57–66

9 Retrieval of Surface Albedo from Satellite Sensors

241

Gao F, Schaaf C, Strahler A, Roesch A, Lucht W, Dickinson R (2005) The MODIS BRDF/Albedo Climate Modeling Grid Products and the variability of albedo for major global vegetation types. J. Geophys. Res. 110: D01104, doi:10.1029/2004JD00519 Govaerts Y (1999) Correction of the Meteosat-5 and -6 VIS vand relative spectral response with Meteosat-7 characteristics. Int. J. Remote Sens. 20:3677–3682 Govaerts Y, Clerici M, Clerbaux N (2004) Operational calibration of the Meteosat radiometer VIS band. IEEE Trans. Geosci. Remote Sens. 43:31900–31914 Govaerts Y, Lattanzio A, Pinty B, Schmertz J (2004) Consistent surface albedo retrieved from two adjacent geostationary satellite. Geophys. Res. Lett. 31:L15201, doi:10.1029/2004GL020418 03 Govaerts Y, Pinty B, Taberner M, Lattanzio A (2006) Spectral conversion of surface albedo derived from Meteosat first generation observations. IEEE Geosci. Remote Sens. Lett. 3:23–27, doi:10.1109/LGRS.2005.854202 Henderson-Sellers A, Wilson MF (1983) Surface albedo data for climatic modeling. Rev. Geophys. Space Phys. 21:1743–1778 Holben BN, Eck TF, Slutsker I, Tanre D, Buis JB, Setzer A, Vermote E, Reagan JA, Kaufman YJ, Nakajima T, Lavenu F, Jankowiak I, Smirnov A (1998) AERONET – A federated instrument network and data archive for aerosol characterization. Remote Sens. Environ. 66:1–16 Hautecoeur O, Leroy MM (1998) Surface bi-directional reflectance distribution function observed at global scale by POLDER/ADEOS. Geophys. Res. Lett. 22:4197–4200 Jin Y, Schaaf CB, Woodcock CE, Gao F, Li X, Strahler AH, Lucht W, Liang S (2003a) Consistency of MODIS surface BRDF/Albedo retrievals: 1. Algorithm performance. J. Geophys. Res. 108(D5):4158, doi:10.1029/2002JD002803 Jin Y, Schaaf CB, Woodcock CE, Gao F, Li X, Strahler AH, Lucht W, Liang S (2003b) Consistency of MODIS surface BRDF/Albedo retrievals: 2. Validation. J. Geophys. Res. 108(D5):4159, doi:10.1029/2002JD002804 Knorr W, Schnitzler KG, Govaerts Y (2001) The role of bright desert regions in shaping North African climate. Geophys. Res. Let. 28:3489–3492 Lattanzio A, Govaerts Y, Pinty B (2006) Consistency of surface anisotropy characterization with Meteosate observations. Adv. Space Res. doi:10.1016/j.asr.2006.02.049 Leroy M, Deuze JL, Breon FM, Hautecoeur O, Herman M, Buriez JC, Tanre D, Bouffies S, Chazette P, Roujean J-L (1997) Retrieval of atmospheric properties and surface bidirectional reflectances over land from POLDER/ADEOS. J. Geophys. Res. 102:17023–17037 Li X, Strahler AH (1992) Geometric-optical bidirectional reflectance modeling of the discrete crown vegetation canopy: effect of crown shape and mutual shadowing. IEEE Trans. Geosci. Remote Sens. 30:276–292 Li Z, Garand L (1994) Estimation of surface albedo from space: a parameterization for global application. J. Geophys. Res. 99:8335–8350 Liang S (2000) Narrowband to broadband conversions of land surface albedo I: algorithms. Remote Sens. Environ. 76:213–238 Liang S, Strahler AH, Walthall CW (1999) Retrieval of land surface albedo from satellite observations: a simulation study. J. Appl. Meteorol. 38:712–725 Lucht W, Schaaf CB, Strahler AH (2000) An algorithm for the retrieval of albedo from space using semiempirical BRDF models. IEEE Trans. Geosci. Remote Sens. 38:977–998 Maignan F, Br´eon FM, Lacaze R (2004) Bidirectional reflectance of Earth targets: evaluation of analytical models using a large set of spaceborne measurements with emphasis with the hot spot. Remote Sens. Environ. 90:210–220 Martonchik JV (1997) Determination of aerosol optical depth and land surface directional reflectances using multi-angle imagery. J. Geophys. Res., 102:17015–17022. Martonchik JV, Diner DJ, Kahn RA, Ackerman TP, Verstraete MM, Pinty B, Gordon HR (1998a) Techniques for the retrieval of aerosol properties over land and ocean using multiangleimaging. IEEE Trans. Geosci. Remote Sens. 36:1212–1227 Martonchik JV, Diner DJ, Pinty B, Verstraete MM, Myneni RB, Knyazikhin Y, Gordon HR (1998b) Determination of land and ocean reflective, radiative, and biophysical properties using multiangle imaging. IEEE Trans. Geosci. Remote Sens. 36:1266–1281

242

C. Schaaf et al.

Martonchik JV, Diner DJ, Crean KA, Bull M (2002a) Regional aerosol retrieval results from MISR. IEEE Trans. Geosci. Remote Sens. 40:1520–1531 Martonchik JV, Pinty B, Verstraete MM (2002b) Note on an improved model of surface BRDFatmospheric coupled radiation. IEEE. Trans. Geosci. Remote Sens. 40:1637–1639 Moody EG, King MD, Platnick S, Schaaf CB, Gao F (2005) Spatially complete global spectral surface albedos: value-added datasets derived from Terra MODIS land products. IEEE Trans. Geosci. Remote Sens. 43:144–158 Myhre G, Kvalevag MM, Schaaf CB (2005a) Radiative forcing due to anthropogenic vegetation change based on MODIS surface albedo data set. Geophys. Res. Lett. 32:L21410, doi:10.1029/2005GL024004 Myhre G, Govaerts Y, Haywood JM, Berntsen TK, Lattanzio A (2005b) Radiative effect of surface albedo change from biomass burning. Geophys. Res. Lett. 32:L20812, doi:10.1029/2005GL022897 Oleson KW, Bonan GB, Schaaf C, Gao F, Jin Y, Strahler A (2003) Assessment of global climate model land surface albedo using MODIS data. Geophys. Res. Lett. 30(8):1443, doi:10.1029/2002GL016749 Pedelty J, Devadiga S, Masuoka E, Brown M, Pinzon J, Tucker C, Vermote E, Prince S, Nagol J, Justice C, Roy D, Ju J, Schaaf C, Liu J, Privette J, Pinheiro, A (2007) Generating a Long-term Land Data Record from the AVHRR and MODIS Instruments, Proceedings, IEEE International Geosciences and Remote Sensing Symposium (IGARSS07), Barcelona, Spain, 23–27 July, 2007 Pinty B, Roveda F, Verstraete MM, Gobron N, Govaerts Y, Martonchik J, Diner D, Kahn R (2000a) Surface albedo retrieval from METEOSAT – Part 1: theory. J. Geophys. Res. 105:18099–18112 Pinty B, Roveda F, Verstraete MM, Gobron N, Govaerts Y, Martonchik J, Diner D, Kahn R (2000b) Surface albedo retrieval from METEOSAT – Part 2. Appl. J. Geophys. Res. 105:18113–18134 Pinty B, Verstraete MM, Gobron N, Roveda F, Govaerts Y (2000c) Do manmade fires affect the Earth’s surface reflectance at continental scales? Eos Trans. American Geophys. Union 81:388–389 Pinty B, Taberner M, Liang S, Govaerts Y, Martonchik JV, Lattanzio A, Schaaf CB, Verstraete MM, Dickinson RE, Gobron N, Widlowski J-L (2004) Intercomparison of surface albedo products from various spaceborne sensors. In: Proc. of the Workshop on Inter-Comparison of Large Scale Optical and Infrared Sensors, ESA ESTEC, Noordwijk, The Netherlands, 12–14 October 2004. ESA ESTEC Privette JL, Eck TF, Deering DW (1997) Estimating spectral albedo and nadir reflectance through inversion of simple BRDF models with AVHRR/MODIS-like data. J. Geophys. Res. 102:29529–29542 Rahman H, Pinty B, Verstraete MM (1993) Coupled surface-atmosphere reflectance (CSAR) model – 2: semiempirical surface model usable with NOAA advanced very high resolution radiometer data. J. Geophys. Res. 98:20791–20801 Roesch A, Schaaf C, Gao F (2004) Use of Moderate-Resolution Imaging Spectroradiometer bidirectional reflectance distribution function products to enhance simulated surface albedos. J. Geophys. Res. 109(D12), doi:10.1029/2004JD004552 Ross JK(1981) The Radiation Regime and Architecture of Plant Stands,Junk W (ed), Norwell, MA: Artech House, p. 392 Roujean J-L, Leroy M, Deschamps PY (1992) A bidirectional reflectance model of the Earth’s surface for the correction of remote sensing data. J. Geophys. Res. 97:20455–20468 Salomon J, Schaaf CB, Strahler AH, Gao F, Jin Y (2006) Validation of the MODIS Bidirectional Reflectance Distribution Function and albedo retrievals using combined observations from the aqua and terra platforms. IEEE Trans. Geosci. Remote Sens. 44:No. 6 Sellers PJ, Los SO, Tucker CJ, Justice CO, Dazlich DA, Collatz CJ, Randall DA (1996) A revised land surface parameterization (SiB2) for atmospheric GCMs, part II, the generation of global fields of terrestrial biospheric parameters from satellite data. J. Clim. 9:706–737

9 Retrieval of Surface Albedo from Satellite Sensors

243

Schaaf CB, Gao F, Strahler AH, Lucht W, Li X, Tsang T, Strugnell NC, Zhang X, Jin Y, Muller J-P, Lewis P, Barnsley M, Hobson P, Disney M, Roberts G, Dunderdale M, Doll C, d’Entremont R, Hu B, Liang S, Privette JL (2002) First operational BRDF, Albedo and Nadir Reflectance Products from MODIS. Remote Sens. Environ. 83:135–148 Strugnell N, Lucht W (2001) An algorithm to infer continental-scale albedo from AVHRR data, land cover class and field observations of typical BRDFs. J. Climate 14:1360–1376 Strugnell N, Lucht W, Schaaf C (2001) A global albedo data set derived from AVHRR data for use in climate simulations. Geophys. Res. Lett. 28:191–194 Tian Y, Dickinson RE, Zhou L, Myneni RB, Friedl M, Schaaf CB, Carroll M, Gao F (2004) Land boundary conditions from MODIS data and consequences for the albedo of a climate model. Geophys. Res. Lett. 31, doi:10.1029/2003GL019104 Vermote EF, Tanre D, Deuze JL, Herman M, Morcrette JJ (1997) Second simulation of the satellite signal in the solar spectrum: an overview. IEEE Trans. Geosci. Remote Sens. 35:675–686 Walthall CL, Norman JM, Welles JM, Campbell G, Blad BL (1985) Simple equation to approximate the bidirectional reflectance from vegetation canopies and bare soil surfaces. Appl. Opt. 24:383–387 Wanner W, Li X, Strahler AH (1995) On the derivation of kernels for kernel-driven models of bidirectional reflectance. J. Geophys. Res. 100:21077–21090 Wanner W, Strahler AH, Hu B, Lewis P, Muller J-P, Li X, Barker Schaaf CL, Barnsley MJ (1997) Global retrieval of bidirectional reflectance and albedo over land from EOS MODIS and MISR data: theory and algorithm. J. Geophys. Res. 102:17143–17162 Woodcock CE, Harward VJ (1992) Nested-Hierarchical Scene Models and image segmentation. Int. J. Remote Sens. 13(16):3167–3187 Zhang X, Friedl MA, Schaaf CB, Strahler AH, Hodges JCF, Gao F, Reed BC, Huete A (2003) Monitoring vegetation phenology using MODIS. Remote Sens. Environ. 84:471–475 Zhou L, Dickinson RE, Tian Y, Zeng X, Dai Y, Yang Z-L, Schaaf CB, Gao F, Jin Y, Strahler A, Myneni RB, Yu H, Wu W, Shaikh M (2003) Comparison of seasonal and spatial variations of albedos from Moderate-Resolution Imaging Spectroradiometer (MODIS) and Common Land Model. J. Geophys. Res. 108(D15):4488, doi:10.1029/2002JD003326

Chapter 10

Modeling and Inversion in Thermal Infrared Remote Sensing over Vegetated Land Surfaces Fr´ed´eric Jacob, Thomas Schmugge, Albert Olioso, Andrew French, Dominique Courault, Kenta Ogawa, Francois Petitcolin, Ghani Chehbouni, Ana Pinheiro, and Jeffrey Privette

Fr´ed´eric Jacob Formerly at Remote Sensing and Land Management Laboratory Purpan Graduate School of Agriculture, Toulouse, France Now at Institute of Research for the Development Laboratory for studies on Interactions between Soils – Agrosystems – Hydrosystems UMR LISAH SupAgro/INRA/IRD, Montpellier, France [email protected] Thomas Schmugge Gerald Thomas Professor of Water Resources College of Agriculture New Mexico State University, Las Cruces, NM, USA Albert Olioso and Dominique Courault National Institute for Agronomical Research Climate – Soil – Environment Unit UMR CSE INRA/UAPV, Avignon, France Andrew French United States Department of Agriculture/Agricultural Research Service US Arid Land Agricultural Research Center, Maricopa, AZ, USA Kenta Ogawa Department of Geo-system Engineering, University of Tokyo Japan Francois Petitcolin ACRI-ST, Sophia Antipolis, France Ghani Chehbouni Institute of Research for the Development Center for Spatial Studies of the Biosphere UMR CESBio CNES/CNRS/UPS/IRD, Toulouse, France Ana Pinheiro Biospheric Sciences Branch, NASA’s GSFC, Greenbelt, MD, USA Jeffrey Privette NOAA’s National Climatic Data Center, Asheville, NC, USA S. Liang (ed.), Advances in Land Remote Sensing, 245–291. c Springer Science + Business Media B.V., 2008

245

246

F. Jacob et al.

Abstract Thermal Infra Red (TIR) Remote sensing allows spatializing various land surface temperatures: ensemble brightness, radiometric and aerodynamic temperatures, soil and vegetation temperatures optionally sunlit and shaded, and canopy temperature profile. These are of interest for monitoring vegetated land surface processes: heat and mass exchanges, soil respiration and vegetation physiological activity. TIR remote sensors collect information according to spectral, directional, temporal and spatial dimensions. Inferring temperatures from measurements relies on developing and inverting modeling tools. Simple radiative transfer equations directly link measurements and variables of interest, and can be analytically inverted. Simulation models allow linking radiative regime to measurements. They require indirect inversions by minimizing differences between simulations and observations, or by calibrating simple equations and inductive learning methods. In both cases, inversion consists of solving an ill-posed problem, with several parameters to be constrained from few information. Brightness and radiometric temperatures have been inferred by inverting simulation models and simple radiative transfer equations, designed for atmosphere and land surfaces. Obtained accuracies suggest refining the use of spectral and temporal information, rather than innovative approaches. Forthcoming challenge is recovering more elaborated temperatures. Soil and vegetation components can replace aerodynamic temperature, which retrieval seems almost impossible. They can be inferred using multiangular measurements, via simple radiative transfer equations previously parameterized from simulation models. Retrieving sunlit and shaded components or canopy temperature profile requires inverting simulation models. Then, additional difficulties are the influence of thermal regime, and the limitations of spaceborne observations which have to be along track due to the temperature fluctuations. Finally, forefront investigations focus on adequately using TIR information with various spatial resolutions and temporal samplings, to monitor the considered processes with adequate spatial and temporal scales.

10.1 Introduction Using TIR remote sensing for environmental issues have been investigated the last three decades. This is motivated by the potential of the spatialized information for documenting the considered processes within and between the Earth system components: cryosphere [1–2], atmosphere [3–6], oceans [7–9], and land surfaces [10]. For the latter, TIR remote sensing is used to monitor forested areas [11–14], urban areas [15–17], and vegetated areas. We focus here on vegetated areas, natural and cultivated. The monitored processes are related to climatology, meteorology, hydrology and agronomy: (1) radiation, heat and water transfers at the soil–vegetation–atmosphere interface [18–24]; (2) interactions between land surface and atmospheric boundary layer [25]; (3) vegetation physiological processes such as transpiration and water consumption, photosynthetic activity and CO2 uptake, vegetation growth and biomass production [26–39]; (4) soil processes such as respiration

10 Modeling and Inversion in Thermal Infrared Remote Sensing

247

and CO2 uptake, evapotranspiration and water depletion, spatio-temporal variability of soil moisture [39–43]; (5) long-term dynamics of land cover [44], land surface radiative budget [45–48], water shortage and drought [49]. TIR remote sensing allow retrieving emissivity and temperature, with various complexity degrees presented in Section 10.2. The remotely sensed information is collected from operational and prospective sensors, listed in Section 10.3. This information is characterized by temporal and spatial dimensions (Section 10.3.1), as well as by spectral and directional dimensions (Section 10.3.2). Then, inferring emissivity and temperature consists of developing and inverting modeling tools, by exploiting the dimensions of the collected information (Section 10.4). Based on TIR fundamentals (Section 10.4.1), simple radiative transfer equations directly link measurements to emissivities and temperatures of interest (Section 10.4.2), and simulation models describe the influence of radiative regime on measurements (Section 10.4.3). However, simple radiative transfer equations must be parameterized, and simulation models require significant information. Further, inversion is not trivial: most of simulation models are not directly invertible, and the numerous parameters to be constrained from remote sensing often make inversion an illposed problem (Section 10.4.4). The several solutions proposed to overcome these difficulties are assessed using validations, intercomparisons, and sensitivity studies (Section 10.5). Current limitations and proposed solutions are presented with an increasing complexity for the temperatures of interest (Section 10.6). Atmospheric perturbations are corrected by inverting modeling tools for atmosphere, and surface brightness temperature measurements are simulated using modeling tools for land surfaces (Section 10.6.1). Surface emissivity effects are removed using simple radiative transfer equations (Section 10.6.2). Reported performances suggest accuracies rather close to requirements, though refinements are necessary. Recovering temperature for the one source modeling of heat transfers is still not trivial, since the required parameterization significantly varies in time and space (Section 10.6.3). Recent studies suggested focusing on more elaborated temperatures: soil and vegetation components, optionally sunlit and shaded, and canopy temperature profile. Their retrieval is a forthcoming challenge, with efforts on measuring, modeling and inversion (Section 10.6.4). The paper ends with forefront investigations about space and time issues in TIR remote sensing: monitoring land processes with adequate spatial scales and temporal samplings, by using available remote sensing observations (Section 10.7).

10.2 Land Surface Emissivity/Temperature from TIR Remote Sensing This section defines the various terms considered in TIR remote sensing, which are related to land surface emissivity and temperature. We focus on their physical definitions and various interests. The corresponding equations are detailed in Section 10.4.

248

F. Jacob et al.

– Surface brightness temperature is equivalent to the radiance outgoing from the target, by assuming a unity emissivity [50], and corresponds to the basic TIR remote sensing measurement. It is recovered from at sensor measurements performing atmospheric corrections. It can be assimilated, using modeling tools for land surface, into process models such as SVAT and crop models [18, 38, 39, 43]. – Ensemble waveband emissivity is needed to derive radiometric temperature from brightness temperature [50, 51]. It is also useful for retrieving ensemble broadband emissivity, a key parameter for land surface radiative budget [52–54]. – Ensemble radiometric temperature is emissivity normalized [50, 51]; and corresponds to kinetic temperature for an homogeneous and isothermal surface [55]. It is used to estimate surface energy fluxes and water status from spatial variability indicators: the vegetation index / temperature triangle [41, 56–58]; or the albedo / temperature diagram [23, 37, 59, 60]. It is also used for retrieving soil and vegetation temperatures from two source energy balance modeling [19, 24]. – Aerodynamic temperature is air temperature at the thermal roughness length [50]. It is the physical temperature to be used with one source models of surface energy fluxes based on excess resistance [61–63]. These can be SVAT models [39, 64]; or energy balance models [22, 23, 37, 59, 60, 65, 66]. – Soil and vegetation temperatures correspond to kinetic [67] or radiometric [68] temperatures. They are often used for two-source modeling. The latter can be SVAT models [43, 67, 69]; or energy balance models [20, 70, 71]. Retrieving these temperatures requires an adequate estimation of directional ensemble emissivity. – Sunlit and shaded components are refinements of soil and vegetation temperatures. They can significantly differ, according to various factors which drive the thermal regime: the water status, the solar exposure resulting from the canopy geometry and the illumination direction. These components are of interest for understanding canopy directional brightness and radiometric temperatures [58, 72–74]. – Canopy temperature profile, from the soil surface to the top of canopy, is the finest temperature one can consider. Similarly to sunlit and shaded components for soil and vegetation temperatures, this thermal regime is considered for understanding canopy directional brightness and radiometric temperatures, in relation with local energy balance within the canopy [75–78]. The seek accuracies vary from one application to another,according to the sensitivities of process models. For temperature, the goal is accuracy better than 1 K [79]. For emissivity, the goal is absolute accuracy better than 0.01 [80]. Recovering both relies on exploiting the dimensions of the TIR remotely sensed information.

10.3 Available Information from TIR Remote Sensing The four dimensions of the remotely sensed information are temporal and spatial (Section 10.3.1), and spectral and directional (Section 10.3.2). Due to orbital

10 Modeling and Inversion in Thermal Infrared Remote Sensing

249

rules and technological limitations, current spaceborne sensors cannot provide full information over these dimensions. Further, the latter can be linked, according to the mission objectives: a daily monitoring with sun-synchronous sensor requires a kilometric resolution with an across track angular sampling. Exploratory missions with airborne and ground-based sensors are under progress, for assessing the potential of original remotely sensed information. Table 10.1 provides an overview of the main operational and prospective sensors. We deal here with recent, current and forthcoming US and EU missions.

10.3.1 Temporal and Spatial Capabilities The temporal dimension corresponds to the time interval between consecutive observations. It is of importance for monitoring land surface temperature and related processes: radiative and convective transfers, soil respiration and vegetation physiological activity. The spatial dimension corresponds to the ground resolution of the measurements. It is of importance for the meaning of surface temperature collected over kilometric size pixels which include different land units. Both dimensions are strongly correlated for current TIR spaceborne sensors: high temporal samplings for finer monitoring correspond to coarse spatial resolutions with larger heterogeneity effects, and reversely. The highest temporal samplings are provided by geostationary sensors: 15–30 min with GOES Imager [81] and MSG/SEVIRI [82], corresponding to ground resolutions between 2 and 4 km. Intermediate scales correspond to kilometric resolution sensors onboard sun-synchronous platforms, providing daily nighttime and daytime observations: NOAA/AVHRR [83], ADEOS/GLI [84], and Terra-Aqua/MODIS [85]. A 3 day temporal sampling with a 1 km resolution has been provided by ERS/ATSR-1 and -2, and ENVISAT/AATSR [86]. The highest spatial resolutions are 60 and 120 m from Landsat/TM & ETM [87], and 90 m from Terra/ASTER [88]; with 16-day temporal samplings. ASTER and Landsat/ETM missions have limited lifetimes, with currently no follow on TIR high spatial resolution missions from space. Regarding current possibilities, new spaceborne sensors are demanded, to monitor land processes with adequate temporal and spatial scales. Past missions IRSUTE and SEXTET proposed 40–60 m spatial resolutions with a 1-day revisit [89, 90] and SPECTRA proposed 50 m with 3 days [91]. MTI mission offers a 20 m resolution with a 7-day revisit [92], but the military context restricts the data access. Airborne prospective observations have allowed studying temporal and spatial issues, with metric resolutions and adjustable revisits. Let us cite the airborne missions TIMS [93], DAIS [94], MAS [95] and MASTER [96]; and the airborne-based ReSeDA program [97–99].

250

F. Jacob et al.

Table 10.1 Nominal characteristics for operational and prospective sensors; in relation with recent, current and forthcoming US and EU missions. VZA means View Zenith Angle, VAA means View Azimuth Angle. Across (respectively along) track means viewing directions in a plan perpendicular (respectively parallel) to the satellite path. Sensor

Daytime Spatial Spectral sampling resolution features

Directional features

Spaceborne MSG SEVIRI

15 mn

3 km

GOES 10 and 12 Imager

30 mn

2–4 km

NOAA 15–17 AVHRR / 3

1 day

Terra-Aqua MODIS

1 MIR: 3.9 µm 1 latitude5 TIR: 8.7, 9.7, 10.8, 12, 13.4 µm dependent VZA 1 MIR: 3.7 µm 2 TIR: 10.8, 12 µm

1 latitudedependent VZA

1 km

1 MIR: 3.8 µm 2 TIR: 11, 12 µm

Across track VZA: ±55◦

1 day

1 km

3 MIR: 3.8, 3.95, 4.1 µm 3 TIR: 8.6, 11, 12 µm

Across track VZA: ±55◦

ADEOS GLI

1 day

1 km

1 MIR: 3.7 µm 3 TIR: 8.6, 10.8, 12 µm

Across track VZA: ±40◦

ERS-ATSR 1 and 2 ENVISAT-AATSR

3 days

1–2 km

1 MIR: 3.7 µm 2 TIR: 10.8, 12 µm

Along track VZA: 0, 55◦

Landsat 5–7 TM and ETM

16 days

120 m

1 TIR: 11.5 µm

Close nadir VZA

Terra ASTER

16 days

90 m

5 TIR: 8.3, 8.6, 9.1, 10.7, 11.3 µm

Close nadir VZA

TIMS (multispectral)

-

1–5 m

6 TIR: 8.4, 8.8, 9.2, . . . . . . 9.9, 10.7, 11.7 µm

Across track VZA: ±38◦

DAIS (multispectral)

-

1–5 m

6 TIR: 8.7, 9.7, 10.5, . . . . . . 11.4, 12.0, 12.7 µm

Across track VZA: ±26◦

MAS / MASTER (multispectral)

-

1–5 m

10 TIR: 7.8, 8.2, 8.6, 9.1, 9.7, . . . Across track . . . 10.1, 10.6, 11.3, 12.1, 12.9 µm VZA: ±40◦

SEBASS (hyperspectral)

-

1–5 m

MIR: [2.5–5.3] µm TIR: [7.6–13.5] µm Spectral resolution > 0.1 µm

Close nadir VZA

Two temperature Box method

-

∼50 cm

1 broadband over [8–13] µm

Nadir VZA

Hyperspectral FTIR BOMEM suite

-

Few cm

Optical spectral range: [2–20] µm Nadir VZA Spectral resolution: 1 cm−1

Goniometric systems

-

Few cm

1 broadband over [8–13] µm

Airborne

Ground-based

VZA ∈ [0–90◦ ] VAA ∈ [0–360◦ ]

10 Modeling and Inversion in Thermal Infrared Remote Sensing

251

10.3.2 Spectral and Directional Capabilities The spectral dimension corresponds to the number and location of sensor wavebands within the TIR and optionally the MIR domains. The directional dimension corresponds to the number and angular distribution of viewing directions. Both dimensions are used for recovering emissivities and temperatures via modeling tools. The basic spectral configuration corresponds to TM and ETM, with 1 channel. Richer information is provided via two channels with GOES Imager, AVHRR and the ATSR suite; three channels with MODIS and GLI; and five channels with SEVIRI and ASTER. Additional MIR information can be combined with TIR information, to be used with continuous observations from geostationary sensors (SEVIRI, GOES Imager), or day night observations from sun-synchronous sensors (AVHRR, MODIS). The basic directional configuration corresponds to SEVIRI, GOES Imager, TM, ETM, and ASTER; with a single viewing direction. Richer information is collected from across track viewing with AVHRR, MODIS, GLI; and along track viewing with the ATSR suite. Across track viewing allows a daily monitoring, while sampling the angular dynamic within a given temporal window (16 days for MODIS). This is of interest for stable surface properties such as emissivity. For surface temperature which fluctuates, capturing the angular dynamic requires almost simultaneous observations. This is possible with ATSR along track bi-angular observations only, which is limited. Future spaceborne missions will pursue current ones for long-term records: the GOES suite [100], NPOESS/VIIRS following AVHRR and MODIS [101]. MTI provides original information: 2 MIR/3 TIR bands, 0◦ and 50◦ along track. At the airborne level, the spectral dimension has been investigated with multispectral (TIMS, DAIS, MAS & MASTER) and hyperspectral (SEBASS [102]) sensors, and the directional dimension has been assessed with video cameras (see [103] with the ReSeDA program). At the ground level, the spectral dimension has been explored with hyperspectral sensors (FTIR BOMEM [104]), or with broadband radiometers [105–107], and the directional dimension has been examined with goniometric systems [58, 108, 109]. In the context of monitoring land processes, the various types of information presented here are valuable for recovering land surface emissivity and temperature. Using this information requires designing modeling tools and inversion methods, either under development for prospective studies or with operational capabilities.

10.4 Developing Modeling Tools and Inversion Methods Modeling tools aim at forwardly simulating, with different complexities, measured brightness temperature from emissivities and temperatures of interest. Table 10.2 provides an overview of the modeling tools currently used. Based on TIR fundamentals (Section 10.4.1), simple radiative transfer equations directly link measurements

252

F. Jacob et al.

Table 10.2 Listing of the modeling tools currently used, with an increasing complexity. The second rightmost column gives the related medium, and the rightmost column gives the types of land surface emissivity and temperature currently investigated with each tool. The modeling of atmospheric radiative transfer is considered here in the context of performing atmospheric corrections. Modeling tools

Literature examples

Related medium

Investigated land surface temperatures and emissivities

Simple radiative transfer equations Atmospheric radiative transfer

Eq. 10.6 [81, 178]

Atmosphere

• Brightness temperature (Atmospheric corrections)

Composite surface radiative transfer

Eq. 10.7 [12, 158]

Land surface

• Ensemble emissivity and radiometric temperature

Split Window and Dual Angle

Eq. 10.9 [125, 126]

Soil and vegetation radiative transfer

Eq. 10.10 [68, 127]

Land surface

• Soil and vegetation temperatures

Kernel-driven radiative transfer

Eq. 10.11 [128, 129]

Land surface

• Ensemble emissivity • Soil and vegetation temperatures

Radiative transfer

MODTRAN [134]

Atmosphere

• Brightness temperature (Atmospheric corrections)

Radiative transfer

Prevot’s [139] SAIL [74, 137]

Land surface

• Brightness temperature • Ensemble emissivity • Soil and vegetation temperatures with sunlit and shaded components

Geometric-optics

Kimes’s [141] Caselles’s [143] Yu’s [73]

Land surface

• Brightness temperature • Soil and vegetation temperatures with sunlit and shaded components

Geometric-optics radiative transfer

CUPID [147] Thermo [148] Jia’s [149] DART [76, 77]

Land surface

• Brightness temperature • Canopy temperature profile

Monte Carlo ray tracing

[127, 150, 151]

Land surface

• Brightness temperature • Ensemble emissivity • Soil and vegetation emissivities

Atmosphere and • Ensemble radiometric temperature land surface (atmospheric corrections)

Simulation models

to emissivities and temperatures of interest (Section 10.4.2), and simulation models describe the influence of radiative regime on measurements (Section 10.4.3). Next, inversion methods aim at backwardly retrieving emissivities and temperatures of interest from measurements (Section 10.4.4).

10 Modeling and Inversion in Thermal Infrared Remote Sensing

253

10.4.1 Fundamentals in TIR Remote Sensing The use of TIR remote sensing to infer the temperatures of interest involves an aerodynamic issue for the related temperature, and a radiative issue for the other temperatures.

10.4.1.1 Aerodynamic Issue Aerodynamic temperature Taero is not radiative-based and cannot be remotely sensed. It is required for one source modeling of surface energy fluxes, since it corresponds to the value of the logarithmic-based air temperature profile Tair (z) at thermal roughness length zoh [110]. For a negligible displacement height, sensible heat flux H is expressed from the air temperature gradient between zoh and reference level zre f : H=

Tair (zoh ) − Tair (zre f ) rah (zoh , zre f )

with

Taero = Tair (zoh )

(10.1)

where rah (zoh , zre f ) is aerodynamic resistance for heat between zoh and zre f [111]. Due to larger resistance for heat transfers, zoh is lower than mechanical roughness length zom [112]. The link between both is the aerodynamic kB−1 parameter [113]:

zom kB−1 = ln (10.2) zoh The physical meanings of Taero and zoh are equivocal. Taero is an effective temperature for heat sources that are soil and vegetation [114]. zoh is an effective level for which Tair = Taero . Their retrieval from remote sensing is not trivial (Section 10.6.3). Nevertheless, Taero can be unequivocally derived from soil and vegetation temperatures Tsoil and Tveg , by merging one source and two source modeling [20, 115]:

Taero =

Tsoil ra,soil

T

Tair (zre f ) rah 1 1 ra,veg + rah

veg + ra,veg +

1 ra,soil

+

(10.3)

where ra,soil (respectively ra,veg ) is aerodynamic resistance from the soil (respectively vegetation) to zom , and rah is aerodynamic resistance from zom to zre f [111].

10.4.1.2 Radiative Issue Apart from aerodynamic temperature, the land surface temperatures inferred from TIR remote sensing are radiative-based. Then, fundamentals deal with the TIR radiative regime within atmosphere and over land surfaces. This includes three mechanisms which drive the wave matter interactions: emission, absorption, and

254

F. Jacob et al.

scattering. Emitted radiance L(λ , θ , T ) from a natural object at a kinetic temperature T is written: L(λ , θ , T ) = ε (λ , θ ) B(λ , T ) = ε (λ , θ )

π

C1 λ −5 exp TC2λ − 1

(10.4)

λ is the monochromatic wavelength. θ is the emission direction. B(λ , T ) is the blackbody emitted radiance, expressed from Planck’s Law. C1 and C2 are first and second radiative constants. Emissivity ε (λ , θ ) is the conversion factor from thermodynamic to radiative energy, lower than 1 for natural objects. This so-called e-emissivity definition is linked to emission mechanisms, since it is the ratio of the actual to the blackbody emitted radiances for the same kinetic temperature. Under local thermodynamic equilibrium, Kirchhoff’s Law assumes emissivity and absorptivity are equal. For opaque elements, emissivity is then linked to hemisphericaldirectional reflectance ρ (λ j , θ ): ε (λ j , θ ) = 1 − ρ (λ j , θ )

(10.5)

ρ (λ j , θ ) is the average of bidirectional reflectance over illumination angles [116]. This so-called r-emissivity definition is derived from Kirchhoff’s Law, and therefore linked to reflection mechanisms. Finally, emitted radiance from a given element can be reflected by other elements, inducing changes in radiation path, called scattering effects. Within the atmosphere, scattering is negligible: the radiative regime is driven by the temperature and density of absorbers and emitters (water vapor, CO2 , O3 , . . .). A clear atmosphere behaves as an horizontally homogeneous medium: the radiative regime primarily depends on vertical profiles for temperature and density of absorbers and emitters (Fig. 10.1). Over heterogeneous land surfaces with structured patterns, the radiative regime is more complex than within atmosphere: soil and vegetation act as emitters, absorbers and scatterers for canopy and atmospheric irradiances. Additional effects are surface and volume scatterings (Fig. 10.2). Surface scattering corresponds to shadowing effects for a geometric medium, with sunlit and shaded areas. Volume scattering corresponds to reflections between soil and vegetation: radiation is trapped within the canopy. TIR remotely sensed measurements result from the processes discussed above. Sensor brightness temperature is driven by vertical profiles for temperatures and densities of atmospheric constituents. Surface brightness temperature results from the radiative regime over a heterogeneous and non isothermal area. Then, emissivity and kinetic temperature are equivocal: the canopy acts as an effective medium with ensemble emissivity and radiometric temperature [50]. Besides, e- and r-emissivities differ according to vegetation amount, since spatial averaging for e-emissivity includes emitted radiance as an additional weighting factor [50, 117]. Due to its simpler formulation, r-emissivity is preferred [68, 71, 74, 118–120]. Further, the measured brightness temperature results from emission, but also from absorption and scattering of canopy and atmospheric irradiances. This induces spectral and directional variations, driven by (1) radiative properties of soil and vegetation

10 Modeling and Inversion in Thermal Infrared Remote Sensing

255

Fig. 10.1 Atmospheric TIR radiative regime for an off nadir propagation. The key processes to be considered for atmospheric corrections are emission and absorption by atmospheric constituents. Within a horizontally homogeneous atmosphere, the radiative regime depends on the vertical fields of temperature and density for emitters and absorbers. Regardless of considered layer (zi or zk ), radiative regime is driven by atmospheric absorption (1), atmospheric emission (2), and surface emission through atmosphere transmission (3). (Adapted from [264].)

(reflectance and emissivity), (2) surface scattering with sunlit and shaded areas, and (3) volume scattering with the cavity effect. These three factors induce ensemble emissivity is anisotropic, with values greater than that of vegetation as the latter quantitatively increases [118, 121, 122]. Various modeling tools have been developed to simulate sensor and surface brightness temperature measurements. The first way is using simple radiative transfer equations for directly linking measurements to emissivities and temperatures of interest. The second way is using simulation models for understanding the influence of the TIR radiative regime on the measured brightness temperature.

10.4.2 Simple Radiative Transfer Equations Simple radiative transfer equations directly link TIR measurements to emissivities and temperatures of interest. Their advantages are linearity and simplicity, but most of them are limited to homogeneous media by assuming turbidity and azimuthal isotropy.

256

F. Jacob et al.

Fig. 10.2 Surface (or geometric) and volume (or volumetric) scattering. Surface scattering induces shadowing effects with hotter and cooler elements. Volume scattering induces an increase of brightness temperature by adding a component to emission. (Adapted from [129].)

Measured brightness temperature at the sensor level Tbrs is linked to surface brightness temperature Tbs via the atmospheric radiative transfer equation: (10.6) B (λ j , Tbrs (θ , λ j )) = B (λ j , Tbs (θ , λ j )) τa (θ , λ j ) + B λ j , Tba↑ (θ , λ j )

θ is the view zenith angle. λ j is the equivalent waveband over the sensor channel j [123]. B(λ , T ) is the blackbody emitted radiance, expressed from Planck’s Law (Eq. 10.4). τa is the atmospheric transmittance, vertically integrated between the surface and the sensor. B(λ j , Tba↑ ) is the atmospheric upward radiance towards the sensor. Surface brightness temperature is expressed as the sum of canopy emission and scattering of atmospheric irradiance, via the composite surface radiative transfer equation:

10 Modeling and Inversion in Thermal Infrared Remote Sensing

257

B(λ j , Tbs (θ , λ j )) = ε (λ j , θ ) B(λ j , Trad (θ )) + (1 − ε (λ j , θ )) B(λ j , Tba↓ (λ j )) (10.7) B(λ j , Tba↓ ) is the hemispherical average of atmospheric downward radiance. ε (λ j , θ ) and Trad (θ ) are ensemble emissivity and radiometric temperature. Ensemble emissivity can be expressed from emissivities of soil εsoil (λ j ) and vegetation εveg (λ j ), with the optional inclusion of a correction term d ε for the cavity effect [124]:

ε (λ j , θ ) = Fsoil (θ ) εsoil (λ j ) + Fveg (θ ) εveg (λ j ) + 4 d ε Fveg (θ ) Fsoil (θ ) (10.8) Fsoil (θ ), Fveg (θ ) are directional gap and cover fractions, with Fsoil (θ ) = 1 − Fveg (θ ). Brightness temperature measured from space can be linked to emissivity and radiometric temperature by merging Eqs. 10.6 and 10.7. Another possibility is simultaneously considering atmospheric and surface effects: Split Window (SW) and Dual Angle (DA) methods directly express radiometric temperature Trad as a spectral or angular difference between two brightness temperatures Tbrs at the sensor level [125, 126]: rs rs rs rs rs 2 + A(Tb1 − Tb2 ) + B(Tb1 − Tb2 ) +C Trad = Tb1

ε 1 + ε2 + D(ε1 − ε2 ) + E (10.9) 2

ε is surface emissivity. A, B, C, D, E are empirical coefficients. Indices 1 and 2 are two spectral channels for SW method, or two view zenith angles for DA method. The angular differencing uses variations in atmospheric transmittance between different paths for two view zenith angles. The spectral differencing uses variations in atmospheric transmittance due to different water vapor absorptions for two spectrally close channels. The emission term of Eq. 10.7 can be split into soil and vegetation components, which yields the soil and vegetation radiative transfer equation [68, 119, 127]: ε (λ j , θ ) B (λ j , Trad (θ )) = τ can (θ ) εsoil (λ j ) B(λ j , Tsoil ) + ω (θ , εveg (λ j )) B(λ j , Tveg )

(10.10)

Tsoil and Tveg are soil and vegetation radiometric temperatures [68]. τ can (θ ) and ω (θ , εveg (λ j )) are vegetation directional transmittance and fraction of emitted radiation. The angular effects can also be described with linear kernel driven approaches, by expressing the directional emission as a linear combination of generic shapes [128]:

ε (λ j , θ ) B (λ j , Trad (θ )) = N

∑ βi (λ j ) Ki (Tveg , Tsoil , εveg (λ j ), εsoil (λ j ), θ , θs , ϕ − ϕs )

(10.11)

i=1

θs is the solar zenith angle. ϕ − ϕs is the relative azimuth between illumination and viewing directions. βi,λ j are weighting coefficients. Kernels Ki describe gray body isotropy, volume scattering, and surface scattering. Various kernel formulations may be proposed, by linearizing different sets of complex equations. Kernel driven approaches are also used to derive ensemble r-emissivity from accurate

258

F. Jacob et al.

hemispherical - directional reflectance (Eq. 10.5): [129] expressed TIR BRDF as a linear combination of generic shapes, following previous works over the solar domain [130–132].

10.4.3 Simulation Models Simulation models mimic the TIR radiative regime within atmosphere and canopies, to understand spatial, spectral and directional behaviors of brightness temperature measurements. These models are classified here via an increasing complexity: radiative transfer, geometric-optics, geometric-optics/radiative transfer, and ray tracing. Radiative transfer models are designed for turbid media (atmosphere, homogeneous canopies). Assuming turbidity and azimuthal isotropy, they split the medium into a finite layer number, and account for volume scattering between layers. For the atmosphere, volume scattering is negligible, and each layer is described with temperature and densities of absorbers and emitters. For canopies, soil and vegetation layers are described with temperature; and with densities of absorbers, emitters and scatterers, derived from LAI and LIDF. Brightness temperature is simulated using the stream concept: transmittance, upward and downward radiances are computed for each layer, and vertically integrated (see MODTRAN for atmosphere [133, 134] and SAIL for canopy [68, 74, 135–137]. Simulations can also be probabilistic calculations for photon interception, deduced from the directional gap fraction of each layer [118, 138, 139]. Geometric-optics models are designed for structured patterns over land surfaces, such as row crops of cotton or maize. Considering vegetation as an opaque medium, they account for surface scattering with shadowing effects. Sunlit and shaded areas are described via their cross sections, derived from canopy geometry (vegetation height, row size, etc.), illumination and viewing directions, and directional gap fraction within and between rows. Canopy brightness temperature is computed from the resulting spatial distribution of temperature [73, 121,140–143]. The finest radiosity models are geometric-optics/radiative transfer models, designed for complex land surfaces. By accounting for both volume and surface scattering, they are appropriate to vegetation patchworks. They can conjugate a radiative transfer and a geometric-optic module [144, 145]. They can be more complex, such as 3-Dimensional mock-ups based models. This allows a finer description of the radiative regime within canopies, but requires significant information about the micro-scale conditions. Examples are CUPID [146, 147]; Thermo [72, 148]; Jia’s model [149], and DART [76, 77]. Further, accounting for convective and energetic transfers allow understanding their influence on the radiative regime, such as with DART-EB [78]. The finest modeling degree is Monte Carlo ray tracing, which stochastically calculates photon trajectories within turbid or geometric atmosphere and canopies. A photon is tracked from birth (emission or penetration within medium) to death (absorption or escape from medium), with scattering based on probabilistic wave

10 Modeling and Inversion in Thermal Infrared Remote Sensing

259

matter interactions. Millions of simulations describe spectral, directional and spatial behaviors. Ray tracing is used to assess the influence of multiple scattering on spatial aggregation and angular dynamics, over heterogeneous and non isothermal land surfaces [127, 150, 151].

10.4.4 Inversion Methods Retrieving variables from measurements is an inverse problem. Given a set of m measurements M for a physical system, with k known parameters K and p unknown parameters P to be retrieved, direct F and inverse F −1 problems are written [152]: ⎞ ⎞ ⎤ ⎛⎡ ⎤ ⎡ ⎡ ⎤ ⎛⎡ ⎤ M1 P1 M1 P1 ⎟ ⎟ ⎢ .. ⎥ ⎢ . ⎥ ⎜⎢ . ⎥ −1 ⎜⎢ . ⎥ ⎣ . ⎦ = F ⎝⎣ .. ⎦ , [K1 · · · Kk ]⎠ ⇐⇒ ⎣ .. ⎦ = F ⎝⎣ .. ⎦ , [K1 · · · Kk ]⎠(10.12) Mm

Pp

Pp

Mm

Inversion is possible if there are more independent equations than unknowns (m ≥ p). Direct inversion analytically writes the inverse problem. This is possible for simple radiative transfer equations (Section 10.4.2), but not for most simulation models (Section 10.4.3). For the latter, indirect inversion numerically sets parameters such as simulations agree with observations [153]. It has been improved for accuracy and rapidity, by calibrating neural networks, lookup tables, genetic algorithms or regression trees [152, 154]. Inversion can be a well-posed problem, when solving an overdetermined equation system using optimization techniques. However, it is usually an ill-posed problem, with several parameters to be constrained from few observations. Proposed solutions use a priori information about soil and vegetation properties, or parameter ranges [152, 155, 156]. Inversion over the TIR domain is not as developed as over the solar domain. This results from (1) additional micrometeorological complex influences, and (2) the lack of high resolution data. Atmospheric simulation models have been inverted calibrating neural networks [51], SW and DA methods (Eq. 10.9) [125, 126], or the atmospheric radiative transfer equation (Eq. 10.6) [157, 158]. Over land surfaces, simulation models have been assessed in the forward mode [72–74, 77, 144]. No investigation was found about their indirect inversion, but they can serve as references for parameterizing simple radiative transfer equations which are directly invertible. Thus, various formulations have been assessed for the soil and vegetation radiative transfer equation (Eq. 10.10), optionally accounting for multiple scattering and non linearities [68, 71, 118, 119, 127]. Further, inverting simple radiative transfer equations is often an ill-posed problem. For instance, inverting the composite surface radiative transfer equation (Eq. 10.7) from N multispectral observations includes N emissivities and radiometric temperature. Similarly, inverting the soil and vegetation radiative transfer equation (Eq. 10.10) or linear kernel driven approaches (Eq. 10.11) from multiangular observations requires angular parameters: ensemble, soil and vegetation emissivities; vegetation transmittance.

260

F. Jacob et al.

10.5 Assessing Modeling Tools and Inversion Methods Modeling tools and inversion methods have been assessed experimentally through validation exercises, and theoretically via sensitivity studies. Validation exercises have been conducted over databases collected in the framework of various international programs such as FIFE [159], EFEDA [160], HAPEX [161], ReSeDA [97], JORNEX [162], FLUXNET [163], DAISEX [164], SALSA [165], SMACEX [166]. Assessments over these various datasets allow accounting for different biomes and climates. Some exercises were ground-based [73, 104, 145]. Most of them were airborne-based [94, 96, 103, 157, 167–169,170–174], for assessments in actual conditions by reducing spatial heterogeneity effects. Few validations were conducted using spaceborne observations with hectometric resolutions [175–179]; and with kilometric ones over areas almost homogeneous [180–182]. Original exercises based on classifications were designed for kilometric scale heterogeneities [126, 183], while new improvements for the solar domain should be implemented over the thermal one [184]. Complementary to validations, intercomparisons are now feasible thanks to multisensor missions such as Terra. This allows accounting for larger panels of environmental situations [158]. Validations and intercomparisons have also been performed using simulated datasets. This allow considering more conditions than measured datasets, and focusing on physics modeling without measurement intrinsic errors [81, 82, 126, 185]. Simulated datasets are necessary when dealing with elaborated temperatures: aerodynamic, soil and vegetation, sunlit and shaded components, and canopy temperature profile [68, 71, 73, 76, 118]. Indeed, validating the latter using measured datasets is not trivial, since the corresponding ground-based measurements are difficult to implement. Additionally to validations and intercomparisons, sensitivity studies allow assessing information requirements such as accuracies on remotely sensed information, medium structural and radiative properties. Examples are (1) accuracy on atmospheric status for retrieving brightness temperature [171, 178], (2) accuracy on observations, atmospheric status and land use for recovering ensemble emissivity and radiometric temperature [12, 157, 158, 169, 182, 186–188], (3) accuracy on canopy structural parameters and radiative properties for deriving soil and vegetation temperatures [68, 118, 189]. Finally, sensitivity studies of simulation models provide valuable information about the pertinent parameters for inversion [73, 76], with innovative approaches over the solar domain based on adjoint models (Baret et al., this issue).

10.6 Current Capabilities and Future Directions From the basic materials presented before, we focus now on current investigations, via an increasing temperature complexity. Success and failures suggest future directions.

10 Modeling and Inversion in Thermal Infrared Remote Sensing

261

10.6.1 Surface Brightness Temperature Surface brightness temperature is derived from that at the sensor level by inverting modeling tools for atmosphere. It is simulated using modeling tools for land surfaces. In both cases, these tools are simple radiative transfer equations or simulation models.

10.6.1.1 Atmospheric Radiative Regime and Related Corrections Atmospheric corrections for the retrieval of surface brightness temperature can be performed inverting simulation models, via the calibration of the atmospheric radiative transfer equation (Eq. 10.6) for a given atmosphere [12, 81, 157, 158, 171, 172, 177, 178]. An operational context faces two challenges: reducing computation time to process millions observations, and accurately characterizing the atmospheric status. To reduce by a third-order computation time for simulation models without accuracy degradation, [190] implemented correlated-K methods, by quickly integrating waveband atmospheric absorption and emission. Predictor-based models accurately compute the latter for a range of reference profiles, to next differencing current ones and nearest predictors [191]. Multilayer computation based on water vapor continuum absorption can replace simulation models, with an accuracy degradation lower than 1 K [81]. Computation time can also be reduced via inversion by including a range of atmospheres into the simulation set. Expressing transmittance and upwelling radiance of Eq. 10.6 from atmosphere water vapor content and mean temperature yields an accuracy degradation lower than 2 K [180]. Neural networks can replace Eq. 10.6 considering atmospheric profiles and view zenith angle, with an accuracy degradation lower than 0.5 K [186, 192]. The atmospheric status can be well documented using ancillary information: measured profiles allow reaching a 1 K accuracy [171, 172, 177, 178], but meteorological networks are not dense enough for regional inversion. One alternative is profile simulation from meteorological models [193, 194]. Such information is soon available with a 3 h sampling, and a 0.25◦ latitude/longitude griding to be re-sampled to sensor resolutions via interpolation procedures [12, 195]. The relief influence is handled using digital elevation models, now available with decametric resolutions and metric accuracies [196]. Also, the TIR observations to be corrected can inform about the atmospheric status. Atmosphere absorption and emission can be retrieved from multispectral and hyperspectral observations, using variabilities of atmospheric properties [80, 197]. Thus, water vapor content was adjusted from ASTER multispectral observations, such as emissivity spectrum is flat over vegetation or water [185]. It was also inferred from the ATSR-2 SW channels with a 0.2 g. cm−2 accuracy, using the SWVCR which relies on the spatial variability of SW surface brightness temperatures [198]. Solar or TIR observations collected onboard the same platform also provide coincident information about the atmospheric status. [199] expressed water vapor

262

F. Jacob et al.

content as a polynomial of MODIS near infrared radiance ratios, with a 0.4 g. cm−2 accuracy. Atmospheric sounders allow inferring profiles of temperature and water vapor density, using Eq. 10.6 or neural networks [3, 4]. Previous sounders such as TOVS permitted to reach a 0.4 g. cm−2 accuracy on water vapor content [200]. New sounders such as IASI [201], with finer spectral samplings and spatial resolutions, should provide accuracies better than 1 K and 10% for atmospheric profiles of temperature and humidity.

10.6.1.2 Land Surface Radiative Regime and Related Measurements Surface brightness temperature is simulated using simple radiative transfer equations or simulation models. The former provide easy and efficient solutions for assimilating TIR remote sensing data into land process models. The latter are fine and accurate solutions for understanding TIR remotely sensed measurements. To constrain land process model parameters, surface brightness temperature can be simulated using simple radiative transfer equations coupled with SVAT models. [39] coupled the composite surface radiative transfer equation (Eq. 10.7) with a crop and a one source SVAT model. The latter calculated ensemble radiometric temperature by closing the surface energy budget. R-emissivity was estimated using the SAIL TIR version of [136], documented by the crop model for vegetation structural parameters. Similarly, [67] coupled the soil and vegetation radiative transfer equation (Eq. 10.10) with a two source SVAT model. The latter calculated soil and vegetation temperatures by closing the energy budget for each, while setting soil and vegetation emissivities to nominal values. Calculating surface brightness temperature from simulation models requires information about vegetation structure (row crop, LAI, LIDF, cover fraction), soil and vegetation radiative properties (emissivity, reflectance), and thermal regime (canopy temperature distribution). The latter can be derived from a SVAT model, which solves local energy budget according to meteorological conditions (solar position, wind speed, air temperature), vegetation status (leaf stomatal resistance), and soil moisture. Then, simulation models mimic the radiative regime using more or less complex descriptions of the thermal regime: a unique vegetation temperature [73], soil and vegetation temperatures with optional sunlit and shaded components [74, 137], additional vegetation layer temperatures for specific crops [145], or canopy temperature profiles [78]. Simulation models are currently under development, verification and analysis [73, 74, 76, 78, 144, 145]. Current investigations focus on spectral behaviors [120], but especially on directional effects which allow normalizing multiangular observations (Fig. 10.3). For instance, [58, 72] angularly normalized water stress indices over row structured crops. Similarly, [202] normalized across track observations from sun-view geometry effects, for a daily monitoring at the continental scale.

10 Modeling and Inversion in Thermal Infrared Remote Sensing

263

0 80 30

330 60

34

40

300

33

60

20

32

90

270

31

30

29 120

240

28 150

210 180

Fig. 10.3 Simulating measured brightness temperature over a maize canopy in row structure, with a resulting angular dynamic about 8 K [73]. Black star indicates the solar direction. The brightness temperature maximum value is located in the solar direction. However, this hot spot effect is not systematical (Section 10.6.4.)

10.6.1.3 Partial Conclusions The various methods developed to perform atmospheric corrections are of interest, since they were designed for optimizing the collected information according to sensor configurations. Measured or simulated profiles are tributary to their representativeness, and coincident information relies on strong assumptions. Despite these limitations, significant progresses were made the last decades, with accuracies now close to 1 K. Current investigations focus on refinements rather than new developments. Simulating brightness temperature is ongoing for describing brightness temperature measurements, according to the various land surface behaviors: geometric like, radiative transfer like, or both. Validation results emphasized good performances with accuracies close to 1 K, though significant documentations are required about thermal regime, medium structure and radiative properties. Such simulation models will be of interest for future designs of inversion methods, conjointly to the solar domain (Section 10.4.4).

264

F. Jacob et al.

10.6.2 Ensemble Emissivity and Radiometric Temperature Ensemble radiometric temperature is derived by directly inverting composite surface radiative transfer equation (Eq. 10.7), or indirectly inverting simulation models via differencing equations (Eq. 10.9). The first way is two-step based and requires previous atmospheric corrections. The second way is one-step based by simultaneously correcting atmosphere and surface effects. In both cases, performances depend on characterizing these effects. Inverting Eq. 10.7 is an ill-posed problem, with N equations from channel measurements and N + 1 unknowns being channel emissivities and radiometric temperature. Proposed solutions consist of adding an N + 1 equation. They are reported here via an increasing amount of information, according to the spectral, directional and temporal dimensions. 10.6.2.1 Single-Channel TIR Instantaneous Observations Radiometric temperature is derived from single channel observations using two step approaches. After atmospheric corrections, inverting the composite surface radiative transfer equation (Eq. 10.7) requires estimating waveband emissivity. The latter is inferred using in-situ observations, nominal values proposed by literature, or solar remotely sensed observations. This have been investigated for ground-based and airborne sensors during field experiments, and for spaceborne sensors such as the Landsat TM series. Considering ensemble emissivity increases with vegetation amount, it can be linked to NDVI [203], or to cover fraction (Eq. 10.8) neglecting spatial variabilities for soil and vegetation emissivities [177, 199]. However, low correlations were observed between AVHRR emissivities and cover fraction [188]; and between ASTER broadband emissivity and MODIS solar albedo [46]. Indeed, the link between emissivity and vegetation amount depends on canopy structure, cavity effect, and optical properties of soil and vegetation [136]. Besides, emissivity may decrease with the vegetation amount, according to the type of soil and the vegetation water status [120]. Good results were reported with TM and DAIS (1 K over semi-arid agricultural areas [174, 177]), but the use of in situ information at the local scale raises the question of method applicability. A promising way is using additional MIR data, which contain information on water content. For optimizing the temporal monitoring, another possibility is deriving single channel emissivity from multispectral ones, by conjointly using different sensors such as Landsat and ASTER. However, this is contingent upon the temporal stability of surface conditions between consecutive satellite overpasses. 10.6.2.2 Dual-Channel and Dual-Angle TIR Instantaneous Observations Radiometric temperature is recovered from dual channel and dual angle observations using SW and DA one step approaches, which require accounting well for

10 Modeling and Inversion in Thermal Infrared Remote Sensing

265

atmospheric and surface effects [125, 126]. Most investigations deal with the SW method, since multispectral observations are more usual than multiangular ones. Various versions have been proposed for Eq. 10.9: linear or quadratic forms (B = 0 or B = 0), optional inclusion of emissivity (C = 0 or C = 0, D = 0 or D = 0), expressing coefficients from atmospheric water vapor content. Larger freedom degrees perform better [125, 126, 174, 183, 204]. Wavebands around 11 and 12 µm are the most appropriate for SW and DA assumptions, since they correspond to low variations of emissivity, spectrally and spatially [172, 174]. Calibration relies on simulations from emissivity spectral libraries and atmospheric radiative transfer [82, 125, 126, 204]. Operational use requires documentation. Atmospheric water vapor content is inferred from climatological database [183], the SWVCR [198] or near infrared radiance ratios [199]. Emissivities are derived from classifications [181, 182, 205], or from Eq. 10.8 with nominal values for soil and vegetation emissivities [172, 199]. Several validation exercises reported accuracies better than 1 K. Excellent results were obtained from TIMS without a priori information [172]. Using classificationbased knowledge of emissivity can perform well [181, 182], though significant subclass variabilities were observed [206, 207]. However, a 1 K accuracy usually requires local information on surface conditions for emissivity effects. Further, the lack of such information can induce errors up to 3 K [125, 126, 174, 183, 199].

10.6.2.3 Multispectral and Hyperspectral TIR Instantaneous Observations Radiometric temperature is derived from multispectral and hyperspectral observations using two step approaches. After atmospheric corrections, the ill-posed problem can be solved using either a priori information, or the spectral variability captured over the whole TIR range. This last possibility is very different from the two channel SW differencing which aims at avoiding emissivity variations. For multispectral observations, the NEM approach sets maximum emissivity to a nominal value [208], where the latter can be derived from Eq. 10.8 using a priori information about soil and vegetation emissivities [173]. The adjusted ANEM relies on land use [168, 170], and the MIR NEM is extended to MIR observations [209]. Rather than using a priori information, other approaches aim at benefiting the variability captured from multispectral and hyperspectral data, the latter providing finer spectral samplings. The TES algorithm derives minimum emissivity from the spectral variations, via an empirical relationship verified over most land surfaces for the TIR and the MIR domains [104, 157, 210–212]. Capturing larger variabilities with hyperspectral data increase TES accuracy up to 0.5 K [104]. To derive absolute emissivity from relative spectral variations, the alpha residuals logarithmically linearize Planck’s equation, with an optional improvement based on Taylor expansion for hyperspectral observations [213]. Taylor expansion also provides derivative approaches, such as the “Grey Body” method [123]. Finally, multispectral and hyperspectral observations are useful for deriving broadband emissivity via NTB conversions [52, 54, 207, 214].

266

F. Jacob et al.

Fig. 10.4 Validation against sample-based laboratory measurements (squares), of ASTER/TES emissivity retrievals (circles), for eleven days over a gypsum site at the White Sands National Monument in New Mexico. TES derived emissivity spectra were averaged over the 11 acquisitions, where the corresponding standard deviation ranges from 10−3 to 10−2 . (From [265].)

Several validation exercises reported 1 K accuracies for NEM, with or without a priori information [168, 170, 173]. Good results were obtained for the TES algorithm (Fig. 10.4), with accuracies better than 0.01 on emissivity [176] and 1 K on temperature [173]. Similar performances were reported by [158] when intercomparing ASTER/TES and MODIS/TISIE retrievals (Fig. 10.5, the TISIE concept is presented below).

10.6.2.4 Multispectral MIR and TIR Consecutive Observations Solving the ill-posed problem to invert Eq. 10.7 is also possible using temporal series from geostationary or sun-synchronous daytime/nighttime observations. Assuming emissivity is stable between consecutive observations yields more equations than unknowns. Then, investigations rely on using TIR observations only [215], or MIR/TIR observations [12, 186, 188, 216]. TTM is a two step approach for inverting Eq. 10.7 over TIR consecutive observations [217]. Assuming surface emissivity is constant yields 2N equations for N + 2 unknowns: N channel emissivities and two radiometric temperatures. Then, two channels are enough for solving the ill-posed problem. TTM performs better

10 Modeling and Inversion in Thermal Infrared Remote Sensing

267

Fig. 10.5 Intercomparison, over a Savannah landscape (Africa) and a semiarid rangeland (Jornada), of surface radiometric temperature retrievals from the MODIS/TISIE and ASTER/TES algorithms. Differences were lower than 0.9 K. (From [158].)

268

F. Jacob et al.

with the SEVIRI finest temporal sampling, and with observations near daily temperature extrema [187, 215]. TISIE is a two step approach for inverting Eq. 10.7 over TIR and MIR daytime/nighttime observations. Raising emissivity ratios to specific powers yields relative variations independent of radiometric temperature. Assuming TISIE are stable between consecutive observations, MIR r-emissivities can be retrieved, and next TIR ones. Various TISIE versions were designed for AVHRR, MODIS, SEVIRI [12, 186, 188, 218]. Day Night Pair is a one step approach for inverting both Eqs. 10.6 and 10.7 over TIR and MIR daytime/nighttime observations. The system of 2N equations with N + 2 unknowns can be solved with k additional unknowns, as long as k ≤ N − 2. Thus, the 7 MODIS channels allow recovering five unknowns about atmospheric and surface effects [216]. The accuracies reported for these methods range from 0.5 to 2.5 K, and are slightly worse for TTM. They correspond to sensitivity studies for TTM and TISIE [186–188], to validation exercises over various study sites for Day Night Pair (Fig. 10.6) [181, 182], and to intercomparisons against ASTER/TES retrievals for TISIE [158].

10.6.2.5 Partial Conclusions Radiometric temperature is derived by indirectly inverting simulation models through differencing equations (Eq. 10.9), or directly inverting simple radiative transfer equations (Eqs. 10.6 and 10.7). Solving the ill-posed problem depends on

1 0.95 0.9 sand1

Emissivity

0.85

sand2 sand3

0.8

sand4 sahara_em20

0.75

sahara_em22 sahara_em23

0.7

sahara_em29 sahara_em31

0.65

sahara_em32

0.6 3

4

5

6

7

8 9 10 Wavelength (um)

11

12

13

14

15

(a)

Fig. 10.6 Validation against sample-based laboratory spectra (lines), of Day Night Pair-based MODIS retrievals over the Sahara Desert (points). (From [182].)

10 Modeling and Inversion in Thermal Infrared Remote Sensing

269

the available remote sensing information. A priori information is needed to obtain good results with instantaneous single channel or dual channel / dual angle data. It can be avoided with more remote sensing information: instantaneous TIR and MIR data, instantaneous multispectral TIR data, or temporal series with dual channel data. Similarly, more surface and atmospheric parameters can be recovered with a larger amount of remotely sensed information. Reported accuracies have increased these last years, and are now closer to that required for further applications, i.e., 1 K. For instance, differences between retrievals from ASTER/TES and MODIS/TISIE were found lower than 0.9 K, though both methods differ in terms of using spectral, directional and temporal information [158]. However, ASTER/TES, MODIS/Split Window and MODIS/Day Night Pair were found very different over Northern America [206]. Current efforts are refinements rather than new concepts: disaggregation methods should allow benefiting the synergy between IASI hyperspectral and AVHRR kilometric sensors onboard METOP. Then, the next challenge is retrieving more elaborated temperatures, discussed below.

10.6.3 Aerodynamic Temperature Aerodynamic temperature Taero and thermal roughness length zoh are equivocal variables which cannot be directly recovered from remote sensing (Section 10.4.1). Therefore, investigations have aimed at substituting aerodynamic temperature by radiometric temperature Trad , by parameterizing a correcting factor in the sensible heat flux expression (Eq. 10.1). The physical meaning of sensible heat flux (Eq. 10.1) can be preserved using the Taero − Tair (zre f ) , empirically expressed from LAI [62]. Howmultiplicative factor Trad − Tair (zre f ) ever, studying this factor from simulations and measurements for growing sparse vegetation showed significant variations according to meteorological and surface conditions [63]. The correction factor can also be included in the kB−1 parameter (Eq. 10.2), which is then called thermal kB−1 [113]. It includes corrections for (1) the difference between thermal and mechanical roughness lengths, and (2) the difference between radiometric and aerodynamic temperatures. According to environmental conditions, thermal kB−1 varies from a vegetation type to another, and up to 100% in relative terms [113, 219]. Parameterizations based on near surface wind speed and temperature gradients depend on sensible heat flux [220]. Overall, formulating the thermal kB−1 seems almost impossible, since it is driven by several factors which vary in time and space: vegetation structure and water stress, meteorological conditions, canopy temperature profile and solar position [221–229]. A potential way for characterizing the kB−1 parameter would be multiangular TIR remote sensing [230]. However, using this information for deriving soil and vegetation temperatures seems more pertinent. First, these temperatures are

270

F. Jacob et al.

functionally equivalent to aerodynamic temperature (Eq. 10.3). Second, they provide information for monitoring vegetation photosynthesis and soil respiration.

10.6.4 Directional Emissivity and Soil/Vegetation Temperatures Single directional radiometric temperature can be split into soil and vegetation components by closing energy balance for each. However, this relies on strong assumptions about vegetation water status [231–233]. The promising approach is then using multiangular TIR observations [20, 68, 71, 118, 119, 189]. No literature was found about retrieving soil and vegetation temperatures by inverting simulation models over multiangular data. Nevertheless, such models have been used for parameterizing the soil and vegetation radiative transfer equation which is directly invertible (Eq. 10.10). These parameterizations were designed considering the [8–14] µm spectral range. 10.6.4.1 Parameterizing the Soil/Vegetation Radiative Transfer Equation Inverting Eq. 10.10 requires estimating the involved parameters (ensemble emissivity ε (λ j , θ ), vegetation transmittance τ can (θ ), and vegetation fraction of emitted radiance ω (θ , εveg (λ j ))), with optional simplifications for easier use. Various complexity degrees have been proposed, listed in Table 10.3 (P1–P8). This introduces two new parameters. Hemispherical gap fraction σ f is directional gap fraction integrated over illumination angles, to account for atmospheric thermal irradiance down to the soil via the vegetation. The cavity effect coefficient α is the ratio of canopy to vegetation hemispherical-directional reflectance, to account for radiation trapping within the canopy. The finest parameterization (P1), proposed by [118], accounts for multiple scattering and cavity effect. The latter, which does not depend on LAI, was previously calculated as a function of LIDF and view zenith angle, by using the probabilistic simulation model from [139]. Similarly, [127] introduced effective directional gap ef f ef f ef f Fsoil (θ ) and cover Fveg (θ ) fractions (P2). Fsoil (θ ) included single scattering of soil ef f emission by vegetation. Fveg (θ ) included single scattering of vegetation emission by soil and vegetation. Half complex parameterizations (P3–P5) account for multiple scattering, with optional linearizations [68, 71]. The simplest parameterizations (P6–P8) do not account for cavity effect nor multiple scattering. They differ by their linearization degrees, and their assumptions about soil and vegetation emissivities [71, 119, 189]. Some of these parameterizations were assessed in direct mode for simulating directional ensemble emissivity and radiometric temperature [68]. Apart from simplest versions, most provided close results for directional canopy emissivity, with discrepancies lower than 0.01. For radiometric temperature, all provided similar results, with differences lower than 1 K. Next, differences decreased with atmospheric irradiance which compensates emission (Eq. 10.7). However, this is minor under

10 Modeling and Inversion in Thermal Infrared Remote Sensing

271

Table 10.3 Listing of existing parameterizations (P1–P8) for the soil and vegetation radiative transfer equation (Eq. 10.10), with a decreasing complexity. The spectral dependence was removed since these parameterizations were designed considering the [8–14] µm spectral range. Labels fi refer to specific functions proposed by the corresponding references. The dependence on LAI and LIDF is implicitly included into directional gap and cover fractions Fsoil (θ ) and Fveg (θ ), the cavity effect coefficient α (θ ), and hemispherical gap fraction σ f [68]. Standard formulation

ε (θ ) B (Trad (θ )) = τ can (θ ) εsoil B(Tsoil ) + ω (θ , εveg ) B(Tveg ) Label

From

Formulations ε (θ ) = f3 Fsoil (θ ), εsoil , εveg , σ f , α (θ ) τ can (θ ) = Fsoil (θ ) ω (θ , εveg ) = f4 Fsoil (θ ), Fveg (θ ), εsoil , εveg , σ f , α (θ )

P1

[118]

P2

[127]

P3

Remarks Accounts for multiple scattering and cavity effect

ef f ef f ε (θ ) = Fsoil (θ ) εsoil + Fveg (θ ) εveg e f f τ can (θ ) = Fsoil (θ ) ef f ω (θ , εveg ) = Fveg (θ ) εveg ef f Fsoil (θ ) = f1 Fveg (θ ), σ f ef f (θ ) = f2 Fsoil (θ ), σ f Fveg

Accounts for multiple scattering and cavity effect Effective parameterization

[68]

ε (θ ) = f5 (Fsoil (θ ), Fveg (θ ), εsoil , εveg , σ f ) τ can (θ ) = f6 Fsoil (θ ), εsoil , εveg , σ f ω (θ , εveg ) = f7 Fsoil (θ ), Fveg (θ ), εsoil , εveg , σ f

Accounts for multiple scattering

P4 P5

[71]

Linearizing P3 considering B(T ) ≈ σ T 4 Linearizing P3 considering B(T ) ≈ σ T

Accounts for multiple scattering

P6

[119]

ε (θ ) = Fsoil (θ ) εsoil + Fveg (θ ) εveg τ can (θ ) = Fsoil (θ ) ω (θ , εveg ) = Fveg (θ ) εveg

P7 P8

[71]

Linearizing P6 considering B(T ) ≈ σ T Simplifying P7 considering εveg = εsoil = 1

clear sky conditions, with irradiance lower than 30 W. m−2 between 8 and 14 µm [234]. Finally, [127] observed analytical formulation P2 significantly diverged from a ray tracing reference when soil and vegetation emissivities were very different.

10.6.4.2 Inverting the Soil/Vegetation Radiative Transfer Equation The various parameterizations reported above have been assessed in inverse mode considering dual angle observations, nadir and 45◦ or 55◦ off nadir. Given soil and vegetation emissivities, dual angle measurements allow retrieving component temperatures. Off nadir angles above 45◦ are required to capture large angular dynamics and reduce observation errors [108, 109, 119, 144, 235–237]. Dual angle observations at 0 and 55◦ correspond to the ATSR suite viewing configuration.

272

F. Jacob et al.

4

4

3

3

Error in leaves temperature retrieval (⬚C)

Error in soil temperature retrieval (⬚C)

Converging conclusions were reported about the documentation requirements for canopy structural parameters. Poorly estimating LIDF can result in errors on temperatures up to 1 K [118]. LAI must be known within 5% (respectively 10%) for a 0.5 K accuracy on vegetation (respectively soil) temperature [118, 119]. Similarly, a 7–8% relative error on directional cover fraction can induce errors on soil and vegetation temperatures from 1 to 3 K [189]. Such recommendations have to be compared with current accuracies on LAI retrievals from solar remote sensing, i.e., around 20% [238]. Diverging conclusions were reported about the documentation requirements for canopy radiative properties, the parameterization degree to be considered, and the performances. First, [118] concluded a 0.01 accuracy is necessary for soil and vegetation emissivities, whereas [71] claimed using unity values has no consequence. Second, [71, 119] concluded simple parameterizations similarly performed than complex ones, while multiple scattering and cavity effect can be neglected. Conversely, [68] reported it is necessary to account for multiple scattering (Fig. 10.7). Third, [68] concluded a 1 K accuracy can be reached on both temperatures, and better for vegetation than soil; whereas [71] reported lower errors for soil (<2 K) than vegetation (<4 K).

2

1

0

−1

2

1

0

−1

SAIL IRT - Mod1 Mod4 - Mod1 Mod3 - Mod1 Mod2 - Mod1

SAIL IRT - Mod1 Mod4 - Mod1 Mod3 - Mod1 Mod2 - Mod1

−2

−2 0

1

2

3

LAI

4

5

0

1

2

3

4

5

LAI

Fig. 10.7 Performance intercomparison for the different parameterizations listed in Table 10.3. Mean errors on soil (left) and vegetation (right) temperature retrievals are plotted as functions of LAI, along with standard deviations (bars). Reference “Mod 1” is the probabilistic simulation model of [139]. “SAIL IRT” is the TIR version of the SAIL radiative transfer model from [68]. “Mod 2” is the parameterization P6, “Mod 3” the P3 and “Mod 4” the P1. (From [68].)

10 Modeling and Inversion in Thermal Infrared Remote Sensing

273

10.6.4.3 Partial Conclusions Parameterizing the soil and vegetation radiative transfer equation still is in debate, as is the required documentation. Additionally to parameters to be well estimated, both structural (LAI, LIDF, directional gap and cover fractions, hemispherical gap fraction, cavity effect), and radiative (ensemble emissivity and radiometric temperature), inversion includes emissivity and radiometric temperature for soil and vegetation. When focusing on finer temperatures such as sunlit and shaded components for soil and vegetation, or canopy temperature profile, complex simulation models with additional unknowns have to be considered (Section 10.4.3). Then, attractive modeling tools are linear kernel driven approaches for accuracy and feasibility (Section 10.4.2). Such tools have been inverted and validated against multidirectional observations, with promising results about the potential for adjusting the captured angular dynamic [103, 128]. Regarding the aforementioned challenges, ongoing multiangular observations are of prime interest. Maximum temperatures were captured in nadir and solar directions [58, 73, 108, 109]. Solar peak corresponds to the hot spot effect, with more sunlit surfaces in the solar direction. Nadir peak corresponds to a larger fraction of hot soil with a cooler vegetation. These measurements at the ground level have to be used along with modeling tools, for designing observation configurations and inversion methods.

10.7 Forefront Investigations: Pertinently Using TIR Remote Sensing This section extends the discussion to its scientific context, with the use of land surface temperature for monitoring exchanges of heat, water and mass between soil, vegetation and atmosphere. For instance, the difficulties faced with aerodynamic temperature can be overcome using space or time differencing energy balance models. New difficulties are related to time and space issues. First, monitoring land surface processes from TIR remote sensing requires accounting for the fluctuating nature of surface temperature. Second, current spaceborne spatial resolutions do not systematically provide the appropriate scales for process modeling. This yields new approaches such as remote sensing data assimilation and aggregation/disaggregation procedures. Regarding the insufficient accuracy on surface temperature and the difficult use of aerodynamic temperature (Sections 10.6.2 and 10.6.3), several studies have suggested using soil and vegetation temperatures with two source models (Section 10.6.4). However, one source modeling with ensemble radiometric temperature still is very pertinent: time and space differencing models have been developed this last decade to minimize errors on surface temperature. Spatially

274

F. Jacob et al.

is it possible adjusting air temperatures from surface temperatures to make the gradient consistent [239]. Similarly, evaporative fraction can be derived from the radiometric temperature dynamic given an albedo range [59]. Temporally, surface fluxes can be estimated regarding the potential extreme situations for the considered surfaces [22]. Similarly, time differencing between morning geostationary observations minimizes error on surface temperature [19]. These approaches have been widely validated for various environmental conditions, showing the pertinence of such differencing concepts [19, 22, 23, 25, 60, 66, 240]. However, they are limited by the agreement between the assumed and captured variabilities. Monitoring land surfaces is performed using dynamic models which simulate the considered processes, such as the functioning of cultivated and natural vegetation [241, 242]. Assimilating remotely sensed information allows constraining model trajectory, to obtain consistent series for the considered processes and related variables. Model parameters and initial variables (respectively state variables) are adjusted (respectively readjusted) to make simulations and observations agreeing over a temporal window (respectively at a given time) [39, 243]. Minimizing differences between simulations and observations relies on stochastic methods and adjoint models [38, 244]. Solar and radar information can be directly assimilated, since the considered variables are almost temporally stable [242, 245–248]. Assimilating TIR observations requires adding a SVAT model, due to surface temperature fluctuations [29, 38, 39, 67, 137]. Then, the SVAT model documents a simple radiative transfer equation for simulating surface brightness temperature (Section 10.6.1). To avoid using complex SVAT models, secondary variables such as energy fluxes or soil moisture can be assimilated in place of primary variables such as surface temperature [34, 249, 250]. Monitoring land surfaces from remote sensing faces the inadequacy of sensor spatial resolutions to match the process scales (Section 10.3.1). TIR remote sensors perform at hectometric resolutions with poor temporal samplings, or at kilometric resolutions with daily revisiting. Then, high spatial resolution observations provide valuable information for understanding heterogeneity effects and aggregation processes [21, 63, 119, 251–260]. This is of interest for hydrology and meteorology, when the considered processes can be monitored at a kilometric resolution. Agricultural issues require higher spatial resolutions, in accordance with the field scale. Since TIR hectometric resolution remote sensors do not provide sufficient temporal sampling, a solution is disaggregating daily kilometric resolution variables [261]. By using the well known vegetation index/temperature triangle, it is possible to disaggregate TIR observations from solar ones, since the latter always have higher resolutions [56]. However, this is limited by soil moisture conditions [41, 258]. Other possibilities are statistical or deterministic disaggregation. Statistical procedures have been applied at the landscape unit scale for solar observations [262]. Deterministic procedures have been proposed for TIR geostationary observations [115, 263].

10 Modeling and Inversion in Thermal Infrared Remote Sensing

275

10.8 Concluding Remarks Significant progresses were achieved for the retrieval of ensemble emissivity and radiometric temperature, with accuracies close to requirements for many applications. Current efforts are refinements rather than new concepts. Regarding the difficult use of aerodynamic temperature, it is recommended focusing on soil and vegetation temperatures which are functionally equivalent. Then, the next challenge is deriving soil and vegetation temperatures with optional sunlit and shaded components, and canopy temperature profile from the soil to the top of canopy via vegetation layers. Soil and vegetation temperatures can be recovered parameterizing and inverting simple radiative transfer equations, where simplicity is preferred for operational applications. More elaborated temperatures require finer approaches, with efforts to be made on measuring and modeling. Ground-based goniometric systems allow capturing complex angular dynamics of brightness temperature. Simulation models currently under improvement are increasingly accurate. The next step is developing inversion methods such as inductive learning and lookup tables, already implemented over the solar domain. Expected difficulties result from the temperature profile which is driven by micro meteorological conditions. Given surface temperature fluctuations require along track observations from space, recovering temperature profile seems limited to few layers, optionally sunlit and shaded. This should allow designing optimal viewing configurations for spaceborne sensors, in terms of observation number and angular distribution. Finally, current focus on spatial and temporal issues are important. TIR spaceborne sensors do not provide optimum temporal monitoring and spatial scales. Adequate spatial resolutions and revisit rates still do not exist, despite the numerous missions proposed the last decade. While waiting for such information, it is necessary to develop aggregation and disaggregation methods, to benefit from the combination of spatial scales from high resolution sensors and daily monitoring from kilometric resolution sensors. Acknowledgements This review article was possible thanks to numerous interactions, discussions and scientific exchanges, in the framework of various collaborations supported by several programs: the US ASTER Project of NASA’s EOS-Terra Program (P.I. Thomas Schmugge), the US NASA EOS Grant 03-EOS-02 (P.I. Andrew French), the French PNTS program (project P.Is. Frederic Jacob and Albert Olioso), the French PNBC program (project P.I. Jean Claude Menaut), the French Inter Region MIP / PACA program (project P.I. Dominique Courault), and finally the program from the Department of Research, Development, and International Relations of Purpan Graduate School of Agriculture. Many thanks to the ISPMSRS 2005 sponsors for permitting this symposium.

Glossary AATSR ANEM ASTER

Advanced ATSR Adjusted NEM Advanced Spaceborne Thermal Emission and Reflection Radiometer

276

F. Jacob et al.

ATSR AVHRR

Along-Track Scanning Radiometer Advanced Very High Resolution Radiometer

DA DAIS DAISEX DART DART-EB

Dual Angle differencing method Digital Airborne Imaging Spectrometer DAIS EXperiment Discrete Anisotropic Radiative Transfer DART-Energy Balance

EFEDA ENVISAT ERS ETM

European Field Experiment in Desertification threatened Areas ENVIronment SATellite European Remote Sensing Enhanced Thematic Mapper

FIFE FLUXNET FTIR

First ISLSCP Field Experiment Flux Network Fourier Transform Infra Red spectroradiometer

GLI

GLobal Imager

HAPEX HIRS

Hydrology Atmosphere Pilot EXperiment High-resolution Infrared Radiation Sounder

IASI IRSUTE ISLSCP

Infrared Atmospheric Sounding Interferometer Infra Red Satellite Unit for the Thermal Environment International Satellite Land Surface Climatology Project

JORNEX

JORNada EXperiment

LAI LIDF

Leaf Area Index Leaf Inclination Distribution Function

MAS MASTER METOP MIR MODIS MODTRAN MSG MTI

MODIS Airborne Simulator MODIS / ASTER airborne simulator METeorological OPerational Middle Infra Red: from 3 to 5 µm MODerate resolution Imaging Spectroradiometer MODerate resolution atmospheric TRANsmission Meteosat Second Generation Multispectral Thermal Imager

NDVI NEM NOAA NPOESS NTB

Normalized Difference Vegetation Index Normalized Emissivity Method National Oceanic and Atmospheric Administration National Polar Orbiting Environmental Sensor Suite Narrowband To Broadband conversion

10 Modeling and Inversion in Thermal Infrared Remote Sensing

ReSeDA

REmote SEnsing Data Assimilation (European research program)

SAIL SALSA

SVAT SW SWVCR

Scattering by Arbitrarily Inclined Leaves Semi Arid Land Surface Atmosphere (Mexico–United States–France joint research program) Spatially Enhanced Broadband Array Spectrograph System Spinning Enhanced Visible and Infrared Imager Soil Moisture Atmosphere Coupling Experiment Surface Processes and Ecosystem Changes Through Response Analysis (proposal for a European Space Agency mission) Soil–Vegetation–Atmosphere Transfer Split Window differencing method Split Window Variance Covariance Ratio

TES TIMS TIR TIROS TISIE TM TOVS TTM

Temperature Emissivity Separation Thermal Infrared Multispectral Scanner Thermal Infra Red: from 7 to 14 µm Television Infra Red Observation Satellite Temperature Independent Spectral Indices of Emissivity Thematic Mapper TIROS Operational Vertical Sounders Two Temperature Method

VIIRS

Visible Infrared Imaging Radiometer Suite

SEBASS SEVIRI SMACEX SPECTRA

277

References 1. Romanov P, Tarpley D (1993) Automated monitoring of snow cover over South America using GOES imager data. Int. J. Remote Sens. 24(5):1119–1125 2. Kay J, Gillespie A, Hansen G, Pettit E (2003) Spatial relationships between snow contaminant content, grain size, and surface temperature from multispectral images of Mt. Rainier, Washington, DC (USA). Remote Sens. Environ. 86(2):216–231 3. Aires F, Ch´edin A, Scott N, Rossow W. (2002) A regularized neural net approach for retrieval of atmospheric and surface temperatures with the IASI instrument. J. Appl. Meteorol. 41:144–159 4. Aires F, Rossow W, Ch´edin A, Scott N (2002) Remote sensing from the infrared atmospheric sounding interferometer instrument 2. Simultaneous retrieval of temperature, water vapor, and ozone atmospheric profiles. J. Geophys. Res. 107(D22):4620 5. Prigent C, Aires F, Rossow W (2003) Land surface skin temperatures from a combined analysis of microwave and infrared satellite observations for an all weather evaluation of the differences between air and skin temperatures. J. Geophys. Res. 108(D10):4310 6. Hong G, Heygster G, Kunzi K (2005) Intercomparison of deep convective cloud fractions from passive infrared and microwave radiance measurements. IEEE Geosci. Remote Sens. Lett. 2(1):18–24 7. Sunlner M, Michael K, Bradshaw C Hindell M (2003) Remote sensing of Southern Ocean sea surface temperature: implications for marine biophysical models. Remote Sens. Environ. 84(2):161–173

278

F. Jacob et al.

8. Gould R, Arnone R (2004) Temporal and spatial variability of satellite sea surface temperature and ocean colour in the Japan/East Sea. Int. J. Remote Sens. 25(7–8):1377–1382 9. Simpson J, Yueh L, Schmidt A, Harris A (2005) Analysis of along track scanning radiometer-2 (ATSR-2) data for clouds, glint and sea surface temperature using neural networks. Remote Sens. Environ. 98(2–3):152–181 10. Quattrochi D, Luvall J (2003) Thermal remote sensing in land surface processes. Taylor & Francis, London 11. Justice CO, Giglio L, Korontzi S, Owens J, Morisette JT, Roy D, Descloitres J, Alleaume S, Petitcolin F, Kaufman Y (2002) The MODIS fire products. Remote Sens. Environ. 83:244– 262 12. Petitcolin F, Vermote E (2002) Land surface reflectance, emissivity and temperature from MODIS middle and thermal infrared data. Remote Sens. Environ. 83:112–134 13. Kasischke ES, Hewson JH, Stocks B, van der Werf G, Randerson J (2003) The use of ATSR active fire counts for estimating relative patterns of biomass burning a study from the boreal forest region. Geophys. Res. Lett. 30(18):1969, doi:10.1029/2003GL017859 14. Manso-Delgado L, Aguirre-Gomez R, Alvarez R (2003) Multitemporal analysis of land surface temperature using NOAA-AVHRR: preliminary relationships between climatic anomalies and forest fires. Int. J. Remote Sens. 20:4417–4423 15. Lo C, Quattrochi D, Luvall J (1997) Application of high-resolution thermal infrared remote sensing and GIS to assess the urban heat island effect. Int. J. Remote Sens. 18(2):287–304 16. Lagouarde JP, Moreau P, Irvine M, Bonnefond JM, Voogt JA, Solliec F (2004) Airborne experimental measurements of the angular variations in surface temperature over urban areas: case study of Marseille (France). Remote Sens. Environ. 93:443–462 17. Mestayer PG, Durand P, Augustin P, Basting S, Bonnefond JM, Benech B, Camp-istron B, Coppalle A, Delbarre H, Dousset B, Drobinski P, Druilhet A, Frejafon E, Grimmond, CSB, Groleau D, Irvine M, Kergomard C, Kermadi S, Lagouarde JP, Lemonsu A, Lohou F, Long N, Masson V, Moppert C, Noilhan J, Offerle B, Oke TR, Pigeon G, Puygrenier V, Roberts S, Rosant JM, Sa¨ıd F, Salmond J, Tal-baut M, Voogt J (2005) The urban boundarylayer field campaign in Marseille (UBL/CLU-escompte): set-up and first results. BoundaryLayer Meteorol. 114(2):315–365 18. Olioso A, Taconet O, Ben Mehrez M (1996) Estimation of heat and mass fluxes from IR brightness temperature. IEEE Trans. Geosci. Remote Sens. 34:1184–1190 19. Norman J, Kustas W, Prueger J, Diak GR (2000) Surface flux estimation using radiometric temperature: a dual-temperature-difference method to minimize measurement errors. Water Res. Res. 36:2263–2274 20. Chehbouni A, Nouvellon Y, Lhomme J, Watts C, Boulet G, Kerr Y, Moran M, Goodrich D (2001) Estimation of surface sensible heat flux using dual angle observations of radiative surface temperature. Agric. Forest Meteorol. 108:55–65 21. Wassenaar T, Olioso A, Hasager C, Jacob F, Chehbouni A (2002) Estimation of evapotranspiration over heterogeneous pixels. In: Sobrino J (ed.) Proceedings of the First International Symposium on Recent Advances in Quantitative Remote Sensing, September 2002, Valencia, Spain, pp 458–465 22. Su Z (2002) The Surface Energy Balance System (SEBS) for estimation of turbulent heat fluxes. Hydrol. Earth Sci. Syst. 6:85–99 23. Jacob F, Olioso A, Gu X, Su Z, Seguin B (2002) Mapping surface fluxes using visible, near infrared, thermal infrared remote sensing data with a spatialized surface energy balance model. Agronomie: Agric. Environ. 22:669–680 24. French A, Jacob F, Anderson M, Kustas W, Timmermans W, Gieske A, Su B, Su H, McCabe M, Li F, Prueger J, Brunsell N (2005) Surface energy fluxes with the Advanced Spaceborne Thermal Emission and Reflection radiometer (ASTER) at the Iowa 2002 SMACEX site (USA). Remote Sens. Environ. 99:55–65 25. Courault D, Lacarr`ere P, Clastre P, Lecharpentier P, Jacob F, Marloie O, Pr´evot L, Olioso A (2003) Estimation of surface fluxes in a small agricultural area using the three-dimensional atmospheric model Meso-NH and remote sensing data. Can. J. Remote Sens. 29:741–754

10 Modeling and Inversion in Thermal Infrared Remote Sensing

279

26. Olioso A, Carlson T, Brisson N (1996) Simulation of diurnal transpiration and photosynthesis of a water stresses soybean crop. Agric. Forest Meteorol. 81:41–59 27. Moran M, Vidal A, Troufleau D, Qi J, Clarke T Jr., Pinter PJ, Mitchell T, Inoue Y, Nealeg MU (1997) Combining Multifrequency Microwave and Optical Data for Crop Management. Remote Sens. Environ. 61:96–109 28. Moran M, Inoue Y, Barnes E (1997) Opportunities and Limitations for Image-Based Remote Sensing in Precision Crop Management. Remote Sens. Environ. 61:319–346 29. Cayrol P, Kergoat L, Moulin S, Dedieu G, Chehbouni A. (2000) Calibrating a coupled SVAT – vegetation growth model with remotely sensed reflectance and surface temperature. J. Appl. Meteorol. 39:2452–2472 30. Bastiaanssen W, Molden D, Makin I (2000) Remote sensing for irrigated agriculture: examples from research and possible applications. Agric. Water Manag. 46:137–155 31. Droogers P, Bastiaanssen W (2002) Irrigation performance using hydrological and remote sensing modeling. J. Irrig. Drain. Eng. 128:11–18 32. Hale R, Duchon C (2003) Use of AVHRR-derived surface temperatures in evaluating a landatmosphere model. Int. J. Remote Sens. 24:4527–4541 33. Jacobs A, Ronda R, Holtslag A (2003) Water vapour and carbon dioxide fluxes over bog vegetation. Agric. Forest Meteorol. 116:103–112 34. Schuurmans J, Troch P, Veldhuizen A, Bastiaanssen W, Bierkens M (2003) Assimilation of remotely sensed latent heat flux in a distributed hydrological model. Adv. Water Res. 26:151–159 35. Nijbroek R, HoogenBoom G, Jones J (2003) Optimizing irrigation management for a spatially variable soybean field. Agric. Syst. 76:359–377 36. Panda R, Behera S, Kashyap P (2003) Effective management of irrigation water for wheat under stress conditions. Agric. Water Manag. 63:37–56 37. Bastiaanssen W, Chandrapala L (2003) Water balance variability across Sri Lanka for assessing agricultural and environmental water use. Agric. Water Manag. 58:171–192 38. Demarty J, Ottl´e C, Braud I, Olioso A, Frangi JP, Gupta H, Bastidas L (2005) Constraining a physically based Soil-Vegetation-Atmosphere Transfer model with surface water content and thermal infrared brightness temperature measurements using a multiobjective approach. Water Resour. Res. 41 (2005) W01011, doi:10.1029/2004WR003695 39. Olioso A, Inoue Y, Ortega-Farias S, Demarty J, Wigneron JP, Braud I, Jacob F, Lecharpentier P, Ottl´e C, Calvet JC, Brisson N (2005) Future directions for advanced evapotranspiration modeling: Assimilation of remote sensing data into crop simulation models and SVAT models. Irrig. Drain. Syst. 19(3–4):355–376 40. Capehart W, Carlson T (1997) Decoupling of surface and near-surface soil water content: A remote sensing perspective. Water Resour. Res. 33:1383–1395 41. Gillies R, Carlson T, Cui J, Kustas W, Humes K (1997) Verification of the ‘triangle’ method for obtaining surface soil water content and energy fluxes from remote measurements of the Normalized Vegetation Index NDVI and surface radiant temperature. Int. J. Remote Sens. 18:3145–3166 42. Inoue Y, Olioso A, Choi W (2004) Dynamic change of CO2 flux over bare soil field and its relationship with remotely sensed surface temperature. Int. J. Remote Sens. 25(10):1881– 1892 43. Coudert B, Ottl´e C, Boudevillain B, Demarty J, Guillevic P (2005) Contribution of thermal infrared remote sensing data in multiobjective calibration of a dual source SVAT model. J. Hydrometeorol. 7:404–420 44. Sobrino J, Raissouni N (2000) Toward remote sensing methods for land cover dynamic monitoring. Application to Morroco. Int. J. Remote Sens. 21:353–366 45. Wilber A, Kratz D, Gupta S (1999) Surface emissivity maps for use in satellite retrievals of longwave radiation. NASA/TP-1999-209362, NASA 46. Zhou L, Dickinson R, Ogawa K, Tian Y, Jin M, Schmugge T (2003) Relations between albedos and emissivities from MODIS and ASTER data over North African Desert. Geophys. Res. Lett. 30(20):2026, doi:10.1029/2003GL018069

280

F. Jacob et al.

47. Zhou L, Dickinson R, Tian Y, Jin M, Ogawa K, Yu H, Schmugge T (2003) A sensitivity study of climate and energy balance simulations with use of satellite derived emissivity data over Northern Africa and the Arabian Peninsula. J. Geophys. Res. 108(D24):4795, doi:10.1029/2003JD004083 48. Jin M, Liang S (2006) Improving Land surface emissivity parameter of land surface model in GCM. J. Climate 19:2867–2881 49. Wan Z, Wang P, Li X (2004) Using MODIS Land Surface Temperature and Normalized Difference Vegetation Index products for monitoring drought in the southern Great Plains, USA. Int. J. Remote Sens. 25(1):61–72 50. Norman J, Becker F (1995) Terminology in thermal infrared remote sensing of natural surfaces. Remote Sens. Rev. 12:159–173 51. Dash P, G¨ottsche FM, Olesen FS, Fischer H (2002) Land surface temperature and emissivity estimation from passive sensor data: theory and practice-current trends. Int. J. Remote Sens. 23(13):2563–2594 52. Ogawa K, Schmugge T, Jacob F, French A (2003) Estimation of land surface window (8– 12 µm) emissivity from multi-spectral thermal infrared remote sensing – a case study in a part of Sahara Desert. Geophys. Res. Lett. 30:1067–1071 53. Ogawa K, Schmugge T (2004) Mapping Surface Broadband Emissivity of the Sahara Desert Using ASTER and MODIS Data. Earth Interactions 8 Paper No. 7 54. Wang K, Wan Z, Wang P, Sparrow M, Liu J, Zhou X, Haginoya S (2005) Estimation of Surface Long Wave Radiation and Broadband Emissivity Using Moderate Resolution Imaging Spectroradiometer (MODIS) Land Surface Temperature/Emissivity Products. J. Geophys. Res. –Atmos. 110(D11, D11109) Paper No. 10.1029/2004JD005566 55. Becker F, Li Z (1995) Surface temperature and emissivity at various scale: definition, measurement and related problem. Remote Sens. Reviews 12:225–253 56. Kustas W, Norman J, Anderson M, French A (2003) Estimating subpixel surface temperatures and energy fluxes from the vegetation index-radiometric temperature relationship. Remote Sens. Environ. 85:429–440 57. Nishida K, Nemani RR, Glassy JM, Running SW (2003) Development of an Evapotranspiration Index From Aqua/MODIS for Monitoring Surface Moisture Status. IEEE Trans. Geosci. Remote Sens. 41(2):493–501 58. Luquet D, B´egu´e A, Vidal A, Clouvel P, Dauzat J, Olioso A, Gu X, Yu T (2003) Using multidirectional thermography to characterize water status of cotton. Remote Sens. Environ. 84:411–421 59. Roerink GJ, Su Z, Menenti M (2000) S-SEBI: a simple remote sensing algorithm to estimate the surface energy balance. Phys. Chem. Earth (B) 25(2):147–157 60. Gomez M, Sobrino J, Olioso A, Jacob F (2005) Retrieval of evapotranspiration over the Alpilles test site using PolDER and thermal camera data. Remote Sens. Environ. 96:399– 408 61. Kustas W, Humes K, Norman J, Moran M (1996) Single and dual source modeling of surface energy fluxes with radiometric surface temperature. J. Appl. Meteorol. 35:110–121 62. Chehbouni A, Lo Seen D, Njoku E, Monteny B (1996) Examination of the difference between radiative and aerodynamic surface temperatures over sparsely vegetated surfaces. Remote Sens. Environ. 58:177–186 63. Chehbouni A, Lo Seen D, Njoku E, Lhomme JP, Monteny B, Kerr Y (1997) Estimation of sensible heat flux over sparsely vegetated surfaces. J. Hydrol. 188–189:855–868 64. Habets F, Noilhan J, Golaz C, Goutorbe JP, Lacarr`ere P, Leblois E, Ledoux E, Martin E, Ottl´e C, Vidal-Madjar D (1999) The ISBA surface scheme in a macroscale hydro-logical model applied to the Hapex-Mobilhy area. Part I: model and database. J. hydrol. 217:75–96 65. Su Z, Menenti M, Pelgrum H, van den Hurk B, Bastiaanssen W (1999) Remote sensing of land surface fluxes for updating numerical weather predictions. In: Nieuwenhuis, Vaughan, Molenaar (eds), Operational Remote Sensing for Sustainable Development, Balkema, Rotterdam

10 Modeling and Inversion in Thermal Infrared Remote Sensing

281

66. Li J, Su Z, van den Hurk B, Menenti M, Moene A, de Bruin H, Baselga Yrisarry J, Ibanez M, Cuesta A. (2003) Estimation of sensible heat flux using the Surface Energy Balance System (SEBS) and ATSR measurements. Phys. Chem. Earth 28:75–88 67. Demarty J, Ottl´e C, Braud I, Olioso A, Frangi JP, Bastidas L, Gupta H (2004) Using a multiobjective approach to retrieve information on surface properties used in a SVAT model. J. Hydrol. 287:214–236 68. Franc¸ois C (2002) The potential of directional radiometric temperatures for monitoring soil and leaf temperature and soil moisture status. Remote Sens. Environ. 80:122–133 69. Braud I, Dantas Antonino A, Vauclin M, Thony J, Ruelle P (1995) A Simple Soil Plant Atmosphere Transfer model (SiSPAT) development and field verification. J. Hydrol. 166:213– 250 70. Boulet G, Chehbouni A, Braud I, Vauclin M (1999) Mosaic versus dual source approaches for modeling the surface energy balance of a semi-arid land. Hydrol. Earth Sci. Syst. 3(2):247–258 71. Merlin O, Chehbouni A (2004) Different approaches in estimating heat flux using dual angle observations of radiative surface temperature. Int. J. Remote Sens. 25:275–289 72. Luquet D, Vidal A, Dauzat J, B´egu´e A, Olioso A, Clouvel P (2003) Using directional TIR measurements and 3D simulations to assess the limitations and opportunities of water stress indices. Remote Sens. Environ. 90:53–62 73. Yu T, Gu X, Tian T, Legrand M, Baret F, Hanocq JF, Bosseno R, Zhang Y (2004) Modeling directional brightness temperature over a maize canopy in row structure. IEEE Trans. Geosci. Remote Sens. 42:2290–2304 74. Verhoef W, Xiao Q, Jia L, Su Z (2006) Earth observation modeling based on layer scattering matrices. IEEE Trans. Geosci. Remote Sens. (submitted). 75. Smith J Jr, Ballard, J.R (1999) Physics Based Modeling and Rendering of Vegetation in the Thermal Infrared. In: 1999 IEEE Workshop on Photometric Modeling for Computer Vision and Graphics 34 76. Guillevic P, Gastellu-Etchegorry JP, Demarty J, Pr´evot L (2003) Thermal infrared radiative transfer within three-dimensional vegetation cover. J. Geophys. Res. -Atm. 108(D8):4248, doi:10.1029/2002JD02247 77. Gastellu-Etchegorry JP, Martin E, Gascon F (2004) DART: a 3D model for simulating satellite images and studying surface radiation budget. Int. J. Remote Sens. 25(1):73–96 78. Belot A, Gastellu-Etchegorry JP, Perrier A (2005) DART-EB, a 3-D model of mass and energy transfers in vegetated canopy. Remote Sens. Environ. (submitted) 79. Seguin B, Becker F, Phulpin T, Gu X, Guyot G, Kerr Y, King C, Lagouarde J, Ottl´e C, Stoll M, Tabbagh T, Vidal A (1999) IRSUTE: a minisatellite project for land surface heat flux estimation from field to regional scale. Remote Sens. Environ. 68:357–369 80. Hernandez-Baquero ED (2000) Characterization of the Earth’s Surface and Atmosphere from Multispectral and Hyperspectral Thermal Imagery. Ph.D. thesis, Air Force Institute of Technology, Wright-Patterson, AFB, OH, 270pp. 81. French A, Norman J, Anderson M (2003) A simple and fast atmospheric correction for space-borne remote sensing of surface temperature. Remote Sens. Environ. 87:326–333 82. Sobrino J, Romaguera M (2004) Land surface temperature retrieval from MSG1-SEVIRI data. Remote Sens. Environ. 92:247–254 83. Petitcolin F, Nerry F, Stoll MP (2002) Mapping temperature independent spectral indice of emissivity and directional emissivity in AVHRR channels 4 and 5. Int. J. Remote Sens. 23:3473–3491 84. Nakajima T, Nakajima T, Nakajima M, Fukushima H, Kuji M, Uchiyama A, Kishino M (1998) Optimization of the Advanced Earth Observing Satellite (ADEOS) II Global Imager (GLI) channels by use of radiative transfer calculations. Appl. Opt. 37(15):3149–3163 85. Justice C, Vermote E, Townshend J, Defries R, Roy D, Hall D, Salomonson V, Privette J, Riggs G, Strahler A, Lucht W, Myneni R, Knyazikhin Y, Running S, Ne-mani R, Wan Z, Huete A, van Leeuwen W, Wolfe R, Giglio L, Muller JP, Lewis P, Barnsley M (1998) The MODerate Imaging Spectroradiometer (MODIS): land remote sensing for global change research. IEEE Trans. Geosci. Remote Sens. 36:1228–1249

282

F. Jacob et al.

86. Llewellyn-Jones D, Edwards M, Mutlow C, Birks A, Barton I, Tait H (2001) AATSR: global-change and surface-temperature measurements from Envisat. Eur Space Agency Bull 105:11–21 87. Goward S, Masek J, Williams D, Irons J, Thompson R (2001) The Landsat 7 mission – Terrestrial research and applications for the 21st century. Remote Sens. Environ. 78:3–12 88. Yamaguchi Y, Kahle A, Tsu H, Kawakami T, Pniel M (1998) Overview of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). IEEE Trans. Geosci. Remote Sens. 36:1282–1289 89. Becker F, Seguin B, Phulpin T, Durpaire J (1996) IRSUTE, a small satellite for water budget estimate with high resolution infrared imagery. Acta Astronautica 39:883–897 90. Lagouarde J (1997) SEXTET: a high resolution and high repetitivity infra red instrument on a mini satellite platform. Proposal to CNES call for ideas 91. Menenti M, Rast M, Baret F, van den Hurk B, Knorr W, Mauser W, Miller J, Moreno J, Schaepman M, Verstraete M.M (2004) Understanding vegetation response to climate variability from space: recent advances towards the SPECTRA Mission. In: Meynart R (ed) Proceedings of SPIE: Sensors, Systems, and Next-Generation Satellites VII. Vol 5234, pp 76–85 92. Weber PG, Brock BC, Garrett AJ, Smith BW, Borel CC, Clodius WB Bender SC, Kay RR, Decker ML (1999) MTI Mission Overview. In: Proceedings of SPIE Conference on Imaging Spectrometry V. Vol 3753, pp 340–346 93. Palluconi F, Meeks G (1985) Thermal Infrared Multi-spectral Scanner (TIMS): an investigator’s guide to TIMS data. JPL Publications N◦ 85–32 94. M¨uller A, Gege P, Cocks T (2001) The airborne imaging spectrometers used in daisex. In: DAISEX Final Results Workshop, ESTEC, Holland, 15–16 March 2001, ESA SP-499, pp 3–6 95. King MD, Menzel WP, Grant PS, Myers JS, Arnold GT, Platnick SE, Gumley LE, Tsay SC, Moeller CC, Fitzgerald M, Brown KS, Osterwisch FG (1996) Airborne scanning spectrometer for remote sensing of cloud, aerosol, water vapor and surface properties. J. Atm. Oceanic Technol. 13:777–794 96. Hook S, Myers J, Thome K, Fitzgerald M, Kahle A (2001) The MODIS/ ASTER airborne simulator (MASTER) – a new instrument for Earth science studies. Remote Sens. Environ. 76:93–102 97. Pr´evot L, Baret F, Chanzy A, Olioso A, Wigneron J, Autret H, Baudin F, Besse-moulin P, Bethenod O, Blamont D, Blavoux B, Bonnefond J, Boubkraoui S, Bouman B, Braud I, Bruguier N, Calvet J, Caselles V, Chauki H, Clevers J, Coll C, Company A, Courault D, Dedieu G, Degenne P, Del´ecolle R, Denis H, Desprats J, Ducros Y, Dyer D, Fies J, Fischer A, Francois C, Gaudu J, Gonzalez E, Gouget R, Gu X, Gu´erif M, Hanocq J, Hautecoeur J, Haverkamp R, Hobbs S, Jacob F, Jeansoulin R, Jongschaap R, Kerr Y, King C, Laborie P, Lagouarde J, Laques A, Larcena D, Laurent G, Laurent J, Leroy M, McAneney J, Macelloni G, Moulin S, Noilhan J, Ottl´e C, Paloscia S, Pampaloni P, Podvin T, Quaracino F, Roujean J, Rozier C, Ruisi R, Susini C, Taconet O, Tallet N, Thony J, Travi Y, Van Leewen H, Vauclin M, Vidal-Madjar D, Vonder O, Weiss M (1998) Assimilation of multi-sensor and multi-temporal remote sensing data to monitor vegetation and soil: the Alpilles ReSeDA project. In: Tsang L (ed) IGARSS ’98 International Geoscience and Remote Sensing Symposium, IEEE, Institute of Electrical and Electronics Engineers, Piscataway (USA), Sensing and managing the environment. Vol 5, pp 2399–2401 98. Olioso A, Pr´evot L, Baret F, Chanzy A, Braud I, Autret H, Baudin F, Bessemoulin P, Bethenod O, Blamont D, Blavoux B, Bonnefond J, Boubkraoui S, Bouman B, Bruguier N, Calvet J, Caselles V, Chauki H, Clevers J, Coll C, Company A, Courault D, Dedieu G, Degenne P, Del´ecolle R, Denis H, Desprats J, Ducros Y, Dyer D, Fies J, Fischer A, Francois C, Gaudu J, Gonzalez E, Gouget R, Gu X, Gu´erif M, Hanocq J, Hautecoeur J, Haverkamp R, Hobbs S, Jacob F, Jeansoulin R, Jongschaap R, Kerr Y, King C, Laborie P, Lagouarde J, Laques A, Larcena D, Laurent G, Laurent J, Leroy M, McAneney J, Macelloni G, Moulin S, Noilhan J, Ottl´e C, Paloscia S, Pampaloni P, Podvin T,

10 Modeling and Inversion in Thermal Infrared Remote Sensing

99.

100.

101. 102. 103. 104. 105.

106.

107.

108.

109. 110. 111. 112. 113.

114.

115. 116. 117. 118.

283

Quaracino F, Roujean J, Rozier C, Ruisi R, Susini C, Taconet O, Tallet N, Thony J, Travi Y, Van Leewen H, Vauclin M, Vidal-Madjar D, Vonder O, Weiss M, Wigneron J (1998) Spatial aspects in the Alpilles-ReSeDA project. In: Marceau D (ed.) Scaling and Modeling in Forestry: Application in Remote Sensing and GIS, Ed. D Marceau, Universit´e de Montr´eal, Qu´ebec pp 92–102 Baret F (2000) ReSeDA, assimilation of multisensor and multitemporal remote sensing data to monitor soil and vegetation functioning. Final Report EC project ENV4CT960326, INRA, France, 59p. National Research Council (2004) Committee on NASA-NOAA Transition from Research to Operations: Satellite Observations of the Earth’s Environment: Accelerating the Transition of Research to Operations. National Academies Press, Washington, DC Yu Y, Privette J, Pinheiro A (2005) Analysis of the NPOESS VIIRS Land Surface Temperature Algorithm Using MODIS Data. IEEE Trans. Geosci. Remote Sens. 43:2340–2350 Vaughan R, Calvin W, Taranik J (2003) SEBASS hyperspectral thermal infrared data: surface emissivity measurement and mineral mapping. Remote Sens. Environ. 85:48–63 Liu Q, Gu X, Li X, Jacob F, Hanocq JF (2000) Study on airborne experiment for directional patterns of thermal infrared radiation. Science in China(E) 30:89–97 Payan V, Royer A (2004) Analysis of the Temperature Emissivity Separation (TES) algorithm applicability and sensitivity. Int. J. Remote Sens. 25:15–37 Rubio E, Caselles V, Badenas C (1997) Emissivity measurements of several soils and vegetation types in the 8–14 µm wave band: analysis of two fields method. Remote Sens. Environ. 59:490–521 Kant Y, Badarinath K (2002) Ground-based method for measuring thermal infrared effective emissivity and perspectives on the measurement of land surface temperature from satellite data. Int. J. Remote Sens. 23(11):2179–2191 Rubio E, Caselles V, Coll C, Valor E, Sospedra F (2003) Thermal infrared emissivities of natural surfaces: improvement on the experimental set-up and new measurements. Int. J. Remote Sens. 24(24):5379–5390 Li ZL, Zhang R, Sun X, Su H, Tang X, Zhu Z, Sobrino JA (2004) Experimental system for the study of the directional thermal emission of natural surfaces. Int. J. Remote Sens. 25(1):195–204 Coret L, Briottet X, Kerr Y, Chehbouni G (2005) Experimental study of directional aggregation over heterogeneous surfaces in the thermal infrared. Remote Sens. Environ. (submitted). Garrat J, Hicks B (1973) Momentum, heat and water vapor transfer to and from natural and artificial surfaces. Quart. J. Roy. Meteorol. Soc. 99:680–687 Ham J, Heilman J (1991) Aerodynamic land surface resistances affecting energy transport in a sparse crop. Agric. Forest Meteorol. 53:267–284 Garrat J (1978) Transfer characteristics for a heterogeneous surface of large aerodynamic roughness. Quart. J. Roy. Meteorol. Soc. 104:491–502 Troufleau D, Lhomme J, Monteny B, Vidal A (1997) Sensible heat flux and radiometric surface temperature over sparse Sahelian vegetation. I. An Experimental analysis of the kB parameter. J. Hydrol. 188–189:815–838 Choudhury B (1989) Estimating evaporation and carbon assimilation using infrared temperature data: vistas in modeling. In: Theory and Applications of Optical Remote Sensing. Wiley Interscience, Paris, France, pp 629–690 Norman J, Divakarla M, Goel N (2000) Algorithms for Extracting Information from Remote Thermal-IR Observations of the Earth’s Surface. Remote Sens. Environ. 51:157–168 Martonchik JV, Bruegge CJ, Strahler AH (2000) A Review of Reflectance Nomenclature for Remote Sensing. Remote Sens. Rev. 19:9–20 Li X, Strahler A, Friedl M (1999) A conceptual model for effective directional emissivity from non-isothermal surfaces. IEEE Trans. Geosci. Remote Sens. 37:2508–2517 Franc¸ois C, Ottl´e C, Pr´evot L (1997) Analytical parameterization of canopy directional emis-sivity and directional radiance in the thermal infrared. Application on the retrieval of soil and foliage temperatures using two directional measurements. Int. J. Remote Sens. 18:2587–2621

284

F. Jacob et al.

119. Chehbouni A, Nouvellon Y, Kerr Y, Moran M, Watts C, Pr´evot L, Goodrich D, Rambal S (2001) Directional effect on radiative surface temperature measurements over a semi-arid grassland site. Remote Sens. Environ. 76:360–372 120. Olioso A, Soria G, Sobrino J, Duchemin B (2006) Evidences of low land surface thermal infrared emissivity in presence of dry vegetation. IEEE Geosci. Remote Sens. Lett. 4(1):112– 116 121. Sutherland R, Bartholic J (1977) Significance of vegetation in interpreting thermal radiation from a terrestrial surface. J. Appl. Meteorol. 16:759–763 122. Colton A (1996) Effective thermal parameters for a heterogeneous lans surface. Remote Sens. Environ. 57:143–160 123. Barducci A, Pippi Y (1996) Temperature and emissivity retrieval from remotely sensed images using the “Grey Body Emissivity” method. IEEE Trans. Geosci. Remote Sens. 34:681– 695 124. Valor E, Caselles V (1996) Mapping land surface emissivity from NDVI: application to European, African and South American areas. Remote Sens. Environ. 57:167–184 125. Sobrino J, S`oria G, Prata AJ (2004) Surface temperature retrieval from Along Track Scanning Radiometer 2 data: Algorithms and validation. J. Geophys. Res. 109D11101, doi:10.1029/2003JD004212 126. S`oria G, Sobrino J (2006) Envisat/AATSR derived land surface temperature over a heterogeneous region. Remote Sens. Environ. 111(4):409–422 127. Chen L, Li ZL, Liu Q, Chen S, Tang Y, Zhong B (2004) Definition of component effective emissivity for heterogeneous and non-isothermal surfaces and its approximate calculation. Int. J Remote Sens. 25(1):231–244 128. Su L, Li X, Friedl M, Strahler A, Gu X (2002) A kernel-driven model of effective directional emissivity for non-isothermal surfaces. Prog. Natural Sci. 12(8):603–607 129. Snyder W, Wan Z (1998) BRDF models to predict spectral reflectance and emissivity in the infrared. IEEE Trans. Geosci. Remote Sens. 36:214–225 130. Roujean J, Leroy M, Deschamps P (1992) A bidirectional reflectance model of the Earth’s surface for the correction of remote sensing data. J. Geophys. Res. 97:20455–20468 131. Wanner W, Li X, Strahler A (1995) On the derivation of kernels for kernel-driven models of bidirectional reflectance. J. Geophys. Res. 100:21077–21089 132. Engelsen O, Pinty B, Verstraete M, Martonchik J (1996) Parametric bidirectional reflectance factor models : evaluation, improvements and applications. Report EUR16426EN, European Commission, Joint Researches Center, Space Application Institute, ISPRA, Italy 133. Berk A, Bernstein L, Anderson G, Acharya P, Robertson D, Chetwynd J, Adler-Golden S (1998) MODTRAN cloud and multiple scattering upgrades with application to AVIRIS. Remote Sens. Environ. 65:367–375 134. Anderson G, Berk A, Acharya P, Matthew M, Bernstein L, Chetwynd J, Dothe H, AdlerGolden S, Ratkowski A, Felde G, Gardner J, Hoke M, Richtsmeier S, Pukall B, Mello J, Jeong L (2000) MODTRAN 4.0: radiative transfer modeling for remote sensing. In: Algorithms for multispectral, hyperspectral and ultraspectral imagery VI: Proceedings of SPIE , pp 176–183 135. Verhoef, W. (1984) Light scattering by leaf layers with application to canopy reflectance modeling: the SAIL model. Remote Sens. Environ. 16:125–141 136. Olioso A (1995) Simulating the relationship between thermal infrared emissivity and the Normalized Difference Vegetation Index. Int. J. Remote Sens. 16:3211–3216 137. Olioso A, Chauki H, Courault D, Wigneron J (1999) Estimation of evapotranspiration and photo-synthesis by assimilation of remote sensing data into SVAT models. Remote Sens. Environ. 68:341–356 138. Kimes D (1980) Effects of vegetation canopy structure on remotely sensed canopy temperatures. Remote Sens. Environ. 10:165–174 139. Pr´evot L (1985) Mod´elisation des e´ changes radiatifs au sein des couverts v´eg´etaux – Application a` la t´el´ed´etection – Validation sur un couvert de Ma¨ıs. Ph.D. thesis, p 185, Universit´e Paris VI, Paris, France

10 Modeling and Inversion in Thermal Infrared Remote Sensing

285

140. Jackson R, Reginato R, Pinter P (1979) Plant canopy information extraction from composite site reflectance of row crop. Appl. Opt. 18:3775–3782 141. Kimes D, Kirchner J (1983) Directional radiometric measurements of row – crop temperatures. Int. J. Remote Sens. 4:299–311 142. Sobrino JA, Caselles V (1990) Thermal infrared radiance model for interpreting the directional radiometric temperature of a vegetative surface. Remote Sens. Environ. 33:193–199 143. Caselles V, Sobrino JA, Coll C (1992) A physical model for interpreting the land surface temperature obtained by remote sensors over incomplete canopies. Remote Sens. Environ. 39:203–211 144. Coret L, Briottet X, Kerr Y, Chehbouni G (2004) Simulation study of view angle effects on thermal infrared measurements over heterogeneous surfaces. IEEE Trans. Geosci. and Remote Sens. 42(3):664–672 145. Du Y, Liu Q, Chen L, Liu Q, Yu, T (2007) Modeling directional brightness temperature of the winter wheat canopy at the ear stage. IEEE Trans. Geosci. Remote Sens. 45(11):3721– 3739 146. Norman J, Campbell G (1983) Application of a plant-environment model to problems in irrigation. In: Hillel D (ed.) Advances in irrigation, Academic Press, New York, pp 155–188 147. Norman J, Arkebauer T (1991) Predicting canopy light-use efficiency from leaf characteristics. In: Ritchie J, Hanks R (eds), Modeling plant and soil systems, ASA, Madison, WI, pp 125–143 148. Dauzat J, Rapidel B, Berger A (2001) Simulation of leaf transpiration and sap flow in virtual plants: description of the model and application to a coffee plantation in Costa Rica. Agric. Forest Meteorol. 109:143–160 149. Jia L (2004) Modeling heat exchanges at the land atmosphere interface using multi-angular thermal infrared measurements. Ph.D. thesis, Wageningen University, Wageningen, The Netherlands 150. Su H, Zhang R, Tang XZ, Sun X (2000) Determination of the effective emissivity for the regular and irregular cavities using Monte-carlo method. Int. J. Remote Sens. 21(11):2313– 2319 151. Su L, Li X, Liang S, Strahler A (2003) Simulation of scaling effects of thermal emission from non-isothermal pixels with the typical three dimension structure. Int. J. Remote Sens. 24(19):3743–3753 152. Kimes D, Knyazikhin Y, Privette J, Abuelgasim A, Gao F (2000) Inversion methods for physically-based models. Remote Sens. Rev. 18:381–439 153. Chen J, Li X, Nilson T, Strahler A (2000) Recent advances in geometrical optical modeling and its applications. Remote Sens. Rev. 18:227–262 154. Qin W, Liang S (2000) Plane-parallel canopy radiation transfer modeling: recent advances and future directions. Remote Sens. Rev. 18:281–305 155. Nolin A, Liang S (2000) Progress in bidirectional reflectance modeling and applications for surface particulate media: snow and soils. Remote Sens. Rev. 18:307–342 156. Lucht W, Roujean J (2000) Considerations in the parametric modeling of BRDF and albedo from multiangular satellite sensor observations. Remote Sens. Rev. 18:343–379 157. Schmugge T, French A, Ritchie J, Rango A, Pelgrum H (2002) Temperature and emissivity separation from multispectral thermal infrared observations. Remote Sens. Environ. 79:189– 198 158. Jacob F, Petitcolin F, Schmugge T, Vermote E, Ogawa K, French A (2004) Comparison of land surface emissivity and radiometric temperature from MODIS and ASTER sensors. Remote Sens. Environ. 83:1–18 159. Hall F, Huemmrich K, Goetz S, Sellers P, Nickeson P (1992) Satellite remote sensing of surface energy balance : success, failures and unresolved issues in FIFE. J. Geophys. Res. 97:19061–19089 160. Bolle H, et al. (1993) EFEDA: European Field in Desertification threatened Area. Annales Geophysicae 11:173–189

286

F. Jacob et al.

161. Goutorbe J, et al. (1994) HAPEX-SAHEL: a large scale study of land-atmosphere interactions in the semiarid tropics. Annales Geophysicae 12: 53–64 162. Havstad K, Kustas W, Rango A, Ritchie J, Schmugge T (2000) Jornada Experimental Range: a unique arid land location for experiments to validate satellite system. Remote Sens. Environ. 74:13–25 163. Baldocchi D, Falge E, Gu LH, Olson R, Hollinger D, Running S, Anthoni P, Bern-hofer C, Davis K, Evans R, Fuentes J, Goldstein A, Katul G, Law B, Lee XH, Malhi Y, Meyers T, Munger W, Oechel W, Paw UKT, Pilegaard K Schmid HP, Valentini R, Verma S (2001) FLUXNET: a new tool to study the temporal and spatial variability of ecosystem-scale carbon dioxide, water vapor, and energy flux densities. Bull. Am. Meteorol. Soc. 82-11:2415– 2434 164. Berger M, Rast M, Wursteisen P, Attema E, Moreno J, Mueller A, Beisl U, Richter R, Schaepman M, Strub G, Stoll MP, Nerry F, Leroy M (2001) The DAISEX campaigns in support of a future land-surface-processes mission ESA Bulletin. Remote Sens. Environ. 105:101–111 165. Goodrich DC, Chehbouni A, Goff BF, Macnish R, Maddock T, Moran MS, Shut-tleworth WJ, Williams DG, Watts CJ, Hipps LJ, Cooper DI, Schieldge J, Kerr YH, Arias H, Kirkland M, Carlos R, Cayrol P, Kepner W, Jones B, Avissar R, Begue A, Bonnefond JM, Boulet G, Branan B, Brunel JP, Chen LC, Clarke T, Davis MR, Debruin H, Dedieu G, Elguero E, Eichinger WE, Everitt J, Garatuza-Payan J, Gempko VL, Gupta H, Harlow C, Hartogensis O, Helfert M, Holifield C, Hymer D, Kahle A, Keefer T, Krishnamoorthy S, Lhomme JP, Lagouarde JP, Lo Seen D, Luquet D, Marsett R, Monteny B, Ni W, Nouvellon Y, Pinker R, Peters C, Pool D, Qi J, Rambal S, Rodriquez J, Santiago F, Sano E, Schaeffer SM, Schulte M, Scott R, Shao X, Snyder KA, Sorooshian S, Unkrich CL, Whitaker M, Yucel I (2000) Preface paper to the Semi-Arid Land-Surface- Atmosphere (SALSA) Program special issue. Agric. Forest Meteorol. 105:320 166. Njoku E, Lakshmi V, O’Neill P (2004) Soil moisture field experiment special issue. Remote Sens. Environ. 92:425–426 167. Schmugge T, Hook S, Coll C (1998) Recovering surface temperature and emissivity from thermal infrared multispectral data. Remote Sens. Environ. 65:121–131 168. Coll C, Caselles V, Rubio E, Sospedra F, Valor E (2000) Temperature and emissivity separation from calibrated data of the Digital Airborne Imaging Spectrometer. Remote Sens. Environ. 76:250–259 169. Buongiorno M, Realmuto V, Doumaz F (2002) Recovery of spectral emissivity from thermal infrared multispectral scanner imagery acquired over a mountainous terrain: a case study from mount Etna, Sicily. Remote Sens. Environ. 79:123–133 170. Coll C, Caselles V, Rubio E, Valor E, Sospedra F, Baret F, Pr´evot L, Jacob F (2002) Temperature and emissivity extracted from airborne multi-channel data in the ReSeDA experiment. Agronomie: Agric. Environ. 22:567–574 171. Jacob F, Gu X, Hanocq JF, Baret F (2003) Atmospheric corrections of single broadband channel and multidirectional airborne thermal infrared data. Application to the ReSeDA Experiment. Int. J. Remote Sens. 24:3269–3290 172. Coll C, Caselles V, Valor E, Rubio E (2003) Validation of temperature-emissivity separation and split-window methods from TIMS data and ground measurements. Remote Sens. Environ. 85:232–242 173. Coll C, Valor E, Caselles V, Nicl`os R (2003) Adjusted Normalized Emissivity Method for surface temperature and emissivity retrieval from optical and thermal infrared remote sensing data. J. Geophys. Res. 108(D23):4739, doi:10.1029/2003JD003688 174. Sobrino J, Jimenez-Munoz J, El-Kharraz J, Gomez M, Romaguera M, S`oria G (2004) Single-channel and two-channel methods for land surface temperature retrieval from DAIS data and its application to the Barrax site. Int. J. Remote Sens. 25(1):215–230 175. Schmugge T, Jacob F, French A, Ogawa K (2002) Quantitative estimates of emissivity using ASTER data. In: Sobrino J (ed.) Proceedings of the First International Symposium on Recent Advances in Quantitative Remote Sensing, September 2002, Valencia, Spain, pp 585–589

10 Modeling and Inversion in Thermal Infrared Remote Sensing

287

176. Schmugge T, Ogawa K, Jacob F, French A, Hsu Y, Ritchie J, Rango A (2003) Validation of emissivity estimates from ASTER data. In: Proceedings of 2003 International Geoscience and Remote Sensing Symposium, Toulouse, France, 21–25 July 2003. Vol III , pp 1873– 1875 177. Sobrino J, Jimenez-Munoz J, Paolini L (2004) Land surface temperature retrieval from Land-sat TM 5. Remote Sens. Environ. 90:434–440 178. Li F, Jackson T, Kustas W, Schmugge T, French A, Cosh M, Bindlish R (2004) Deriving land surface temperature from Landsat 5 and 7 during SMEX02/SMACEX. Remote Sens. Environ. 92:521–534 179. Jimenez-Munoz JC, Sobrino JA, Gillespie A, Sabol D, Gustafson TG (2006) Improved land surface emissivities over agricultural areas using ASTER NDVI. Remote Sens. Environ. 103:474–487 180. Jimenez-Munoz JC, Sobrino JA (2003) A generalized single-channel method for retrieving land surface temperature from remote sensing data. J. Geophys. Res. 108(D22):4688, doi:10.1029/2003JD003480 181. Wan Z, Zhang Y, Zhang Q, Li ZL (2002) Validation of the land surface temperature products retrieved from Terra Moderate Resolution Imaging Spectroradiometer data. Remote Sens. Environ. 83:163–180 182. Wan Z, Zhang Y, Zhang Q, Li ZL (2004) Quality assessment and validation of the MODIS global land surface temperature. Int. J. Remote Sens. 25:261–274 183. Coll C, Caselles V, Galve JM, Valor E, Nicl`os R, S´anchez J, Rivas R (2005) Ground measurements for the validation of land surface temperatures derived from AATSR and MODIS data. Remote Sens. Environ. 97:288–300 184. Weiss M, Baret F, Garrigues S, Lacaze R (2007) LAI and fAPAR CYCLOPES global products derived from VEGETATION. Part2: validation and comparison with MODIS collection 4 products. Remote Sens. Environ. 110(3):317–331 185. Tonooka H (2001) An atmospheric correction algorithm for thermal infrared multispectral data over land – a water vapor scaling method. IEEE Trans. Geosci. Remote Sens. 39:682– 692 186. Dash P, G¨ottsche FM, Olesen FS (2002) Potential of MSG for surface temperature and emissivity estimation: considerations for real-time applications. Int. J. Remote Sens. 23(20):4511–4518 187. Peres L, DaCamara C (2004) Land surface temperature and emissivity estimation based on the two-temperature method: sensitivity analysis using simulated MSG/SEVIRI data. Remote Sens. Environ. 91:377–389 188. Dash P, G¨ottsche FM, Olesen FS, Fischer H (2005) Separating surface emissivity and temperature using two-channel spectral indices and emissivity composites and comparison with a vegetation fraction method. Remote Sens. Environ. 96:1–17 189. Jia L, Li Z, Menenti M, Su Z, Verhoef W, Wan Z (2003) A practical algorithm to infer soil and foliage component temperatures from bi-angular ATSR-2 data. Int. J. Remote Sens. 24:4739–4760 190. Kratz DP, Chou MD, Yan MMH, Ho CH (1998) Minor Trace Gas Radiative Forcing Calculations Using the k Distribution Method with One-Parameter Scaling. J. Geophys. Res. 103(D24):31647–31656 191. Saunders R, Brunel P, English S, Bauer P, OKeeffe U, Francis P, Rayer P (2005) Rttov-8, science and validation report. Technical report, EUMESAT (2005) report NWPSAF-MOTV-007, v1.2, 02/03/2005 192. Gottsche FM, Olesen F (2002) Evolution of neural networks for radiative transfer calculations in the terrestrial infrared. Remote Sens. Environ. 80:157–164 193. Derber J, Parrish D, Lord S (1991) The new global operational analysis system at the National Meteorological Center (NMC). Weather Forecast. 6:538–547 194. Derber J, Wu W (1998) The use of TOVS cloud-cleared radiances in the NCEP SSI analysis system. Monthly Weather Rev. 126:2287–2299

288

F. Jacob et al.

195. Schroedter M, Olesen FS, Fischer H (2003) Determination of land surface temperature distributions from single channel IR measurements: an effective spatial interpolation method for the use of TOVS, ECMWF, and radiosonde profiles in the atmospheric correction scheme. Int. J. Remote Sens. 24:1189–1196 196. Hirano A, Welch R, Lang H (2002) Mapping from ASTER stereo image data: DEM validation and accuracy assessment. ISPRS J. Photogramm. Remote Sens. 55:1–15 197. Gu D., Gillespie AR, Kahle AB, Palluconi FD (2000) Autonomous atmospheric compensation (AAC) of high resolution hyperspectral thermal infrared remote sensing imagery. IEEE Trans. Geosci. Remote Sens. 38:2557–2569 198. Li ZL, Li J, Su Z, Wan Z, Zhang R (2003) A new approach for retrieving precipitable water from ATSR2 split-window channel data over land area. Int. J. Remote Sens. 24(24):5095– 5117 199. Sobrino J, El Kharraz J, Li ZL (2003) Surface temperature and water vapor retrieval from MODIS data. Int. J. Remote Sens. 24(24):5161–5182 200. De Felice T, LLoyd D, Meyer D, Baltzer T, Piraino P (2003) Water vapour correction of the daily 1 km AVHRR global land dataset: part I – validation and use of the water vapour input field. Int. J. Remote Sens. 24(11):2365–2375 201. Rabier F, Fourri´e N, Chafa¨ı D, Prunet P (2001) Channel selection methods for infrared atmospheric sounding interferometer radiances. Quart. J. Roy. Meteorol. Soc. 128:1–17 202. Pinheiro A, Privette J, Mahoney R, Tucker C (2004) Directional effects in a daily AVHRR land surface temperature dataset over Africa. IEEE Trans. Geosci. Remote Sens. 42(9):1941–1954 203. van de Griend A, Owe M (1993) On the relationship between thermal infrared emissivity and the Normalized Difference Vegetation Index for natural surfaces. Int. J. Remote Sens. 14 :1119–1131 204. Kerr Y, Lagouarde J, Nerry F, Ottl´e C (2004) Land surface temperature retrieval techniques and applications: case of the AVHRR. In: Thermal remote sensing in land surface processes. Boca Raton, FL: CRC, pp 33–109 + 14 colour plates 205. Snyder W, Wan Z, Zhang Y, Feng YZ (1998) Classification-based emissivity for land surface temperature measurement from space. Int. J. Remote Sens. 19:2753–2774 206. Payan V, Royer A (2004) Spectral emissivity of northern land cover types derived with MODIS and ASTER sensors in MWIR and LWIR. Research Note, Can. J. Remote Sens. 30(2):150–156 207. Peres L, DaCamara C (2005) Emissivity maps to retrieve land-surface temperature from MSG/SEVIRI. IEEE Trans. Geosci. Remote Sens. 43(8):1834–1844 208. Gillespie A (1985) Lithologic mapping of silicate rocks using TIMS. In: The TIMS data user workshop, June 18–19, 1985. JPL Publication N◦ 86-38, Jet Propulsion Laboratory, Pasadena, CA (1985), pp 29–44 209. Mushkin A, Balick LK, Gillespie AR (2005) Extending surface temperature and emissivity retrieval to the mid-infrared (3–5 µ m) using the Multispectral Thermal Imager (MTI). Remote Sens. Environ. 98 141–151 210. Gillespie A, Rokugawa S, Matsunaga T, Cothern S, Hook S, Kahle A (1998) A temperature and emissivity separation algorithm for Advanced Spaceborne Thermal Emission and Reflection radiometer (ASTER) images. IEEE Trans. Geosci. Remote Sens. 36:1113–1126 211. Liang S (2001) An optimization algorithm for separating land surface temperature and emissivity from multispectral thermal infrared imagery. IEEE Trans. Geosci. Remote Sens. 39:264–274 212. Mushkin A, Balick LK, Gillespie AR (2002) Temperature/emissivity separation of MTI data using the Terra/ASTER TES algorithm. In: Shen SS, Lewis PE (eds) (2002) Proceedings of SPIE – Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery VIII. Vol 4725, pp 328–337 213. Gu D, Gillespie AR (2000) A new approach for temperature and emissivity separation. Int. J. Remote Sens. 21:2127–2132

10 Modeling and Inversion in Thermal Infrared Remote Sensing

289

214. Ogawa K, Schmugge T, Jacob F, French A (2002) Estimation of broadband land surface emissivity from multispectral thermal infrared remote sensing. Agronomie: Agric. Environ. 22:695–696 215. Peres L, DaCamara C (2004) Inverse problems theory and application: analysis of the twotemperature method for land-surface temperature and emissivity estimation. Geosci. Remote Sens. Lett. 1(3):206–210 216. Wan Z, Li ZL (1997) A physics-based algorithm for retrieving land-surface emissivity and temperature from EOS/MODIS data. IEEE Trans. Geosci. Remote Sens. 35:980–996 217. Watson K (1992) Two-temperature method for measuring emissivity. Remote Sens. Environ. 42:117–121 218. Boyd D, Petitcolin F (2004) Remote sensing of the terrestrial environment using middle infrared radiation (3.0–5.0 µ m). Int. J. Remote Sens. 25(17):3343–3368 219. Huntingford C, Verhoef A, Stewart J (2000) Dual versus single source models for estimating surface temperature of African Savannah. Hydrol. Earth Sci. Syst. 4(1):185–191 220. Kustas W, Choudhury B, Moran M, Reginato R, Jackson R, Gay L, Weaver H (1989) Determination of sensible heat flux over sparse canopy using thermal infrared data. Agric. Forest Meteorol. 44:197–216 221. Moran M, Kustas W, Vidal A, Stannard D, Blanford J, Nichols W (1994) Use of groundbased remotely sensed data for surface energy balance evaluation of a semiarid rangeland. Water Res. Res. 30:1339–1349 222. Stewart J, Kustas W, Humes K, Nichols W, Moran M, de Bruin H (1994) Sensible heat flux – radiometric surface temperature relationship for eight semiarid areas. J. Appl. Meteorol. 33:1110–1117 223. Sun J, Mahrt L (1995) Determination of surface fluxes from the surface radiative temperature. J. Atmos. Sci. 52(8):1096–1106 224. Cahill A, Parlange M (1997) On the Brutsaert temperature roughness length model for sensible heat flux estimation. Water Res. Res. 10:2315–2324 225. Lhomme J, Troufleau D, Monteny B, Chehbouni A, Bauduin S (1997) Sensible heat flux and radiometric surface temperature over sparse Sahelian vegetation II. A model for the kB−1 parameter. J. Hydrol. 188–189:839–854 226. Verhoef A, De Bruin H, Van Den Hurk B (1997) Some practical notes on the kB−1 . J. Appl. Meteorol. 36:560–572 227. Qualls R, Yates D (2001) Directional radiometric temperature profiles within a grass canopy. Adv. Water Res. 24:145–155 228. Molder M, Lindroth A (2001) Dependence of kB−1 factor on roughness Reynolds number for barley and pasture. Agric. Forest Meteorol. 106:147–152 229. Molder M, Kellner E (2002) Excess resistance of bog surfaces in central Sweden. Agric. Forest Meteorol. 112:23–30 230. Pr´evot L, Brunet Y, Paw UK, Seguin B (1993) Canopy modeling for estimating sensible heat flux from thermal infrared measurements. In: Workshop on Thermal Remote Sensing of the Energy Balance Over Vegetation in Conjunction with Other Sensor, La Londe les Maures, 20–23 September. Vol 93, pp 17–30 231. Norman J, Kustas W, Humes K (1995) Source approach for estimating soil and vegetation energy fluxes in observations of directional radiometric surface temperature. Agric. Forest Meteorol. 77:263–293 232. Kustas W, Norman J (1999) Evaluation of soil and vegetation heat flux predictions using a simple two-source model with radiometric temperatures for partial canopy cover. Agric. Forest Meteorol. 94:13–29 233. Kustas W, Norman J (2000) A two source energy balance approach using directional radiometric temperature observations for sparse canopy covered surfaces. Agron. J. 92:847–854 234. Olioso A (1995) Estimating the difference between brightness and surface temperatures for a vegetal canopy. Agric. Forest Meteorol. 72:237–242 235. Lagouarde JP, Kerr Y, Brunet Y (1995) An experiment study of angular effects on surface temperature for various plant canopies and bare soils. Agric. Forest Meteorol. 77:167–190

290

F. Jacob et al.

236. Verbrugghe M, Cierniewski J (1998) Influence and modeling of view angles and micro-relief on surface temperature measurements of bare agricultural soils. ISPRS J. Photogramm. Remote Sens. 53:166–173 237. Sobrino J, Cuenca J (1999) Angular variation of thermal infrared emissivity for some natural surfaces from experimental measurements. Appl. Opt. 38(18):3931–3936 238. Weiss M, Baret F, Leroy M, Hautecoeur O, Bacour C, Pr´evot L, Bruguier N (2002) Evaluation of Neural Network techniques to estimate canopy biophysical variables from remote sensing data. Agronomie: Agric. Environ. 22:547–553 239. Bastiaanssen W, Menenti M, Feddes R, Holtslag A (1998) A remote sensing surface energy balance algorithm for land (SEBAL). I: formulation. J. Hydrol. 212–213:198–212 240. Su Z, Yacob A, Wen J, Roerink G, He Y, Gao B, Boogaard H, van Diepen C (2003) Assessing soil moisture with remote sensing data: theory, experimental validation, and application to drought monitoring over the North China Plain. Phys. Chem. Earth 28:89–101 241. Brisson N, Gary C, Justes E, Roche R, Mary B, Ripoche D, Zimmer D, Sierra J, Bertuzzi P, Burger P, Bussi`ere F, Cabidoche Y, Cellier P, Debaeke P, Gaudill`ere J, H´enault C, Maraux F, Seguin B, Sinoquet H (2003) An overview of the crop model STICS. Eur. J. Agron. 18:309– 332 242. Mougin E, Lo Seen D, Rambal S, Gaston A, Hiernaux P (2004) A regional sahelian grassland model to be coupled with multispectral satellite data. I: model description and validation. Remote Sens. Environ. 52:181–193 243. Pellenq J, Boulet G (2004) A methodology to test the pertinence of remote-sensing data assimilation into vegetation models for water and energy exchange at the land surface. Agronomie 24:197–204 244. Lauvernet C, Baret F, Le Dimet FX (2003) Assimilating high temporal frequency SPOT data to describe canopy functioning: the ADAM project. In: Geoscience and Remote Sensing Symposium, 2003. IGARSS ’03. Proceedings. 2003 IEEE International. Vol 5, pp 3184– 3186 245. Del´ecolle R, Maas S, Gu´erif M, Baret F (1992) Remote sensing and crop production model: present trends. ISPRS J. of Photogramm. Remote Sens. 47:145–161 246. Weiss M, Troufleau D, Baret F, Chauki H, Pr´evot L, Olioso A, Bruguier N, Brisson N (2001) Coupling canopy functioning and radiative transfer models for remote sensing data assimilation. Agric. Forest Meteorol. 108:113–128 247. Moulin S, Kergoat L, Cayrol P, Dedieu G, Pr´evot L (2002) Calibration of a coupled canopy functioning and SVAT model in the ReSeDA experiment. Towards the assimilation of SPOT/HRV observations into the model. Agronomie: Agric. Environ. 22:681–686 248. Pr´evot L, Chauki H, Troufleau D, Weiss M, Baret F, Brisson N (2003) Assimilating optical and radar data into the STICS crop model for wheat. Agronomie: Agric. Environ. 23:297– 303 249. Ottl´e C, Vidal-Madjar D (1994) Assimilation of soil moisture inferred from infrared remote sensing in a hydrological model over the HAPEX-MOBILHY region. J. Hydrol. 158:241– 264 250. van den Hurk B, Bastiaanssen W, Pelgrum H. van Meijgaard E (1997) A new methodology for assimilation of initial soil moisture fields in weather prediction models using METEOSAT and NOAA data. J. Appl. Meteorol. 36:1271–1283 251. Chehbouni A, Njoku E, Lhomme JP, Kerr Y (1995) Approaches for averaging surface parameters and fluxes over heterogeneous terrain. J. Climate 8:1386–1393 252. Njoku E, Hook S, Chehbouni A (1996) Effects of surface heterogeneity on thermal remote sensing of land parameters, Chapter 2. In: Scaling up in hydrology using remote sensing. John, West Sussex, pp 19–31 253. Su Z, Pelgrum H, Menenti M (1999) Aggregation effects of surface heterogeneity in land surface processes. In: Su Z, Menenti M (eds), Hydrology and Earth Science System. Vol 3, pp 549–563 254. Bouguerzaz FA, Olioso A, Raffy M (1999) Modeling radiative and energy balance on heterogeneous areas from measured radiances. Can. J. Remote Sens. 25(4):412–424

10 Modeling and Inversion in Thermal Infrared Remote Sensing

291

255. Kustas W, Norman J (2000) Evaluating the effects of subpixel heterogeneity on pixel average fluxes. Remote Sens. Environ. 74:327–342 256. Chehbouni A, Watts C, Kerr Y, Dedieu G, Rodriguez JC, Santiago F, Cayrol P, Boulet G, Goodrich D (2000) Methods to aggregate turbulent fluxes over heterogeneous surfaces: application to SALSA data set in Mexico. Agric. Forest Meteorol. 105:133–144 257. French A, Schmugge T, Kustas W, Brubaker K, Prueger J (2003) Surface energy fluxes over El Reno, Oklahoma using high resolution remotely sensed data. Water Res. Res. 39:1164 258. Brunsell N, Gillies R (2003) Scale issues in land atmosphere interactions: implications for remote sensing of the surface energy balance. Agric. Forest Meteorol. 117:203–221 259. Kustas W, Li F, Jackson T, Prueger J, MacPherson J, Wolde M (2004) Effects of remote sensing pixel resolution on modeled energy flux variability of croplands in Iowa. Remote Sens. Environ. 92:535–547 260. Lyons T, Halldin S (2004) Surface heterogeneity and the spatial variation of fluxes. Agric. Forest Meteorol. 121:153–165 261. Norman J, Anderson M, Kustas W, French A, Mecikalski J, Torn R, Diak G, Schmugge T, Tanner B (2003) Remote sensing of surface energy fluxes at 10−1 m pixel resolutions. Water Res. Res. 39:1221 262. Cardot H, Faivre R, Goulard M (2003) Functional approaches for predicting land use with the temporal evolution of coarse resolution remote sensing data. J. Appl. Stat. 30(10):1185– 1199 263. Anderson M, Norman J, Diak G, Kustas W, Mecikalski J (1997) A two-source time integrated model for estimating surface fluxes using thermal infrared remote sensing. Remote Sens. Environ. 60:195–216 264. Guillevic P (1999) Mod´elisation des bilans radiatif et e´ nerg´etique des couverts v´eg´etaux. Ph.D. thesis, Universit´e Paul Sabatier – Toulouse III, 181pp. 265. Schmugge T, Ogawa K (2006) Validation of Emissivity Estimates from ASTER and MODIS Data. In: Geoscience and Remote Sensing Symposium, 2006. IGARSS ’06. Proceedings. 2006 IEEE International. Vol I, pp 260–262

Chapter 11

Spectrally Consistent Pansharpening Ari Vesteinsson, Henrik Aanaes, Johannes R. Sveinsson, and Jon Atli Benediktsson

Abstract Several widely used methods have been proposed for fusing high resolution panchromatic data and lower resolution multi-channel data. However, many of these methods fail to maintain spectral consistency of the fused high resolution image, which is of high importance to many of the applications based on satellite data. Additionally, most conventional methods are loosely connected to the image forming physics of the satellite image, giving these methods an ad hoc feel. In this chapter, a method for image fusion of satellite images is given. The method is based on the properties of imaging physics in a statistically meaningful way. Based on our analysis, it is seen that spectral consistency is a direct consequence of imaging physics and hence guaranteed by our method. This is achieved while exploiting the high resolution single-channel data in what can be seen as a statistical optimal way, yielding a framework to which additional constraints can be added in a straight forward manner. In this chapter we exploit this framework and add some simple optimization terms for smoothing the fused image. Specifically, the method is based on the observation that any given channel of the satellites imaging device can be seen as an inner-product between the radiated light arriving at the sensor and the spectral response function of that channel. This gives a simple inner product space encompassing the relationship between the different channels as well as imposing spectral consistency. Normal distributed statistics – inducing the same norm as the above mentioned inner product – are used for regularization. This yields a framework to which additional constraints are added in a straight forward manner.

Ari Vesteinsson, Johannes R. Sveinsson, and Jon Atli Benediktsson Department of Electrical and Computer Engineering, University of Iceland, Hjardarhaga 2-6, 107 Reykjavik, Iceland {sveinsso, benedikt}@hi.is Henrik Aanaes Informatics and Mathematical Modelling, Technical University of Denmark, Denmark {haa}@imm.dtu.dk S. Liang (ed.), Advances in Land Remote Sensing, 293–311. c Springer Science + Business Media B.V., 2008

293

294

A. Vesteinsson et al.

Here we add a simple term to the discussed framework for smoothing the image, although more elaborate terms might be preferable. Computationally, we achieve a solution for the derived objective function via stochastic optimization, i.e., using the Metropolis algorithm in conjunction with simulated annealing. Apart from contributing with a novel analysis giving more insight into the image fusion problem, the method proposed in this chapter has been applied to images from the IKONOS satellite. These experimental results validate the proposed method.

11.1 Introduction In this chapter “Remote Sensing” will be understood to mean “taking images from a great distance.” In the modern context remote sensing started with the aerial photographs, originally taken from balloons and kites, but later from airplanes and later still from satellites. Remote sensing satellites are usually on sun-synchronous orbits (as opposed to communication satellites which are usually geostationary) but differ in the layout of their orbits, resolution of their sensors and their spectral sensitivity (i.e., the width and number of their frequency-bands). The first photographs were black and white (gray-scale), in the 1940s the color photograph was invented and later it became possible to record electromagnetic radiation outside the visible spectrum. In order to take color images1 the satellite’s photosensors restrict the incoming radiation with band selective filters, filtering through the different frequency components of the spectrum: Red, green and blue. These images are called multispectral images, as opposed to panchromatic images which cover the whole visible spectrum, and hyperspectral images which cover many more frequency bands. Image fusion is the subset of data fusion dealing with merging images, data fusion has been defined by Wald [1] as a formal framework in which are expressed means and tools for the alliance of data originating from different sources. It aims at obtaining information of greater quality; the exact definition of ‘greater quality’ will depend upon the application.

In the context of image fusion “different sources” simply means multiple images and any other a priori information that is available. The most common meaning of “quality” in image fusion has been visual improvement, i.e., taking into account how the human perceives the fusion product. Another frequently used meaning of “greater quality” is improved classification accuracy, in particular automated classification. Many other objectives have also been identified, see [2, 3]. Image fusion can be done on several levels: Pixel level, feature level, object level and decision level, depending on the intended use of the fused image. Pixel fusion is 1

Note that in this chapter the terms “color image” and “multispectral image” will be used interchangeably.

11 Spectrally Consistent Pansharpening

295

the combination of images pixel-by-pixel, feature fusion is the merging of features that have been extracted from the image (e.g., long, bright, square), object fusion is used for finding objects in the images (e.g., roads, lights, houses) and decision fusion is the use of multiple images in order to make a decision (e.g., “go to house”). This chapter is only concerned with pixel level fusion and when the terms “image fusion” or “fusion” are used below, pixel level fusion is intended. In the current context, fusion is the next step after preprocessing and the step before further processing, such as segmentation and classification. Historically, a great number of methods have been utilized in image fusion. Pohl et al. [2] provides a good overview of many such methods. There are many ways to categorize image fusion methods. In this chapter four categories will be looked at: Frequency methods, color transformation methods, statistical methods, and hierarchial methods. Frequency methods split the images into frequency parts and then join frequency parts from different images together. Color transformation methods transform the images to a new coordinate system which is more easily comprehended in terms of the human visual experience, and then combine those parts that enhance this experience. Statistical methods use the correlation between features of the image and the fact that some features are more dominant than others. Finally, hierarchial or scale space methods, create a pyramid of scaled down images and traverse this pyramid in reverse to scale up images, thus, in effect, fusing the images. Where applicable, the focus of this chapter will be on pansharpening rather than on image fusion in general. Frequency Methods: The idea behind the frequency methods is that the high frequency part of panchromatic image contains the spatial details of the image while the (low) frequency part of the multispectral image contains its spectral details. By combining the two frequency parts an image containing details and color can be created. This idea is heuristically supported by the fact that a low resolution image can be considered to have been created by the averaging (Low Pass Filtering (LPF)) of a high resolution image. As a consequence of the averaging process details of the high resolution image become less pronounced or smoothed out. Belonging to this category are methods like High Pass Filtering (HPF), High Frequency Modulation (HFM) and the Modified Brovey Transform (MBT). HPF can be subdivided into High Pass Filtering Substitution (HPFS) and High Pass Filtering Addition (HPFA). In both cases the high-frequency part of the panchromatic and the low-frequency part of the multispectral image are isolated. Then, for HPFS, the fused image is constructed by adding the two isolated parts together. On the other hand, for HPFA, the high-frequency component of the panchromatic image is added to the high-frequency component of the multispectral image before combining the two parts. The high and low pass operations can either be achieved by kernel methods, e.g., Gaussian or Laplacian kernels, or by using the Fast Fourier Transform (FFT). HPFS and HPFA are commonly used to evaluate new fusion methods, see, e.g., [4].

296

A. Vesteinsson et al.

With HFM, each low resolution channel is “boosted” with the ratio of the highfrequency to low-frequency filtered versions of the high resolution panchromatic image. The MBT is similar to the HFM, but instead of fusing each channel separately, the intensity component of the color image is fused with the high resolution panchromatic image in MBT. Actually, this is a product of combining the frequency and color transformation methods, which will be discussed shortly. For the frequency category of methods, it is important to find the best cut-off frequency for the high and low pass filters. One way of accomplishing this is by using an entropy measure. However, due to the imperfections of the filters, such as the windowing effect, this filtering will not be perfect. Also, there is an inherent problem due to the difference in resolutions, i.e., the frequency spectrum of the lower resolution image will contain aliasing components that are not present in the higher resolution image. Additionally, these methods are very sensitive to any high frequency noise that could be present in the panchromatic image. Color Transformation Methods: The Intensity-Hue-Saturation (IHS) method and its close relatives belong to the color transformation methods. Since it was first applied to image fusion, by Carper et al. [5], the IHS method has been the most popular fusion technique. This is mainly due to how simply it can be implemented, how fast it can be executed, and how visually pleasing the fusion result is. A color image is natively represented by three color channels: The red, green and blue channels (i.e., RGB). The IHS methods describes how to transform these three channels into a new coordinate system, where a coordinate is described by three values: The intensity, hue and saturation. The intensity measures the strength of the color, the hue is a measure of the dominant wavelength in the spectrum, and the saturation is a measure of the purity of the color. The idea of the IHS methods is to substitute the intensity component with the panchromatic image and transform the result back into the RGB representation, producing a high resolution color (RGB) image. There exist several algorithms for transforming a color image from the RGB color space to the IHS color space, not all giving the same results. In fact, IHS fusion sometimes goes under different names that reflect how the transformation is done, e.g., Hue-Saturation-Value2 (HSV) and Hue-Saturation-Lightness3 (HSL). Two IHS algorithms are given by ⎤ ⎡ 1 1 1 ⎤⎡ ⎤ R I 3 3 3 √ √ √ ⎢ ⎥ ⎢ − 2 − 2 2 2 ⎥⎢ ⎥ ⎣ v1 ⎦ = ⎣ 6 6 6 ⎦⎣ G ⎦ −1 √1 √ B v2 0 2 2

v2 and S = v12 + v22 , H = arctan v1 ⎡

2 3

Value is defined as max(R, G, B). Lightness is defined as (max(R, G, B) − min(R, G, B))/2.

(11.1)

11 Spectrally Consistent Pansharpening

297

and R+G+B % 3 arccos(a) if G ≥ R H= 2π − arccos(a) if G < R I=

a=

(2B − G − R)/2

(11.2)

(B − G)2 + (B − R)(G − R) 3 · min(R, G, B) . S = 1− R+G+B The major shortcoming of the IHS method is that it produces color distortion in the fused image. In [6] it is demonstrated how this color distortion results in an incorrect saturation value. There it is in fact shown that the distortion at a given pixel is proportional to the difference of the intensity values of the images. Zhang [7] points out that for IKONOS images, the panchromatic spectrum extends into the infra-red. Therefore, the difference between the intensity values is even greater than in the fusion of earlier Landsat and SPOT images. Furthermore, the color degradation is even more pronounced. To counteract the differences in the intensity values, histogram equalization methods have been used. These methods bring the intensities closer together on the average. Potentially, on the other hand, they leave out some distortion at the individual pixels. Statistical Methods: One example of a statistical method is the Principle Components Substitution (PCS) method. It applies principle component analysis to the multivariate image data, transforming them into a set of uncorrelated components. The PCS method is similar to the IHS method, except that in the PCS, the first principle component is replaced by the intensity component of the panchromatic image. Then, by reverting to a RGB representation, a fused high resolution color image is created. Shettigara was amongst the first to use the PCS method for image fusion [8]. Zhang [7] explains that the problem with the PCS method is that the first principal component usually has the greatest variance. Therefore, the first component dominates in the fused image, possibly resulting in color distortion. Hierarchial Methods: The common idea for the hierarchial methods is to reduce stepwise the resolution of the panchromatic image, reaching the low resolution of the multispectral image. Then, this ‘path’ is retracted with the multispectral image, producing a high resolution multispectral image. There are several ways to perform the reduction, e.g., using Laplacian and Gaussian kernels, morphological operators, the wavelet transform, or the curvelet transform. The current focus in data fusion algorithms seems to be concentrated on the hierarchical family of methods [9], and the wavelet transform in particular. The main drawback of the hierarchical methods is the assumption that the spectral content of the panchromatic and multispectral images coincide exactly, i.e., the same “path” would have been generated with an original high resolution multispectral image. This is not necessarily always the case.

298

A. Vesteinsson et al.

Other Fusion Methods: Many other fusion methods exist, e.g., the Color Normalized transform (CN), the Spectral Balance Preserving fusion (SPB), and more. All the methods that have been discussed above are derived from a certain perspective of what an image is, rather than serving a given objective. In this chapter, these two approaches will be combined, i.e., an algorithm will be derived from how an image is formed and then improved to suite a given objective. In this chapter we present a fusion algorithm, called SCP (Spectrally Consistent Pansharpening), which utilizes the known spectral responses of the photosensors to fuse multispectral and panchromatic satellite images at the pixel level into a highresolution multispectral image. The intention is to create an image using as much information as feasible from both all the available images and all the available a priori knowledge. The created image maintains spectral consistency with the original multispectral image. Then, a more elaborate model for the imaging physics is introduced and a fusion framework is developed in which it is possible to, simultaneously, keep spectral consistency and smooth the final image. It is noteworthy that the smoothing will be used as a constraint, but any other constraint or constraints could have been selected. Scene smoothness, however, is a common assumption in remote sensing. Finally, the fusion problem is modeled as a minimization problem where the objective function is defined as some energy functions (defined in Section 11.3). The SCP method is tested in experiments by fusing a low resolution hyperspectral image with a high resolution panchromatic image. The chapter is organized as follows. In Section 11.2, the spectrally consistent pansharpening method is introduced. A fusion framework in which it is possible to, simultaneously, keep spectral consistency and smooth the final image is introduced in Section 11.3. Experimental results are given in Section 11.4, and conclusions drawn in Section 11.5.

11.2 Spectrally Consistent Pansharpening (SCP) Spectral consistency of a high resolution multispectral image with respect to a low resolution multispectral image is the requirement that the integration of the spectral density over the same geographical area gives the same spectral value. This amounts to saying that the sum of pixel values over the same geographical area is the same for both images. Put yet another way: down-sampling the high resolution image, by averaging, down to the resolution of the low resolution image should give an image identical to the low resolution image. The maintenance of spectral consistency between the fused image and the original multispectral image is of great importance for many applications. For example, it is very important in thematic classification, where objects are being classified on the bases of their spectral values, both relatively within the image, as well as absolutely with respect to some know spectral signatures. Maintaining spectral consistency is a non-trivial task and most conventional fusion algorithms fail in doing so. It is important to base the fusion of images on solid mathematical and physical foundations. In this chapter, a model of the image formation process will be derived.

11 Spectrally Consistent Pansharpening

299 q*F

F

q

=

* A w

w

w

Fig. 11.1 Construction of pixel value

This model will help to understand how the observation channels can be combined to produce a fused image of full spectral integrity with respect to the original color channels. In deriving the image formation model, it will be assumed to be noise free and deterministic.

11.2.1 Image Formation Model A sensor’s observation is captured to a digital media in the form of a set of pixels, each having a given value (often referred to as a Digital Number (DN)). The number of stored pixels depends on the resolution of the sensors. Furthermore, the value of the pixel depends upon the amount of light that comes through the sensor. The output of the sensor due to the accumulated incident spectral energy can be expressed in terms of the spectral energy density, q, and the sensor’s spectral response, F, integrated over the pixel area, A, and all wavelengths, w, as && & A w

F · q = pixel value,

(11.3)

see Figure 11.1 [3]. It is reasonable to assume that the sensor’s spectral response is constant for all the pixels. Therefore, (11.3) can equally be expressed as &

&

w

&

q=

F A

w

F · S = pixel value,

(11.4)

where S is the total spectral energy density for A.

11.2.2 Spectral Inner Product Space (W) Considering functional analysis, it is seen that the integral in (11.4) defines an inner product between a spectral response and the spectral density, i.e.4 4 For readers not familiar with functional analysis, consider the discrete version/approximation of the spectral response and the spectral density. In this case the integral is substituted by a summation, and (11.4) reduces to the standard inner product between two real vectors as know from linear algebra.

300

A. Vesteinsson et al.

F, S =

& w

F · S.

(11.5)

In this context, the spectral responses for the red (R), green (G), blue (B), nearinfrared (NIR) and panchromatic (P) channels along side the spectral density (S) can be thought of as vectors in an infinite dimensional inner product space, W. Please note that, in this setting there is also no difference between the spectral density and the spectral responses, hence, e.g., the inner product R, G =

& w

R · G,

(11.6)

between the spectral responses for the red and the green channel makes perfect sense. As seen from the above the pixel value of a given pixel and a given channel, can be expressed as an inner product, i.e., F, S. This, however, implies that even though S is infinite dimensional, we can only observe the parts of S spanned by the spectral response functions. That is, Wobs = span{R, G, B, NIR, P},

(11.7)

with Wobs denoting the observable part of W. Hence for the purpose at hand we can just as well operate in the four-dimensional space Wobs , which is a subset of W. The interesting here is that the inner product defined in (11.4) on elements in W induces an inner product on Wobs . Hence from the above we can calculate the inner product between the different spectral response functions, which intuitively tells us how much information one channel has about another. 1. Projections in W: The problem at hand is to find out how the high-resolution panchromatic measurements (P) can be transformed into high-resolution RGBmeasurements. Figure 11.2 shows how the spectral density Si of a pixel i projects

FP P1 PL

SP1

P2 SR SP2

θ

R2

R1

FR

R

Fig. 11.2 Projections resulting from the assumption that SP1 and SP2 lie as close as possible to SR

11 Spectrally Consistent Pansharpening

301

onto the spectral response function of the panchromatic channel FP giving the value Pi of pixel i in the panchromatic image. In the same way the spectral density projects into R, G, and B measurements of the corresponding spectral channels. Note that when Pi is known all that can be said about Si is that it lies in a hyperplane perpendicular to Fi , intersecting Fi at the point Pi . In order to derive the relationship between panchromatic measurements and R, G, and B measurements, it is simplest to start with a high to low resolution ratio of 2 and focus on one of the spectral channels, for instance, the red. Figure 11.2 shows three spectral densities, SR , SP1 and SP2 , and their corresponding measurements R, P1, and P2, respectively. The relationship between SR and {SP1 , SP2 } is simply SR = SP1 + SP2 ,

(11.8)

where linearity of the sensors has been assumed. As a side note: The extension of this idea to a ratio of M high resolution pixels for each low resolution pixel is simply M

SR = ∑ SPi . i=1

Now, it is important to remember that the measurements have been normalized and it can be assumed that the normalizing constant for the high-resolution image is twice that of the low-resolution image (given the 2-to-1 resolution ratio). So the normalized version of (11.8) is SR =

SP1 + SP2 . 2

(11.9)

From the above, the possible spectral densities corresponding to a measurement have been restricted to a hyperplane in W. This, however, does not fully specify where the true spectral density lies in W, for that additional knowledge is needed. 2. Problem Regularization: In order to extract a unique solution out of the set of possible solutions for three given measurements (P1, P2, and R), a priori assumptions are added to the problem specification. One possible such a priori assumption is that the distance from SP1 and SP2 to SR should be as small as possible. The justification for this assumption is that if the spectrum of the neighbors of the low resolution pixel can be assumed related by the normal distribution, then it is most likely that the high resolution pixels lie as close as possible to the low resolution pixel. This can also be called a smoothness assumption. This assumption results in the two intersections shown with closed-circles in Figure 11.2. The projection of these two intersections onto FR gives the estimated high resolution R values, R1 and R2 (see Figure 11.2). The assumption amounts to the same as saying that the difference in the spectral values lies only in their intensity component (not hue or saturation). 3. Fusion Formula and Results: Using the assumption displayed in Figure 11.2, the high resolution R values can be calculated by

302

A. Vesteinsson et al.

R1 = R + (P1 − PL ) cos(θ ) R2 = R + (P2 − PL ) cos(θ ),

(11.10)

where the cos(θ ) is calculated from the inner-product, using FP , FR = |FP | · |FR | · cos θ

FP , FR θ = arccos |FP | · |FR |

' FP FR dw . = arccos |FP | · |FR |

(11.11)

11.3 SCP Fusion Framework This section introduces a more elaborate model for the imaging physics as well as developing a fusion framework in which it is possible to, simultaneously, keep spectral consistency and smooth the final image. Note that here smoothing will be used as an constraint, but any other constraint or constraints could have been selected; scene smoothness, however, is common assumption in remote sensing. The fusion problem is modeled as a minimization problem where the objective function is the energy of a Gibbs distribution [11]. In this implementation the energy consists of three energy terms: One for the spectral consistency constraint, one for the smoothness constraint, and one for the image formation, or imaging physics (IP) constraint. The minimization of the objective function is sought using stochastic optimization.

11.3.1 Objective Function The Gibbs distribution, P( f ), is defined as 1 P( f ) = Z −1 exp( U( f )), T

(11.12)

where T is called the temperature, Z is a normalizing constant, called the partition function: 1 (11.13) Z = ∑ exp( U( f )), T f ∈F and U( f ) is called the energy function, defined by U( f ) =

∑ Vc ( f ).

c∈C

(11.14)

11 Spectrally Consistent Pansharpening

303 aspectral

Pspectral asmooth

1/T

Psmooth

Σ aIP

PIP

Fig. 11.3 Calculating the energy function

over the set C. As mentioned above, the energy function is specified in terms of three energies, namely, Pspectral , Psmooth , and PIP , i.e., U( f ) = αspectral Pspectral ( f ) + αsmooth Psmooth ( f ) + αIP PIP ( f ),

(11.15)

where αspectral , αsmooth , and αIP are the weights for the respective energy functions, chosen such that energy function is minimized. Figure 11.3 displays the above schematically. The derivation and application of these three energies is the subject of the following subsections. 1. Spectral Constraint: The reason for applying the spectral constraint is to enforce the spectral consistency between the original low resolution multispectral image and the new, constructed, high resolution image. Pspectral ( f ) measures the difference between the average of the pixels in the fused image and the corresponding pixel in the original low resolution multispectral image, i.e., Pspectral ({CH ∈ CL }) = (CL −

1 CH )2 , n2 {C ∑ ∈C } H

(11.16)

L

where CL is the low resolution pixel, the set {CH ∈ CL } contains the corresponding high resolution pixels, and n2 is the number of elements in that set, i.e., the resolution ratio between the high and low resolution images is n. 2. Smoothness Constraint (SP): The smoothness constraint is applied to enforce that the new constructed image is smooth. This is to counteract the blockyness remaining in the image after the SCP fusion. The SP is defined by the smoothness energy, Psmooth , which is given as Psmooth ( fi ) =

1 ( fi − fi )2 k i ∑ ∈N

∀i ∈ S,

(11.17)

i

where fi is the value of the pixel i in the fused image, Ni is the neighborhood for pixel i, k is the number of pixels in the neighborhood, and S contains all the image’s

304

A. Vesteinsson et al.

pixels. Here a simple 4-neighborhood will be used, i.e., k = 4. On the other hand, the Euclidian three-dimensional distance is used as a measure of the distance measure between fi and fi . The larger objects should become smoothed without loss of detail. 3. Imaging Physics (IP) Constraint: By using (1) it can been shown that the IP constraint is the extension of the theory presented earlier. Consider the subspace of the frequency space, W, spanned by the red (R), green (G), blue (B), and panchromatic (P) spectral values, and denote the mean of the P-channel for a corresponding low resolution pixel as Pµ . Now, by assuming that the pixels are normally distributed in the frequency space, with a mean and variance ⎡ ⎤ RL ⎢ GL ⎥ Σ Σ CP C ⎥ µC = ⎢ ⎣ BL ⎦ , ΣC = Σ T σ 2 , P CP Pµ respectively, where

⎡

⎤ σR2 σR σG R, G σR σB R, B ⎢ ⎥ σG2 σG σB G, B ⎦ , ΣC = ⎣ σG σR G, R σB σR B, R σB σG B, G σB2 ⎡ ⎤ σR σP R, P ΣCP = ⎣ σG σP G, P ⎦ , σB σP B, P

σx2 is the variance for channel x ∈ {R, G, B, P}, and ., . denotes the inner product. Then the conditional distribution of a high resolution RGB pixel given P, is a normal distribution with the mean given as ⎡ ⎤ RL ΣCP µC|P = ⎣ GL ⎦ + 2 (P − Pµ ) (11.18) σP B L

and variance

T ΣCP ΣCP . σP2 This distribution gives that the panchromatic energy can be expressed as

ΣC|P = ΣC −

PIP ( fi ) =

−1 ( fi − µC|P ) log(det(ΣC|P )) + ( fi − µC|P )T ΣC|P

2

(11.19)

(11.20)

for all i ∈ S. By applying the simplifying assumption that the standard deviations of all channels are the same, i.e.,

σR = σG = σB = σP = σ , then µC|P becomes the new pixel value in the SCP fusion algorithm, namely

11 Spectrally Consistent Pansharpening

⎡

305

⎤

⎡

⎤

R, P RL µC|P = ⎣ GL ⎦ + ⎣ G, P ⎦ (P − Pµ ). B, P BL

(11.21)

However, contrary to earlier definition, the distribution’s variance will be utilized here as well. 4. Selection of a Search Algorithm: The selection of a quadratic potential function and a Gaussian observation model, guarantees that the a posteriori energy function is convex and, consequently, has only one extremum. Therefore, in the search for a global minimum there is no risk of getting trapped in a local minimum. Consequently, there is no need to apply global minimization algorithms. In this case, the choice of a local minimization algorithm is between the ICM and Metropolis algorithms. For the images that are going to be used, it is known that the value of each pixel is stored as 11 bits, i.e., a pixel can take 211 = 2,048 possible values. Since the ICM algorithm tests every possible configuration on each pass, it will test LS labels, or ≈1275 million labels in the current case. This immediately rules out the use of ICM, since it would be impossibly slow to execute. The Metropolis algorithm, on the other hand, only tests S number of labels on each pass. Therefore, it is chosen in this case.

11.4 Experimental Results The data set used in this paper consists of two images of Reykjavik, Iceland acquired by the IKONOS earth imaging satellite on 9 August 2001. One image is a multispectral image of 4 m resolution (see Figure 11.4a). The other is a panchromatic image of 1m resolution (see Figure 11.4b). The panchromatic image is 976-by-640 pixels. It consists of a single channel (i.e., an intensity image) and is free from speckle noise. From the histogram of the image, it is evident that there is some compression of a curtail around 0 and a slight saturation in the intensity values, indicating that the normalization procedure included compressing the data inside the available range. Individual cars are distinguishable on the road and in parking lots. The multispectral image is 244-by-160 pixels. It consists of the R, G and B channels, i.e., red, green and blue, and it is also free from speckle noise. There is a slight peak in the intensity values at 0 suggesting a compression of a curtail. Most individual cars on the road are distinguishable. On the other hand, that is often not the case for cars in parking lots. Visually, some adjacent houses are more easily distinguished on the multispectral images than on the panchromatic one, which is due to the distinct colors of their rooftops. IKONOS can produce imagery of the same geographical area every 3 days. Its spectral range covers the visible band into the infrared band, see Table 11.1. Notice that the spectral range of the panchromatic channel is significantly larger than the combined spectral range for the red (R), green (G), blue (B) and near-infrared (NIR) channels, see Table 11.1 and Figures 11.5 and 11.6. The SCP algorithm was tested and compared to the IHS method [5]. Results are shown in Figure 11.7 and Table 11.2.

306

A. Vesteinsson et al.

Fig. 11.4 (a) The multispectral image. (b) The panchromatic image

Table 11.1 IKONOS’s spectral bands 1 m black-and-white Panchromatic

0.45–0.90 µm.

4 m multispectral Blue Green Red Near-infrared

0.45–0.52 µm 0.51–0.60 µm 0.63–0.70 µm 0.76–0.85 µm

The SCP fusion framework algorithm was used on the data. First, the SCP algorithm was performed to get the initial image for the optimization problem, shown in Figure 11.7. The results after 500 iterations using the spectral, IP, and smoothness constraints and a fixed non-zero temperature for a 256-by-256 copped out area is shown in Figure 11.8 and numerically results are given in Table 11.3. The energies for each constraint for this experiment are given in Figure 11.9. Comparing the results from all the experiments with the SCP fusion framework shows that none of them actually produce better result than the SCP algorithm, at

11 Spectrally Consistent Pansharpening

307

Fig. 11.5 IKONOS multispectral sensors’ spectral responses

Fig. 11.6 IKONOS panchromatic sensor’s spectral response

least not with respect to the metrics used. Two conclusions can be drawn from this fact. First, that a new minima of the objective function was not found, and perhaps did not exist. Second, this could simply imply that the smoothness constraint (the additional ingredient) was not a good a priori assumption about the ground truth, eventhough it did produce a more visually pleasing image. However, all of the experiments did compare favorably to the results of the IHS fusion algorithm.

308

A. Vesteinsson et al.

Fig. 11.7 The SCP fusion image, which is also used as the initial image in the SCP framework algorithm

Table 11.2 Quality metrics calculated for individual channels, fused image down-sampled RMSE

Correlation

SCP IHS SCP Red 0.006 0.029 1.000 Green 0.006 0.037 1.000 Blue 0.004 0.034 1.000

IHS 0.955 0.930 0.886

11 Spectrally Consistent Pansharpening

309

Fig. 11.8 The fused image after 500 iterations using the spectral, IP, and smoothness constraints and a fixed nonzero temperature

Table 11.3 Quality metrics for the image using smoothness, spectral and IP constraints using a fixed non-zero temperature Metric

Result

Correlation with original red Correlation with original green Correlation with original blue Root-mean-squared-difference with original red Root-mean-squared-difference with original green Root-mean-squared-difference with original blue Correlation with original panchromatic

0.997 0.996 0.995 0.008 0.009 0.008 0.879

150 IP Spectral Smooth

energy

100

50

0

0

50

100

150

200

250

300

350

400

450

500

iteration

Fig. 11.9 After 500 iterations using the spectral, IP, and smoothness constraints and a fixed nonzero temperature

310

A. Vesteinsson et al.

11.5 Conclusion The underlying objective of this chapter has been to approximate an image produced by multispectral sensors with the characteristics (spectral response) of the original sensors and the resolution of a panchromatic sensor. What has been achieved is an image with the spectral integrity of the original color image intact and with some details of the panchromatic image included. Unfortunately, the resulting image from the SCP method is blocky. However, that is as good as it gets using the available raw information. In order to further improve the fused image, additional knowledge (a priori) has to be introduced. By introducing this information into the fusion process, it might no longer be possible to guarantee spectral consistency with respect to the original color image. However, given that the new information is correct, the spectral correctness with respect to the ground truth should in fact have increased. This introduction of new information leads to a construction of SCP fusion framework where multiple constraints can be combined together to make up the objective function for the fusion problem. The underlying algorithm uses known spectral response functions to get a good initial image, as well as utilizing them to calculate the IP constraint. Furthermore, the algorithm employs stochastic optimization to locate a minimum of the objective function. The alpha version of the SCP fusion framework was shown to be promising for combining both data and a priori assumptions. The Focus of our future work in SCP and SCP fusion framework could be to make the framework more generally useful by for example automatically determining the model parameters and including the weights for the objective functions. Possible solutions could use the resulting classification accuracy as an input to the parameter estimation procedure. Acknowledgements The work was partially supported by the Icelandic Research Council and the Research Fund of the University of Iceland.

References 1. Wald L (1999) Some terms of reference in data fusion. IEEE Trans. Geosci. Remote Sens. 37(3):190–1193 2. Phol C, Genderen JLV (1998) Multisensor image fusion in remote sensing: concepts, methods and applications. Int. J. Remote Sens. 19(5):823–854 3. Munechika CK, Warnick JS, Salvaggio C, Schott JR (1993) Resolution enhancement of multispectral image data to improve classification accuracy. Photogramm. Eng. Remote Sens. 59(1):67–72. 4. Tsai VJD (2003) Frequency-based fusion of multiresolution images. Geoscience and Remote Sensing Symposium, July 2003, pp. 3665 –3667 5. Carpter W, Lillesand T, Kiefer R. The use of intensity-hue-saturation transformations for merging SPOT panchromatic and multispectral image data. Photogramm. Eng. Remote Sens. 56(4):459–467 6. Tu T-M, Su S-C, Shyu H-C, Huang PS (2001) A new look at IHS-like image fusion methods. Inform. Fusion 3(2):177–186

11 Spectrally Consistent Pansharpening

311

7. Zhang Y (2002) Problems in the fusion of commercial high-resolution satellite as well as Landsat 7 images and initial solutions. Proceedings of the ISPRS Technical Commission IV Symposium 2002 8. Shettigara VK (1992) A generalized component substitution technique for spatial enhancement of multispectral images using a higher resolution data set. Photogramm. Eng. Remote Sens. 58(5):561–567 9. Schowengerdt RA (1997) Remote sensing models and methods for image processing, 2nd ed. Academic Press, San Diego, CA 10. Space Imaging. IKONOS [online]. Available at: http://www.spaceimaging. com/products/ ikonos/ 11. Geman S, Geman D (1984) Stochastic relaxation, Gibbs distributions, and Bayesian restoration of images. IEEE Trans. Pattern Anal. Machine Intell. PAMI-6(6):721–741 12. Vesteinsson A, Sveinsson JR, Aanes H, Benediktsson JA (2005) Spectral consistent satellite image fusion: using a high resolution panchromatic and low resolution multi-spectral images. IEEE International Geoscience and Remote Sensing Symposium (IGARSS’05), vol 4, Seoul, Korea, 25–29 July 2005, pp 2834–2837

Chapter 12

Data Assimilation Methods for Land Surface Variable Estimation Shunlin Liang and Jun Qin

Abstract Estimating land surface variables from remote sensing data is an ill-posed problem. Integration of observations from multiple satellite sensors with different spectral, spatial, temporal and angular signatures is now an important research frontier. Data assimilation (DA), integrating not only remotely sensed data products, but also other measurements and land dynamic models, is an advanced set of techniques for innovative parameter estimation. After a brief introduction, we describe the basic principles of DA, and then provide in-depth discussions of some relevant issues while using DA. The latest applications of DA for estimation of soil moisture, energy balance, carbon cycle and agricultural productivity are summarized.

12.1 Introduction Despite the abundance and variety of remote sensing measurements, land surface characterization from satellite observations is still very challenging. There are multiple sources of surface information, such as remote sensing data and derived products, in situ measurements, and land surface model outputs. Innovative techniques are needed to merge these information sources and optimize the use of satellite measurements for robust surface products and greater predictability. Data assimilation (DA) is a mathematical approach that enables use of all available information within a given time window to estimate various unknowns. The information that can be incorporated includes observational data, existing a priori information, and, very importantly, a dynamic model that describes our system and encapsulates current theoretical understanding. The model brings consistency to the observational Shunlin Liang Department of Geography, University of Maryland, College Park, USA [email protected] Jun Qin Institute for Geographical Science and Natural Resource Research, Beijing, China S. Liang (ed.), Advances in Land Remote Sensing, 313–339. c Springer Science + Business Media B.V., 2008

313

314

S. Liang, J. Qin

data, and interpolates or extrapolates data into data-devoid regions in space and time. The observational data, representing the actual state of the system, corrects the trajectory of the imperfect model through adjusting model parameters. DA is also called model–data synthesis or data–model fusion in different disciplines. The benefits of DA for maximizing the scientific and economic value of remote sensing data is summarized as follows (O’Neill et al., 2004; Raupach et al., 2005): 1. Forecasting and error tracking. By regularly comparing forecasts with observations, extremely valuable error statistics can be built up, which in turn can be used to improve the quality of the observations (e.g., by revealing biases in instrument calibration) as well as the quality of the models. It serves as a model testing and data quality control procedure. 2. Combining multiple data sources. Different observing systems (both in situ and remote sensing data) have varying virtues and deficiencies. Such variety can be preferentially exploited or contrasted to optimize the value of the resulting data set. 3. Interpolating spatially and temporally sparse observations. The model provides a way to propagate information in a consistent manner in space and time from data-rich regions to data-poor regions. This capability is vital to successfully utilize satellite observations, which due to limited and sequential sampling provide only an incomplete picture of the Earth. DA fills in “missing pieces” to achieve a full global picture. 4. Inferring, from available observations, quantities not directly observable. Through relationships expressed in the model’s governing equations, measured parameters convey knowledge of those that are inadequately measured or completely lacking. For example, soil moisture vertical profile can be inferred from the surface skin temperature or surface top-layer soil moisture content. 5. Forecasting. Predicting forward in time on the basis of past and current observations. 6. Designing observing systems. Decisions to deploy new satellite-borne instruments require critical assessment of the incremental value or benefit of the data to be acquired by the new sensors. With careful design, DA experiments provide an objective, quantitative way to contribute to such assessment. In addition, DA can optimize the sampling pattern from an observing system, and can target observations to capture features of concern, such as a rapidly developing storm. DA methods exploit data streams not only to validate model outputs, or directly to infer fluxes, but principally to constrain internal model parameters to optimize values (i.e., parameter estimation). Different data sets constrain different components of a model which is able to assimilate data across a range of space and time scales. Another distinctive characteristic of DA is that uncertainties associated with the observation, techniques, processing, representation, and accuracy are as important in determining the final outcome as the measured values themselves. Thus, uncertainty estimates, for both measurements and model parameters, take on even greater importance (Canadell et al., 2004). The meteorological and oceanographic communities have been at the forefront in developing and using DA methods. In recent years, meteorologists and

12 Data Assimilation Methods for Land Surface Variable Estimation

315

oceanographers have tended to view DA as a model state estimation problem. The land community has aggressively endeavored to catch up and apply DA methods in recent years, and some example applications are shown in Section 12.4. The remainder of this chapter presents a brief introduction to the basic principles of land DA as this is a relatively new field. Following the introduction is a discussion of some critical issues pertinent to land DA. Section 12.4 introduces recent DA applications in several major disciplines.

12.2 Principles of Land DA In this chapter, land DA is considered as essentially an estimation problem: that is to acquire optimal estimates of model state variables of the land surface and its parameters given a set of remote sensing products, a land surface process model, and any available a priori information. Various land DA schemes have different characteristics, but they may have the following common features: (1) a forward land dynamic model that describes the time evolution of state variables such as surface temperature, soil moisture and carbon stocks; (2) an observation model that relates the model estimates of state variables to satellite observations and vice versa; (3) an objective function that combines model estimates and observations along with any associated prior information and error structure; (4) an optimization scheme that adjusts forward model parameters or state variables to minimize the discrepancy between model estimates and satellite observations; and (5) error matrices that specify the uncertainty of the observations, model and any background information (these are usually included in the objective function).

12.2.1 Dynamic Model Land surface process models are often structured as a discrete-time nonlinear statespace model with additive noise as: xt+1 = f (xt , ut , θ ) + wt

(12.1)

where x denotes the model state vector (e.g., soil moisture), u the external forcing data (e.g., meteorological data), θ the model parameter vector (e.g., soil texture), wt the model noise, and f (·) the model operator mapping the previous state xt to the next state xt+1 . The differences among various land surface models are reflected in different specifications of x, u and f (·).

12.2.2 Measurement Process DA methods assimilate the measured values into a dynamic model, so the measured quantities have to be linked with model variables. The measurement process

316

S. Liang, J. Qin

is a mathematical model that relates the model state vector (xt ) to the observation vector (yt ): (12.2) yt = h(xt , θ ) + et where et the observation noise, h(·) the observation operator, and θ is the parameter set of observation operator. If remote sensing data products are used as the observations, the remote sensing reflectance, or emittance model of land surfaces, and the coupled land and atmosphere system are the observation operator and θ is its parameters.

12.2.3 Objective Function The underlying principle of the DA method is to estimate the parameters and variables of the dynamic models by minimizing the differences between the predictions of the model and the assimilated data products. The difference is often characterized as the objective (or cost) function, and the common measure is the least squares. DA methods usually attempt to minimize the objective function J 1 1 J(x) = (x − xb )T B−1 (x − xb ) + (H(x) − y)T R−1 (H(x) − y) = Jb + Jo 2 2

(12.3)

where y is the observation vector, x is the extended model state variables, xb is the background field (or first guess), H is the model operator, and R is the observationerror covariance matrix and B is background-error covariance matrix. The DA problem now becomes: vary x to minimize J(x), subject to the constraint that the state variables must satisfy the dynamic model. The value of x at the minimum is the a posteriori estimate of x, including information from the observations as well as the background. In Eq. (12.3), the first term Jb is to force the optimal parameters as closely as possible to background fields, and the second term Jo is to adjust parameters so that model outputs will be as close to the observations as possible. Specifying R and B depends on the relative accuracy of background information and remote sensing data products. In extreme cases, if the errors of the “first-guess” values are extremely large, the final estimates will be decided from the fitting of the observations and will be close to the “first-guess” values.

12.2.4 Assimilation Algorithm Since land DA is considered to be an estimation problem, an assimilation or estimation algorithm is needed to estimate model parameters or states by assimilating observations into the dynamic model. This process provides three kinds of output: optimal estimates for the model properties to be adjusted, uncertainty statements about these estimates, and an assessment of how well the model fits the data, given

12 Data Assimilation Methods for Land Surface Variable Estimation

317

the data uncertainties. There are two types of assimilation algorithms popular in current research and applications, namely cost function-based methods and sequential methods. These two methods have many different forms and will be discussed in Section 12.3.5.

12.2.5 Error Model Both dynamic models and measurements are not perfect and contain uncertainties to some degree. These uncertainties or errors have to be characterized in the DA system. Consequently statistical properties of wt in Eq. (12.1) and et in Eq. (12.2) or R and B in Eq. (12.3) need to be estimated. The error terms in the dynamic and observation models can, in principle, be quite general in form, including biases, drifts, temporal correlations, extreme outliers and so on. Many extant methods assume these error terms (e.g., wt and et ) have Gaussian distributions with zero mean and no temporal correlation. Detailed discussions of observational errors are given in Section 12.3.6.

12.3 Critical Issues in Land DA 12.3.1 Dynamic Models Land DA uses land surface process models that describe the exchanges of momentum, energy and mass between soil, vegetation and atmosphere. These models, often called soil–vegetation–atmosphere transfer (SVAT) models, have similar structures, whether land surface, hydrological, ecological or crop growth models. How to select an appropriate dynamic model is largely related to the objectives of the DA study. For the same objective, multiple models might be available. For example, if we aim to estimate land surface energy balance components such as latent and sensible heat fluxes by assimilating remotely sensed land surface temperature (LST) products, almost any land surface process models may be appropriate. However, some models are rather complicated since they may describe other processes such as those in the carbon cycle in addition to water and energy budgets and thus contain more model parameters. Often a good strategy is to use a relatively simple land surface model if possible. A more general and important issue is the uncertainty of the model. All mathematical models are abstractions and approximations of a complex physical environment. When dealing with ecosystem models, for example, keep in mind that their designs are not completely derived from rigorous natural laws and hence the mathematical descriptions of the biological processes are not universal. Thus, we are uncertain not only about values of the numerous model parameters, but also about the model parameterizations and errors.

318

S. Liang, J. Qin

DA aims to incorporate measured observations into a dynamic system model to produce accurate estimates of the system’s current (and future) state variables. The neglected model uncertainties are interpreted, in the course of strongly constraining parameter estimation, as variations in model parameters and this may result either in unrealistic parameter estimates or in solutions deviating far from the data. In reality, the model does not exactly reproduce system behavior. Significant errors can arise due to a lack of resolution, and inaccuracies in physical parameters, boundary conditions and forcing terms. Many studies have been reported to address these issues in terms of weak constraints (Liaqat et al., 2003a; Losa et al., 2004; Natvik et al., 2001; Qin et al., 2007b). As a rule, weakly constrained DA has been used for state estimation. However, if the model parameters are poorly known, state estimation with fixed model parameters can produce unacceptable results. Weakly constrained parameter estimation results from a combination of the adjoint method and the generalized inversion; consequently it can take into account errors inherent in the model and data, as well as find optimal values of the poorly known model parameters. In addition, weakly constrained DA makes it possible to derive valuable supplementary information about the model itself. This information is completely lost when the strong-constraint scheme is used. Chepurin et al. (2005) recently summarized four solutions to correct biases due to model error in the meteorological community. The most straightforward approaches to handling time-mean bias involve computing climatologies, and then introducing correction terms into the equations of motion. The second type of approach is to correct for rapidly changing bias in a data-rich environment. It involves examination of the previous few updating cycles for a systematic trend, which is then corrected. A third class of approaches, useful in linear one-dimensional problems, involves pre-whitening the errors so that their frequency spectrum resembles white noise. A fourth class of approaches is referred to as “two-stage estimation.” The two-stage estimation algorithm begins with the assumption that a reasonable estimate of the bias may be made prior to estimating the state of the system itself, thus allowing the estimation procedures for bias and state to be carried out successively.

12.3.2 Observation Vector Which remotely sensed data products and other measurements are the most valuable for DA? There are two major schemes for assimilating remotely sensed data products into dynamic models (Liang, 2004). The first approach assimilates the derived high-level land products (e.g., leaf area index (LAI), fractional photosynthetically active radiation absorbed by green vegetation (FPAR), LST, gross primary production (GPP), biomass) from remote sensing observations (see Fig. 12.1). It is equivalent to simplification of measurement process through static remote sensing inversion. So the observation vector is directly the components of the model state and h(·) is simplified as a matrix H with non-diagonal elements equal to 0:

12 Data Assimilation Methods for Land Surface Variable Estimation

319

Fig. 12.1 Illustration of assimilating remotely sensed high-level products into a land dynamic model

Fig. 12.2 Illustration of assimilating the direct remote sensing observations to the coupled radiative transfer and land surface dynamic model

yt = Hxt + et

(12.4)

The second approach assimilates the direct observations (radiance, reflectance, brightness temperature, or vegetation indices) (see Fig. 12.2). In this approach, h(·) is a radiative transfer (RT) model relating the state vector x to the remote sensing signal (e.g., radiance) or its simple transformation (e.g., reflectance).

320

S. Liang, J. Qin

Both approaches have strengths and weaknesses. If the high-level product retrieval is very accurate and its uncertainties well-characterized, then, use of such data is preferable. However, most inversion processes are ill-posed (see Chapters 7 and 8), and the resulting products may be poor. This is particularly so if their error characteristics cannot be specified accurately for DA (see Section 12.3.6). Another issue is that most current land products are not continuous both spatially and temporally due to cloud contamination or the inversion failure due to an inadequate number of good observations. Since many biogeophysical variables are “continuous” spatially and temporally, “gaps” need to be filled using various spatiotemporal modeling techniques (Fang et al., 2007a; Fang et al., 2007c; Julien et al., 2006; Moody et al., 2005). On the other hand, there is no need for an inversion if direct observations are used, but the forward RT models must be coupled with land surface process models. With physically based RT models, an additional advantage of the second approach is that consistency can be maintained between the direct physical meaning of model parameters in both the RT and land surface process models. This is not necessarily so for high-level products that may be derived under a different set of assumptions (e.g., spatial distribution of foliage) to those assumed in the process model. For using high-level products, not only those from optical and thermal observations (e.g., LAI/FPAR, albedo, LST) are considered but also those obtained from LIDAR (e.g., canopy structure parameters and above-ground biomass) and RADAR (above-ground biomass and moisture conditions) observations. Other products, such as forest biomass inventories, soil carbon survey, and Eddy covariance flux measurements, atmospheric CO2 and tracers concentrations, nutrient fluxes and stocks (particularly N) can also be used for assimilation. A highly related issue is the compatibility of remotely sensed products and variables in land surface process models. Although substantial efforts have been made from both remote sensing and modeling communities, the gaps still exist. For example, most land surface models separate broadband albedos into direct and diffuse components (Dai et al., 2003), but remote sensing albedo products are usually total broadband albedo (see Chapter 9). Most land surface models normally partition surface temperature into sunlit/shadow leaf and soil temperatures, but remote sensing skin temperature is an effective temperature for each pixel (see Chapters 4 and 10). A particular advantage of DA schemes is that they can be used to evaluate the “advantage gained” (assessed through reduction in uncertainty) by the addition of new observations. This will be of great use in assessing the value of different types and sample availability of observations, as well as performing synthetic experiments to evaluate new types of observations (e.g., from proposed monitoring networks or space missions).

12.3.3 Target Variables The target variables are the properties of the model to be adjusted or estimated in the optimization process. They include any model property considered to be sufficiently

12 Data Assimilation Methods for Land Surface Variable Estimation

321

uncertain as to benefit from constraint by the data. Model properties which can be target variables include: (1) model parameters (θ); (2) forcing variables, if there is substantial uncertainty about them; (3) initial conditions on the state variables; and (4) time-dependent components of the state vector. Land surface models are becoming more and more sophisticated with numerous parameters. Some parameters may be easily estimated, while others may not be very sensitive to the cost (objective) function. These target variables must be identified based on extensive sensitivity experiments. Selecting the target variables depends on the land surface model used in the DA system, along with how many and which data products are assimilated. For example, Kaminski et al. (2002) optimized 24 parameters (light use efficiency and Q10 for heterotrophic respiration for each of 12 biomes) in a terrestrial biosphere model coupled with an atmospheric transport model using CO2 data. Barrett (2002) optimized a set of parameters (turnover times, C allocation ratios, humification ratios, and light use efficiency) in a terrestrial C cycle model. Rayner et al. (2005) estimated 56 process parameters plus an initial condition through terrestrial carbon cycle DA system with a coupled ecosystem and atmospheric transport model. Williams et al. (2005) estimated nine unknown parameter constants in the C box model and the initial values of the five C pools using the ensemble Kalman filter. For a given land surface model, a rigorous sensitivity study is absolutely required, which enables determination of the parameters/state variables that are sensitive to the assimilated data. The automatic differentiation (AD) tool called the Tangent linear and Adjoint Model Compiler (Giering and Kaminski, 1998) as well as its successor, Transformation of Algorithms in Fortran (Giering and Kaminski, 2002) used to generate the adjoint code of the process model may provide an effective way for sensitivity studies.

12.3.4 Multi-Objectives Optimization Practical experience suggests that any single objective (cost) function, no matter how carefully chosen, is often inadequate to properly measure all of the characteristics of the assimilated data sets deemed to be important (Demarty et al., 2005; Vrugt et al., 2003). One strategy to circumvent this problem is to define several objective functions that measure different (complementary) aspects of the system behavior and to use multi-objective optimization to identify the set of non-dominated, efficient, or Pareto optimal solutions. Such an approach makes it more efficient to assimilate multiple data products simultaneously. The choice of fitting criterion or “estimator” is crucial. The very popular weighted least-squares estimator is the simplest, and is amenable to very efficient mathematical algorithms. However it may be seriously inappropriate in certain circumstances. For example, the estimator of maximum likelihood may be a much more reasonable choice when errors in the data or dynamical information are non-Gaussian.

322

S. Liang, J. Qin

Generally speaking, the multi-objective problem can be stated as the following minimizing formulation: Min{J1 (θ ), J2 (θ ), . . . , Jm (θ )}

(12.5)

a single objective function with i = 1, . . . , m and θ = where Ji symbolizes ( ) θ1 , θ2 , . . . , θ p a particular set of p parameters included in the feasible parameter space. This issue is highly related to those discussed in Section 12.3.2. A single objective function corresponds to each high-level product or directly observed variable, thus, we can combine both high-level products and direct observations. Even for the same product, a single objective function can be a measure of errors (Gupta et al., 1998), such as root mean square error, mean absolute error, maximum absolute error, and so on. The solution of the multi-objective formulation given in Eq. (12.5) does not lead to a unique solution, but to a set of solutions, generally named Pareto set or “behavioral” set (Gupta et al., 1998). There are many different algorithms to address this problem (Marler and Arora, 2004), such as the Multiobjective Shuffled Complex Evolution Metropolis (MOSCEM) algorithm that is capable of solving the multi-objective optimization problem for hydrologic models (Vrugt et al., 2003). MOSCEM is available to the public. Other assimilation algorithms will be discussed in the following Section 12.3.5.

12.3.5 Assimilation Algorithms The target variables characterizing the land surface properties are estimated through the assimilation algorithms. On one hand, there are many different assimilation algorithms available in the literature from meteorological and oceanographic DA communities, such as three- or four-dimensional variational algorithms (see Section 12.3.5.1), and Kalman filters (Section 12.3.5.2). On the other hand, land DA has different characteristics, such as handling only dozens of unknowns, while the meteorological DA system typically manages 106 –109 unknowns. Neural network and particle filtering techniques are briefly introduced in Sections 12.3.5.3 and 12.3.5.4. Before moving on, the measurement sequence is defined as y1:t ≡ {yi }ti=1 and the state sequence as x0:t ≡ {xi }ti=0 . With these definitions, the whole DA process as an estimation problem can be stated as: xˆ0:t = g(y1:t )

(12.6)

where xˆ0:t represents estimated state sequence, and g(·) the estimator, which can be viewed as the assimilation algorithm. When a measurement sequence y1:t is inserted into the estimator, we obtain a realization of the estimator (xˆ0:t ).

12 Data Assimilation Methods for Land Surface Variable Estimation

323

12.3.5.1 Cost Function-Based Methods This type of method is also called non-sequential or batch one. The most general form of cost function-based methods can be expressed as the following constrained optimization problem: min J = x0 − x¯0 2B−1 + ∑i=0 wi 2Q−1 + ∑i=0 ei 2R−1 t−1

t

x0:t

i

i

s.t. xi+1 = f (xi , ui , θ ) + wi , i = 0, . . . ,t − 1 yi = h(xi , ui , θ ) + ei i = 1, . . . ,t

(12.7)

where the weighting matrices Qi , Ri , and B can be regarded as covariances of the distribution functions p(wi ), p(ei ), and p(x0 ). If the parameter vector θ needs to * *2 be estimated simultaneously, a term *θ − θ¯ *P−1 can be added into Eq. (12.7) and θ x0:t . the control vector becomes θ This problem is the so-called weak constraint DA approach (see Section 12.3.2) when the dynamic model is considered to be imperfect. The traditional strong constraint DA problem is just its special case in which the model is assumed to be 2 is removed in Eq. (12.7). The formulation perfect and thus the term ∑t−1 i=0 wi Q−1 i statements can then be expressed as: min J = x0 − x¯0 2B−1 + ∑i=0 ei 2R−1 t

x0:t

i

s.t. xi+1 = f (xi , ui , θ ) i = 0, . . . ,t − 1 yi = h(xi , ui , θ ) + ei i = 1, . . . ,t

(12.8)

In the cost function-based problems, all data are treated simultaneously and the minimization problem is solved only once, as presented in Eq. (12.7) or (12.8). The problem becomes the usual optimization problem. It is computationally expensive when the number of unknowns is large. There are many non-sequential algorithms depending on whether the derivative information is used, but the key issue is to incorporate an effective global searching algorithm. The typical algorithms include the shuffled complex evolutionary (SCE) method (Duan et al., 1993; Duan et al., 1992; Duan et al., 1994), a very fast simulated annealing (SA) algorithm (Ingber, 1989; Li et al., 2004), the differential evolutionary (DE) method (Storn and Price, 1997; Storn and Price, 1996), and the genetic algorithm (GA) (Goldberg, 1989). Codes for these methods are available to the public. The common weakness of these methods is their slow computational speed. In order to solve both weak and strong constraint problems stated above, some descent algorithms, such as the conjugate gradient method, can be used. These approaches require the first-order derivative or even Hessian matrix of the cost function. To this end, the adjoint of the dynamic model has to be developed. However, this process is tedious. It is encouraging that automatic differentiation (AD) techniques have been developed and applied to automatically generate the adjoint model

324 Fig. 12.3 Illustrations of evaluating the derivatives of the cost function with AD techniques. F denotes the whole codes to evaluate the cost function. F ∗ represents the adjoint codes of the original codes, which can be used to evaluate derivates easily.

S. Liang, J. Qin

x

F

J

(a) ∂J ∂x

F '*

dJ = 1

(b)

at the level of computer codes (Bischof et al., 1996; Bischof et al., 2002; Carmichael et al., 1997; Dobmanin et al., 1995; Giering and Kaminski, 2002; Verma, 2000). AD is very effective and easy to use. This dramatically conserves time and energy of DA practitioners. The principle of AD is simple and based on two facts. First, any computer code statement can be regarded as a composition of elementary functions. Second, chain rule can be used to differentiate this composition of elementary functions. Many software packages have been developed in accordance with the principles described above for FORTRAN and C computer languages. They are given at the web site www.autodiff.org. The illustration of AD running process is presented in Fig. 12.3. We have recently applied this method in estimating LAI from satellite data (Qin et al., 2007a). The advantages of cost function-based method are twofold. First, all data in a batch window are used to estimate the state. Second, inequality, equality, and bound constraints can be included explicitly. Its disadvantages are also apparent. First, the adjoint model is generated to evaluate the derivative of the cost function if the highly efficient descent optimization method is used. However, the development of the adjoint model requires that the dynamic model should be differentiable. This condition can not usually be met in the land surface process modeling because of many discontinuous parameterizations. Instead, SCE, SA, DE, or GA could be used, but they are computationally very slow. Second, the cost function-based method just uses the inverses of covariance matrices as the weights, as seen in Eqs. (12.3) or (12.7) and (12.8). Since the covariance is the second moment of one distribution, more information included in the distributions p(wi ), p(ei ), and p(x0 ) is not used and therefore wasted. If p(wi ), p(ei ), and p(x0 ) are normal distributions, no information is discarded since Gaussian distributions are completely characterized with the first and second moments.

12.3.5.2 Sequential-Based Methods To derive the “optimal” sequential assimilation scheme, assume that the background states are represented by prior estimates. The sequential approaches are based on the Bayesian inference that combines the prior knowledge of the state vector and the measurement to obtain the posterior distribution of the state vector. Once obtaining the posterior distribution, everything is known about the state vector. This process can be expressed as follows.

12 Data Assimilation Methods for Land Surface Variable Estimation

p(x0:t |y1:t ) ∝ p(y1:t |x0:t )p(x0:t )

325

(12.9)

where p(x0:t ) represents the prior distribution of the state vector, p(y1:t |x0:t ) the measurement distribution, and p(x0:t |y1:t ) the posterior distribution. Typically, a Markov assumption is applied to the prior. So the state vector at time t only depends on the state vector at time t − 1: p(x0:t ) = p(x0 ) ∏i=1 p(xt |xt−1 ) t

(12.10)

where p(xt |xt−1 ) is the evolution distribution, and p(x0 ) is the distribution of the initial state vector (background or “first guesses”). Another important assumption is that measurements are independent given the true state: p(y1:t |x0:t ) = ∏i=1 p(yt |xt ) t

(12.11)

Substituting Eqs. (12.9) and (12.10) into (12.8), we obtain: p(x0:t |y1:t ) ∝ p(x0 ) ∏i=1 p(yt |xt )p(xt |xt−1 ) t

(12.12)

This equation implies that once new data is available, the previous estimate of the state process could be sequentially updated without having to calculate from scratch. However, this also means we have to store all state vectors up to time t, and the size of x0:t will expand as time goes by, becoming very large. In fact, there often is an interest in the filtering distribution p(xt |y1:t ), that is to estimate the probability density of the current state vector conditioned on the measurements up to now. The whole filtering process is straightforward. The filtering density, p(xt |y1:t ), and the one step prediction, p(xt+1 |y1:t ), density are recursively given by a measurement update according to p(xt |y1:t ) = p(yt |y1:t−1 ) =

p(yt |xt )p(xt |y1:t−1 ) p(yt |y1:t−1 )

&

(12.13)

p(yt |xt )p(xt |y1:t−1 )dxt

(12.14)

p(xt+1 |xt )p(xt |y1:t )dxt

(12.15)

and a time update according to p(xt+1 |y1:t ) =

&

and the recursion is initiated by p(x0 |y0:−1 ) = p(x0 )

(12.16)

In the general case, one is normally unable to obtain an analytical expression of the filtering density except under the assumption of linear model and observation, and Gaussian error distributions. This leads to the prominent Kalman filter (KF). If the dynamical model and measurement process are characterized as follows

326

S. Liang, J. Qin

xt+1 = Fxt + wt , wt ∼ N(0, Qt ) yt = Hxt + et , et ∼ N(0, Rt )

(12.17)

the KF can be expressed as xˆt+1|t = F xˆt|t Pt+1|t = FPt|t F T + Qt xˆt+1|t+1 = xˆt+1|t + G(yt+1 − H xˆt+1|t )

(12.18)

Pt+1|t+1 = (I − GH)Pt+1|t G = Pt|t−1 H T (H T Pt|t−1 H + Rt )−1 where xˆ denotes the estimate of the state vector. There are two popular variants of Kalman filter. One is the Extended Kalman filter (EKF) and the other is the Ensemble Kalman filter (EnKF). EKF handles the cases where either H or F is nonlinear. Note that most land surface models (H) are nonlinear. In this case, they can be defined as its tangent linear. The EnKF is a sophisticated sequential DA method. The EnKF applies an ensemble of model states to represent the error statistics of the model estimate, uses ensemble integrations to predict the error statistics forward in time, and employs an analysis scheme which operates directly on the ensemble of model states when observations are assimilated. The EnKF efficiently manages strongly nonlinear dynamics and large state spaces and is now used in realistic applications with primitive equation models for the ocean and atmosphere. Originally proposed by Evensen (1994), the EnKF is more recently reviewed by Evensen (2003), providing detailed information on the formulation, interpretation and implementation of the EnKF. Other sequential assimilation algorithms widely used in meteorological or oceanographic DA communities include successive correction, optimal or statistical interpolation, analysis correction, 3DVAR and 3DPSAS. All these methods have been widely applied in a variety of fields (Evensen, 2003; Houtekamer and Mitchell, 1998; Jones et al., 2004; Kumar and Kaleita, 2003; Qin et al., 2006; Reichle et al., 2002; Wade and Eric, 2003).

12.3.5.3 Neural Network-Based Methods In recent years, people have attempted to combine artificial neural network (ANN) and DA method (Liaqat et al., 2003a, b; Tang and Hsieh, 2001; Wu et al., 2005; Yu et al., 1997). The primary objective of these investigations is to complete or approximate the dynamic models from measurements by estimating weights and biases of ANN using DA method. The advent of the feed-forward neural network model opens the possibility of hybrid neural-dynamical models via variational DA. Such a hybrid model may be used in situations where some variables, difficult to model dynamically, have sufficient data for empirical modeling with ANN. Liaqat et al. (2001) used a neural network for constructing an arbitrary mapping function.

12 Data Assimilation Methods for Land Surface Variable Estimation

327

A neural network is trained by optimizing an object function composed of squared residuals of differential equations at collocation points and squared deviations of the observation data from the computed values. An assimilation problem is solved even if the model differential equations do not express the observed phenomena exactly. Since the dynamic model can be constructed directly from measurements and can be approximated, the computational speed greatly improves.

12.3.5.4 Particle Filtering Particle filtering (PF) is also called sequential Monte Carlo filtering based on point mass (or “particle”) representations of probability densities, which can be applied to any state-space model and which generalize the traditional KF methods (Ristic et al., 2004; Smith et al., 2005). It is an important technique to manage the DA with elements of nonlinearity and non-Gaussianity such that the underlying dynamics of a physical system are modeled accurately. It has been applied in many engineering fields and attracted attention from some DA practitioners since the posterior distribution of state vector can be represented with Monte Carlo samples. However, KF and its variants just evaluate the mean and covariance of the posterior distribution. PF better grasps the filtering density evolution of the nonlinear system in time than KF and its variants do. PF itself also has many variants, among which the bootstrap filter, also called sampling importance resampling filter (SIR), is the most easily used. Its main steps include: (i)

Step 1: for t = 0, sample {x0 }Ni=1 ∼ p(x0 ); (i)

(i)

(i)

(i)

(i)

(i)

Step 2: draw {x˜t+1 }Ni=1 ∼ p(xk+1 |xk ). That is x˜t+1 = f (xt , ut , θ ) + wt , wt ∼ p(wt ); (i)

Step 3: compute weights ut+1 =

(i)

p(yt+1 |xt+1 ) N

(i)

∑ p(yt+1 |xt+1 )

i=1

(i)

(i)

where p(yt+1 |xt+1 ) denotes the value of p(yt+1 − h(x˜t+1 , θ )); (i)

(i)

Step 4: resample {x˜t+1 }Ni=1 with replacement according to weights {ut+1 }Ni=1 in + (i) order to get {x˜t+1 }Ni=1 with weights {1 N }Ni=1 ; Step 5: set t = t + 1 and go to step 2. Observe that SIR is easy to be implemented in DA practice.

12.3.6 Observation Error Matrix Characterizing the errors in DA is extremely important since it significantly affects the estimates (Daley, 1991). The observation error covariance matrix R in Eqs. (12.3), (12.6) and (12.16) can in principle be quite general in form, including biases, temporal correlations, extreme outliers and so on. Many extant methods

328

S. Liang, J. Qin

assume the error to be Gaussian with zero mean and no temporal correlation. However, more general error structures are very common, and the development of methods for dealing with such errors is an active area of current research (Evensen, 2003; Raupach et al., 2005). All NASA Earth Observing System (EOS) land surface high-level products have been claimed to be “validated,” but most were based on limited “ground truths” (Morisette et al., 2006). The error magnitudes, and their spatial and temporal distributions, have never been well specified. Assigning uncertainties to these high level products is complex. There are a few studies reporting the accuracies of individual products, but comprehensive modeling of spatial and temporal error structures, and correlation among errors of different products has not yet been done. It raises the challenge of evaluating the uncertainty properties of major products, and it is evident that this is an enormous goal. A range of issues identified by Raupach et al. (2005) as needing to be addressed include: • The error magnitude rii for each high-level product, inclusive of all error sources (in other words, the diagonal elements of the covariance matrix R) • The correlations among errors in different products, quantified by the offdiagonal elements of the covariance matrix R • The temporal structure of the errors: whether they are random in time or temporally correlated, and the possible presence of unknown long-term drifts or biases • The spatial structure of errors (random, slowly varying or bias as for temporal structure) • The error distribution: normal (Gaussian), lognormal, skewed or the sum of multiple error sources with different distributions, such as a small Gaussian noise together with occasional large outliers because of measurement corruption events • Possible mismatches between the spatial and temporal averaging implicit in the model and the measurements (the “scaling problem”). This same set of challenging issues (to define and specify the error models) pertains to other data sets, such as the error properties of direct flux measurements and direct measurements of carbon stores in addition to remote sensing of land surface properties. It is more straightforward to assign uncertainties to lower-level products than to higher-level ones. This is a major factor for considering the assimilation of direct observations (radiance/reflectance or vegetation indices), as noted above.

12.4 Recent DA Applications DA is a powerful way to consistently combine measurements and dynamical models for the accurate estimation of model parameters and the state vector. Since DA has been explored and implemented in many applications, progress in the following selected fields is summarized.

12 Data Assimilation Methods for Land Surface Variable Estimation

329

12.4.1 Soil Moisture Estimation While microwave remote sensing provides the opportunity to map global soil content, the revisiting frequency is limited. Moreover, the L band brightness temperature is related to only the surface soil moisture (top 5 cm) and yields little information about the root zone. A land surface process model forced by atmospheric data can produce soil moisture and temperature profiles at the model time resolution. It is obvious that just running the model without any constraints can lead to large errors due to uncertainties in the model structure, model parameters, and external forcing data. DA offers a means to consistently take advantage of both modeling and observations. Two DA algorithms have been applied to estimate not only soil moisture, but also latent heat flux (tightly associated with the root-zone moisture) using L band microwave remote sensing data. Entekhabi et al. (1994, 1998) first applied EKF to assimilate microwave remote sensing data into a land surface model to subsequently retrieve soil water moisture and temperature profiles. Although synthetic data was used and the land surface model was relatively simple, their research opened up prospects for land surface DA. Many improvements have been made since then. Walker and Houser (2001) and Walker et al. (2001a, b) compared direct insertion and EKF assimilation methods, conducted retrieval experiments with field data, and then applied their algorithm on the regional scale for initialization of climate and hydrological models. But surface soil moisture data rather than satellite observations was assimilated in their studies. Reichle et al. (2002) used EnKF to assimilate L-band (1.4 GHz) microwave radiobrightness observations into a land surface model, compared it with the variational method, and investigated the influence of the ensemble size on the retrieved results. Their research indicates that the EnKF is a flexible and robust DA option that gives satisfactory estimates even for moderate ensemble sizes although the updating process is suboptimal. Crow (2003) and Crow and Wood (2003) applied EnKF to assimilate L-band microwave data to correct for the impact of poorly sampled rainfall on land surface model predictions of root-zone soil moisture and surface energy fluxes. The results suggest the EnKF-based assimilation system is capable of correcting a substantial fraction of model error in root zone (40 cm) soil moisture and latent heat flux predictions associated with the use of temporally sparse rainfall measurements as forcing data. The recent studies also applied EnKF to estimation of soil moisture profiles (e.g., Merlin et al., 2006; Zhou et al., 2006). Margulis et al. (2002) explained why the EnKF technique is so appealing for soil moisture estimation: (i) its sequential nature is well suited to real-time data streams and forecasting; (ii) it not only provides an estimate of surface and profile soil moisture, but information about the statistical confidence of estimation; (iii) it is sufficiently straightforward to use with “off-the-shelf” models; and (iv) it is relatively efficient, making its application to large-scale problems feasible. The variational cost function-based approach has also been successfully applied to hydrological studies in recent years. The key point in using the variational assimilation method is to develop the adjoint of a dynamic model. This requires the model should be differentiable, but land surface schemes usually are not.

330

S. Liang, J. Qin

Consequently some approximations have to be made. Reichle (2000) and Reichle et al. (2001) first used the variational method to assimilate L-band microwave data into a complicated land surface model to retrieve the soil moisture profile. The results showed that the state estimates obtained from the assimilation algorithm improve significantly over prior model predictions derived without assimilating radiobrightness data. Soil moisture estimation is also one of the major goals in the regional and global land surface DA systems, such as the global land DA system (Rodell et al., 2004), the North American Land DA system (Mitchell et al., 2004), and the South America Land DA system (de Goncalves et al., 2006).

12.4.2 Energy Balance Fluxes Estimation Accurate estimation of energy and momentum fluxes, especially sensible and latent heat fluxes, between the land surface and the atmospheric boundary layer, is required in a wide variety of agricultural, hydrological, and meteorological applications (Courault et al., 2005; Su, 2002). Many methods, such as eddy correlation and Bowen ratio, can be used to measure these fluxes at the field level. Their applicability, however, is limited on the fine spatial scale. Currently, the only way to achieve this goal of mapping fluxes regionally is to use remote sensing techniques that provide various spatial and temporal imageries covering large areas. Estimation of surface energy balance components using remote sensing data can be roughly divided into three categories: empirical, residual and DA. Empirical methods directly build on the relationship between remote sensing products, such as various vegetation indices and retrieved LST for the estimation. Residual methods of the energy budget couple some empirical formulas and physical mechanisms to realize the estimation of evapotranspiration (ET or LE) and sensible heat flux (H) by using remote sensing to directly estimate input parameters, such as the surface energy balance system (SEBS) (Su, 2002). In recent years, DA methods have integrated remote sensing data and soil–vegetation–atmosphere transfer models for estimating surface fluxes, and have achieved encouraging results. This new method has been receiving considerable attention from researchers in recent years (Boni et al., 2001a, b; Caparrini and Castelli, 2004; Caparrini et al., 2003; Castelli et al., 1999) since it combines dynamic models and temporal remote sensing data based on the control theory in order to accurately retrieve critical parameters for flux estimation. The first two methods more or less use only the instantaneous data and empirical relationships. Castelli et al. (1999) developed a simple land surface scheme and its adjoint, defined a moisture index as control variables, used radiometric surface temperature as observations, and performed DA experiments. Results indicated that this algorithm can retrieve land surface energy balance components effectively. Boni et al. (2001a, b) investigated the impact of land surface temperature sampling frequency on assimilation results and suggested that satellite remote sensing of land surface

12 Data Assimilation Methods for Land Surface Variable Estimation

331

temperature may be used to provide estimates of components of the surface energy balance and land surface control on evaporation. Caparrini et al. (2004, 2003) replaced a moisture index with evaporative fraction as control variables, made many assimilation experiments, and applied this algorithm on the regional scale using AVHRR land surface temperature as observations.

12.4.3 Carbon Fluxes Estimation It is increasingly recognized that DA (model–data synthesis, data–model fusion or many other names) is one of the methods that can meet such a challenge. The US Climate Change Science Program strategic plan and the North American Carbon Program (NACP) plan repeatedly call for the development of DA methods for carbon cycle studies. A near-term priority identified by the US Carbon DA Program Workshop Report (Fung et al., 2002) was urgent support “for interdisciplinary teams to develop component and coupled prototype carbon assimilation models.” Canadell et al. (2004) identified three fundamental research areas that require major development in order to provide policy relevant knowledge for managing the carbon-climate system over the next few decades, with the first area being “carbon observations and multiple constraint data assimilation.” For carbon cycle studies, some efforts have been made to develop DA systems, but most of these are based on “top-down” strategies that begin with measured changes in atmospheric carbon concentrations and attempts to infer the spatial distribution and magnitude of the net exchange. For example, the Carbon DataModel Assimilation (C-DAS) project at NCAR (National Center for Atmospheric Research) is developing a carbon DA system based on their atmospheric transport model (http://dataportal.ucar.edu/CDAS). An equally important approach in carbon cycle science is the “bottom-up” approach (Cihlar et al., 2000) that starts with a specific land parcel to account for the various pathways of carbon exchange between the ecosystem and the atmosphere, and then scales up to much larger regions. It relies on both ecosystem process models and spatial data sets. The model can be developed at the local level and validated using conventional measurements. Satellite observations provide spatial distribution and frequent up-to-date information on the rate of change of the variables driving the model. These models can in turn be used to estimate spatial and temporal variations in CO2 uptake and release over large areas, if appropriate “input” data sets (e.g., vegetation and soils maps, weather data) are available. Several obstacles stand in the way of extensive use of ecosystem process models for extrapolation. The major issue is the requirement of much more information on vegetation characteristics than is readily available or even known for many areas of the globe. Many parameters are not measurable at earth system scales, because of high small-scale spatial variability (e.g., any soil property), high temporal variability (e.g., stomatal conductance), and physical inaccessibility (e.g., most roots, deep soil). Moreover, from the modeling perspective, some processes are well understood

332

S. Liang, J. Qin

(e.g., photosynthesis, decomposition of litter and soil organic matter), while other processes are not (C-allocation among plant tissues, T-sensitivity of humus decomposition). As a result, large uncertainties arise in calculating terrestrial carbon cycle using the “bottom-up” approach (Barrett, 2002). Wang and Barrett (2003) developed a modeling framework that synthesizes various types of field measurements at different spatial and temporal scales to estimate monthly means (and their standard deviations) of gross photosynthesis, total ecosystem production, net primary production (NPP), and net ecosystem production (NEP) for eight regions of the Australian continent between 1990 and 1998. Williams et al. (2005) developed a DA approach combining stock and flux observations with a dynamic model to improve estimates of ecosystem carbon exchanges. Rayner et al. (2005) developed a terrestrial carbon cycle DA system (CCDAS) for determining the space-time distribution of terrestrial carbon fluxes for the period 1979–1999. Hazarika et al. (2005) integrated the MODIS LAI product with an ecosystem model for accurate estimation of NPP. Validations of results in Australia and the USA show that NPP estimated using the DA method to be more accurate than that generated by the data “forcing” method. Their research demonstrates the utility of combining satellite observations with an ecosystem process model to achieve improved accuracy in estimates and monitoring global net primary productivity. Barrett et al. (2005) demonstrated that a “multiple-constraints” model–data assimilation scheme using a diverse range of data types offers improved predictions of carbon and water budgets at regional scales. Xu et al. (2006) applied the Bayesian probability inversion and a Markov chain Monte Carlo (MCMC) technique to a terrestrial ecosystem model to analyze uncertainties of estimated carbon (C) transfer coefficients and simulated C pool sizes. Their study shows that the combination of a Bayesian approach and MCMC inversion technique effectively synthesizes information from various sources for assessment of ecosystem responses to elevated CO2 . Sacks et al. (2006) used a model–data synthesis approach with a simplified carbon flux model to extract process-level information from 5 years of eddy covariance data at an evergreen forest in the Colorado Rocky Mountains. Including water vapor fluxes, in addition to carbon fluxes, in the parameter optimization did not yield significantly more information about the partitioning of the net ecosystem exchange of CO2 into gross photosynthesis and ecosystem respiration. Sacks et al. (2007) used the model–data synthesis method to address fundamental questions about climate effects on terrestrial ecosystem net CO2 exchange.

12.4.4 Crop Yield Estimation Advance information on crop yield during the crop growing season is vital for effective crop management and for national food security policy. Agricultural harvested grain yield is a reliable way to estimate crop yields by sampling field measurements of standing crops. However, this method is both time consuming and costly, with results that are not available until after harvest. In the last three decades, satellite

12 Data Assimilation Methods for Land Surface Variable Estimation

333

remote sensing data have been used to estimate crop yields over large areas because these methods are more cost effective and more timely than traditional survey procedures (MacDonald and Hall, 1980). Earlier studies were mostly based on empirical regression methods that relate crop yield to remotely sensed surface reflectance and their combinations (i.e., vegetation indices). These relationships could be described with linear, cubic polynomial, or exponential regression (Jiang et al., 2003). Essentially a statistical model, this method cannot predict the time-dependent processes of crop growth. Additionally, the relationship between yield and NDVI may not be accurate under extreme weather conditions. Mathematical crop growth models simulate fundamental processes such as photosynthesis, respiration, biomass partitioning, and water and nitrogen transfers (Baret et al., 2000). This allows researchers to evaluate a wide array of alternatives, and to assemble processes in an integrated package. Along with supporting better crop management decisions, mechanical crop growth models can simulate the dynamics of LAI and other structural properties of the crop fields (e.g., height and biomass). The combination of remote sensing and crop growth simulation models is increasingly recognized as a promising approach for monitoring growth and estimating yield (Bauman, 1992). The use of crop models is often limited by uncertainties in their input parameters such as soil conditions, sowing date, planting density and initial field conditions. Except in some controlled experimental fields, many of these parameters are poorly known. Remote sensing can play a critical role in helping identify the field and crop status from estimated biophysical parameters (Clevers and Leeuwen, 1996). Remote sensing data, therefore, can be assimilated with crop growth models to improve their overall performance. Several assimilation schemes, of various degrees of complexity and integration, have been developed in the last 10 years (Moulin et al., 1998). Various methods for integrating a crop growth model with remote sensing data were described by Mass (1988a, b) and were also reviewed by Fischer et al. (1997) and Moulin et al. (1998). Mass (1993) compared the results of calibrating a crop simulation model using LAI observations obtained either from field sampling or remote sensing. Winter wheat yields were modeled more accurately using remotely sensed LAI observations than field-sampled LAI observations (Mass, 1993). This difference appeared to result from the apparent ability of the remotely sensed LAI observations to better represent the photosynthetically active plant area in the crop canopy. Bach et al. (2003) experimented with coupling a raster-based PROMET-V model with the radiative transfer model GeoSAIL to predict biomass and yield. In their study, LAI, fraction of brown leaves, and surface soil moisture were used as free variables; surface reflectance was used as the control variable. Their assimilation procedure produced improved biomass and yield results. Gu´erif and Duke (2000) combined the SUCROS crop model with the SAIL canopy reflectance model for accurate estimation of sugar beet yield. Ground measured reflectance was used to match the predicted reflectance. One limitation of their study is that many crop and soil parameters need to be obtained from field measurements. The SAIL model

334

S. Liang, J. Qin

has also been integrated with the EPIC crop model to estimate the yield of spring wheat in North Dakota (Doraiswamy et al., 2003). Planting date is the only adjustable variable in this model. The estimated yields are mostly within 10% of the NASS (National Agricultural Statistical Service) reports. In this work, climate data are based on interpolation of weather station measurements. The crop area mask is based on the 1 km AVHRR classification. NDVI is also calculated from AVHRR. Although the AVHRR data set is easily accessible, using it will compromise the precision of analyses owing to outdated calibrations and the application of partial atmospheric corrections. Doraiswamy et al. (2004) used a look-up table (LUT) method to estimate LAI from 250 m MODIS reflectance data. The crop modeled LAI was adjusted to fit the MODIS simulated LAI by changing planting time, time when maximum LAI is attained, and the beginning of leaf senescence. Jongschaap and Schouten (2005) estimated the regional wheat yield by assimilating SPOT data into a crop model. Microwave remote sensing data (ERS SAR C-band) were used to estimate regional wheat flowering dates to calibrate a wheat growth simulation model. Pauwels et al. (2007) assessed to what extent the results of a fully coupled hydrology–crop growth model can be optimized through the assimilation of observed LAI and soil moisture values using the EnKF. A practical procedure using the variational optimization method to predict crop yield at the regional scale from MODIS data was recently developed by Fang et al. (2007b). This method outputs agronomic variables (yield, planting, emergence and maturation dates) and biophysical parameters (e.g., LAI).

12.5 Summary Though DA has reached maturity in meteorological and oceanographic applications, the land community has just begun to employ it for estimation of land surface variables. Herein, basic DA principles have been described, critical issues in land DA have been identified, and many of the latest applications in hydrology, carbon cycle, and agriculture have been introduced. Because of the continuous improvement in DA methods and computational technology, along with an available wealth of remote sensing observations and extensive ground observation networks, DA is likely to become the best technique to monitor and map land surface environments by integrating a priori knowledge with an enormous variety and sheer volume of data. Revisiting the key issues addressed in this chapter, the following questions are put forth: 1. Which remotely sensed data products and other measurements are the most valuable for land DA? 2. How are the differences between the constrained data sets and predictions of a dynamic model characterized? 3. How can the model parameters/variables be estimated effectively? 4. How will the errors of the assimilated data be specified? 5. Which land surface properties can be estimated? 6. What types of dynamic models are suitable for DA?

12 Data Assimilation Methods for Land Surface Variable Estimation

335

To answer these questions, “explorers” from the community are needed to traverse this new frontier, and we hope they will do so. The computational issues of DA have not been broached since DA algorithms are usually computationally expensive. However, the fast pace of computer science advances promises to minimize such obstacles. Community efforts are needed to build the practical tools so that few researchers have to start from scratch. A good example of such endeavor is the land information system developed by NASA (Kumar et al., 2006). Ideally, more educational tools for this enterprise will follow. Acknowledgements We are very grateful for Dr. John Townshend for providing valuable comments. S. Liang is partially funded by NASA under grant NNG04GL85G.

References Bach H, Mauser W, Schneider K (2003) The use of radiative transfer models for remote sensing data assimilation in crop growth models. In: J Stafford, A Werner (eds), Precision agriculture: Papers from the 4th European Conference on Precision Agriculture, Berlin, Germany, 15–19 June 2003 Baret F, Weiss M, Troufleau D, Prevot L, Combal B (2000) Maximum information exploitation for canopy characterisation by remote sensing. Asp. Appl. Biol. 60(71–82) Barrett DJ (2002) Steady state turnover time of carbon in the Australian terrestrial biosphere. Glob. Biogeochem. Cycles 16(4):Art. No. 1108 Barrett DJ, Hill MJ, Hutley LB, Beringer J, Xu JH, Cook GD, Carter JO, Williams RJ (2005) Prospects for improving savanna biophysical models by using multiple-constraints model-data assimilation methods. Aust. J. Bot. 53(7):689–714 Bauman BAM (1992) Linking physical remote sensing models with crop growth simulation models applied for sugarbeet. Int. J. Remote Sens. 14:2565–2581 Bischof C, Khademi P, Mauer A, Carle A (1996) Adifor 2.0: Automatic Differentiation of Fortran 77 Programs. IEEE Comp. Sci. Eng. 3(3):18–32 Bischof CH, Hovland PD, Norris B (2002) Implementation of automatic differentiation tools. Proceedings of the 2002 ACM SIGPLAN Workshop on Partial Evaluation and Semantics-Based Program Manipulation (PEPM-02), vol. 37, pp. 98–107 Boni G, Castelli F, Entekhabi D (2001a) Sampling Strategies and Assimilation of Ground Temperature for the Estimation of Surface Energy Balance Components. IEEE Trans. Geosci. Remote Sens. 39(1):165–172 Boni G, Entekhabi D, Castelli F (2001b) Land data assimilation with satellite measurements for the estimation of surface energy balance components and surface control on evaporation. Water Resour. Res. 37(6):1713–1722 Canadell JG, Ciais P, Cox P, Heimann M (2004) Quantifying, understanding and managing the carbon cycle in the next decades. Climatic Change 67(2–3):147–160 Caparrini F, Castelli F (2004) Estimation of surface turbulent fluxes through assimilation of radiometric surface temperature sequences. J. Hydrometeorol. 5:145–159 Caparrini F, Castelli F, Entekhabi D (2003) Mapping of land-atmosphere heat fluxes and surface parameters with remote sensing data. Bound. Lay. Meteorol. 107:605–633 Carmichael GR, Sandu A, Potra FA (1997) Sensitivity analysis for atmospheric chemistry models via automatic differentiation. Atmos. Environ. 31(3):475–489 Castelli F, Entekhabi D, Caporali E (1999) Estimation of surface heat flux and an index of soil moisture using adjoint-state surface energy balance. Water Resour. Res. 35(10):3115–3125

336

S. Liang, J. Qin

Chepurin GA, Carton JA, Dee D (2005) Forecast model bias correction in ocean data assimilation. Monthly Weather Rev. 133(5):1328–1342 Cihlar J, Denning AS, Gosz J (2000) Global Terrestrial Carbon Observation: Requirements, Present Status, and Next Steps, Report of a Synthesis Workshop, 8–11 February 2000, Ottawa, Canada Clevers JPGW, Leeuwen V (1996) Combined use of optical and microwave remote sensing data for crop growth monitoring. Remote Sens. Environ. 56:42–51 Courault D, Seguin B, Olioso A (2005) Review on estimation of evapotranspiration from remote sensing data: from empirical to numerical modeling approaches. Irrig. Drain. Syst. 19:223–249 Crow WT (2003) Correcting land surface model predictions for the impact of temporally sparse rainfall rate measurements using an ensemble Kalman filter and surface brightness temperature observations. J. Hydrometeorol. 4(5):960–973 Crow WT, Wood EF (2003) The assimilation of remotely sensed soil brightness temperature imagery into a land surface model using Ensemble Kalman filtering: a case study based on ESTAR measurements during SGP97. Adv. Water Resour. 26(2):137–149 Dai Y, Zeng X, Dickinson RE, Baker I, Bonan GB, Bosilovich MG, Denning AS, Dirmeyer PA, Houser PR, Niu G, Oleson KW, Schlosser CA, Yang Z.L (2003) The Common Land Model (CLM). Bull. Am. Meteorol. Soc., 84(8):1013 Daley R (1991) Atmospheric data analysis. Cambridge University Press, New York. de Goncalves LGG, Shuttleworth WJ, Burke EJ, Houser P, Toll DL, Rodell M, Arsenault K (2006) Toward a South America Land Data Assimilation System: Aspects of land surface model spinup using the simplified simple biosphere. J. Geophys. Res. -Atmos. 111(D17) Demarty J, Ottle C, Braud I, Olioso A, Frangi JP, Gupta HV, Bastidas LA (2005) Constraining a physically based Soil-Vegetation-Atmosphere Transfer model with surface water content and thermal infrared brightness temperature measurements using a multiobjective approach. Water Resour. Res. 41(1):Art. No. W01011 Dobmanin M, Liepelt M, Schittkowski K (1995) Algorithm 746: PCOMP: a Fortran Code for Automatic Differentiation. ACM Trans. Math. Software 21(3):233–266 Doraiswamy PC, Hatfield JL, Jackson TJ, Akhmedov B, Prueger J, Stern A (2004) Crop condition and yield simulation using Landsat and MODIS. Remote Sens. Environ. 92(4):548–559 Doraiswamy PC, Moulin S, Cook PW, Stern A (2003) Crop yield assessment from remote sensing. Photogramm. Eng. Remote Sens. 69(6):665–674 Duan QY, Gupta VK, Sorooshian S (1993) Shuffled complex evolution approach for effective and efficient global minimization. J. Optim. Theor. Appl. 76(3):501–521 Duan QY, Sorooshian S, Gupta VK (1992) Effective and efficient global optimization for conceptual rainfall-runoff models. Water Resour. Res. 28(4):1015–1031 Duan QY, Sorooshian S, Gupta VK (1994) Optimal use of sce-ua global optimization method for calibrating watershed models. J. Hydrol. 158:265–284 Entekhabi D, Nakamura H, Njoku EG (1994) Solving the inverse problem for soil moisture and temperature profiles by sequential assimilation of multifrequency remotely sensed observations. IEEE Trans. Geosci. Remote Sens. 32(2):438–448 Evensen G (1994) Sequential data assimilation with a non-linear quasi-geostrophic model using Monte Carlo methods to forecast statistics. J. Geophys. Res. 99(C5):10143–10162 Evensen G (2003) The Ensemble Kalman Filter: theoretical formulation and practical implementation. Ocean Dyn. 53:343–367 Fang H, Kim H, Liang S, Schaaf C, Strahler A, Townshend GRG, Dickinson R (2007a) Spatially and temporally continuous land surface albedo fields and validation. J. Geophys. Res. 112, doi: 10.1029/2006JD008377 Fang H, Liang S, Hoogenboom G, Teasdale J, Cavigelli M (2007b) Crop yield estimation through assimilation of remotely sensed data into DSSAT-CERES. Int. J. Remote Sens. doi: 10.1080/ 01431160701408386 Fang H, Liang S, Townshend J, Dickinson R (2007c) Spatially and temporally continuous LAI data sets based on an new filtering method: examples from North America. Remote Sen. Environ. doi:10.1016/j.rse.2006.07.026

12 Data Assimilation Methods for Land Surface Variable Estimation

337

Fischer A, Kergoat L, Dedieu G (1997) Coupling satellite data with vegetation functional models: review of different approaches and perspectives suggested by the assimilation strategy. Remote Sens. Rev. 15:283–303 Fung I, Kalnay E, Schimel D, Denning S, Doney S, Pawson S (2002) A US Carbon Data Assimilation Program Workshop Report, University of Maryland, College Park Giering R, Kaminski T (1998) Recipes for Adjoint Code Construction. ACM Trans. Math. Software 24(4):437–474 Giering R, Kaminski T (2002) Generating recomputations in reverse mode AD. In: G Corliss, A Griewank, FCL Hascoet, U Naumann (eds), Automatic differentiation of algorithms: from simulation to optimization. Springer, Heidelberg, pp 283–291 Goldberg DE (1989) Genetic algorithms in search, optimization and machine learning. AddisonWesley, Reading, MA Gu´erif M, Duke CL (2000) Adjustment procedures of a crop model to the site specific characteristics of soil and crop using remote sensing data assimilation. Agric. Ecosyst. Environ. 81:57–69 Gupta HV, Sorooshian S, Yapo PO (1998) Toward improved calibration of hydrologic models: multiple and noncommensurable measures of information. Water Resour. Res. 34(4):751–763 Hazarika MK, Yasuoka Y, Ito A, Dye D (2005) Estimation of net primary productivity by integrating remote sensing data with an ecosystem model. Remote Sen. Environ. 94(3):298–310 Houtekamer PL, Mitchell HL (1998) Data assimilation using an Ensemble Kalman Filter technique. Monthly Weather Rev. 126(3):796–811 Ingber L (1989) Very fast simulated re-annealing. J. Math. Comput. Model. 12(8):967–973 Jiang D, Wang NB, Yang XH, Wang JH (2003) Study on the interaction between NDVI profile and the growing status of crops. Chin. Geogr. Sci. 13(1):62–65 Jones JW, Graham WD, Wallach D, Bostick WM, Koo J (2004) Estimating soil carbon levels using an Ensemble Kalman filter. Trans. ASAE 47(1):331–339 Jongschaap REE, Schouten LSM (2005) Predicting wheat production at regional scale by integration of remote sensing data with a simulation model. Agron. Sustain. Dev. 25(4):481–489 Julien Y, Sobrino JA, Verhoef W (2006) Changes in land surface temperatures and NDVI values over Europe between 1982 and 1999. Remote Sens. Environ. 103:43–55 Kaminski T, Knorr W, Rayner PJ, Heimann M (2002) Assimilating atmospheric data into a terrestrial biosphere model: a case study of the seasonal cycle. Glob. Biogeochem. Cycles 16(4):Art No. 1066 Kumar P., Kaleita AL (2003) Assimilation of near-surface temperature using Extended Kalman Filter. Adv. Water Resour. 26:79–93 Kumar SV, Peters-Lidard CD, Tian Y, Houser PR, Geiger J, Olden S, Lighty L, Eastman JL, Doty B, Dirmeyer P, Adams J, Mitchell K, Wood EF, Sheffield J (2006) Land information system: an interoperable framework for high resolution land surface modeling. Environ. Model. Software, 21(10):1402–1415 Li X, Koike T, Pathmathevan M (2004) A very fast simulated re-annealing (VFSA) approach for land data assimilation. Comput. Geosci. 30(3):239–248 Liang S (2004) Quantitative remote sensing of land surfaces. Wiley, New York Liaqat A, Fukuhara M, Takeda T (2001) Application of neural network collocation method to data assimilation. Comput. Phys. Commun. 141(3):350–364 Liaqat A, Fukuhara M, Takeda T (2003a) Applying a neural network collocation method to an incompletely known dynamical system via weak constraint data assimilation. J. Mar. Syst. 131(8):1696–1714 Liaqat A, Fukuhara M, Takeda T (2003b) Optimal estimation of parameters of dynamical systems by neural network collocation method. Comput. Phys. Commun. 150(3):215–234 Losa SN, Kivman GA, Ryabchenko VA (2004) Weak constraint parameter estimation for a simple ocean ecosystem model: what can we learn about the model and data? J. Mar. Syst. 45(1–2): 1–20 MacDonald R, Hall F (1980) Global crop forecasting. Science 208:670–679 Margulis S, McLaughlin DB, Entekhabi D, Dunne S (2002) Land data assimilation and soil moisture estimation using measurements from the Southern Great Plains 1997 Field Experiment. Water Resour. Res. 38(12):1299, doi:10.1029/2001WR001114

338

S. Liang, J. Qin

Marler RT, Arora JS (2004) Survey of multiobjective optimization method for engineering. Struct. Multidisc. Optim. 26(6):369–395 Mass S (1988a) Using satellite data to improve model estimates of crop yield. Agron. J. 80:662– 665 Mass SJ (1988b) Use of remotely-sensed information in agricultural crop growth models. Ecol. Model. 41:247–268 Mass SJ (1993) Within-season calibration of modeled wheat growth using remote sensing and field sampling. Agron. J. 85(3):669–672 Merlin O, Chehbouni A, Boulet G, Kerr Y (2006) Assimilation of disaggregated microwave soil moisture into a hydrologic model using coarse-scale meteorological data. J. Hydrometeorol. 7(6):1308–1322 Mitchell KE, Lohmann D, Houser PR, Wood EF, Schaake JC, Robock A, Cosgrove BA, Sheffield J, Duan QY, Luo LF, Higgins RW, Pinker RT, Tarpley JD, Lettenmaier DP, Marshall CH, Entin JK, Pan M, Shi W, Koren V, Meng J, Ramsay BH, Bailey AA (2004) The multi-institution North American Land Data Assimilation System (NLDAS): utilizing multiple GCIP products and partners in a continental distributed hydrological modeling system. J. Geophys. Res. -Atmos. 109(D7):Art. No. D07S90 Moody EG, King MD, Platnick S, Schaaf CB, Gao F (2005) Spatially complete global spectral surface albedos: value-added datasets derived from terra MODIS land products. IEEE Trans. Geosci. Remote Sens. 43(1):144–158 Morisette JT, Baret F, Privette JL, Myneni RB, Nickeson JE, Garrigues S, Shabanov NV, Weiss M, Fernandes RA, Leblanc SG, Kalacska M, Sanchez-Azofeifa GA, Chubey M, Rivard B, Stenberg P, Rautiainen M, Voipio P, Manninen T, Pilant AN, Lewis TE, Iiames JS, Colombo R, Meroni M, Busetto L, Cohen WB, Turner DP, Warner ED, Petersen GW, Seufert G, Cook R (2006) Validation of global moderate-resolution LAI products: a framework proposed within the CEOS Land Product Validation subgroup. IEEE Trans. Geosci. Remote Sens. 44(7):1804– 1817 Moulin S, Bondeau A, Del´ecolle R (1998) Combining agricultural crop models and satellite observations; from field to regional scales. Int. J. Remote Sens. 19(6):1021–1036 Natvik LJ, Eknes M, Evensen G (2001) A weak constraint inverse for a zero-dimensional marine ecosystem model. J. Mar. Syst. 28(1–2):19–44 O’Neill A, Mathieu PP, Zehner C (2004) Making the most of earth observation with data assimilation. ESA Bulletin 118:33–38 Pauwels VRN, Verhoest NEC, De Lannoy GJM, Guissard V, Lucau C, Defourny P (2007) Optimization of a coupled hydrology-crop growth model through the assimilation of observed soil moisture and leaf area index values using an ensemble Kalman filter. Water Resour. Res. 43(4) Qin J, Liang S, Li X, Wang J (2007a) Development of the adjoint model of a canopy radiative transfer model for sensitivity study and inversion of leaf area index. IEEE Trans. Geosci. Remote Sens. (revised) Qin J, Liang S, Liu R, Zhang H, Hu B (2007b) A weak-constraint based data assimilation scheme for estimating surface turbulent fluxes. IEEE Geosci. Remote Sens. Lett. 4(4):649–653 Qin J, Yan G, Liu S, Liang S, Zhang H, Wang J, Li X (2006) Application of Ensemble Kalman Filter to geophysical parameter retrieval in remote sensing: a case study of kernal-driven BRDF model inversion. Sci. China, Series D, 49(6):632–640 Raupach MR, Rayner PJ, Barrett DJ, DeFries RS, Heimann M, Ojima DS, Quegan S, Schmullius CC (2005) Model-data synthesis in terrestrial carbon observation: methods, data requirements and data uncertainty specifications. Glob. Change Biol., 11(3):378–397 Rayner PJ, Scholze M, Knorr W, Kaminski T, Giering R, Widmann H (2005) Two decades of terrestrial carbon fluxes from a carbon cycle data assimilation system (CCDAS). Glob. Biogeochem. Cycles 19(2):Art. No. GB2026 Reichle RH (2000) Variational Assimilation of Remote Sensing Data for Land Surface Hydrologic Applications, MIT, Boston, MA Reichle RH, McLaughlin DB, Entekhabi D (2001) Variational data assimilation of microwave radiobrightness observations for land surface hydrology applications. IEEE Trans. Geosci. Remote Sens. 39(8):1708–1718

12 Data Assimilation Methods for Land Surface Variable Estimation

339

Reichle RH, McLaughlin DB, Entekhabi D (2002) Hydrologic data assimilation with the Ensemble Kalman Filter. Month. Weather Rev. 130:103–114 Ristic B, Aruampalam S, Gordon N (2004) Beyond the Kalman Filter. Artech House, Boston/London Rodell M, Houser PR, Jambor U, Gottschalck J, Mitchell K, Meng CJ, Arsenault K, Cosgrove B, Radakovich J, Bosilovich M, Entin JK, Walker JP, Lohmann D, Toll D (2004) The global land data assimilation system. Bull. Am. Meteorol. Soc. 85(3):381 Sacks WJ, Schimel DS, Monson RK, 2007. Coupling between carbon cycling and climate in a high-elevation, subalpine forest: a model-data fusion analysis. Oecologia 151(1):54–68 Sacks WJ, Schimel DS, Monson RK, Braswell BH (2006) Model-data synthesis of diurnal and seasonal CO2 fluxes at Niwot Ridge, Colorado. Glob. Change Biol. 12(2):240–259 Smith A, Doucet A, Freitas ND, Gordon N (eds) (2005) Sequential Monte Carlo methods in practice. Springer, New York Storn R, Price K (1997) Differential Evolution – a simple and efficient heuristic for global optimization over continuous space. J. Global Optim. 11:341–359 Storn R, Price KV (1996) Minimizing the real function of the ICEC’96 contest by differential evolution, IEEE conference on evolutionary computation, pp 842–844 Su Z (2002) The Surface Energy Balance System (SEBS) for estimation of turbulent heat fluxes. Hydrol. Earth Syst. Sci. 6(1):85–99 Tang YM, Hsieh WW (2001) Coupling neural networks to incomplete dynamical systems via variational data assimilation. Month. Weather Rev. 129(4):818–834 Verma A (2000) An introduction to automatic differentiation. Curr. Sci. 78(7):804–807 Vrugt JA, Gupta HV, Bastidas LA, Bouten W, Sorooshian S (2003) Effective and efficient algorithm for multiobjective optimization of hydrologic models. Water Resour. Res. 39(8):Art No. 1214 Wade TC, Eric FW (2003) The assimilation of remotely sensed soil brightness temperature imagery into a land surface model using Ensemble Kalman Filtering: a case study based on ESTAR measurements during SGP97. Adv. Water Resour. 26(2):137–149 Walker JP, Houser PR (2001) A methodology for initializing soil moisture in a global climate model: assimilation of near-surface soil moisture observations. J. Geophys. Res. -Atmos 106(D11):11761–11774 Walker JP, Willgoose GR, Kalma JD (2001a) One-dimensional soil moisture profile retrieval by assimilation of near-surface measurements: a simplified soil moisture model and field application. J. Hydrometeorol. 2(4):356–373 Walker JP, Willgoose GR, Kalma JD (2001b) One-dimensional soil moisture profile retrieval by assimilation of near-surface observations: a comparison of retrieval algorithms. Adv. Water Resour. 24(6):631–650 Wang YP, Barrett DJ (2003) Estimating regional terrestrial carbon fluxes for the Australian continent using a multiple-constraint approach I. Using remotely sensed data and ecological observations of net primary production. Tellus Ser B Chem. Phys. Meteorol. 55:270–289 Williams M, Schwarz PA, Law BE, Irvine J, Kurpius MR (2005) An improved analysis of forest carbon dynamics using data assimilation. Glob. Change Biol. 11(1):89 Wu J, Fukuhara M, Takeda T (2005) Parameter estimation of an ecological system by a neural network with residual minimization training. Ecol. Model. 189(3–4):289–304 Xu T, White L, Hui DF, Luo YQ (2006) Probabilistic inversion of a terrestrial ecosystem model: analysis of uncertainty in parameter estimation and model prediction. Glob. Biogeochem. Cycles 20(2) Yu TW, Iredell M, Keyser D (1997) Global data assimilation and forecast experiments using SSM/I wind speed data derived from a neural network algorithm. Weather Forecast. 12(4):859–865 Zhou YH, McLaughlin D, Entekhabi D (2006) Assessing the performance of the ensemble Kalman filter for land surface data assimilation. Month. Weather Rev. 134(8):2128–2142

Chapter 13

Methodologies for Mapping Land Cover/Land Use and its Change Nina Siu-Ngan Lam

13.1 Introduction Mapping and identifying land cover/land use and its change is the most important, as well as the most widely researched, topic in remote sensing. Land cover/land use has been used extensively to derive a number of biophysical variables, such as vegetation index, biomass, and carbon content (see other chapters). More importantly, land cover/land use pattern and its change reflect the underlying natural and/or social processes, thus providing essential information for modeling and understanding many different phenomena on the Earth. Knowledge of land cover/land use and its change is also critical to effective planning and management of natural resources. Mapping land cover/land use accurately and efficiently via remote sensing requires good image classification methods. Unfortunately, there are numerous factors (e.g., image resolution and atmospheric condition) that could affect the effectiveness and accuracy of the classification algorithms. Different land cover/land use classification methods may be needed for different problems under different environmental conditions, making generalization and hence automation of the image classification process across time and space extremely difficult. As a result, new and sophisticated classification methods designed to improve the classification process continue to appear in the literature (e.g., Jensen, 2005; Gong, 2006). Newer approaches such as fuzzy classification, artificial neural network, and object-based classification have been developed and successfully applied (Definiens, 2004; Benz et al., 2004). However, these methods require extensive training and human supervision. We are still far from being able to develop a common framework to successfully identify a variety of features in different landscapes and to generalize and automate the classification process.

Nina Siu-Ngan Lam Department of Environmental Studies, Louisiana State University, Baton Rouge, USA [email protected] S. Liang (ed.), Advances in Land Remote Sensing, 341–367. c Springer Science + Business Media B.V., 2008

341

342

N.S.-N. Lam

Extending the mapping and modeling of land cover/land use at one time period to multiple time periods to analyze change will undoubtly add more complexity and challenges. In addition to the above image classification issues, efficient methods are needed to ensure comparability and compatibility of images taken in different time periods. Many studies on change detection using remote sensing imagery have already been reported in the literature. Lunetta and Elvidge (1998) provided an excellent summary of the state of the science. Many others examined the performance of various techniques in various applications (e.g., Coppin and Bauer, 1996; Jensen et al., 1993, 1997; Lu et al., 2005; Mas, 1999; Nackaerts et al., 2005; Yuan et al., 1998). Among the new techniques for land cover/land use classification and change analysis, textural (spatial) analyses are gaining increasing attention from the remote sensing community (e.g., Briggs and Nellis, 1991; Dunn et al., 1991; Estreguil and Lambin, 1996; Frank, 1984; Jupp et al., 1986; Lambin, 1996; Lambin and Strahler, 1994; Pickup and Foran, 1987; Smits and Annoni, 2000; Crews-Meyer, 2002). We have seen new applications of old textural measures such as the spatial co-occurrence matrix, local variance, and others (Haralick et al., 1973; Clausi and Jobanputra, 2006), as well as development of new textural analytical techniques such as wavelets (Daubechies, 1990; Muneeswaran et al., 2005). A number of textural measures which had not been used for remote sensing applications before have recently been utilized for more accurate land cover/land use classification, such as fractals, variograms, lacunarity, and spatial autocorrelation statistics (Lam, 1990; Lam and De Cola, 1993; Plotnick et al., 1993; Lam et al., 1998; Carr and de Miranda, 1998; Carr, 1999; Dale, 2000; Dong, 2000; Franklin et al., 2000). Although applications of these newer textural measures in change analysis have seldom been reported, we expect that the same textural and spatial methods that can be used for identification of land cover/land use can also be used for change detection. The purpose of this chapter is to introduce the use of textural/spatial measures in land cover/land use classification and its potential for change analysis. Our main notion is that the utilization of textural/spatial measures, in combination with original spectral information, will increase classification accuracy and have great potential for rapid change detection. The chapter is organized into four main sections. A summary of the major textural/spatial methods is first provided in Section 13.2, with a focus on those measures that can be applied directly to unclassified images. This property is important for rapid image segmentation, classification, and change detection. Section 13.3 describes a number of examples that have utilized these measures in land cover/land use classification. In Section 13.4, a framework for classifying the various land cover/land use change detection methods is introduced. We argue for the need to develop innovative methods for rapid and reliable change detection especially during disastrous and unexpected events. We further argue that such need could be best served by utilizing a metric of textural/spatial measures, in conjunction with the original spectral information of the images. Section 13.5 describes a real example of change analysis using textural measures. The prospect of utilizing textural measures in combination with other approaches for better and faster mapping of land cover/land use and its change is summarized in the conclusion.

13 Methodologies for Mapping Land Cover/Land Use and its Change

343

13.2 Major Textural/Spatial Measures 13.2.1 Terminology First, some clarification of the terms textural vs spatial measure is in order. The term “textural” is more commonly used in pattern recognition and the general field of image processing for raster/pixel data, while the term “spatial” is usually adopted in geography, economics, statistics and other related disciplines and is derived mostly for vector/polygonal data. Texture is an important characteristic in many types of images. Despite its importance, a formal definition of texture does not exist. Haralick (1979) attempted to characterize texture using two properties, the tonal primitive properties as well as the spatial interrelationships among them. Under this two-layered tone-texture concept, when an image has little variation of tonal primitives, the dominant property of that image is tone. On the contrary, when an image has wide variation of tonal primitives, the dominant property of that image is texture. Haralick described eight statistical approaches to measure image texture, which include autocorrelation functions, optical transforms, digital transforms, textural edgeness, structural elements, spatial gray tone co-occurrence probabilities, gray tone run lengths, and autoregressive models. In remote sensing, texture is the spatial relationship exhibited by gray levels in a digital image. Therefore, textural measures are measures that capture the spatial relationship among pixels. Spatial measures, which refer to measures mostly derived from spatial statistics, have been used largely in geospatial applications for characterizing and quantifying spatial patterns and processes Although traditionally the two fields of studies, textural analysis and spatial analysis, refer to quite different sets of methods and analyses, they do intersect to a great extent in remote sensing. In this chapter, we adopt the notion that the two terms, textural and spatial measures, are interchangeable.

13.2.2 Criteria for Evaluating and Classifying Textural Measures There are numerous textural measures in the literature, and it is beyond the scope of this chapter to exhaust and evaluate each of them. Table 13.1 lists some of the commonly used measures, which illustrates clearly the diversity, as well as redundancy, of these various measures. It is important to note that some measures are based on vigorous statistical theory and mathematical derivation (e.g., fractals, wavelets, spatial autocorrelation) (Lam et al., 1998), while others are based on simple geometric measurements with unknown statistical properties and/or theoretical minimum or maximum (e.g., edge density). Some metrics measure one aspect of the landscape (e.g., landscape composition), while others measure another (e.g., landscape configuration) (McGarigal and Mark, 1995). Baskent and Jordan (1995) classified

344

N.S.-N. Lam

Table 13.1 Some commonly used textural measures Textural measures

References

First-order metrics (computed on the original data matrix) using traditional statistical measures: mean, standard deviation, variance, correlation First-order metrics using traditional texture measures: entropy, energy, contrast, homogeneity, angular second moment, Shannon diversity First-order metrics using adapted spatial measures (from ICAMS): fractal dimension, lacunarity, spatial autocorrelation, variogram Second-order metrics: same set of metrics calculated on matrix derived from the original matrix (e.g., gray-level co-occurrence matrix, wavelet decomposed images) Landscape indices designed for classified images (from FRAGSTATS): area, density, edge, shape, proximity, connectivity, contagion/interspersion, diversity

Jensen (2005); Gong et al. (1992); Woodcock and Strahler (1987) Haralick (1979); Haralick et al. (1973); Jensen (2005); Gong et al. (1992) Lam et al. (1998, 2002); Quattrochi et al. (1997); Myint and Lam (2005a, b) Haralick (1979); Haralick et al. (1973); Myint et al. (2002, 2004) McGarigal and Mark (1995); McGarigal (2002)

landscape indices into areal, linear, and topological. Yet another classification of textural measures is based on whether the measures are applied directly to the original matrix (first-order textural measures), or to matrices derived from the original matrix such as the gray-level co-occurrence matrix or wavelet decomposed images (second-order textural measures) (Jensen, 2005). Hence, it is imperative to develop useful criteria to evaluate and/or classify these textural measures. What is a good textural measure? Although there may be no definite answers until each measure is tested extensively for their discriminating and explanatory power, we suggest the following criteria to evaluate and guide our understanding of these various measures. Ideally, a good textural measure should have the following properties: 1. The textural measure should be conceptually simple and easy to calculate. For example, statistical mean and standard deviation are concepts easily grasped by most researchers and their statistical properties are well known. Extension of these basic statistical measures in a spatial domain with some modifications may provide a useful first approximation towards an understanding of the land cover/land use pattern being studied. On the contrary, some measures may be conceptually simple but require additional steps for calculation, such as edge density, which needs an additional step to find the edges. 2. The textural measure should have theoretical maximum and minimum. For example, Moran’s I, a most commonly used spatial autocorrelation statistic, has a range of ±1. A Moran’s I value of 1 indicates a maximum positive spatial autocorrelation; on the contrary, a −1 indicates a maximum negative spatial autocorrelation (Cliff and Ord, 1973). 3. The textural measure should reflect clearly and intuitively the characteristics of the image pattern in a consistent manner. For example, a lower fractal dimension value means a less spatially complex image, therefore, given an image

13 Methodologies for Mapping Land Cover/Land Use and its Change

345

computed with a fractal dimension value of 2.3, we should be able to infer that this image is far less complex than an image computed with a fractal dimension value of 2.9, and a visual display of the two images should be able to reveal the difference (Lam, 1990). Fractal dimension (D) also has the second property, where D is expected to range from 2 to 3.0 for surfaces and 1.0–2.0 for lines (Mandelbrot, 1982). 4. The statistical properties of the textural measure should be known to provide statistical confidence of the computed value. For example, theoretically a Moran’s I value of 0 indicates a random pattern. If a pattern yields a computed value of 0.2, can we determine if this value is statistically the same or different from 0 to conclude if the pattern is random or not? Fortunately, the statistical properties of Moran’s I are relatively well known and hypothesis testing of whether a computed I value is significant can be conducted. Under the assumption of randomization, the first and second moments of the Moran’s I value can be computed and the statistical significance of the value determined (Goodchild, 1986). On the contrary, the statistical properties of fractal dimension are still not clear, though it has well-defined theoretical minimum and maximum. Hence, it is difficult to judge, for example, if an image with a fractal dimension of 2.3 is significantly different from another image with a fractal dimension of 2.4. It is noted that the statistical properties of most spatial measures are very difficult to derive and therefore remain unclear; many researchers have resorted to the Monte Carlo approach to develop empirical probability functions for statistical hypothesis testing (e.g., Openshaw, 1989). 5. The textural measure should be computable globally for the entire study area or locally for a local neighborhood. For example, some landscape metrics developed in FRAGSTATS are only computable at the landscape level, instead of at all levels (patch, class, and landscape) (McGarigal, 2002), whereas mean, variance, Moran’s I, and fractals can be applied both globally and locally to capture local change (Lam, 2004; Emerson et al., 2005). This property refers only to whether the measure can be computed at all levels; it does not necessarily imply that the measure is useful in describing the landscape at all levels. 6. Finally, the textural measure should be applicable directly to both classified and unclassified images. For example, the landscape metrics in FRAGSTATS were developed exclusively for categorical maps (O’Neill et al., 1998; McGarigal, 2002), or in other words, classified images, though some of the metrics can be modified and applied to unclassified images. On the contrary, fractals, Moran’s I, local variance (Woodcock and Strahler, 1987), variogram, lacunarity, and wavelet measures can be applied to both unclassified and classified images. This last property is considered very important to automated land cover/land use classification and change detection for two reasons. First, if they can be applied directly to unclassified images, land cove change could be detected first before the tedious classification process. Only after the change is determined to be significant, then there is a need to identify or classify what the changes are. This is considered a more efficient approach, especially for continuous environmental monitoring. Second, since these textural methods measure the spatial variations among

346

N.S.-N. Lam

pixels instead of comparing pixel by pixel, they are more likely to reflect dominant changes rather than spurious changes that might have resulted from using images taken in different time periods (Lambin, 1996; Smits and Annoni, 2000). If there are only small and insignificant changes in land cover, it is expected that the spatial relationship will not alter and the spatial index values will remain the same. On the contrary, if there are significant land-cover changes, then it is expected that the spatial properties will be altered, and the spatial indices that are designed to measure the spatial properties should be able to capture these changes.

13.2.3 Description of Selected Textural Measures We describe below four textural measures, including spatial autocorrelation (Moran’s I), fractal, lacunarity, and wavelet transform. All four methods have already been implemented in a software module called ICAMS (Image Characterization And Modeling System), developed previously by the author and collaborators (Quattrochi et al., 1997; Lam et al., 1998, 2002; Emerson et al., 1999). ICAMS was developed mainly to provide spatial analytical tools, such as fractals, variograms, and spatial autocorrelation, to visualize, measure, and characterize landscape patterns. Detailed descriptions of ICAMS can be found in a number of publications (e.g., Quattrochi et al., 1997; Lam et al., 1997, 1998). The four methods were selected for illustration because they at least have properties 5 and 6 (can apply locally and to unclassified images), which are important properties for rapid classification and change detection. Furthermore, Moran’s I possesses all properties, fractals have all but property 4. Lacunarity is an old measure but its use in land cover/land use analysis is relatively new, its properties remain to be thoroughly studied (Myint and Lam, 2005a, b). The mathematics of wavelet transformation is well defined. However, the properties of the textural measures computed on the wavelet transformed images have seldom been studied, and will be a subject of our ongoing research.

13.2.3.1 Spatial Autocorrelation Spatial autocorrelation statistic has been commonly used to measure the degree of clustering, randomness, or fragmentation of a spatial pattern. The two most common spatial autocorrelation measures for interval-ratio data are Moran’s I and Geary’s C (Cliff and Ord, 1973). Moran’s I is generally preferred over Geary’s C, because the values of the former are more intuitive (i.e., positive values for positive autocorrelation and vice versa). Moran’s I was also found to be generally more robust (Goodchild, 1986). These two measures were originally developed for measuring polygonal data (e.g., geographical regions) where the number of units or points measured (n) is often smaller. It is only recently that these two measures were applied to raster data such as remotely sensed images, where the number of units being

13 Methodologies for Mapping Land Cover/Land Use and its Change

347

measured (n × n) is much larger (Emerson et al., 1999; Lam et al., 2002). Moran’s I is calculated from the following formula: I(d) =

∑ni ∑nj wi j zi z j w ∑ni z2i

(13.1)

Z-value

−0 90 181

where wi j is the weight at distance d so that wi j = 1 if point j is within distance d from point i; otherwise, wi j = 0; z’s are deviations from the mean for variable y, and w is the sum of all the weights where i = j. Moran’s I varies from +1 for perfect positive correlation (a clumped pattern) to −1 for perfect negative correlation (a checkerboard pattern). Figure 13.1 shows three simulated surfaces and their corresponding Moran’s I and fractal dimension values. Moran’s I and fractal

5

5

4

4

2 2

Z-value

lu Co

−0 72 145 217

0

ns

lum

Co

0

1

1

2

1

ws

3

3

Ro

5

5

4

4 2 1

Z-value

−0 72 145

0

s

mn

0

ws

3

3

Ro

5

5

4

4

1

1 0

ns

lum

2

2

Co

0

ws

3

3

Ro

Fig. 13.1 Three simulated surfaces mapped in three-dimensional and image forms. From top to bottom, D = 2.1, 2.5, 2.9, and I = 1.0, 0.99, 0.82. (Modified from Lam et al., 2002.)

348

N.S.-N. Lam

dimension (D) have an inverse relationship, whereby a spatial pattern with a high degree of fragmentation will have a low Moran’s I but a high fractal dimension.

13.2.3.2 Fractal Dimension Fractals were derived mainly to overcome the difficulty in analyzing spatial forms and processes by classical Euclidean geometry. The key parameter in fractals is fractal dimension D, which is used to represent the complexity of spatial forms and processes. The higher the D, the more complex is the curve or surface. The D value of a curve can be any non-integer value between 1 and 2, and a surface between 2 and 3. For example, coastlines have dimension values typically around 1.2, and topographic surface dimensions around 2.3. Dimension values for satellite image surfaces have been reported to be much higher, and depending on the type of landscapes examined, they can be as high as 2.7–2.9 (Lam, 1990; Jaggi et al., 1993). Fractal dimension is derived from the concept of self-similarity, where a curve or a surface is made up of copies of itself in a reduced scale (Mandelbrot, 1982). The number of copies (m) and the scale reduction factor (r) can be used to determine the dimensionality of the curve or surface, where D = − log(m)/ log(r) (Falconer, 1988). Practically the D value of a curve is estimated by measuring the length of the curve using various step sizes. The more irregular the curve, the greater increase in length as step size decreases. Such an inverse relationship between total line length and step size can be captured by a regression: log(L) = C + B log(G)

(13.2)

where L is the line length, G is the step size, B is the slope of the regression, and C is a constant. D can be calculated by: D = 1 − B. For surfaces, D = 2 − B. Because of its attractive theoretical foundation, literally every major discipline has found applications using the fractal concept in the past two decades, with numerous algorithms developed for computing the fractal dimension. Unfortunately, a major problem in applying fractals is that different fractal measurement algorithms yield different results. Often times, empirically computed fractal dimensions may exceed the theoretical ranges. Moreover, fractal dimension is defined in various ways in different algorithms. For example, some algorithms use only a single measurement to derive the dimension, instead of using multiple step sizes to derive the dimension through regression analysis. FRAGSTATS defines fractal dimension as the ratio between perimeter and area of a patch, which is very different from the algorithms described below (Lam, 1990). The former definition of fractal dimension, though simple and easy to calculate, applies only to images that have already been classified, whereas the algorithms described below (e.g., the triangular prism algorithm) follow closely the original definition by Mandelbrot and can be applied directly to unclassified images for textural comparison. Lam (1990) demonstrated the use of three methods, including isarithm, triangular prism, and variogram methods, in measuring the spatial complexity of the reflectance surfaces from remote sensing imagery (Goodchild, 1980; Clarke, 1986;

13 Methodologies for Mapping Land Cover/Land Use and its Change

349

Mark and Aronson, 1984). In a subsequent benchmark study, Lam et al. (2002) found that the modified triangular prism method was the most accurate and reliable method for estimating the fractal dimension of surfaces. Hence, the modified triangular prism method is described as follows. The modified triangular prism method (Clarke, 1986; Jaggi et al., 1993; Lam et al., 2002) constructs triangles by connecting the heights or z-values at the four corners of a grid cell to its center, with the center height being the average of the pixels at the four corners. These triangular “facets” of the prism are then summed to represent the surface area. In the second step, the algorithm increases the step size from one pixel to two pixels, and the z-values at the four corners of the 2 × 2 composites are used to construct the prism. It is expected that as step size increases, the prism surface area will increase, but at a decreasing rate, which can then be used to determine the fractal dimension by a regression equation similar to Eq. (13.2): Log A = K + (2 − D) Log S, where A is the prism surface area, K is a constant, D is the fractal dimension, and S is the pixel size.

13.2.3.3 Lacunarity Despite the potential of fractals, Mandelbrot (1982) realized that fractal dimensions are very far from providing a complete characterization of spatial forms. He introduced the term lacunarity (lacunar in Latin means gap) to further describe the gappiness or texture of a spatial pattern. In other words, different fractal sets may have the same fractal dimension values, but they may look different because they have different lacunarities (Myint and Lam, 2005a). Lacunarity represents the distribution of gap sizes; low lacunarity implies homogeneity as all gap sizes are the same, whereas high lacunarity implies heterogeneity (Dong, 2000). Unfortunately, lacunarity is highly sensitive to scale, and depending on the size of the gliding box used in computing the lacunarity value, the same pattern can return with very different values, as objects that are homogeneous at a small scale can be heterogeneous at a large scale (Plotnick et al., 1993). Myint and Lam (2005a, b) compared several hypothetical patterns; three of them are shown in Fig. 13.2. When a smaller gliding box of 3 × 3 is used, the small gap pattern (Fig. 13.2) results in low lacunarity (1.05), the big gap pattern yields the highest (1.40), and the random pattern yields a value in between (1.15). But when a bigger gliding box of 11 × 11 is used, the results are reverse, with the big gap pattern yielding the lowest lacunarity (1.02) and the small gap pattern the highest (1.10). Lacunarity value of the random pattern decreased slightly from 1.15 to 1.08. It is also observed that the range of difference between the two patterns is much smaller with bigger gliding box. An algorithm for computing lacunarity using the gray-scale approach is described as follows (Voss, 1986; Myint and Lam, 2005a, b). Let P(m, L) be the probability that there are m intensity points within a cube size of L centered at an arbitrary point in an image. Intensity points are points that fill the cube in each step. Hence, we have

350

N.S.-N. Lam

Fig. 13.2 Three hypothetical binary patterns with different lacunarity values, see text for explanations. (Modified from Myint and Lam, 2005a.) N

∑ P(m, L) = 1

(13.3)

m=1

where N is the number of possible points in the cube of L. Suppose that the total number of points in the image is M. If one overlays the image with cubes of side L, then the number of cubes with m points inside the cube is (M/m)P(m, L). Hence N

M(L) =

∑ mP(m, L)

(13.4)

m=1

and M 2 (L) =

N

∑ m2 P(m, L)

(13.5)

m=1

Lacunarity Λ(L) can be computed from the same probability distribution P(m, L), and is defined as: M 2 (L) − (M(L))2 Λ(L) = (13.6) (M(L))2 Unlike fractals, lacunarity has no theoretical maximum or minimum. The performance of the index, especially its high scale dependency, will need to be further studied. However, a few studies have shown that adding a lacunarity layer in image classification has dramatically improved accuracy (Myint and Lam, 2005a, b), indicating a promising approach towards more accurate, automated land cover/land use mapping. 13.2.3.4 The Wavelet Transform Method Pioneered by Mallat (1989) and Daubechies (1990), the wavelet method has found numerous applications in a wide range of disciplines. The method has also recently been demonstrated as a promising approach to increasing accuracy in image classification and image retrieval using remote sensing imagery (Manjunath and Ma, 1996; Zhu and Yang, 1998; Bian, 2003; Myint et al., 2004). In brief, wavelets are translated and dilated versions of a common mathematical function, called the mother wavelet. In the case of images, the translation refers to

13 Methodologies for Mapping Land Cover/Land Use and its Change

351

Fig. 13.3 Multiresolution wavelet decomposition of a remote sensing image. (a) Original image, (b) level-1 decomposition: upper left is approximate sub-image, clockwise from upper left are horizontal, diagonal, and vertical detailed sub-images, (c) level-2 decomposition image (From Myint et al., 2004, reprint with permission from the American Society for Photogrammetry and Remote Sensing)

the geographic location, and the dilation relates to different scales. By adjusting the translation and dilation parameters, we can study the texture and scale locally. For the 2D discrete wavelet transform, which is used for remote sensing image analysis, the wavelet method will decompose an image into four sub-images: an approximate image (low frequency) and three detailed images (high frequency – horizontal, vertical, and diagonal). The approximate image can further be decomposed into another level, resulting in a multi-resolution wavelet analysis. Figure 13.3 shows an example of multiresolution wavelet decomposition. The coefficients of the four subimages are computed by Eqs. (13.7)–(13.10): A(i, j) = ∑ ∑ h(k − 2i)h(l − 2 j) f (k, l)

(13.7)

H(i, j) = ∑ ∑ h(k − 2i)g(l − 2 j) f (k, l)

(13.8)

V (i, j) = ∑ ∑ g(k − 2i)h(l − 2 j) f (k, l)

(13.9)

D(i, j) = ∑ ∑ g(k − 2i)g(l − 2 j) f (k, l)

(13.10)

k

k

k

k

l

l

l

l

where f is the original image, A is the approximate image, H is the horizontal detailed image, V is the vertical detailed image, and D is the diagonal detailed image. h(k), g(k) are the scaling filter and the wavelet filter, respectively, and k, l are the number of rows and columns (Mallat, 1989; Daubechies, 1990). After decomposition, indices can be computed for each sub-image at each level to represent the texture of the image. In addition to mean and standard deviation, Eqs. (13.11)–(13.14) show other commonly used measures (Myint et al., 2002), including log energy, Shannon index (SHAN), angular second moment (ASM), and entropy:

352

N.S.-N. Lam K

energy = ∑

K

∑ log(P(i, j)2 )

(13.11)

i=1 j=1 K

SHAN = − ∑

K

∑ P(i, j) ∗ log(P(i, j))

(13.12)

i=1 j=1

K

ASM = ∑

K

∑ P(i, j)2

(13.13)

i=1 j=1 K

entropy = − ∑

K

|P(i, j)|2

∑ Q(i, j) ∗ log |Q(i, j)|; Q(i, j) = ,

i=1 j=1

2

∑ |P(i, j)|

(13.14)

i, j

where P(i, j) is the (i, j)th pixel wavelet coefficient value of a decomposed image at a particular level. These computed textural indices are then used to discriminate different types of land cover/land use.

13.2.4 Scale and Uncertainty in Textural Analysis Textural measures must be computed from a group of pixels or objects. Hence, texture is very scale-dependent. The size of the moving window combined with the resolution of the imagery plays a big part in determining what features are highlighted by these techniques. This scale and uncertainty issue has long been a central concern across a number of disciplines that involve geospatial/environmental data, textural analysis is not an exception. Scale variations are well known to constrain the detail with which information can be observed, represented, analyzed, and communicated (Lam et al., 2004). It should be noted that the term “scale” has different meanings, and depending on the field of study, its meaning could be opposite. Lam and Quattrochi (1992) outlined four different meanings of scale (Cao and Lam, 1997). Cartographic scale refers to the degree of reduction in spatial dimensions that occurs when real-world measurements are represented in hard-copy maps or computer screens. Operational scale is an expression of the spatial or temporal dimensions over which a process operates, while observational scale refers to the dimensions within which a particular phenomenon or process is observed. Measurement scale, commonly called resolution, refers to the smallest observable unit, such as pixels in remote sensing imagery. In landscape ecology, observational scale is the extent, whereas measurement scale is referred to as grain. Radiometric scales also exist within digital imagery. Older Landsat Multispectral Scanner (MSS) images have a 6-bit grayscale depth, while IKONOS images have a 16-bit depth. One of the basic goals of scale-related research is to be able to move up and down spatial scales, within disciplines and across disciplines, so that the results concluded at one scale can be inferred to another

13 Methodologies for Mapping Land Cover/Land Use and its Change

353

scale. Extrapolation of results across broad spatial scales remains the most difficult problem in environmental research (O’Neill et al., 1989; Turner et al., 1989; Lam and Quattrochi, 1992; Quattrochi and Goodchild, 1997; Tate and Atkinsons, 2001). Scale affects change detection. The myriad spatial, spectral, radiometric, and temporal scales of remotely sensed imagery pose a real challenge to change detection, as techniques developed for imagery with a pixel resolution of 1 m (IKONOS imagery) may not be applicable to imagery with a pixel resolution of 1 km (AVHRR imagery). Since changes may occur at different scales, globally, regionally, or locally, and they may also occur rapidly or slowly, it is important to examine how change detection methods and indices perform at different spatial scales. Scale-related uncertainty in modeling results has significant impacts on decision making, and basic research on decision making under uncertainty is necessary. Effective scale-related research requires interdisciplinary efforts of social, physical, and computer scientists. Scale and scale-related uncertainty is a difficult problem to tackle. Increasingly, it has been recognized that scale effects exist and can never be eliminated, therefore strategies must be developed to understand and mitigate the scale effects rather than to eliminate them. Two interrelated approaches were suggested to mitigate the scale effects (Lam et al., 2004). The first approach is to develop techniques to detect the scale ranges within which levels of observation are phenomena scale-dependent. Techniques such as geographic variance, variograms, correlograms, fractal analysis, and a number of textural methods have been proposed to detect the range of scales that yield the most information (Emerson et al., 1999). The second approach is to develop a multi-scale assessment module so that the same analysis can be conducted in multiple scales to compare the results and estimate the uncertainty. A thorough benchmark study is very much needed to examine how textural methods perform at different spatial scales and resolutions in land cover classification and change detection.

13.3 Land Cover Classification Using the Textural Approach: Examples 13.3.1 Characterizing Land Cover in the Tropics Read and Lam (2002) compared the performances of selected textural measures to characterize different land covers using two sets of unclassified Landsat-TM data (1986 and 1996/97) for a site in north-eastern Costa Rica. The purpose was to determine whether landscape complexity can be captured by these methods, and whether these methods can be used to reflect the degree of human disturbance. The hypothesis was that complexity in a natural landscape decreases with increasing intensity of human activities. The methods evaluated were: (1) fractal dimension using the isarithm method, (2) fractal dimension using the modified triangular prism method, (3) spatial autocorrelation using Moran’s I, (4) Shannon’s diversity index, (5) contagion, and (6) fractal dimension from perimeter/area. The first three methods

354

N.S.-N. Lam

were available from ICAMS, and the last three were landscape metrics available from FRAGSTATS. The results revealed that fractal dimension using the triangular prism method and Moran’s I could serve as indices for characterizing spatial complexity of LandsatTM data, whereas the landscape indices were not consistent. The fractal dimension decreased along a gradient of increasing human disturbance: forest–scrub–pasture– agriculture. This study is among the first to examine how spatial indices can be used to examine hypotheses related to land cover/land use and human disturbance in the tropics.

13.3.2 Improving Urban Land Cover Classification Emerson et al. (2005) examined the utility of local variance, fractal dimension, and Moran’s I in improving urban land cover classification. Landsat ETM+ imagery of Atlanta, Georgia obtained in 1999 was examined. Using the routines in ICAMS, texture images were computed from the 15m panchromatic band using a 21 × 21 pixel moving window for every other row and column. This two-pixel increment between rows and columns produced a 30m resolution texture image. The real number local variance, Moran’s I, and fractal dimension values (computed using the modified triangular prism method) were converted to 8-bit image, and they were added to the stack of multispectral bands for classification using a supervised maximum likelihood classification technique. Five land cover classes based on the USGS Anderson Level 1 classification scheme were used, including low intensity urban, high intensity urban, pasture/grassland, forest, and water. Results show that classification accuracy improved with additional texture layer, with the fractal dimension band being the most effective. By adding the fractal dimension band to the multispectral bands, the overall percent correctly classified increased from 67.1% to 77.3%. Although not as effective as the fractal band, addition of local variance and Moran’s I still yielded an improved accuracy of 72.4% and 69.4%, respectively. The results show great promise, but further research is needed to better utilize these indices.

13.3.3 Urban Feature Discrimination Using Wavelets Based on a high-resolution ATLAS (Advanced Thermal Land Application Sensor) image of Baton Rouge, Louisiana, USA, Myint et al. (2002, 2004) introduced the wavelet approach for urban land cover classification. The ATLAS image had a 2.5 m pixel resolution and was acquired with 15 channels (0.45–12.2 µm) from a NASA LearJet on May 7, 1999. Six urban land cover/land use classes with different texture appearances were selected, including agriculture, commercial, woodland, water body, single-family homes with less than 50% tree canopy (residential-1), and single-family homes with more than 50% tree canopy (residential-2). These land

13 Methodologies for Mapping Land Cover/Land Use and its Change

355

cover classes were based on classification by Lo et al. (1997), which was designed for the purpose of urban planning, as information on surface vegetation and water availability are crucial for city officials and environmental agencies in developing better urban infrastructure. Based on previous studies, band 2 (0.52–0.60 µm), band 6 (0.76–0.90 µm), and band 12 (9.60–10.20 µm) were selected. In addition to these three bands, principal component analysis band 1 (PCA1) was also examined to see if a composite band could produce better accuracy. Two segmented regions of each class were identified, and five training pixels were then randomly selected from each region, leading to a total of 10 samples for each class. Windows of 65 × 65, 33 × 33, and 17 × 17 pixels were selected using these 10 pixels as centers. Textural measures of these samples were computed and a linear discriminate analysis was applied to evaluate which measure is the most effective in discriminating the different land covers. Four different textural approaches were evaluated, including the wavelet transform, spatial autocorrelation, spatial cooccurrence matrix, and fractals. It was found from both studies that the wavelet approach was the most accurate among all approaches considered (Myint et al. 2002, 2004). When 65 × 65 samples were used, the wavelet approach yielded 100% accuracy. The overall accuracy, however, decreased with smaller window sizes, with an accuracy of 93% and 78%, respectively, for 33 × 33 and 17 × 17 samples. These studies demonstrated the great potential of using the texture approach. They also highlighted the importance of different scale parameters such as window size in affecting its performance. Future studies that systematically examine the effects of scale on the certainty, or rather uncertainty, of the results are needed.

13.3.4 Lacunarity-Based Urban Classification Myint and Lam (2005a, b) introduced the use of lacunarity in urban land cover/land use classification. As mentioned in Section 13.2.3, lacunarity measures the gappiness of a pattern; low lacunarity implies homogeneity, whereas high lacunarity suggests heterogeneity. An IKONOS image at 4 m spatial resolution with four channels acquired over Norman, Oklahoma, on March 20, 2000 was used. The selected landuse and land-cover classes in this study included single-family houses with less than 50% tree canopy (residential-1), single-family houses with more than 50% tree canopy (residential-2), commercial, woodland, agriculture, golf course, and water body. Lacunarity measures for band 4 (0.76–0.90 µm), band 3 (0.63–0.69 µm), and band 2 (0.52–0.60 µm) were computed for the entire image using the gray-scale method with a gliding cube of size 3 and a moving widow size of 29 × 29. The lacunarity-transformed bands were then stacked as additional bands for maximum likelihood classification. The results show that adding three lacunarity transformed bands to the original three spectral bands (band 4, 3, 2) increased the overall classification accuracy from 55% to 92%, a drastic improvement. If only the three lacunarity transformed bands were used for classification, the overall accuracy still increased, but not substantially, to 68%. This study further confirmed the study by Emerson et al. (2005)

356

N.S.-N. Lam

discussed above and demonstrated clearly that an integrated textural and spectral approach is needed for more accurate land cover/land use classification and mapping.

13.4 Land Cover/Land Use Change Analysis Obviously, extending the mapping and modeling of land cover/land use at one time period to multiple time periods to analyze change adds lot more complexity and challenges. We outline in this section the inherent difficulties of change detection, summarize the existing methods into a framework, and then argue that the textural approach has potential for rapid change detection.

13.4.1 Change Detection Issues There are inherent difficulties involved in using time-series remote sensing data for land cover/land use change detection. Ideally, same type of images that have the same spectral, radiometric, spatial, and temporal resolutions should be used. However, this may not be possible especially for change studies that involve a longer time span. For example, Landsat-MSS with a pixel resolution of 80 m started in 1972, whereas Landsat-TM with a 30 m pixel resolution became available in 1982. Using digital imagery for change analysis before 1972 would be difficult. Often times, old aerial photographs before these dates are the only image sources to be used. The imminent danger of discontinuing global coverage due to sensor malfunction or budget constraints, such as Landsat-7 ETM+, will definitely hamper land cover/land use research and make long-term change analysis impossible. When two very different types of images are involved, the only viable approach to change detection is to conduct detailed image classification of individual images and then overlay the two classified images to assess the changes. However, even with the same sensor, images taken in different dates may be affected by several factors, causing them to be different even if there is no real land cover change. Therefore, the following factors need to be considered for more accurate change detection: 1. The dates of the two images should be approximately the same to avoid seasonal difference in vegetation, soil moisture, sun-angle, and other response that are not real land cover change. 2. Even with the same date but in different years, individual atmospheric condition can obscure our ability to uncover real changes, such as cloud cover or rainfall. An image taken right after rainfall in a desert will look very different from an image of the same desert before rainfall. Atmospheric correction of each image may need to be applied before the analysis. For coastal landscape, tidal stage is a crucial factor in conducting change detection. Significant changes can be observed simply because the images were taken at different tidal stages (Jensen, 2005).

13 Methodologies for Mapping Land Cover/Land Use and its Change

357

3. Pre-image processing steps may also contribute errors. Extra caution is needed to ensure no pixel mis-registration between the two images. A single pixel shift will shift the entire image and that could lead to substantial error in assessing change. Another point that has seldom been mentioned in the literature refers to the algorithm used to convert pixel values from analog to digital scale. Assuming an 8-bit scale (0–255) is used, some algorithms will convert the continuous signal using the image’s minimum and maximum values as the limit, whereas others use the 99% or 95% interval. The result is that the same digital number in different images may have very different actual radiometric value, and the value is only true relative to the rest of the pixel values in its own image. Hence, change detection methods that involve direct pixel-by-pixel spectral comparison could be misleading, whereas change detection methods that are based on ratios among bands within its own image are more reliable. By the same token, it is expected that comparing the textural difference between two images, instead of pixel-by-pixel spectral comparison, would yield more accurate change analysis.

13.4.2 A Classification of Change Detection Methods New methods for change detection using remote sensing imagery have been continuously reported in the literature, and it is not the scope of this chapter to exhaust the list and provide evaluation of each approach. However, the following framework may be useful to classify the various change detection methods. Figure 13.4 shows how the various commonly used methods are placed in this framework.

Fig. 13.4 A framework for classifying change detection methods

358

N.S.-N. Lam

Change detection methods can be differentiated into two main groups, depending on whether the method requires classification before or after the changes are detected. As shown in Fig. 13.4, the first group of change detection methods, which is also the most traditional approach to change detection, will first classify individual images of two dates using a statistical maximum likelihood classifier and then compare the classified images to provide an assessment of change. This traditional approach generally requires extensive human supervision for classifying the images. However, new image classification methods, other than the traditional maximum likelihood classifier, can be applied to increase accuracy and efficiency. These methods include, for example, fuzzy classification, artificial intelligence based classifier, Bayesian approach, and even the textural approach (Moller-Jensen, 1990; Gopal and Woodcock, 1996; Jensen, 2005; Gong, 2006). Recently, object-based image segmentation and classification has gained increasing attention, with new software such as eCognition (Definiens, 2004; Benz et al., 2004) and Feature Analyst (Visual Learning Systems, Inc.) made available to general users. These methods use both the spectral (or color) information and various spatial metrics to define homogenous areas (called objects). Despite its potential, this group of change detection methods is not germane to rapid change detection, as extensive human supervision is needed to pre-classify the images. The second group of change detection methods does not require images to be pre-classified. Image differencing, change vector method, and multidate comparison methods (Fig. 13.4 – box d) can be applied directly to original pixel values or indirectly to modified values from the spectral bands (e.g., band ratios, principal components, chi-squared transformed, and texture transformed) (Fig. 13.4 – boxes b and c). The main advantage of this group of methods is that pre-classification is not necessary until significant changes are detected, hence avoiding the tedious classification process at the beginning. The problem remains to be that of determining the threshold value at which the difference between the two images is considered significant. Continuous monitoring of land cover/land use and rapid identification of their changes is crucial to providing timely decision support and risk assessment especially during extreme events (e.g., hurricanes, earthquakes, forest fires, terrorist attacks, disease spread). There is a need to develop efficient and reliable change detection methods that can be automated, easy to use, and applicable to different land covers observed by different sensors at different scales, times, and places. Although it is difficult to achieve fully automated change detection, we expect that an integrated approach that incorporates both textural and spectral indices could alleviate some of the existing change detection problems for two reasons (which have also been elaborated in Section 13.2.2). First, the texture measures that have property 5 and 6 (e.g., fractals, lacunarity, wavelets, and spatial autocorrelation statistics) can be applied directly to pre-classified images without the need to go through the image classification process, thereby reducing the need for extensive human supervision upfront. Second, since the spatial/texture methods measure the spatial variations across the image instead of comparing brightness values on a pixel by pixel basis, they are more likely to reflect dominant changes rather than spurious changes that

13 Methodologies for Mapping Land Cover/Land Use and its Change

359

occur due to noise, clouds, or illumination differences. However, whether a combination of these methods can be successfully applied to reliably characterize land covers and identify changes remains to be studied and is part of our ongoing research.

13.5 Change Detection Using the Textural Approach: An Example In this section, we demonstrate the use of textural measures only (not including spectral information) for change detection in New Orleans before and after Hurricane Katrina. The utility of fractal dimension and spatial autocorrelation statistics in this capacity is compared. Hurricane Katrina hit New Orleans on August 29, 2005. Two Landsat-TM images, dated November 7, 2004 and September 7, 2005, were obtained from the US Geological Survey/National Wetlands Research Center at Louisiana State University and the LSU FEMA GIS Store project (http://www.cadgis.lsu.edu). Although not exactly the same anniversary dates, these two images are the best available so far for change detection. Both images have already been registered and geometrically rectified with a pixel resolution of 28.5 m prior to this study. For this study, we created a subset of 512 × 512 pixels from both images. Figure 13.5a and b display the subsets using band 4 (near-infrared band), which is the most effective band in discriminating water and non-water features. The subsets were mainly confined to the Orleans Parish, with a small part of the Lower Ninth Ward in St. Bernard Parish shown at the east edge of the image (east of the Industrial Canal breach). The post-Katrina image (Fig. 13.5b) clearly shows that most of New Orleans was flooded 9 days after Katrina, except in the natural levee area along the Mississippi River. The northwest corner of the image is Lake Pontchartrain, where significant storm surge has destroyed properties around the Lake. The three levee breaches at Industrial Canal, the 17th Street Canal, and the London Avenue Canal are also marked in Fig. 13.5b. Both Landsat-TM images used in this example have previously been normalized to minimize sensor calibration offsets and differences in atmospheric effects. However, other factors may change the pixel values even though there are no real changes on the ground. A random check on the pixels at the northwest corner (Lake Pontchartrain) shows that a fair amount of difference in pixel values occurred in the same location between the two time periods (e.g., 32 vs 72 in pre- and postKatrina image, respectively), even though no significant real change is expected at this location. This shows that change detection using the original spectral value pixel-by-pixel comparison approach alone could be problematic, especially regarding the determination of the threshold value to identify real changes. Using ICAMS, we computed the local fractal dimension using the modified triangular prism algorithm and the local Moran’s I for each image. For the local fractal dimension method, the following parameters were used: moving window size of 17 × 17, step size of 5, stretch option, and arithmetic progression. For Morans’ I,

360

N.S.-N. Lam

Fig. 13.5 (a and b) Display of pre-Katrina (November 7, 2004) and post-Katrina (September 7, 2005) Landsat-TM images using band 4. (c) and (d) are Moran’s I transformed pre- and postimages; (e) and (f) are fractal-transformed pre- and post-images

the only parameter needed to be input was the moving window size, which was also set to 17 × 17. The 17 × 17 window was chosen because a previous study on the impacts of Hurricane Hugo along the South Carolina’s coast by Kulkarni (2004) found that this window size was the best in representing land cover features. Figure 13.5c–f

13 Methodologies for Mapping Land Cover/Land Use and its Change

361

Table 13.2 Summary statistics of band 4 for pre- and post-Katrina Landsat – TM images

Min Max Mean SD CV

Pre

Original Post

24.00 255.00 161.50 52.28 0.32

1.00 255.00 45.02 25.51 0.57

Diff. −250.00 96.00 −116.49 44.95 −0.38

Fractal-transformed Pre Post Diff. 1.86 4.11 2.77 0.22 0.08

1.75 4.13 2.74 0.21 0.08

−1.23 1.15 −0.03 0.15 5.00

Moran’s I-transformed Pre Post Diff. −0.07 0.97 0.61 0.17 0.28

−0.05 −0.63 0.97 0.63 0.66 0.05 0.14 0.11 0.21 2.20

SD – standard deviation; CV – coefficient of variation (= SD/mean); difference image = (post – pre)

show the Moran’s I-transformed and the fractal-transformed images. Brighter pixels refer to higher values in fractal dimension or Moran’s I. It should be stressed that since fractal dimension and Moran’s I have an inverse relationship, features with low fractal dimension, such as the Mississippi River (darker pixels in the fractaltransformed images), will be shown as brighter pixels in the Moran’s I transformed images. The difference images were computed by subtracting the pre-Katrina image from the post-Katrina image, and the summary statistics of all images are listed in Table 13.2. In general, the post-image had lower spectral values than the pre-image, and the mean difference between the two images (band 4) was as high as −116.49. This is expected as most of the study area was flooded after Katrina, resulting in lower spectral reflectance value in the near-infrared band. The fractal-transformed summary statistics show that the mean spatial complexity, as represented by fractal dimension, slightly decreased from 2.77 to 2.74. Conversely, the mean Moran’s I increased from 0.61 to 0.66, which also indicates that the overall spatial complexity decreased slightly for the post-image. The fractal and Moran difference images were first mapped in a continuous mode (with a two-standard deviation stretch) using ICAMS. The fractal difference image (Fig. 13.6a) shows that increases in fractal dimension (positive changes), as represented by brighter pixels, occurred in areas close to the Industrial Canal (east side of the image) and the areas between the 17th Street Canal and London Avenue Canal (middle part of the image). Areas with decrease in fractal dimension (negative changes) are represented by darker pixels and they scattered over the image. With a 17 × 17 window size, the general features of the study area, such as the Mississippi River, can still be recognized. The Moran’s I difference image (Fig. 13.6b) shows the same pattern; the darkest pixels represented the highest decrease in spatial autocorrelation, implying an increase in spatial complexity. It can be observed from the two difference images that the location of the darkest pixels in the Moran’s I difference image generally coincided with the brightest pixels in the fractal different image, and vice versa. The difference images can also be mapped using standard deviation unit as class intervals in ICAMS. One of the display options is to map the changes in three class intervals using two-standard deviations as class boundaries. The first interval, which contains pixels that have highest positive changes in spatial complexity

362

N.S.-N. Lam

Fig. 13.6 Display of the fractal-difference and Moran-difference images in a continuous mode (a) and (b) and three-class mode (c) and (d). In (c) and (d), the brightest pixels indicate the highest positive changes in spatial complexity (>2 standard deviations), the darkest pixels indicate highest negative changes (<−2 standard deviations), and the gray pixels are values in between. Both brightest and darkest pixels should be of interest, which may point to areas that are most “affected”

(>2 SD), is shaded as the brightest. The second class, which is in middle gray, is for pixels that have difference values falling between ±2 standard deviations. The third class, which has the darkest shade, is for pixels that have the greatest negative changes in spatial complexity (<−2SD). Using this mapping method, both the brightest and the darkest pixels in the images (Fig. 13c and d) signal the greatest changes and hence attention is most warranted for these pixels and their surrounding pixels. This method should guide resources to the most “affected” areas, and in this case, greatest change in spatial complexity in both positive and negative directions. Figure 13.6d, the fractal difference image, shows only a few concentrated spots belonging to the first and third classes (the brightest and darkest pixels), and they were generally located close to the three canal breach areas. The rest of the secondclass pixels were scattered throughout the image. For the Moran difference image (Fig. 13.6c), because of its inverse relationship with fractal dimension (i.e., the higher the fractal dimension, the lower the Moran’s I value), one should expect that

13 Methodologies for Mapping Land Cover/Land Use and its Change

363

the brightest spot in the fractal difference image would coincide with the darkest spot in the Moran’s I difference image. A visual comparison between the two images (Fig. 13.6c and d) shows that this is generally true, with the Moran’s I difference image portraying a wider area of brightest and darkest spots than the fractal difference image. Based on the Moran’s I difference image (Fig. 13.6c), the greatest decrease in Moran’s I values (greatest increase in spatial complexity – darkest pixels) were also found near the three canal breach areas. Areas that showed greatest increase in Moran’s I values (greatest decrease in spatial complexity – brightest pixels) were scattered in the mid city and the area surrounding Lake Pontchartrain. In summary, this example shows that the textural approach alone could be useful in pinpointing the areas that need the most attention. With additional information layers, these maps could serve as a useful guide to focus our efforts in detecting largest and meaningful changes in a rapid and reliable manner. It is expected that combining spectral and spatial layers, as well as combining different textural measures, will increase the accuracy of this approach. Other mapping methods could also be employed to further enhance the visualization of these changes.

13.6 Conclusions Efficient methods for rapid monitoring of land cover/land use and their changes through remote sensing imagery are urgently needed to provide timely decision support and risk assessment especially during extreme events (e.g., terrorist attacks, hurricanes, forest fires, earth quakes, disease spread). Although there is a huge literature on land cover classification and change detection, we are still far from being able to automate these tasks via remote sensing and GIS. The high variability of ground conditions as manifested in individual as well as time-series imagery makes it very difficult to generalize and automate. The search for useful approaches and methods for rapid land cover identification and change detection remains a very challenging task. This chapter introduced the use of textural and spatial metrics as a promising approach to automated land cover/land use classification and change detection. We identified in this chapter the major criteria for selecting textural measures and then illustrated through several examples from previous studies how textural metrics, in combination with the original spectral bands, have greatly improved the classification accuracy. For change detection analysis, we developed a framework for classifying the numerous change detection approaches. Then, using a recent example of evaluating the impacts of Hurricane Katrina on New Orleans land cover change, we illustrated the use of local fractal dimension and local Moran’s I to detect the largest changes that might need further attention. More research is needed to determine the effectiveness of the various textural metrics with different types of remote sensing imagery, different scales and resolutions, different land cover features, and different environments. These issues are currently being examined in our ongoing research.

364

N.S.-N. Lam

Acknowledgements John Barras, Geographer at the USGS/CR/BRD National Wetlands Research Center, Coastal Restoration Field Station, Louisiana State University, provided the Landsat-TM 2004 image data. Mary Lee Eggart assisted in preparing the figures. This research is supported in part by a NASA Intelligent Systems research grant (NCC-2-1246) and in part by a US National Science Foundation grant (BCS-0554937).

References Baskent EZ, Jordan GA (1955) Characterizing spatial structure of forest landscapes. Can. J. Forest Res. 25:1830–1849 Benz UC, Hofmann P, Willhauck G, Lingenfelder I, Heynen M (2004) Multi-resolution, objectoriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS J. Photogramm. Remote Sens. 58:239–258 Bian L (2003) Retrieving urban objects using a wavelet transform approach. Photogramm. Eng. Remote Sens. 69(2):133–141 Briggs JM, Nellis JM (1991) Seasonal variation of heterogeneity in the tallgrass prairie: A quantitative measure using remote sensing. Photogramm. Eng. Remote Sens. 57:407–411 Cao C, Lam NSN (1997) Understanding the scale and resolution effects in remote sensing and GIS. In: DA Quattrochi, MF Goodchild (eds), Scale in remote sensing and GIS. Lewis Publishers, Boca Raton, FL, pp 57–72 Carr JR (1999) Classification of digital image texture using variograms. In: PM Atkinson, NJ Tate (eds), Advances in remote sensing and GIS analysis. Wiley, London, pp 135–146 Carr JR, de Miranda FP (1998) The semivariogram in comparison to the co-occurrence matrix for classification of image texture. IEEE Trans. Geosci. Remote Sens. 36(6):1945–1952 Clarke KC (1986) Computation of the fractal dimension of topographic surfaces using the triangular prism surface area method. Comput. Geosci. 12(5):713–722 Clausi DA, Jobanputra R (2006) Preserving boundaries for image texture segmentation using grey level co-occurring probabilities. Pattern Recog. 39(2):234–245 Cliff AD, Ord JK (1973) Spatial autocorrelation. Methuen, New York Coppin PR, Bauer ME (1966) Digital change detection in forest ecosystems with remote sensing imagery. Remote Sensing Reviews 13:207–234 Crews-Meyer KA (2002) Characterizing landscape dynamism using paneled-pattern metrics. Photogramm. Eng. Remote Sens. 68(10):1031–1040 Dale MRT (2000). Lacunarity analysis of spatial pattern: a comparison. Landscape Ecol. 15(5):467–478 Daubechies I (1990) The wavelet transform, time/frequency localization and signal analysis. IEEE Trans. Inf. Theory 36:961–1005 De Pietri DE (1995) The spatial configuration of vegetation as an indicator of landscape degradation due to livestock enterprises in Argentina. J. Appl. Ecol. 32:857–865 Definiens AG (2004) eCognition User Guide (accessed May 2006) Dong P (2000) Test of a new lacunarity estimation method for image texture analysis. Int. J. Remote Sens. 21(17):3369–3373 Dunn CP, Sharpe DM, Guntenspergen GR, Stearns F, Yang Z (1991) Methods of analyzing temporal changes in landscape pattern. In: MG Turner, RH Gardner (eds), Quantitative methods in landscape ecology. The analysis and interpretation of landscape heterogeneity. Springer, New York Emerson CW, Lam NSN, Quattrochi DA (1999) Multiscale fractal analysis of image texture and pattern. Photogramm. Eng. Remote Sens. 65(1):51–61 Emerson CW, Lam NSN, Quattrochi DA (2005) A comparison of local variance, fractal dimension, and Moran’s I as aids to multispectral image classification. Int. J. Remote Sens. 26(8):1575– 1588

13 Methodologies for Mapping Land Cover/Land Use and its Change

365

Estreguil C, Lambin E (1996) Mapping forest disturbances in Papua New Guinea with AVHRR data. J. Biogeogr. 23:757–773 Falconer K (1988) Fractal geometry: mathematical foundations and applications. Wiley, New York Frank TD (1984) The effect of change in vegetation cover and erosion patterns on albedo and texture of Landsat images in a semiarid environment. Ann. Assoc. Am. Geogr. 74:393–407 Franklin SE, Hall RJ, Moskal LM, Maudie AJ, Lavigne, MB (2000) Incorporating texture into classification of forest species composition form airborne multispectral images. Int. J. Remote Sens. 21(1):61–79 Gong P (2006) Information extraction. In: M Ridd, JD Hipple (eds), Remote sensing of human settlements. ASPRS, Bethesda, MD, pp 275–334 Gong P, Marceau DJ, Howarth PJ (1992) A comparison of spatial feature extraction algorithms for land use classification with SPOT HRV data. Remote Sens. Environ. 40:137–151 Goodchild MF (1980) Fractals and the accuracy of geographical measures. Mathematical Geology 12:85–98 Goodchild MF (1986) Spatial Autocorrelation. CATMOG (Concepts and Techniques in Modern Geography) No. 47. Geo Books, Norwich, England Gopal S, Woodcock C (1996) Remote sensing of forest change using artificial neural networks. IEEE Trans. Geosci. Remote Sens. 34(2):398–404 Haralick RM (1979) Statistical and structural approaches to texture. Proc. IEEE 67(5):786–804 Haralick RM, Shanmugan K, Dinstein J (1973) Textural features for image classification. IEEE Trans. Syst. Man Cybern. 3(6):610–621 Jaggi S, Quattrochi D, Lam NSN (1993). Implementation and operation of three fractal measurement algorithms for analysis of remote sensing data. Comput. Geosci. 19(6):745–767 Jensen J, Cowen D, Althausen J, Narumalani S, Weatherbee O (1993) An evaluation of the CoastWatch change detection protocol in South Carolina. Photogrammetric Engineering and Remote Sensing 59(6):1039–1046 Jensen J, Cowen D, Narumalani S, Halls J (1997) Principles of change detection using digital remote sensor data. In: JL Star, JE Estes, KC McGwire (eds), Integration of geographic information systems and remote sensing. Cambridge University Press, Cambridge, pp 37–54 Jensen J (2005) Introductory digital image processing: a remote sensing perspective, 3rd edn. Prentice-Hall, New Jersey Jupp DLB, Walker J, Pendridge LK (1986) Interpretation of vegetation structure in Landsat MSS imagery: a case study in disturbed semi-arid eucalypt woodland. Part 2. model-based analysis. J. Environ. Manag. 23:35–57 Kulkarni A (2004) Evaluation of the Impacts of Hurricane Hugo on the Land Cover of Francis Marion National Forest, South Carolina Using Remote Sensing. M.S. thesis, Louisiana State University, Baton Rouge, Louisiana Lam NSN (1990) Description and measurement of Landsat TM images using fractals. Photogramm. Eng. Remote Sens. 56(2):187–195 Lam NSN (2004) Fractals and scale in environmental assessment and monitoring. In: E Sheppard, R McMaster R (eds), Scale and Geographic Inquiry: Nature, Society, and Method. Blackwell, Oxford, pp 23–40 Lam NSN, Quattrochi DA (1992) On the issues of scale, resolution, and fractal analysis in the mapping sciences. Prof. Geogr. 44(1):89–99 Lam NSN, De Cola L (eds) (1993) Fractals in geography. Prentice-Hall, Englewood Cliffs, NJ, 308p Lam NSN, Quattrochi DA, Qiu HL, Zhao W (1998) Environmental assessment and monitoring with image characterization and modeling system using multiscale remote sensing data. Appl. Geogr. Stud. 2(2):77–93 Lam NSN, Qiu HL, Quattrochi DA, Emerson CW (2002) An evaluation of fractal methods for measuring image complexity. Cartogr. Geogr. Inform. Sci. 29:25–35 Lam NSN, Catts C, Quattrochi DA, Brown D, McMaster R (2004) Scale. In: R McMaster, L Usery (eds), A Research Agenda for Geographic Information Science. CRC Press, Bacon Raton, FL, Chapter 4, pp 93–128

366

N.S.-N. Lam

Lambin EF (1996) Change detection at multiple temporal scales: seasonal and annual variations in landscape variables. Photogramm. Eng. Remote Sens. 62:931–938 Lambin EF, Strahler AH (1994) Indicators of land-cover change for change-vector analysis in multitemporal space at coarse spatial scales. Int. J. Remote Sens. 15:2099–2119 Lo CP, Quattrochi DA, Luvall JC (1997) Application of high-resolution thermal infrared remote sensing and GIS to assess the urban heat island effect. International Journal of Remote Sensing 18(2):287–304 Lu D, Mausel P, Brondizio E, Moran E (2005) Land-cover binary change detection methods for use in the moist tropical region of the Amazon: a comparative study. Int. J. Remote Sens. 26(1):101–114 Lunetta R, Elvidge C (1998) Remote sensing change detection: environmental monitoring methods and applications. Sleeping Bear Press, Ann Arbor, MI Mallat SG (1989) A theory for multi-resolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Anal. Machine Intell. 11:674–693 Mandelbrot B (1982) The fractal geometry of nature. Freeman, New York Manjunath BS, Ma WY (1996) Texture features for browsing and retrieval of image data. IEEE Trans. Pattern Anal. Machine Intell. 18(8):837–842 Mark DM, Aronson PB (1984) Scale-dependent fractal dimensions of topographic surfaces: an empirical investigation, with applications in geomorphology and computer mapping. Math. Geol. 11:671–684 Mas JF (1999) Monitoring land-cover changes: a comparison of change detection techniques. Int. J. Remote Sens. 20:139–152 McGarigal K (2002) Landscape pattern metrics. In AH El-Shaarawi, WW Piegorsch (eds), Encyclopedia of environmentrics, vol 2. Wiley, Sussex, England, pp 1135–1142 McGarigal K, Mark BJ (1995) FRAGSTATS: spatial pattern analysis program for quantifying landscape structure. USDA Forest Service General Technical Report PNW-351, Portland, Oregon Mallat, S (1989) A theory of multi-resolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Anal. Machine Intel. 11:674–693 Moller-Jensen L (1990) Knowledge-based classification of an urban area using texture and context information in Landsat-TM imagery. Photogramm. Eng. Remote Sens. 56(6):899–904 Muneeswaran K, Ganesan L, Arumugam S, Sounda KR (2005) Texture classification with combined rotation and scale invariant wavelet features. Pattern Recogn. 38(10):1495–1506 Myint S, Lam NSN (2005a) A study of lacunarity based texture analysis approaches to improve urban image classification. Comput. Environ. Urban Syst. 29:501–523 Myint S, Lam NSN (2005b) Examining lacunarity approaches in comparison with fractal and spatial autocorrelation techniques for urban mapping. Photogramm. Eng. Remote Sens. 71(8):927– 937 Myint S, Lam NSN, Tyler J (2002) An evaluation of four different wavelet decomposition procedures for spatial feature discrimination in urban areas. Trans. GIS 6(4):403–429 Myint S, Lam NSN, Tyler J (2004) Wavelets for urban spatial feature discrimination: Comparisons with fractals, spatial autocorrelation, and spatial co-occurrence approaches. Photogramm. Eng. Remote Sens. 70(8):803–812 Nackaerts K, Vaesen K, Muys B, Coppin P (2005) Comparative performance of a modified change vector analysis in forest change detection. Int. J. Remote Sens. 26(5):839–852 O’Neill RV, Johnson AR, King AW (1989). A hierarchical framework for the analysis of scale. Landscape Ecol. 7(1):55–61 O’Neill RV, Krummel JR, Gardner RH, Sugihara G, Jackson B, DeAngelis DL, Milne BT, Turner MG, et al. (1998) Indices of landscape pattern. Landscape Ecol. 1:153–162 Openshaw S (1989) Automating the search for cancer clusters. Prof. Statistician 8:7–8 Pickup G, Foran BD (1987) The use of spectral and spatial variability to monitor cover change on inert landscapes. Remote Sens. Environ. 23:351–363 Plotnick RE, Gardner RH, O’Neill RV (1993) Lacunarity indices as measures of landscape texture. Landscape Ecol. 8:201–211

13 Methodologies for Mapping Land Cover/Land Use and its Change

367

Quattrochi DA, Goodchild MF (1997) Scale in remote sensing and GIS. CRC/Lewis Publishers, Boca Raton, FL Quattrochi DA, Lam NSN, Qiu HL, Zhao W (1997) Image characterization and modeling system (ICAMS): a geographic information system for the characterization and modeling of multiscale remote sensing data. In: D Quattrochi, M Goodchild (eds), Scaling in remote sensing and GIS. CRC/Lewis Publishers, Boca Raton, FL, pp 295–307 Read JM, Lam NSN (2002) Spatial methods for characterizing land-cover changes for the tropics. Int. J. Remote Sens. 23(12):2457–2474 Smits PC, Annoni A (2000) Towards specification-driven change detection. IEEE Trans. Geosci. Remote Sens. 38:1484–1488 Tate N, Atkinsons P (eds) (2001) Modelling scale in geographical information science. Wiley, New York Turner MG, Dale VH, Gardner RH (1989) Predicting across scales: theory development and testing. Landscape Ecol. 3(3/4):245–252 Voss R (1986) Random fractals: characterization and measurement. In: R Pynn, A Skjeltorp (eds), Scaling phenomena in disordered systems. Plenum, New York Woodcock CE, Strahler (1987) The factor of scale in remote sensing. Remote Sens. Environ. 21:311–332 Yuan D, Elvidge C, Lunetta R (1998) Survey of multispectral methods for land cover change analysis. In: R Lunetta, D Elvidge (eds), Remote sensing change detection. Sleeping Bear Press, Ann Arbor, MI Zhu C, Yang X (1998) Study of remote sensing image texture analysis and classification using wavelet. Int. J. Remote Sens. 19(16):3197–3203

Chapter 14

Methodologies for Mapping Plant Functional Types Wanxiao Sun and Shunlin Liang

Abstract Plant functional type (PFT) is a crucial variable needed in studies of global climate, carbon cycle and ecosystem change. Using remote sensing techniques to extract PFTs is a relatively recent field of research. To date, only a very few methods for mapping PFTs have been reported. This chapter provides an overview of recent developments in this evolving field and discusses future research needs. A brief survey of existing methods for mapping PFTs is presented, followed by a discussion of several methodological issues pertaining to the development of robust remote sensing techniques for mapping of PFTs at regional to global scales. The chapter also outlines a multisource data fusion framework for improved mapping of PFTs from satellite observations.

14.1 Introduction Vegetation plays a vital role in the exchange of energy, carbon, water, and momentum between the land surface and the atmosphere (Sellers and Schimel, 1993; Bonan, 1996; Sellers et al., 1997; Lawton et al., 2001; Marland et al., 2003; Nair et al., 2003; Gamon et al., 2004; Feddema et al., 2005). Reliable information about the geographic distribution and extent of major types of vegetation around the globe is required to initiate and validate various land surface models that provide the boundary conditions for the simulation of climate, carbon cycle and ecosystem change. Traditionally, land surface models represent vegetation as discrete biomes such as evergreen broadleaf forest, shrub, grass, and savanna. These biomes then set surface biogeophysical variables such as albedo, LAI, f PAR, canopy roughness and stomatal physiology for each grid cell (Sellers et al., 1986; Running and Coughlan, 1988; Bonan, 1993; Prince and Goward, 1995). Wanxiao Sun Department of Geography and Planning, Grand Valley State University, USA Shunlin Liang Department of Geography, University of Maryland, USA S. Liang (ed.), Advances in Land Remote Sensing, 369–393. c Springer Science + Business Media B.V., 2008

369

370

W. Sun, S. Liang

A major problem with the biome-based land classification approach is that biomes are not natural vegetation units but are products of classification. Many of the vegetation parameters needed by land models are leaf-level and whole-plant parameters that are difficult to parameterize in the case of mixed life-form biomes such as mixed forests and savannas (Bonan et al., 2002). For example, how does one obtain the necessary leaf physiological and whole-plant allocation parameters for a savanna, which consists of physiologically distinct grasses and trees? Land models are expanding beyond their traditional biogeophysical roots to include biogeochemistry, especially photosynthesis and the carbon cycle (Bonan, 1995; Foley et al., 1996; Dickinson et al., 1998; Kucharik et al., 2000). Inclusion of photosynthesis and the carbon cycle in land models makes the mixed life-form problem even more acute. To address this problem, the land modeling community has started using plant functional types to represent land surface. Plant functional types (PFT) are groups of plant species that share similar functioning at the organismic level, similar responses to environmental factors and/or similar effects on ecosystems (Smith et al., 1997). Deciduous broadleaf trees, evergreen needleleaf trees, grasses and broadleaf crops are examples of PFTs, whereas savannas, mixed forests and cropland/natural vegetation mosaics are not PFTs but biomes. Representing land surface in terms of PFTs offers several important advantages over the biome approach (Smith et al., 1997; Bonan et al., 2002). First, PFT provides a direct link to leaf-level physiological measurements, making it possible to more accurately set ecological parameters in land models. Second, PFT allows modelers to more accurately represent the land surface by separately altering the vegetation composition (i.e., the number of PFTs and their abundance) and structure (e.g., LAI, canopy height) within a grid cell. Third, representing landscapes as patches of PFTs also allows land surface models to better interface with ecosystem models, because the latter typically simulate vegetation change in terms of the abundance of PFTs. Reliable PFT information is increasingly needed by the global change research community, especially the carbon, climate and ecosystem modeling community. For example, the carbon models used to scale carbon fluxes typically require specification of PFTs (Denning et al., 1996; Sellers et al., 1997). The National Center for Atmospheric Research land surface model (NCAR LSM) has recently shifted from using land cover information to using satellite-derived PFT maps (Bonan et al., 2002; Tian et al., 2004). Plant functional types have also been advocated in dynamic global vegetation models (DGVM) to predict the composition and functioning of ecosystems in a changing environment (Running and Coughlan, 1988; Prentice et al., 1992; Woodward and Cramer, 1996; Smith et al., 1997; Kucharik et al., 2000). As such, accurate mapping of PFTs over large areas can contribute to improved predictive capabilities of global and regional carbon cycle, climate and ecosystem models. Remote sensing is the only practical means by which land surface biogeophysical variables can be obtained at the temporal and spatial scales required by global change research (Townshend et al., 1991; Roughgarden et al., 1991; Sellers et al., 1995; Myneni et al., 1997; Liang, 2004). In fact, land surface parameter estimation is

14 Methodologies for Mapping Plant Functional Types

371

a major direction in quantitative remote sensing (Liang, 2004). There is a wealth of literature on how to use remotely sensed data to extract land cover information and certain biogeophysical variables (e.g., Townshend and Justice, 1981; Townshend et al., 1987; Lloyd, 1990; Kimes et al., 1991; Myneni et al., 1995, 1997; DeFries et al., 1998, 1999; Loveland et al., 1991, 1995, 2000; Liang, 2001, 2003; Sun, 2004; Fang and Liang, 2005). However, using remote sensing techniques to extract PFTs is a relatively recent field of research. A review of the literature shows that to date only a very few methods for mapping PFTs have been reported (Strahler et al., 1999; Bonan et al., 2002; Sun et al., 2005, 2007). Furthermore, the accuracies of the PFT maps generated with existing methods have yet to be validated. Studies have demonstrated that the use of different PFT data sets generated with different methods has a significant effect on global climate modeling results (e.g., Oleson and Bonan, 2000; Bonan et al., 2002; Tian et al., 2004). These results highlight the importance of the quality of PFT data in global change research. Landscape can be thought of as a mixture of PFTs. As such, spectral unmixing techniques can be used to separate different types of vegetation within a geographic area. Substantial progress has been made in this context such that prototype data are now available. For example, a linear mixture model was employed to estimate a global land cover with continuous fields of vegetation characteristics (DeFries et al., 1999, 2000a). This global data set of continuous fields was able to capture the heterogeneity of vegetation within a grid cell at subpixel level. The main advantage of using continuous distributions of vegetation in land models is to avoid abrupt boundaries and unrealistic homogeneity that can be introduced into parameter estimates as is the case with discrete biome-based data sets. Linear unmixing was also used to derive a “pure PFT” NDVI data set using every 1 km AVHRR pixel with its corresponding PFT abundances (Bonan et al., 2002). In addition to the linear unmixing, another approach was also utilized to generate the pure PFT NDVI (Bonan et al., 2002). In that approach, the pure NDVI was extracted by averaging the NDVI over 1 km pixels in which the abundance of the PFT was greater than 60%. In fact, the pure PFT NDVI obtained in this way is not actually pure but rather a mix of co-occurring PFTs. Although the linear unmixing method is theoretically preferable and this method was expected to give a more accurate representation of the pure PFT NDVI, it was unstable (Bonan et al., 2002). Possible reasons may include calibration problems of the AVHRR data. When applied to EOS era data such as higher spatial and spectral resolution MODIS data, spectral unmixing techniques hold promises for exacting subpixel PFT fractions with improved accuracy and for estimating the individual PFT LAI of mixed pixels (Oleson and Bonan, 2000). Moreover, more sophisticated methods to allow for multiple endmembers and nonlinear statistical unmixing will potentially improve the accuracy of vegetation characteristics (DeFries et al., 2000a). The objective of this chapter is to provide an overview of recent developments in this evolving field and discuss future research needs in developing new methodologies for mapping of PFTs. The next section provides a brief summary of the existing methods for mapping PFTs reported in the literature. Section 14.3 discusses several methodological issues pertaining to the development of effective and robust

372

W. Sun, S. Liang

remote sensing techniques for extracting PFTs over large areas (e.g., continental to global scales). Section 14.4 outlines a multisource data fusion framework for inferring PFTs from MODIS. We conclude the chapter with some general remarks.

14.2 Existing Methods for Mapping PFTs: State of the Art 14.2.1 Extracting PFTs from Existing Land Cover Data Sets The land modeling community pioneered in the development of PFT data sets over large areas for use in land models (e.g., Bonan, 1996; Oleson and Bonan, 2000; Bonan et al., 2002). Typically, the methods used by modelers to extract PFT information rely on preexisting land cover maps. An early example of representing vegetation as PFTs is the National Center for Atmospheric Research land surface model (NCAR LSM) (Bonan, 1996). In this model, the vegetation in a grid cell is represented as mixtures of 12 PFTs. Up to three PFTs can form distinct patches in a grid cell, with lakes and wetlands forming additional patches. In the standard LSM, the PFT composition for each grid cell is obtained by classifying grid cells as one of 28 possible biomes. In other words, each grid cell is assigned a biome type, which in turn determines the patch fractions for each PFT. For example, savanna grid cells are represented as consisting of 70% C4 grasses and 30% tropical trees. All needleleaf evergreen forests are 25% bare ground and 75% needleleaf evergreen tree. The types of biomes and their geographic distributions used in the standard LSM are based on Olson et al. (1983). In their study of boreal forest surface fluxes, Oleson and Bonan (2000) prepared a PFT data set for the BOREAS region. This PFT data set was derived from the 1 km AVHRR BOREAS land cover map of Steyaert et al. (1997). The land cover classes defined at 1 km resolution were used to calculate the number of PFTs and their abundance (i.e., fractional areas) of each grid cell. Up to six PFTs can coexist within each grid cell. Fractions of each grid cell occupied by water and wetlands were also determined. Using a methodology similar to Oleson and Bonan (2000), Bonan et al. (2002) developed a global PFT map from 1 km satellite data. In this method, each model grid cell is divided into four primary land cover types: glacier, lake, wetland, and vegetation. The vegetated portion of a grid cell is further divided into patches of up to four of the model’s 15 PFTs. The geographic distribution and abundance of these 15 PFTs were derived from the 1 km IGBP DISCover data set (Loveland et al., 2000) and the 1 km University of Maryland tree cover data set (DeFries et al., 1999, 2000a, 2000b). Specifically, each 1 km pixel was assigned the percentage of needleaf evergreen, needleleaf deciduous, broadleaf evergreen, and broadleaf deciduous trees as given in the tree cover data. The non-tree-covered portion of the 1 km pixel was determined from the IGBP DISCover data. Figure 14.1 is an example of the PFT map for North America prepared by Bonan et al. (2002).

14 Methodologies for Mapping Plant Functional Types

373 Bare ground Needleleaf evergreen temperate trees Needleleaf evergreen boreal trees Needleleaf deciduous trees Broadleaf evergreen tropical trees Broadleaf evergreen temperate trees Broadleaf deciduous tropical trees Broadleaf deciduous temperate trees Broadleaf deciduous boreal trees Broadleaf evergreen temperate shrubs Broadleaf deciduous temperate shrubs Broadleaf deciduous boreal shrubs C3 Arctic grasses C3 Non-arctic grasses C4 Grasses Crops

Fig. 14.1 PFT map for North America from Bonan et al. (2002)

Studies done by the land research community have demonstrated that satellitederived PFT data has the potential to allow for more accurate representation of land surface processes and properties. However, the approach used by modelers to derive PFTs has several limitations. First, the accuracies of the existing land cover maps used to extract PFTs are generally unknown and, in many cases, appear poor (Townshend et al., 1991). Second, virtually all land cover maps produced in the past represent the land surface in terms of biomes. As a result, the number of PFTs and their abundance within each grid had to be prescribed or estimated from the land cover types depicted on the maps. The accuracy of the PFT data generated this way may be a concern. Third, this approach often involves using several separate and not necessary compatible land cover data sets. Consequently, modelers had to make lots of assumptions in places where existing land cover maps conflict with each other. Fourth, existing land cover data sets often lack detailed information needed for accurate parameterization of the land surface. Investigators were frequently forced to make assumptions about missing information. For example, Bonan et al. (2002) noted that, because consistent information on nonvegetated cover was not available, they had to assume that non-tree-covered land in forests, savannas, and grasslands was covered by grasses, in shrub lands by shrubs, in croplands by crops. Finally, as Fig. 14.1 illustrates, the PFT maps prepared by modelers are often of very coarse resolutions (e.g., 0.5◦ × 0.5◦ ). Such PFT data sets are also hard to update.

374

W. Sun, S. Liang

14.2.2 MODIS PFT: Supervised Decision Tree Classifiers In response to the needs of the land research community, the Moderate Resolution Imaging Spectroradiometer (MODIS) Land Team is producing a global PFT map (i.e., MODIS Land Cover 5) for use in the Community Land Model (CLM). This PFT product is generated as one of the five MODIS Land Cover products (MOD12Q1). Technical details of the methodology for producing MODIS Land Cover products are given in Strahler et al. (1999). We provide here a brief description of the key components of the methodology pertaining to the production of the MODIS PFT map. MODIS Land Cover is a global database of land cover types. The Land Cover product is prepared at 1 km spatial resolution and includes five different sets of land cover labels (Table 14.1). • Land Cover Type 1 includes the 17 classes of land cover defined in the International Geosphere-Biosphere Programme (IGBP) global vegetation classification scheme. • Land Cover Type 2 uses the University of Maryland land cover classification scheme, which is a version of the IGBP scheme modified to exclude wetlands, vegetation mosaic, and snow information. • Land Cover Type 3 uses the MODIS LAI/fPAR scheme, which is input to the MODIS LAI/fPAR product (i.e., MODIS15A2) and emphasizes vegetation classes. • Land Cover Type 4 uses the MODIS Net Primary Production scheme (Biome BGC Model classes). • Land Cover Type 5 uses the Plant Functional Types (PFT) scheme. The IGBP land cover classification scheme provides the primary layer of MODIS Land Cover data sets. This scheme includes 11 natural vegetation classes broken down by life form, three classes of developed and mosaic lands, and three classes of nonvegetated lands (Table 14.1). Strahler et al. (1999) note that the IGBP classes can be re-labeled or “cross-walked” to provide compatibility with current and future land cover classification systems used by the modeling community. Detailed information on how the IGBP land cover classes are “cross-walked” to the PFT classes currently produced by the MODIS Land Team is not released. A comparison of the IGBP scheme and the PFT scheme shows that for 10 of the 11 PFT classes, there is a direct mapping of one or more IGBP classes to their PFT equivalents. However, the PFT classes of cereal crop and broadleaf crop do not have their equivalents in the IGBP scheme, because in the latter scheme, crops are either lumped together as croplands or as part of cropland/natural vegetation mosaic. Information regarding what procedures are used to separate cereal crop from broadleaf crop in the production of the current MODIS PFT map is not available. Figure 14.2 shows a MODIS PFT map for North America produced by the MODIS Land Team. MODIS land cover classes, and by extension the PFT classes, are distinguished with a decision tree classification method. A decision tree is a supervised classifier

14 Methodologies for Mapping Plant Functional Types

375

Table 14.1 The five classification schemes used in MODIS land cover Land Cover Types Class

Type 1 IGBP

Type 2 UMD

Type 3 LAI/FPAR

Type 4 NPP

Type 5 PFT

0 1

Water Evergreen needleleaf forest

Water Evergreen needleleaf forest

Water Grasses/cereal crops

Water Evergreen needleleaf trees

2

Evergreen broadleaf forest

Evergreen broadleaf forest

Shrubs

3

Deciduous needleleaf forest

Broadleaf crops

4

Deciduous broadleaf forest

Deciduous needleleaf forests Deciduous broadleaf forest

5

Mixed forests

Mixed forests

Broadleaf forest

6 7

Closed shrublands Open shrublands

Closed shrublands Open shrublands

Needleleaf forest Unvegetated

8 9

Woody savannas Savannas

Woody savannas Savannas

Urban

Water Evergreen needleleaf vegetation Evergreen broadleaf vegetation Deciduous needleleaf vegetation Deciduous broadleaf vegetation Annual broadleaf vegetation Annual grass vegetation Non-vegetated land Urban

10 11

Grasslands Permanent wetlands Croplands Urban and built-up Cropland/natural vegetation mosaic Permanent snow and ice Barren or sparsely vegetated

Grasslands

12 13 14

15 16

Savanna

Evergreen broadleaf trees Deciduous needleleaf trees Deciduous broadleaf trees Shrub

Grass Cereal crop Broadleaf crop Urban and built up Snow and Ice Barren or sparse vegetation

Croplands Urban and built-up

Barren or sparsely vegetated

that recursively partitions a data set into smaller sub-divisions via binary rules and a heterogeneity-minimization function (Breiman et al., 1984). The tree is composed of a root node, a set of intermediate notes (splits), and a set of terminal nodes (leaves). In this framework, a data set is classified by sequentially sub-dividing it according to the decision framework defined by the tree, and class labels are assigned to each observation according to the leaf node into which the observation falls. Several studies have demonstrated the utility of decision trees in land cover classification at

376

W. Sun, S. Liang Evergreen needleleaf tree Evergreen broadleaf tree Deciduous needleleaf tree Deciduous broadleaf tree Shrub Grass Cereal crop Broadleaf crop Urban and built-up Snow and ice Barren or sparsely vegetated

Fig. 14.2 An example of MODIS PFT map for North America

regional to global scales (e.g., Hansen et al., 1996, Friedl and Brodley, 1997; DeFries et al., 1998; Friedl et al., 1999). Decision tree classifiers have several advantages over traditional supervised classification procedures used in remote sensing such as maximum likelihood classification (Hansen et al., 1996; Friedl and Brodley, 1997). First, decision trees are nonparametric and do not require any assumptions regarding the distributions of the input data. Second, decision trees can handle noisy or missing features and capture nonlinear and hierarchical relationships between the input variables. Third, decision trees have significant intuitive appeal because the classification structure is explicit and therefore easily interpretable. The MODIS land classification method exploits spectral and temporal information from MODIS. Key inputs include Nadir BRDF-Adjusted Reflectances (NBARs) derived from the MODIS BRDF/Albedo product (MOD43B4) in the MODIS Land Bands (1–7), MODIS Enhanced Vegetation Index (EVI) (MOD13), etc. (Strahler et al., 1999; Friedl et al., 2002). These data are composited over a 32-day period to produce a globally consistent, multitemporal database on a 1 km grid as input to the classification algorithm. Land cover classes are assigned by processing 12 (annual) 32-day composites using a decision tree classifier trained by site data. The success of decision tree classifiers requires extensive, high quality training site data base. The System for Terrestrial Ecosystem Parameterization (STEP) database is used to train the MODIS decision tree classifier. The STEP database is based on the information interpreted from Landsat and ancillary data. Key STEP parameters include vegetation life form, cover fraction, leaf types phenology, elevation, moisture regime, disturbance as well as descriptions of site and type. A more detailed discussion of the decision tree algorithm and its implementation in the production of MODIS Land Cover product is beyond the scope of this chapter.

14 Methodologies for Mapping Plant Functional Types

377

The reader is referred to Strahler et al. (1999), Hansen et al. (1996), Friedl and Brodley (1997), and Friedl et al. (1999) for more technical details.

14.2.3 Multisource Evidential Reasoning Sun et al. (2005, 2007) recently presented a multisource evidential reasoning method for mapping PFTs from MODIS data. Their method first utilizes a suite of improved and standard MODIS products to generate evidence measures for each PFT class. The multiple lines of evidence computed from input data are then combined using Dempster–Shafer theory of evidence (Shafer, 1976) to make classification decisions. In the Sun et al. (2007) study, the evidential reasoning method is implemented in three steps. Step 1: Generating evidence measures for each PFT class from each input data source. The PFT classification scheme used in the Sun et al. (2007) study is the same as the one used in the MODIS PFT product, which consists of 12 PFTs (Table 14.1). In Dempster–Shafer theory of evidence, a set of these 12 PFT classes constitute the frame of discernment, denoted by Θ. The degree of belief in the evidence from an input data source (e.g., LAI data) in support of a PFT class (e.g., grass) is expressed in terms of a mass function (m). A mass function has the following property:

∑

m(X) = 1

(14.1)

m(ø) = 0

(14.2)

X∈P(Θ)

where P(Θ) is the power set consisting of 2C subsets for a set of size C; ø is the empty set. The size of a set is the number of singleton classes in the frame of discernment. In the Sun et al. (2006) study, they used C instead of 2C because their study focused only on singleton hypotheses. Due to its generality, Dempster–Shafer theory of evidence does not specify how to compute evidence measures. Sun et al. (2007) developed a three-step procedure to derive masses of evidence. First, a mean vector over a whole year or the growing seasoning (April–October) is computed for each PFT class using each input data source. An example of the computed mean values of EVI for 10 PFT classes over the year 2001 for the state of Illinois, USA is given in Fig. 14.3. Second, the spectral distances of each pixel to the mean vector of each PFT class are then calculated for each input data source. Finally, these spectral distances are converted to probabilities of class membership (i.e., mass functions) using the sigmoidal fuzzy set membership function or a weighing function (Liang, 2004). Figure 14.4 is an example showing the masses or probabilities of each pixel belonging to three PFT classes computed from EVI data for Illinois, USA. Note that the higher the grayscale value of a pixel, which appears darker in the image, the higher the probability of that pixel belonging to a certain PFT.

378

W. Sun, S. Liang

0.7 Evergreen Needleaf Tree

0.6

Evergreen Broadleaf Tree

0.5

Deciduous Needleleaf Tree Deciduous Broadleaf Tree

0.4

Shrub

0.3 Grass

0.2

Cereal Crop

0.1

Broadleaf Crop Urban and Build-up

3

1

9

7

5

Barren or Sparsely Vegetated

35

32

28

25

22

1

3 19

16

9 12

97

65

−0.1

33

1

0

Fig. 14.3 Mean values of EVI for each PFT class over the year 2001 for Illinois, USA (X axis = day of the year, Y axis = EVI value)

Fig. 14.4 Evidence measures computed from EVI data for Illinois, USA showing the masses (or probabilities) of each pixel belonging to (a) deciduous broadleaf trees, (b) grass, and (c) broadleaf crop

Step 2: Combining evidence from all sources. Once the masses of evidence from all sources for each PFT have been determined, they are combined using Dempster’s rule of combination (Dempster, 1967). The equation for computing the orthogonal sum (⊕) of source 1 (with mass m1 over a set of labels X) and source 2 (with m2 over a set of labels Y ) is as follows:

14 Methodologies for Mapping Plant Functional Types

m1 ⊕ m2 (Z) = where: k=

∑

∑

X∩Y =Z

379

m1 (X)m2 (Y ) 1−k

m1 (X)m2 (Y )

(14.3)

(14.4)

X∩Y =Φ

k indicates the extent of conflict between the two sources considered (Shafer, 1976). Orthogonal summation of additional sources is achieved by repeated application of Eqs. (14.3) and (14.4). Step 3: Making classification decisions. To classify a pixel into one of the PFT classes, a decision rule is applied to the measure of support and/or plausibility. Support or belief function (Bel) is the total belief of a set and all its subsets. It is defined in terms of the mass: (14.5) Bel(X) = ∑ m(H) H⊆X

where H represents any subset of a set X. Plausibility (Pls) is defined as the degree to which a proposition, X, cannot be disbelieved or refuted. It is calculated as one minus the support for all other propositions (Shafer, 1976). (14.6) Pls(X) = 1 − Bel(X ) where (X ) is not (X). In the context of remote sensing image classification, Bel defines the lower boundary of the support committed to a class labeling proposition, Pls defines an upper boundary, and the range [Bel, Pls] is referred to as evidential interval. In the Sun et al. (2007) study, a maximum support decision rule is used, that is, the class with the highest support is assigned as the pixel label. The reader is directed to Sun et al. (2007) for a more detailed discussion of the theoretical background and key steps in the implementation of the evidential reasoning algorithm. The method of Sun et al. (2007) was tested over the states of Illinois, Indiana, Iowa and North Dakota, USA. They supplied the evidential reasoning algorithm with nine MODIS data sets, that is, improved Leaf Area Index (LAI) product obtained from a NASA funded data assimilation project, MODIS Enhanced Vegetation Index (EVI) (MOD13A2), and seven bands of “black-sky” spectral albedos (MOD43B3). Sun et al. (2007) reported that for several major PFTs in the study areas, the PFT maps generated with the evidential reasoning classifier represent significant improvements over the MODIS PFT product. Figure 14.5c shows an example of the classification results from evidential reasoning for the state of Illinois, USA. As a first step in the validation of their results and the MODIS PFT products, Sun et al. (2007) used the Cropland Data Layer (CDL) data provided by the National Agricultural Statistics Services (NASS) of the United States Department of Agriculture (USDA) as reference data (Fig. 14.5b). The NASS data was derived from high-resolution Landsat ETM+ data and its spatial resolution is 30m. As can be seen from Fig. 14.5, the PFT map generated with evidential reasoning gives much more spatial details of the PFTs in Illinois compared

380

W. Sun, S. Liang

Fig. 14.5 Comparison of MODIS PFT, USDA NASS data, and PFT classification results generated with evidential reasoning for Illinois, USA: (a) MODIS PFT map, (b) USDA NASS data used as “ground truth,” and (c) evidential reasoning classification results using evidence measures calculated from Illinois state growing season means plus weighing factors (weights used: LAI = 0.001, EVI = 0.6, albedo 1, 2, 3, 4 = 0.2, albedo 5, 6, 7 = 0.1) Table 14.2 Percent area of each PFT class identified in USDA NASS data (NASS), MODIS PFT (MODIS), and evidential reasoning classifications (EV) for Illinois, USA Class Trees Grass and shrub Crop Urban and built-up Water and wetland Clouds Snow and ice

NASS 11.63 19.54 59.04 5.64 1.61 2.53 0

MODIS 2.63 0.87 92.11 3.13 1.25 0 0

MODIS – NASS −9 −18.67 33.07 −2.51 −0.36 −2.53 0

EV

EV – NASS

11.61 12.01 72.88 2.65 0.86 0 0

−0.02 −7.53 13.84 −2.99 −0.75 −2.53 0

to the MODIS PFT map (Fig. 14.5a). In terms of percent area of each PFT class, the results obtained from evidential reasoning are much closer to the NASS data (“ground truths”), whereas large discrepancies exist between the MODIS PFT map and the NASS data (Table 14.2). It appears that the MODIS PFT map grossly overestimated the crop class, while it underestimated other important PFT classes such as trees and grass and shrub. This observation applies to all of the four states examined in their study.

14 Methodologies for Mapping Plant Functional Types

381

Sun et al. (2007) also experimented the evidential reasoning algorithm with various combinations of input data to examine the sensitivity of classification results to input data. A main conclusion from their experiment is that careful selection of input data is critical to obtaining satisfactory results from the evidential reasoning method. They also argued that despite the encouraging results from their study, more work is needed to validate their method over other regions, at other geographic scales (e.g., continents to global) and using additional sources of information including ancillary data such as climate and terrain. Overall, the work of Sun et al. (2005, 2007) demonstrates that multisource evidential reasoning is a promising approach to improved mapping of PFTs from MODIS data.

14.3 Key Issues In Mapping of PFTs Over Large Areas The MODIS PFT product (MODIS Land Cover 5) is the only global PFT data set produced on a regular basis at the present time. The current MODIS PFT is a land cover-derived product in that it is produced by “cross-walking” the land cover classes in the IGBP scheme to MODIS PFT classes. This means that, for the most part, the procedures currently used to produce the MODIS PFT map are not developed specifically for extracting PFT information from MODIS. Rather, the overall methodology employed by the MODIS Land Team is geared toward distinguishing land cover classes based on the IGBP scheme. Although the IGBP scheme has gained wide acceptance by the remote sensing community, it is fundamentally a biome-based land cover classification system. It seems arguable that methodologies effective for extracting biome-based land cover classes may not be optimal for mapping PFTs. Thus, at the present time it appears that no satisfactory methodology exists for the extraction of PFT information from satellite observations. An important reason why the land research community has requested for a PFT product is that reliable PFT data would give modelers more confidence in determining land surface processes and parameters for use in carbon cycle, climate and ecosystem models. Validation of the current MODIS PFT product is a demanding task and science quality information about the accuracy of the MODIS PFT data appears scarce. Errors and uncertainties in PFT data can propagate and compromise the credibility of global change research. As such, it seems desirable to develop and test new approaches to generate new PFT products. Such new products can be used as an independent validation method for the MODIS PFT, in addition to the validation activities undertaken by the MODIS Land Team. Due to the enormous diversity of terrestrial plant species and the spatial and temporal variability in the morphological and spectral characteristics of PFTs across regions, climates and years, accurate mapping of major PFTs around the globe is a difficulty task (Prentice et al., 1992; Box, 1996; Smith et al., 1997; Semenova and van der Maarel, 2000). The amount and quality of the data being generated by EOS sensors and other sensors are truly unprecedented (Justice et al., 2002). The increased availability and information content of remotely sensed data

382

W. Sun, S. Liang

has provided considerable potential for generating global PFT data sets. To date, however, research on how to effectively utilize multi-sensor data to accurately map PFTs remained quite limited. We provide below a brief discussion of several methodological issues that appear to be critical in the design and implementation of robust methodologies for extracting PFTs over large areas. (1) PFT classification scheme. The search for plant functional types represents a long standing desire of ecologists to seek simplified explanatory variables for understanding patterns in the richness of plants and the complexity of ecosystems. A review of the literature shows that there is a noticeable lack of consistency regarding how to define and group the major PFTs around the globe (Box, 1996; Woodward and Cramer, 1996; Smith et al., 1997). The PFT scheme used in the production of MODIS PFT is adopted from a scheme described in Bonan et al. (2002). The PFT scheme of Bonan et al. (2002) itself is adapted from the logic of Running et al. (1995) and Nemani and Running (1996). Briefly, the PFT scheme of Bonan et al. (2002) consists of seven primary PFTs (i.e., needleleaf evergreen or deciduous tree, broadleaf evergreen or deciduous tree, shrub, grass, and crop) and these primary PFTs are expanded to 15 physiological variants based on climate rules. McIntyre et al. (1999) call for simplified functional species classifications on the premise that the fewer groups used the greater the chance that broad patterns can be discerned. However, it is likely that the utility of PFT classifications will be scaledependent. Broad categories are likely to be most useful when considering patterns over large geographic areas such as continents or global, while more narrowly defined types may be necessary at more local scales. In order to accommodate the needs of environmental research at varying geographic scales, it seems desirable to develop a system of PFT classifications that can be used in remote sensing of PFTs at a hierarchy of scales. (2) How to characterize PFTs. PFTs can and should be characterized by a variety of variables in domains such as plant physiognomy, vegetation structure, phenology, and environmental conditions (Running et al., 1995). However, the synthesis we have done of the research on PFTs to date shows that the question of what variables are most crucial to distinguishing major PFTs using remote sensing techniques remains unanswered. As such, there is an urgent need to identify a set of key variables that together can effectively characterize major PFTs and therefore should be used to extract PFTs using remote sensing techniques. (3) Limitation of remote sensing instruments. While some of the characteristics exhibited by individual PFTs such as their phenologies and stand structure are observable by remote sensing, others may not. For example, certain site-specific environmental and ecological conditions such as climate and terrain appear to be less observable, but they are among the most important factors determining the geographic distribution of PFTs. As such, it seems that the use of remotely sensed data alone is inadequate to distinguish PFTs. This is especially true of mapping PFTs over large areas and using moderate resolution satellite data.

14 Methodologies for Mapping Plant Functional Types

383

(4) Limitation of spectral information. Remote sensing scientists have long recognized that the use of spectral information alone to interpret remote sensing data is seriously inadequate (Campbell, 1978; Townshend and Justice, 1981; Sun et al., 2003). Many studies, though not directly related to mapping of PFTs, have demonstrated that, when mapping land cover types over large areas and using moderate resolution satellite data, ancillary data such as climate and elevation contributed essential evidence for postclassification refinement and/or labeling of land cover classes where differing types exhibited similar spectral-temporal signatures (e.g., DeFries et al., 1998; Loveland et al., 2000; Hansen et al., 2000). The classification method used in the production of MODIS Land Cover products relies primarily on spectral and temporal information in MODIS. As such, how to integrate a variety of ancillary data into PFT classification procedures is an issue that deserves further research. (5) Environmental complexity of large geographic areas. When mapping PFTs over large areas (e.g., continents and global), the huge differences in climate and terrain across vast landmasses will almost certainly complicate image interpretation (Brown et al., 1993). This again suggests that integration of both satellite data and other relevant spatial data describing the environmental conditions of major PFTs will be critical to reliable mapping of PFTs, especially at regional to global scales. Ideally it would be useful for the remote sensing community to synthesize and encode a system of environmental and ecological constraints in some sort of hierarchy from major constraints operating at the global scale down to constraints operating at the regional and local level. Such knowledge may prove extremely valuable in improving the mapping of PFTs. (6) Need to integrate a priori knowledge. Although remote sensing of PFTs is a relatively recent field, there are multiple sources of knowledge that can be used to infer PFTs. For example, knowledge exists on what climatic conditions or thresholds (e.g., minimum temperature, occurrence of frost or freezing, and water moisture) are required for a plant species to occur on the earth’s surface. Such knowledge can be encoded in the form of a system of climatic envelopes and utilized, for example, to differentiate tropical, subtropical, temperate and boreal varieties of PFTs. There is also considerable knowledge about the spatial associations of major PFTs. For example, evergreen needleleaf trees and evergreen broadleaf trees rarely intermix geographically and, therefore, they could be separated by simple climate rules. Models of elevation-plant species relationships (altitudinal zonation) can also be incorporated into PFT classification procedures. Knowledge of the spectral characteristics of major PFTs can be generated from training sites. For example, the mean and variance of albedo, LAI, EVI, etc. can be calculated for each PFT class. The phenologies of each PFT class can be modeled using statistical techniques and the parameters derived from phenological models, such as maximum EVI, minimum EVI, greenup, maturity, senescence and dormancy onset dates, duration of growing seasons, etc., can be used to distinguish PFTs. Certain statistical rules can also be established, such as the EVI of broadleaf evergreen

384

W. Sun, S. Liang

trees has to be larger than a threshold value, or shrub albedo values have to be larger than certain threshold values, or the LAI/albedo values of grass are within an envelop for a given time. The above discussion suggests that considerable work is still needed before PFTs can be reliably mapped over large areas and across years. An important direction in the development of new methodologies for extracting PFTs appears to be the need to bring much more information, including both satellite observations and ancillary data, to bear on the PFT identification problem than is currently used in the MODIS product.

14.4 Developing a Multisource Data Fusion Approach to Extracting PFTs In this section, we outline a methodology that we are currently developing to extract PFTs from MODIS data. A key feature of our approach is that it utilizes not only spectral data but also ancillary information to infer PFT from satellite observations. It should be noted that our methodology is not specific to PFT mapping. It can also be applied to mapping of traditional land cover types.

14.4.1 Overall Framework The methodology described here is a fully automated PFT classification system capable of fusing multiple sources of data and information using a Dempster–Shafer belief system. Our methodology is a hybrid approach in that it combines both physically based methods and rule-based methods. Past research has demonstrated that purely physically based methods using spectral information alone often lead to poor results (e.g., Campbell, 1978; Townshend and Justice, 1981; DeFries et al., 1998; Loveland et al., 2000; Hansen et al., 2000). At the other end of the spectrum, purely empirical rule-based approaches have also been tried in the past with marginal success. As such, integrating physically based and rule-based methods into a single classification system appears to be a sound strategy for mapping PFTs over large areas. Our multisource data fusion framework consists of three main components: a knowledge base, evidence generation, and evidential reasoning algorithms (see Fig. 14.6). Each of these components is described in more details in the following sections. 1. The knowledge database contains a set of key parameters and rules essential to discerning major PFTs. 2. The evidence generation module is used to generate masses of evidence for PFT classification from each input data source using a variety of techniques.

14 Methodologies for Mapping Plant Functional Types

MODIS data & products

Evidence generation

385

Other spatial data

Knowledge base

Evidential reasoning algorithm

PFT product

Validation

Fig. 14.6 Flowchart of the multisource data fusion approach

3. The evidential reasoning classifier collects masses of evidence and combines them to compute the belief and plausibility of a pixel belonging to a particular PFT class.

14.4.2 Developing a Knowledge Database Our methodology relies greatly on the quality and amount of evidence and knowledge that we can provide to the evidential reasoning classifier. As discussed above, PFTs can be characterized in multiple domains including plant physiognomy, vegetation structure, phenology, and environmental conditions (Box, 1995; Running et al., 1995). An effective approach to extracting PFTs should therefore take into account the manifestations of PFTs in all of the above domains. To encode existing knowledge about the physiological, phenological, and ecological properties of each PFT and to organize various parameters and decision rules in a structured manner, we are currently constructing a comprehensive knowledge base. Our knowledge database contains a large number of parameters and rules that are essential to discerning major PFTs using remote sensing techniques. The main areas from which knowledge is being extracted include:

386

W. Sun, S. Liang

Table 14.3 Examples of the key items in the knowledge database Parameters/rules Plant physiognomy Growth and life form Leaf longevity Canopy structure/architecture Vegetation structure Fractional vegetation cover Phenology Intra-annual change Inter-annual change Environmental conditions Temperature Precipitation Terrain

Spatial Texture Spatial association

Used to separate

Input data/knowledge

Perennial versus annual Evergreen versus deciduous Broadleaf versus needleleaf

MODIS albedo, LAI, etc. MODIS albedo, LAI, etc. MODIS EVI, NDVI, LAI, etc.

Woody versus nonwoody

MODIS LAI, etc.

Growing activity Vegetation change, climatic impacts

EVI, NDVI, LAI AVHRR NDVI, MODIS NDVI, etc.

Limit to vegetation growth

MOD LST, long-range global monthly temperature data Limit to vegetation growth Long-rang global monthly precipitation data Spatial context, specially Elevation-vegetation models, elevation zonation of vegetation Global digital elevation models (DEM) Different forms of plants MOD albedo, etc. Spatial vegetation associations Meta-analysis of known associations published in the literature

• Physiological and phenological characteristics of each PFT class • Environmental conditions (e.g., climatic envelopes and elevation-PFT models) determining the geographic distribution of PFTs • Spatial associations between PFTs • Image characteristics of each PFT class in terms of its textures, etc. Examples of the key items used in the knowledge database are given in Table 14.3. It should be noted that the knowledge database can be easily updated and refined when new knowledge becomes available. This will allow us to continuously improve the accuracies of PFT product.

14.4.3 Generating Evidence for PFT Classification Integrating a wide array of data and information from multiple sources in a carefully structured manner is a key feature of the methodology described here. In our framework, we ingest a variety of both remotely sensed and other spatial data into the data system for generation of evidence to support PFT classification decisions. Besides standard MODIS products and the data acquired by other sensors, we use a suite of improved higher-level MODIS products (e.g., albedo and LAI) generated from a NASA funded data assimilation project. A range of ancillary data pertaining to the environmental conditions associated with each PFT class is also fused into the data system.

14 Methodologies for Mapping Plant Functional Types

387

Technically, the evidence generation module involves transforming the input data into measures of evidence for each PFT class with the aid of the knowledge database described above. The measures of evidence generated in this step will be input to the evidential reasoning algorithm. In our framework, we employ three techniques to accomplish this transformation. 1. Distances of each pixel to the mean vectors of each PFT class are first calculated for each source of information used in the classification system. These distances are then converted into probability images, with a higher probability indicating a higher level of “belief” of a pixel belonging to a particular PFT class. A variety of weighing functions and fuzzy set membership functions are used to convert distances into measures of evidence. Below are examples of the three weighing functions currently used in our work (Liang, 2004): . R2 − di,2 j (14.7) w(ri , r j ) = max 0, 2 R + di,2 j . di,2 j (14.8) w(ri , r j ) = exp − 2 2R

di, j di, j exp − (14.9) w(ri , r j ) = 1 + R R where: w(ri , r j ) is weighing function dependent on the distance di, j between point ri and r j ; R defines the spatial location in the one-, two-, or three-dimensional domain. 2. Measures of evidence computed from training data within a supervised classification framework. In this approach, evidential support is computed with respect to the frequency of occurrence of pixel values within training samples (Peddle, 1995a). This technique is considered more objective and involves identifying representative areas within an image for each PFT class. The underlying premises to this method are that training data contain evidence for a set of classes, and that the frequency of occurrence of a given value in the training set represents the magnitudes of support for those classes. 3. Rule-based measurements. Our knowledge database also contains a set of rules that are encoded to generate evidence from certain data sources (climate, elevation, etc.) for use in the evidential reasoning algorithm. Each rule has an antecedent, a consequent, and a weight. The antecedent consists of certain attribute values describing a pixel in the image. The consequent is the resultant assertion that the pixel belongs to a certain PFT class. The weight is the confidence in the assertion. Briefly, the rule-based measurements are obtained in the following way. Given an image, the method computes a set of attribute values for each pixel. Next, the classifier feeds pixels with the corresponding attribute values into a rule-based system. A pixel may trigger the firing of multiple rules, asserting complimentary or conflicting PFT classifications with a confidence value or weight.

388

W. Sun, S. Liang

14.4.4 Evidential Reasoning Classifiers For each pixel, the evidence generation module described above generates a set of basic probability assignments (or masses of evidence) from all sources of information used. A basic probability assignment ranges between 0 and 1 for positive belief, and between −1 and 0 for negative belief. Basic probability assignment of 1.0 and −1.0 indicate absolute certainty that a feature belongs or does not belong to a class, respectively. To determine which PFT class (if any) is supported by the strongest evidence, we need to combine these masses of evidence in a consistent manner. In our framework, this is done with an evidential reasoning algorithm. In the evidential reasoning method, the masses of evidence from all data sources for each PFT are combined using Dempster’s rule of combination (Eqs. (14.3) and (14.4)). Once the evidence from each source of information is combined by repeated application of the orthogonal summation rule, a decision rule is applied to the mass function to classify the pixel into one of the classes within the frame of discernment. In our classification system, the decision rule is based on maximum support, where the class with the highest support is assigned as the pixel label. A number of studies have demonstrated that evidential reasoning is a powerful approach to data fusion (Peddle, 1995a, b; Lein, 2003; Le H´egarat-Mascle et al., 2003; Soh et al., 2004). The method is built on the mathematical theory of evidence originally developed by Dempster (1967) and extended and refined by Shafer (1976). Evidential reasoning is based on the recognition that the knowledge and information we use in making decisions such as image classification is often uncertain, incomplete, and imprecise. As such, the method mimics human decisionmaking processes and utilizes many pieces of evidence to support a decision. As evidence increases, the algorithm will narrow the hypothesis set down to a smaller number of possibilities and improve its confidence in making a decision. Because the algorithm bases its decision on many pieces of evidence, it can usually generate better results compared to those classification procedures that rely on a single data source. A fuller discussion of the evidential reasoning method is beyond the scope of this chapter. It suffices here to note that evidential reasoning offers several important advantages over traditional image classifiers (Peddle, 1995a, b). First, it is a nonparametric classifier and therefore can handle data which may violate the Gaussian assumption of parametric classifiers. Second, it can handle data from any number of sources at any measurement scale. Third, the method has an explicit mechanism for handling information uncertainty. In situations of missing or incomplete information, evidential reasoning provides a more accurate representation of the available knowledge compared with the arbitrary assignment of probability used in Bayesian theory. Fourth, it can provide several interpretive measures such as support, ignorance and conflict that can be used to assess uncertainties in classification results. For example, the evidential reasoning method possesses the ability to express ignorance, which describes the incompleteness of one’s knowledge as a measure of the degree to which we cannot distinguish between any of the PFT classes. As such,

14 Methodologies for Mapping Plant Functional Types

389

ignorance can be viewed as a measure of uncertainty in the resultant classification. From a user’s point of view, information about the degree of ignorance (class confusion) may prove valuable.

14.5 Conclusion The production of an improved global PFT product at the MODIS scale should be of great interest to the global change research community. Improved mapping of PFTs can contribute to improved ability to simulate carbon cycle, climate and ecosystem change at regional to global scales. Using remote sensing techniques to map PFTs over large areas is a relatively recent and therefore evolving field of research. In this chapter, we briefly reviewed the methodologies developed by MODIS Land Team, Sun et al. (2005, 2007), and the land modeling community to extract PFTs from satellite observations. It seems clear that to date a limited number of methods for mapping PFTs have been reported in the literature. As such, developing and testing a variety of data analysis methods and extraction techniques for inferring PFTs from remotely sensed data constitutes a fertile field of research for the remote sensing community. In this chapter, we also discussed a number of methodological issues that deserve special attention in the development of new approaches to extracting PFTs over large areas. An important conclusion that we have drawn from this discussion is that incorporation of a wide array of information including both satellite observations and ancillary data into PFT classification procedures is indispensable to reliable mapping of PFTs. Integrating both satellite and other digital spatial data into a single classification procedure is indeed very challenging, because data from different sources often differ in spatial and temporal resolution, scale of measurement, accuracy, completeness, etc. Our ongoing research has shown that evidential reasoning is a promising method capable of fusing a large number of data sets and is effective in extracting and utilizing information from various data sources to generate reliable results. It is our hope that the multisource data fusion framework outlined in this chapter could spark further interest in the development of innovative and robust extraction techniques for inferring PFTs from remotely sensed data.

References Benediktsson JA, Swain PH, Ersoy OK (1990) Neural network approaches versus statistical methods in classification of multisource remote sensing data. IEEE Trans. Geosci. Remote Sens. 28:540–552 Bonan GB (1993) Importance of leaf area index and forest type when estimating photosynthesis in boreal forests. Remote Sens. Environ. 43:303–314 Bonan GB (1995) Land-atmosphere CO2 exchange simulated by a land surface process model coupled to an atmospheric general circulation model. J. Geophys. Res. 100:2817–2831

390

W. Sun, S. Liang

Bonan GB (1996) A land surface model (LSM version 1.0) for ecological, hydrological, and atmospheric studies: technical description and user’s guide. NCAR Technical Note NCAR/TN417 + STR. National Center for Atmospheric Research, Boulder, Colorado, 150pp Bonan GB, Levis S, Kergoat L, Oleson KW (2002) Landscapes as patches of plant functional types: an integrating concept for climate and ecosystem models. Global Biogeochem. Cycles 16(2), doi:10.1029/2000GB001360 Box EO (1995) Factors determining distributions of tree species and plant functional types. Vegetatio 121(1/2):101–116 Box EO (1996) Plant functional types and climate at the global scale. J. Veg. Sci. 7:309–320 Breiman L, Friedman JH, Olshen RA, Stone CJ (1984) Classification and regression trees. Wadsworth, Belmont, CA. Brown JF, Loveland TR, Merchant JW, Reed BC (1993) Using multisource data in global land cover characterization: concepts, requirements and methods. Photogramm. Eng. Remote Sens. 59:977–987 Campbell JB (1978) A geographical analysis of image interpretation methods. Prof. Geogr. 30:264–269 DeFries RS, Hansen M, Townshend JGR, Sohlberg R (1998) Global land cover classification at 8 km spatial resolution: the use of training data derived from Landsat imagery in decision tree classifiers. Int. J. Remote Sens. 19:3141–3168 DeFries RS, Townshend JRG, Hansen MC (1999) Continuous fields of vegetation characteristics at the global scale at 1-km resolution. J. Geophys. Res. 104D:16911–16923 DeFries RS, Hansen MC, Townshend JRG (2000a) Global continuous fields of vegetation characteristics: a linear mixture model applied to multi-year 8 km AVHRR data. Int. J. Remote Sens. 21:1389–1414 DeFries RS, Hansen MC, Townshend JRG, Janetos AC, Loveland TR (2000b) A new global 1-km dataset of percentage tree cover derived from remote sensing. Global Change Biol. 6:247–254 Dempster AP (1967) Upper and lower probabilities induced by multivalued mappings. Ann. Math. Stat. 38:325–329 Denning AS, Collatz GJ, Zhang C, Randall DA, Berry JA, Sellers PJ, Colello GD, Dazlich DA (1996) Simulations of terrestrial carbon metabolism and atmospheric CO2 in a general circulation model. Part 1: surface carbon fluxes, Tellus 48B:521–542 Dickinson RE, Shaikh M, Bryant R, Graumlich L (1998) Interactive canopies for a climate model. J. Climate 11:2823–2836 Fang H, Liang S (2005) A hybrid inversion method for mapping leaf area index from MODIS data: experiments and application to broadleaf and needleleaf canopies. Remote Sens. Environ. 94:405–424 Feddema JJ, Oleson KW, Bonan GB, Mearns LO, Buja LE, Meehl GA, Washington WM (2005) The importance of land-cover change in simulating future climates. Science 310:1674–1678 Foley JA, Prentice CI, Ramankutty N, Levis S, Pollard D, Sitch S, Haxeltine A (1996) An integrated biosphere model of land surface processes, terrestrial carbon balance, and vegetation dynamics. Global Biogeochem. Cycles 10:603–628 Friedl MA, Brodley CE (1997) Decision tree classification of land cover from remotely sensed data. Remote Sens. Environ. 61:399–409 Friedl MA, Brodley CE, Strahler AH (1999) Maximizing land cover classification accuracies produced by decision trees at continental to global scales. IEEE Trans. Geosci. Remote Sens. 37:969–977 Friedl MA, McIver DK, Hodges JCF, Zhang X, Muchoney D, Strahler AH, Woodcock CE, Gopal S, Schnieder A, Cooper A, Baccini A, Gao F, Schaaf C (2002) Global land cover mapping from MODIS: algorithms and early results. Remote Sens. Environ. 83(1–2):287–302 Gamon JA, Huemmrich KF, Peddle DR, Chen J, Fuentes D, Hall FG, Kimball JS, Goetz S, Gu J, McDonald KC, Miller JR, Moghaddam M, Rahman AF, Roujean JL, Smith EA, Walthall CL, Zarco-Tejada P, Fernandes R, Cihlar J (2004) Remote sensing in BOREAS: lessens learned. Remote Sens. Environ. 89:139–162

14 Methodologies for Mapping Plant Functional Types

391

Hansen M, Dubayah R, DeFries R (1996) Classification trees: an alternative to traditional land cover classifiers. Int. J. Remote Sens. 17:1075–1081 Hansen MC, DeFries RS, Townshend JRG, Sohlberg R (2000) Global land cover classification at 1 km spatial resolution using a classification tree approach. Int. J. Remote Sens. 26:1331–1364 Justice CO, Townshend JRG, Vermote E, Masuoka E, Wolfe R, Saleous N, Roy D, Morisette J (2002) An overview of MODIS land data processing and product status. Remote Sens. Environ. 83:3–15 Kimes DS, Harrison PR, Ratclife PA (1991) A knowledge-based expert system for inferring vegetation characteristics. Int. J. Remote Sens. 12:1987–2020 Kucharik CJ, Foley JA, Delire C, Fisher VA, Coe MT, Lenters JD, Young-Molling C, Ramankutty N, Norman JM, Gower ST (2000) Testing the performance of a dynamic global ecosystem model: water balance, carbon balance, and vegetation structure. Global Biogeochem. Cycles 14:795–825 Lawton RO, Nair US, Pielke Sr. RA, Welch RM (2001) Climatic impact of tropical lowland deforestation on nearby montane cloud forest. Science 294:584–587 Le H´egarat-Mascle S, Richard D, Ottl´e C (2003) Multi-scale data fusion using Dempster-Shafer evidence theory. Int. Comput.-Aid. Eng. 10:9–22 Lein JK (2003) Applying evidential reasoning methods to agriculture land cover classification. Int. J. Remote Sens. 24:4161–4180 Liang S (2001) Land cover classification methods for multiyear AVHRR data. Int. J. Remote Sens. 22:1479–1493 Liang S (2003) A direct algorithm for estimating land surface broadband albedos from MODIS Imagery. IEEE Trans. Geosci. Remote Sens. 41:136–145 Liang S (2004) Quantitative remote sensing of land surfaces. Wiley, New York, 534pp Lloyd D (1990) A phenological classification of terrestrial vegetation cover using shortwave vegetation index imagery. Int. J. Remote Sens. 11:2269–2279 Loveland TR, Merchant JW, Ohlen DO, Brown JF (1991) Development of a land cover characteristics database for the counterminous US. Photogramm. Eng. Remote Sens. 57:1453–1463 Loveland TR, Merchant JW, Reed BC, Brown JF, Ohlen DO, Olson P, Hutchinson J (1995) Seasonal land cover regions of the United States. Ann. Assoc. Am. Geogr. 85:339–355 Loveland TR, Reed BC, Brown JF, Ohlen DO, Zhu Z, Yang L, Merchant JW (2000) Development of a global land cover characteristics database and IGBP DISCover from 1 km AVHRR data. Int. J. Remote Sens. 21:1303–1330 Marland G, Pielke Sr. RA, Apps M, Avissar R, Betts RA, Davis KJ, Frumhoff PC, Jackson ST, Joyce LA, Kauppi P, Katzenberger J, MacDicken KG, Neilson RP, Niles JO, Niyogi DS, Norby RJ, Pena N, Sampson N, Xue Y (2003) The climatic impacts of land surface change and carbon management, and the implications for climate-change mitigation policy. Climate Policy 3:149–157 McIntyre S, Diaz S, Lavorel S, Wolfgang C (1999) Plant functional types and disturbance dynamics – introduction. J. Veg. Sci. 10:604–608 Myneni RB, Maggion S, Iaquinta J, Privette JL, Gobron N, Pinty B, Verstraete MM, Kimes DS, Williams DL (1995) Optical remote sensing of vegetation: modeling, caveats and algorithms. Remote Sens. Environ. 51:169–188 Myneni RB, Keeling CD, Tucker CJ, Asrar G, Nemani RR (1997) Increased plant growth in the northern high latitudes from 1981–1991. Nature 386:698–702 Nair US, Lawton RO, Welch RM, Pielke Sr. RA (2003) Impact of land use on Costa Rican tropical montane cloud forests: sensitivity of cumulus cloud field characteristics to lowland deforestation. J. Geophys. Res. 108:D7, 4206, doi:10.1029/2001JD001135 Nemani R, Running SW (1996) Implementation of a hierarchical global vegetation classification in ecosystem function models. J. Veg. Sci. 7:337–346 Oleson KW, Bonan GB (2000) The effects of remotely sensed plant functional type and leaf area index on simulations of boreal forest surface fluxes by the NCAR land surface model. J. Hydrometeorol. 1:431–446

392

W. Sun, S. Liang

Olson JS, Watts JA, Allison LJ (1983) Carbon in live vegetation of major world ecosystems. ORNL-5862, Oak Ridge National Laboratory, Oak Ridge, TN Peddle DR (1995a) Knowledge formulation for supervised evidential classification. Photogramm. Eng. Remote Sens. 61:409–417 Peddle DR (1995b) MERCURY: an evidential reasoning image classifier. Comput. Geosci. 21:1163–1176 Prentice IC, Cramer W, Harrison SP, Leemans R, Monserud RA, Solomon AM (1992) A global biome model based on plant physiology and dominance, soil properties, and climate. J. Biogeogr. 19:117–134 Prince SD, Goward SN (1995) Global primary production: a remote sensing approach. J. Biogeogr. 22:815–835 Roughgarden J, Running SW, Matson PA (1991) What does remote sensing do for ecology? Ecology 72:1918–1922 Running SW, Coughlan JC (1988) A general model of forest ecosystem processes for regional applications. I. Hydrological balance, canopy gas exchange and primary production processes. Ecol. Model. 42:125–154 Running SW, Loveland TR, Pierce LL, Nemani RR, Hunt ER Jr. (1995) A remote sensing based vegetation classification logic for global land cover analysis. Remote Sens. Environ. 51:39–48 Sellers PJ, Schimel DS (1993) Remote sensing of the land biosphere and biogeochemistry in the EOS era: science priorities, methods and implementation – EOS land biosphere and biogeochemical cycles panels. Global Planet. Change 7:279–297 Sellers PJ, Mintz Y, Sud YC, Dalcher A (1986) A simple biosphere model (SIB) for use within general circulation models. J. Atmos. Sci. 43:505–531 Sellers PJ, Meeson BW, Hall FG, Asrar G, Murphy RE, Schiffer RA, Bretherton FP, Dickinson RE, Ellingson RG, Field CB, Huemmrich KF, Justice CO, Melack JM, Roulet NT, Schimel DS, Try PD (1995) Remote sensing of the land surface for studies of global change: models – algorithms – experiments. Remote Sens. Environ. 51:3–26 Sellers PJ, Dickinson RE, Randall DA, Betts AK, Hall FG, Berry JA, Collatz GJ, Denning AS, Mooney HA, Nobre CA, Sato N, Field CB, Henderson-Sellers A (1997) Modeling the exchanges of energy, water, and carbon between continents and the atmosphere. Science 275:502–509 Semenova GV, van der Maarel E (2000) Plant functional types – a strategic perspective. J. Veg. Sci. 11:917–922 Shafer G (1976) A Mathematical theory of evidence. Princeton University Press, Princeton, NJ Smith TM, Shugart HH, Woodward FI (Eds.) (1997) Plant functional types: their relevance to ecosystem properties and global change. Cambridge University Press, New York, 369 pp Strahler A, Coauthors (1999) MODIS Land Cover Product Algorithm Theoretical Basis Document (ATBD). Version 5.0. Boston University Steyaert LT, Hall FG, Loveland TR (1997) Land cover mapping, fire regeneration, and scaling studies in the Canadian boreal forest with 1 km AVHRR and Landsat TM data. J. Geophys. Res. 102:29581–29598 Soh L, Tsatsoulis C, Gineris D, Bertoia C (2004) ARKTOS: an intelligent system for SAR sea ice image classification. IEEE Trans. Geosci. Remote Sens. 42(1):229–247 Sun W, Heidt V, Gong P, Xu G (2003) Information fusion for rural land-use classification with high resolution satellite imagery. IEEE Trans. Geosci. Remote Sens. 41:883–890 Sun, W. (2004) Land-Use Classification Using High Resolution Satellite Imagery: A New Information Fusion Method – An Application in Landau, Germany. University of Mainz Press, 95pp Sun W, Liang S, Xu G, Fang H (2005) Improving MODIS PFT product using multisource evidential reasoning. Proceedings of the 9th International Symposium on Physical Measurements and Signatures in Remote Sensing 1, pp. 225–227 Sun W, Liang S, Xu G, Fang H, Townshend, JR, Dickinson R (2007) Mapping plant functional types from MODIS data using multisource evidential reasoning. Remote Sensing of Environment (in press)

14 Methodologies for Mapping Plant Functional Types

393

Tian Y, Dickinson RE, Zhou L, Myneni RB, Friedl M, Schaaf CB, Carroll M, Gao F (2004) Land boundary conditions from MODIS data and consequences for the albedo of a climate model. Geophys. Res. Lett. 31:L05504, doi:10.1029/2003GL019104 Townshend JRG, Justice CO (1981) Information extraction from remotely sensed data – a user view. Int. J. Remote Sens. 2:313–329 Townshend JRG, Justice CO, Kalb V (1987) Characterization and classification of South American land cover types. Int. J. Remote Sens. 8:1189–1207 Townshend JRG, Justice CO, Li W, Gurney C, McManus J (1991) Global land cover classification by remote sensing: present capabilities and future possibilities. Remote Sens. of Environ. 35:243–255 Woodward FI, Cramer W (1996) Plant functional types and climatic changes: Introduction. J. Veg. Sci. 7:306–308

Part III

Remote Sensing Applications

Chapter 15

Monitoring and Management of Agriculture with Remote Sensing Zhongxin Chen, Sen Li, Jianqiang Ren, Pan Gong, Mingwei Zhang, Limin Wang, Shenliang Xiao, and Daohui Jiang

The intrinsic characteristics of agriculture make remote sensing an ideal technique for its monitoring and management (Chen et al., 2004). These characteristics include: (a) Agricultural activities are usually carried out in large spatial regions, which makes the conventional field survey or census time-consuming and usually costly; (b) the per-unit-area economic output from agriculture is not so significant in comparison with other industries; (c) most of the crops are annual herbs having different growth and development stages in different seasons which means that agricultural activities have obvious phenological rhythms and the intra-annual change may be very drastic; (d) agriculture is strongly affected by human activities and management where timely and accurate monitoring information is required. These intrinsic characteristics of agriculture demand novel techniques in the monitoring of crop growth and agricultural productions. Remote sensing technology meets these requirements by its rapidness, accuracy, economy, timing, dynamic and repetitive monitoring capacity. Remote sensing technology has been applied in agriculture extensively since its early stage in the 1960s. Now several global and national operational systems of monitoring agriculture with remote sensing have been operated. The number of similar operational systems at regional scale is much more. These systems provide timely and valuable information for agricultural production, management and policy-making. On the other hand, the demands arising from the applications in agricultural sectors have also enhanced the progress and innovation in remote sensing technology. The main applications of remote sensing in agriculture management and monitoring include: crop identification and cropland mapping, Zhongxin Chen, Jianqiang Ren, and Limin Wang Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry of Agriculture, Beijing, China [email protected] Zhongxin Chen, Sen Li, Jianqiang Ren, Pan Gong, Mingwei Zhang, Limin Wang, Shenliang Xiao, and Daohui Jiang Institute of Agricultural Resources & Regional Planning, Chinese Academy of Agricultural Sciences, Beijing, China S. Liang (ed.), Advances in Land Remote Sensing, 397–421. c Springer Science + Business Media B.V., 2008

397

398

Z. Chen et al.

crop growth monitoring and yield estimation/prediction, inversion of key biophysical, biochemical and environmental parameters, crop damage/disaster monitoring, precision farming, etc. In this paper, crop mapping, yield prediction, soil moisture monitoring and crop phenology monitoring with remote sensing are reviewed.

15.1 Crop Identification and Crop Mapping Crop distribution and acreage, as well as yield, is the basic information necessary for agricultural management and policy-making (Pradhan, 2001). Remote sensing of the extent and distribution of individual crop types has proven useful to a wide range of end-users, including governments, farmers, and scientists (Allen et al., 2002). The traditional method of crop identification is supervised classification of multiple land resource satellite (including Landsat MSS, TM, ETM+, SPOT-XS, IRS-LISS, CBERS, etc.) remotely sensed images throughout the growing season. This approach usually requires a lot of manual interpretation and cloud-free imagery for critical phenological stages which could be barriers for operational implementation over large areas and in multiple years (Broge and Mortensen, 2002; Doraiswamy et al., 2005; Kressler and Steinnocher, 1999). The Advanced Very High Resolution Radiometer (AVHRR) series of sensors onboard the National Oceanic and Atmospheric Administration (NOAA) satellites and the Moderate Resolution Imaging Spectroradiometer (MODIS) offer a unique combination of spectral and temporal resolutions, making them alternatives for large scale crop type mapping with novel classification techniques such as time series analysis and so on (Broge and Mortensen, 2002; Doraiswamy et al., 2005). The former is more accurate, but may be expensive and time-consuming when applied to a large area. The latter has the characters of low-cost and rapidness, but is low in accuracy because of the restriction of spatial resolution. Furthermore, SAR data is used in the discrimination of agricultural land use classes which can eliminate the influence of cloud and time (Soares et al., 1997; Chakraborty et al., 1997; Shao et al., 2001). Confusion between natural vegetation and cropland is a major source of error in remote sensing-based crop mapping with low-resolution imagery. Sometimes this is true even with high-resolution imagery. This kind of confusion is also obvious in regions with very complicated crop-planting patterns which can be generated from complicated topography or land ownership. For example, Loveland et al. (1999) reported that nearly 60% of the problems addressed in the post classification process for the International Geosphere Biosphere Programme’s Data and Information System (IGBP-DIS) global land cover data set arose from confusion between natural vegetation and cropland. This problem was also noted in the MODIS classification of agricultural land use. The most obvious confusion in this regard arose because of seasonal variation in the Normalized Differentiated Vegetation Index (NDVI) signals caused by seasonal variation in illumination geometry that mimicked a phenological cycle (Spanner et al., 1990; McIver and Friedl, 2002). The accuracy of a method to estimate the variability of cropland is affected by lots of factors, such

15 Monitoring and Management of Agriculture with Remote Sensing

399

as the data used, scale, crop type, etc. There are lots of researches on it in the past years. Agricultural land use has as specific characteristic that the surface reflectance changes regularly in time with the growth of a crop. This may cause it difficult to calculate accurately the total sown area of a specific crop in case of different types of cropping systems. Satellite data must cover the key phenological phase of the cropping system (Thenkabail et al., 2000).

15.1.1 Traditional Classification Methods The frequently adopted crop classification methods include unsupervised classification, supervised classification, and decision tree classifier. In the case where there is less information for a study area, only the image characteristics are used. Multiple groups, from randomly sampled data, will be automatically divided into homogeneous spectral classes using a clustering technique. The clustered classes are then used for estimating the population statistics. This method is only used under special conditions. The accuracy of this case is usually lower than that of supervised classification for shortage of knowledge on the study area. In order to determine a decision rule for the classification, it is necessary to know the spectral characteristics or features with respect to the population of each class. The spectral features can be measured using ground-based spectrometers or by sampling training data from clearly identified training areas, corresponding to defined classes that are usually made for estimating the population statistics. Statistically unbiased sampling of training data should be made in order to represent the population correctly. Maximum likelihood classification has been the most common method used for supervised classification of remotely sensed data (Wessel et al., 2004). The premise of this case is enough field samples in the study area. The date of collecting samples should be close to the date of satellite images, because the surface reflectance is changing with crop growth. The accuracy of crop identification is dependent on the quality of the training data. Many authors report classification accuracies exceeding 85% for crop mapping (Van Niel and McVicar, 2004; Baban and Luke, 2000).

15.1.2 Non-Parametric Classification Methods Increasingly, nonparametric classification algorithms are being used, which make no assumptions regarding the distribution of the data being classified, because the frequently adopted normal distribution hypothesis is usually not true for the remotely sensed data sets. The decision tree classifier is a hierarchically based classifier which compares the data with a range of properly selected features. The selection of features is determined by an assessment of the spectral distributions or separability of the classes (Pal and Mather, 2003). There is no generally established procedure and each decision tree or set of rules should be designed by an expert. When a decision

400

Z. Chen et al.

tree provides only two outcomes at each stage, the classifier is called a binary decision tree classifier (BDT). This method was used in many cases for its flexible characteristic. The rules of a decision tree are obtained through analyzing the particular attributes of different crop types. Usually auxiliary factors are also added to distinguish crop types, including climate, topography, spectral characteristics, etc. McIver and Friedl (2002) present a method for incorporating prior probabilities in remote sensing-based land cover classification using a supervised decision tree classification algorithm. Conventional classification has limited success for farm lands with high fragmentation and high spatial or/and temporal ecological heterogeneity. A knowledgebased approach combined with imagery and geographical data within a framework of an intelligent recognition system can improve the accuracy of crop identification. Experts interpret remote sensing images with knowledge based on experiences. However, computer assisted classification utilizes only very limited expert knowledge. The following two types of knowledge are required for an expert system in remote sensing. (a) Knowledge about image analysis: a feedback system should be introduced for checking and evaluating the objectives and the results. (b) Knowledge about the objects to be analyzed should be introduced in addition to the ordinary classification method. The fact that cropland does not exist over 60◦ slope, is one example of the type of knowledge that can be introduced. Cohen and Shoshany (2002) utilize the “split-and-merge” rules derived from the entire data set of imagery and auxiliary data to enable the formalization of different interpretation keys for each crop in Israel. This research area contains eight crop types that represent 70% of Israeli agricultural production. Multi-date Landsat TM images representing seasonal vegetation cover variations were converted to NDVI layers. Field boundaries were delineated by merging Landsat data with SPOT-panchromatic images. The difference of this method is based on auxiliary geographical and expert knowledge in the post-classification phase. Crop classification mapping with time series data was developed in the 1980s, and applied in remote sensing analysis. The basic theory of this method is that the spectral reflectance or vegetation index from satellite data is changing with crop growth (Lobell and Asner, 2004). Badhwar (1984) identified corn, soybean, and other land cover classes with a multi-temporal classification approach based on the greenness profile derived from Landsat MSS spectral bands. B¨uttner and Csillag (1989) mapped crop and soil inhomogeneities for a complex area through analyzing spring, summer, and autumn SPOT images. Chakraborty et al. (1997) classified the crop types using ERS-1 SAR data due to its independence from cloud cover, and results showed more than 90% classification accuracy for all types of wetland rice using three-date SAR data. Low spatial resolution data was extensively used in case of a large study area, e.g., NOAA-AVHRR and MODIS time series data, because of their advantages of high-temporal resolutions. Jakubauskas et al. (2002) identified crop types based on temporal changes in NDVI values of AVHRR. Harmonic analysis, or Fourier analysis, decomposed a time-dependent periodic phenomenon into a series of constituent sinusoidal functions, or terms, which were defined by a unique amplitude and phase

15 Monitoring and Management of Agriculture with Remote Sensing

401

value. Amplitude and phase angle images were produced by analysis of the timeseries NDVI data and used within a discriminant analysis to develop a methodology for crop type identification. Doraiswamy et al. (2005) separated the cropland area for crop yield estimation with MODIS 8-day surface reflectance images of bands 1 and 2. Areas of Interest (AOI) were created from the Landsat TM classification and standard procedures were followed to develop a land use classification using the parallelepiped method as the non-parametric rule, and the Mahalanobis Distance as the parametric rule. Xiao et al. (2005) developed a paddy rice mapping algorithm that uses time series of three vegetation indices (including NDVI, Enhanced Vegetation Index, EVI and the Land Surface Water Index, LSWI) derived from MODIS images to identify the initial period of flooding and rice transplanting in China. This research indicated that MODIS-based paddy rice mapping could potentially be applied at large spatial scales to monitor paddy rice agriculture on a timely and frequent basis. Because the pixel size of the space-born remotely sensed data is usually larger than the small parcel of cropland, the mixel problem is nearly ubiquitous in cropland mapping. Mixels can be the main source of errors in agricultural monitoring and agricultural statistics by remote sensing. A spectral mixture model method was used in cropland mapping research from microcosmic to macroscopic scale. Spectral mixture analysis (SMA) has become the basic tool for land cover analysis with remote sensing (Broge and Mortensen, 2002; Doraiswamy et al., 2005; Tompkins et al., 1997; Fitzgerald et al., 2005; Theau et al., 2005). SMA methods are typically utilized to calculate the fraction of each endmember in a mixed pixel using an inverse least square devolution method and endmember spectra. One basic assumption of SMA models is that the spectrum for each pixel is a linear or nonlinear combination of endmembers’ spectra dependent on the significance of multiple scattering of light of land cover types (Shimabukuro and Smith, 1991; Wu, 2004). Neural networks have proven to be the most significant improvement in information extraction in remote sensing in the last 15 years (Del Frate et al., 2003). This method compares favorably with an optimal Bayesian classifier. The classification performance of the method is proven superior compared with other statistical and neural classifiers (Moshou et al., 2001). The classification of remote sensing data using artificial neural networks appeared in remote sensing interpretation about 10 years ago (Park et al., 2005). Afterwards, this method was used in crop type identification and other agricultural research fields (Del Frate et al., 2003). In crop identification and cropland mapping, the accuracy has been improved by innovation of new classification algorithms, integrating various classification methods, data fusion of remotely sensed data from different types of sensors, incorporating other auxiliary data and expert knowledge, etc. To guarantee the precision of crop area monitoring in a large area, it is necessary to develop the technology of field sampling. There is a trend to integrate high-resolution and low-resolution remotely sensed imagery in large-scale crop area estimation. That is, the low-resolution imagery can be used to estimate the distribution of crops and this information can be used as sampling frame. Then high-resolution imagery can be sampled for accurate crop mapping.

402

Z. Chen et al.

15.2 Crop Yield Estimation and Prediction Crop yield and production data are key indicators for national food security and sustainable development of society. Early and accurately gathering knowledge of crop yield and production or the changes of production is very important. Regional and national crop yield estimation and prediction with remote sensing are of great interest for scientists, policy-makers and the general public. Crop yield estimation with remote sensing is not only for the staple crops such as wheat, maize, rice, cotton, soybean, but also for some marginal crops. The techniques and methods to estimate crop yield and production include statistical sampling methods, agro-climate models, crop growth models, remote sensing, and some integrated methods. Because of the specific character of remote sensing, for example, the field of view and swath width is wide and the period of detecting the earth surface is short, remote sensing has been used to estimate crop yield and production at a large scale in many countries (Fuller, 1998; Huang et al., 2002; Hochheim and Barber, 1998; Jiao et al., 2005; Potdar, 1993; Groten, 1993; Weissteiner et al., 2004). Crop yield estimation with space-born remotely sensed data can be traced back to mid-20th century. The National Aeronautics and Space Administration (NASA), NOAA and the United States Department of Agriculture (USDA) implemented a large area crop inventory experiment (LACIE) project in the years 1974–1978 and agricultural and a resources inventory surveys though aerospace remote sensing (AgRISTARS) project in the years 1980–1985 (Wang, 1996; Allen et al., 2002). Important information and many experiences about crop monitoring with remote sensing were gained from these tasks. From then on, European and other countries also developed crop yield estimation and crop monitoring systems. Remote sensing-based crop yield estimation is a prominent example of the macro-research of remote sensing of vegetation (Boogaard et al., 2002). Because of the occurrence of a seasonal rhythm of vegetation, the micro-structure of plant cells and the macro-structure of vegetation canopies changes accordingly and the spectral response of individual vegetation types or of a population also changes periodically. So human can carry out many researches such as vegetation growth monitoring and biomass estimation depending on the multi-spectral response of vegetation to derive them and learn their changing information. With the development of remote sensing, the researches of remote sensing of vegetation have developed practicably. Scientists have put forward Vegetation Index (VI) models which are indicators to vegetation and are linear or non-linear combinations of multi-spectral data (Dadhwal and Ray, 2000). The different combinations of measured reflectance in the visible (Rv ) and near-infrared (Rn ) parts of the spectrum compose the core of VI (Carlson and Ripley, 1997; Gitelson and Kaufman, 1998). Remote sensing-based crop yield estimation is a series of techniques and methods to forecast crop yield before harvesting of the target crops. Based on the theories of biology and spectroscopy and cognition of crop foliage and canopy spectral response, the crop classes are identified and the spectral data of different crops in different spectral band are acquired through the sensors. Then using the spectra data,

15 Monitoring and Management of Agriculture with Remote Sensing

403

we can monitor the crop growth and establish various models for crop yield forecasting. Generally, there are three categories of models based on remote sensing: empirical models, physiological models and crop growth models.

15.2.1 Empirical Models The empirical approach to estimate crop yield is based on the hypothesis that biological/physical crop parameters or ecological environmental parameters of the land surface such as biomass, LAI, temperature are correlative to final crop yield or production and remote sensing data are obviously correlative to these above parameters of critical growth stages (Dadhwal and Ray, 2000; Mkhabela et al., 2005). Of course, the empirical relationships may be simple linear or non-linear and may be multiple (Tucker et al., 1980; Sridhar et al., 1994). The crop parameters may be in one critical growth stage or may be in more (Barnett and Thompson, 1983). In some cases, there may be some non-remote sensing-based parameters such as weather data used in crop yield monitoring and prediction (Kalubarme et al., 1995). There are some advantages using empirical relationships for yield estimation. Firstly, it is simple to estimate crop yield. A few ground-truth measurements are sufficient for the model validation. Single or several images can meet the need of yield estimation. Secondly, the accuracy of estimation is higher especially in a homogenous region and it is enough to meet the information need of agricultural management. There are also some disadvantages for this method. For example, it is based on a link between crop biomass and remote sensing data, which sometimes may be occasional but not intrinsic. It is lack of flexibility and stability, so the results at different stages vary drastically.

15.2.2 Physiology-Based Models A physiology-based model is mainly based on crop physiological functions and assumes that crop production results from photosynthesis through which a fraction of incident solar radiation is converted into biomass. The most popular formulation of this approach is the efficiency equation of Monteith (1972), which can be applied to remote sensing data. The simple model for crop yield estimation can be written as follows: DM =

& t2 t1

ε ∗ fPAR ∗ PAR · dt

Yield = DM ∗ HI

(15.1) (15.2)

404

Z. Chen et al.

where DM is dry matter production in a time period t2 − t1 , ε is light use efficiency, PAR (MJ m−2 ) is the incoming photosynthetically active radiation for the wave bands between 0.4 and 0.7 µm. PAR is part of the short wave solar radiation (0.3–3 µm) and is absorbed by chlorophyll for photosynthesis in the crop. fPAR is the fraction of the photosynthetical radiation absorbed by the canopy. HI is the harvest index which means the ratio of grain mass to aboveground biomass. Judging from Eqs. (15.1) and (15.2), we know that using this method to estimate crop yield we only need to learn the parameters of ε , HI, fPAR and PAR. Researchers have done a lot of work on these parameters. ε (g MJ−1 m−2 ) differs between C3 and C4 plants and is affected by climate factors such as temperature and rainfall (Hanan et al., 1995; Bastiaanssen and Ali, 2003; Field et al., 1995). As to the maximum ε for each kind of crop, different researchers often get different results. Many studies have shown that ε is a relative constant property of plants and only varies over a relatively narrow range for crop ecosystems but over a wider range for natural ecosystems. fPAR is a key variable in the assessment of vegetation production. The fPAR can be calculated depending on the relationships between LAI, VI and fPAR which has been studied by many authors (Hu et al., 2003; Shabanov et al., 2003; Myneni et al., 2002). PAR can be obtained from observation data of weather stations or can be derived from remote sensing such as TOMS (Total Ozone Mapping Satellite) reflectance (Goldberg and Klein, 1980; Eck and Dye, 1991). As to the strong points of this method, it is more universal and suitable for more crop types and it only needs a relatively simple data set such as solar radiation and fPAR. But light use efficiency depends on the phenological stage and environment conditions such as temperature and rainfall.

15.2.3 Crop Growth Models A crop growth model describes the primary physiological mechanisms of crop growth such as phenological development, photosynthesis and dry matter partitioning, and their interactions with the underlying environmental factors using mechanistic and sometimes empirical equations (Delecolle et al., 1992). Using crop growth models to estimate crop yield requires a lot of inputs that are specific to the crop, soil characteristics, management practices and local climate conditions. So it has limitations to use this kind of models in large regions because fewer inputs are generally available at this scale. Remote sensing has shown to be capable of providing certain crop characteristics and some other parameters. So a crop growth model can be combined with remote sensing and use input parameters which are derived from remote sensing for a larger region (Guerif et al., 1993; Maas, 1988). When integrating remotely sensed data with a crop growth model, we can use several methods (Doraiswamy et al., 2003). Firstly, model initialization can be done by the parameters estimated from remote sensing and/or taking these parameters

15 Monitoring and Management of Agriculture with Remote Sensing

405

directly as input data. This kind of parameters is used successfully including LAI and crop canopy cover. The other method is data assimilation of data derived from remote sensing that is used to calibrate the crop growth model. There are many successful researches on mechanistic crop growth model incorporating with remote sensing (Clevers and van Leeuwen, 1996; Gu´erif and Duke, 2000; Doraiswamy et al., 2003). Clevers and van Leeuwen (1996) used optical and microwave remote sensing data to estimate LAI for calibrating the SUCROS (simplified and universal crop growth simulator) model and the improved the accuracy of sugar beet yield estimation. Gu´erif and Duke (2000) studied the method of coupling the radiative transfer model SAIL (Scattered by Arbitrary Inclined Leaves) and the crop growth model SUCROS. Doraiswamy et al. (2003) derived LAI data from Landsat TM and NOAA AVHRR to calibrate the EPIC (Erosion Productivity-Impact Calculator) model depending on the link provided by the radiative transfer model SAIL and applied this method to simulate spring wheat yield for the state of North Dakota of America. Recently, Yang (2005) integrated LAI data from Landsat TM with the EPIC model and calibrated the crop growth model including cropping system, planting and harvest data, and applied this method to estimate winter wheat yield in North China. There have been great progresses in crop yield monitoring and prediction because of novel remote sensing sensors, improved algorithms and models. The progress in quantitative remote sensing made the accurate inversion of land surface and crop parameters possible, which strengthened the data input for various crop monitoring models and systems. Different kinds of crop models have been applied for different purposes. Empirical relationships for crop yield monitoring and prediction are widely used for its simplicity. But its shortcoming is also obvious, the unstableness and sometimes site-specific relationship between yield and remote sensing data. Physiology-based models are mainly based on crop physiological functions, which is its strong point. But some parameters are not consistent over a large region and/or for different crops and sometimes not easily acquired by remote sensing or an in situ survey. Crop growth models have a long history and extensive use for crop growth monitoring, yield prediction and farm management around the world. Huge work of data collection and preparation hampered its good performance at a regional scale. During recent years, crop growth models with remote sensing data assimilation have been improved greatly at regional scales for better estimation of crop parameters.

15.3 Crop Phenology Monitoring Crop phenology is important in crop monitoring because it can have great impact on the monitoring accuracies of crop yield and acreage change. Accurate monitoring of crop development patterns (i.e., phenology and growth) is an important component of farm management since it enables us to assess crop growth under various regional weather conditions. Monitoring seasonal changes in vegetation activity and crop phenology over wide areas is essential for many applications, such as estimation of

406

Z. Chen et al.

net primary production (Kimball et al., 2004), deciding on the time period for crop yield modeling (Bouman et al., 2002) and supporting decisions about water supply (Digkuhn and Gal, 1996). Because of the synoptic coverage and repeated temporal sampling that satellite observations afford, remotely sensed data possess significant potential for monitoring vegetation dynamics at regional to global scales (Myneni et al., 1997). Various methods have been developed to monitor seasonal vegetation changes using time series of normalized difference vegetation index (NDVI) data. These methods have employed a variety of different approaches including the use of specific threshold-based methods (Viovy et al., 1992; White et al., 1997), Fourier-based fitting methods (Moody and Johnson, 2001), asymmetric function fitting methods (Jonsson and Eklundh, 2002; Zhang et al., 2003), backward-looking moving averages method (Reed et al., 1994), and so on. The data collected by MODIS onboard NASA’s Terra/Aqua spacecrafts are useful for monitoring crop phenology (Zhang et al., 2003; Sakamoto et al., 2005). Satellite vegetation index (VI) data such as the NDVI are correlated with green leaf area index (LAI), green biomass, and vegetation cover (Baret and Guyot, 1991). In most agricultural remote sensing applications, monitoring spectral characteristics of the crops at one particular stage is more popular than those over the entire growing season. So, the spectral characteristics of vegetation and the background surface (soil or water) should be well understood. Monitoring crop phenologies can serve the purpose of forecasting grain harvest (yield prediction), collecting crop production statistics, facilitating crop rotation records, mapping soil productivity, identification of factors influencing crop stress, assessment of crop damage due to storms and drought, and monitoring farming activity. Crop phenology controls the temporal changes observed from satellite data. Crop development is shown in remote sensing by integrating space and time. Therefore, incorporating crop development stages into remote sensing analyses is conducive to crop type identification and area measurement. Different crops (winter wheat, maize, soybeans, alfalfa, etc.) exhibit a distinctive seasonal pattern and period of maximal greenness. This information (phenology) may be used in the classification process to accurately discriminate vegetation types. Vegetation indices (VIs) have been extensively used for identifying crop types. For instance, NDVI time series analysis was applied to identify several common crop types occurring within the Western Great Plains (Jakubauskas et al., 2002). Overall, the temporal NDVI images provide the crop cycle and the cropping system pattern of agricultural land use. Observed changes in the NDVI through time are generally thought to reflect vegetation type, phenology and local environmental conditions. Therefore, compared to land cover classification using single date data, multi-temporal datasets are often found to improve the accuracy of classification. With the use of NDVI/EVI time-series we can detect and monitor vegetation cycles and timings over large areas. This offers many opportunities for more complete vegetation descriptions than could be achieved with only single-date images. The discrimination of NDVI/EVI time-series is based on their characterization of seasonal dynamics of vegetation growth (phenology).

15 Monitoring and Management of Agriculture with Remote Sensing

407

Phenological knowledge plays a critical role in identifying optimal acquisition dates for selection of satellite imagery for agricultural monitoring. For example, greenup is a particularly important vegetative stage (greenup is the date of onset of photosynthetic activity). Greenup of winter wheat occurs earlier than for other crops. At this stage, winter wheat can be easily identified from other vegetation because of its greater greenness. In general, paddy fields are plowed and flooded before rice planting. Planting date of rice is when rice is often distinctive from other crops since the water background of rice can easily be distinguished from both soil and green vegetation (Xiao et al., 2002). Crop phenology is generally divided into vegetative and reproductive stages. The vegetative stage is largely defined by the part of the growth cycle where the crop develops and grows, starting emergence to tasseling. The reproductive starts at anthesis and ends after maturity. For dryland crops, several transitions are important in terms of management: emergence, tasseling, and initiation of senescence. For example, unfavorable conditions occurring between corn emergence and leaf development will limit the size of the leaves and thus the amount of photosynthetic biomass (Vina et al., 2004). Harmful conditions at the beginning of the reproductive cycle will result in impairment of pollination and decrease of the number of fertilized kernels that are destined to be filled. The water stress or temperature or disease stress during the grain-filling period directly affects the crop yield. Therefore, identifying stress-induce during the key transitions is important for influencing farm management practice. For different locations, the timing of the growth cycle of rice varies depending on local climate, management, and cultivar planted. Traditional methods of obtaining information on crop phenology are census and ground surveying. When planning a monitoring campaign that extends over large areas it is often difficult to obtain precise crop calendars. In general these crop calendars only record specific dates (or ranges) of activities (i.e., field preparation, sowing, emergence, and harvest). These calendars may be applied to the selection of remotely sensed data to discriminate crop types, but they have limitations in supporting decisions about farm management practice. Remote sensing can provide relevant data collection and information extraction strategies.

15.4 Soil Moisture Monitoring Soil moisture is an important variable controlling biogeochemical cycles, heat exchange and infiltration rates at the land/atmosphere boundary (Li and Islam, 1999). Soil moisture information is critical in agricultural water management and drought monitoring. However, soil moisture observations are often inadequate, due to an inability to economically monitor spatial variation in soil moisture from traditional in situ measurement techniques such as weighting and TDR (Time Domain Reflectometry) method. In contrast, the development of remote sensing technology makes it possible to monitor soil water in nearly real time at large scale.

408

Z. Chen et al.

Monitoring soil moisture and drought with remote sensing is based on two basic principles. The first is that soil water content change leads to variations of soil spectral reflectance; the other is that soil water content change can cause a physiological change of a plant, thus changes the spectral characteristics of a leaf and affects the reflectance of the canopy. There are three categories of remote sensing systems used in soil moisture monitoring: optical (including reflective near-infrared), thermal and microwave systems. Optical systems acquire data from visible wavelengths up to, and including, near infrared (∼400–1, 000 nm). They measure the amount and wavelength of the sun’s radiant energy reflected from a target object. The inclusion of wavelengths visible to the human eye often allows an intuitive interpretation of images, especially when combined with expert knowledge of soil properties. Applications for the determination of soil moisture have been investigated but were limited by the need for bare soil surfaces and also by the fact that soil moisture and soil organic carbon properties give similar signals when they change. Thermal systems refer, in general, to sensors that measure energy emitted within the wavelength range from 3 to 13 µm. According to different approaches, microwave systems include passive systems, i.e. radiometers such as SMMR (Scanning Multichannel Microwave Radiometer), SSM/I (Special Sensor Microwave Imager), TMI (TRMM Microwave Imager), AMSR /AMSR-E (Advanced Microwave Scanning Radiometer); and active systems, i.e., radar systems such as ERS-1/2, RADARSAT, ENVISAT ASAR, and so on. Research work on soil moisture remote sensing was initiated in the late 1960s to early 1970s. Watson et al. (1971) used a thermal model for the geologic interpretation of infrared (IR) images. Bijleveld (1978) did further research based on Watson’s work and raised the model of thermal inertia and daily evaporation calculation. Pratt and Ellyett (1979) presented the thermal inertia approach for mapping soil moisture. All this initiative work founded a base for multi-method research on soil moisture remote sensing. Great achievements were made in the field of soil moisture remote sensing in the 1980s. Soil moisture data can be obtained by means of ground, space and satellite remote sensing. The spectrum covers visible light, near-, middle- and far-infrared bands, thermal infrared band and L, C and X microwave bands. Methods and indices used for soil moisture mapping were developed step by step such as regional evaporation estimation, crop surface temperature, soil heat capacity, drought indices, crop water stress and leaf water content, and so on. Pratt and Ellyett (1979) and Price (1985) developed the thermal inertia approach and the mechanism of thermal infrared imagery. Attempts for large scale soil moisture remote sensing using meteorological satellite data were starting at the same time. Carlson (1986) used NOAA /AVHRR data to estimate surface moisture and thermal inertia at regional scale. Owe (1988) used NOAA /AVHRR data and satellite microwave data to derive vegetation indices and estimate surface soil moisture. Although visible, near-infrared and thermal infrared bands can be used to monitor soil moisture, they are not effective in case of cloud cover. Microwaves have a good penetration ability of clouds, so it has an advantage over the other methods in soil moisture monitoring, especially in cloudy regions. Soil moisture remote sensing by microwaves was initiated in the 1970s (Schmugge et al., 1974; Choudhury

15 Monitoring and Management of Agriculture with Remote Sensing

409

et al., 1979) and developed after the 1980s (Schmugge et al., 1986; Njoku and O’Neill, 1982; Wang, 1985). With the launch of a series of microwave sensors such as ERS-1& 2, RADARSAT, ENVISAT, TMI, AMSR and AMSR-E, soil moisture monitoring by microwaves developed rapidly in recent years (Gu et al., 2002; Wickel and Jackson, 2001; Blumberg and Freilikher, 2001; Glenna and Carr, 2003). Approaches or indices used in soil moisture monitoring mainly include thermal inertia approaches, crop water stress index (CWSI), transformed normalized difference vegetation index (TNDVI), vegetation temperature condition index (VTCI), thermal infrared approaches and microwave remote sensing.

15.4.1 Thermal Inertia Approaches Water has a large heat capacity and heat conductivity, so the thermal inertia of wet soil is larger than that of dry soil. Thermal inertia can be observed from the monitoring of ground surface temperature by optical remote sensing, such as NOAA /AVHRR or EOS / MODIS data. The thermal inertia method is commonly used in soil water monitoring research. Soil thermal inertia is a heat characteristic of soil, it is the internal factor which causes the change of soil surface temperature. Thermal inertia has very close relation with soil water content, and it affects the diurnal temperature range of soils at the same time. The thermal inertia method is suitable in case of bare soil and the early stages of crop growth. Thermal inertia can be written as: (15.3) P = λ ρc where P is thermal inertia (J · m−2 · K−1 · s−1/2 ), λ is heat conductivity (J · m−1 · s−1 · K−1 ), ρ is the density of soil (kg · m−3 ), c is specific heat (J · kg−1 · K−1 ). The classical thermal inertia method raised by Price in 1985 needs input of a large number of in situ data, thus leads to poor practicability of this method. So, thermal inertia is usually used in practice as apparent thermal inertia (ATI). ATI can be obtained from satellite data, which greatly enhances the applications of thermal inertia. Apparent thermal inertia is written as AT I =

1−A Tmax − Tmin

(15.4)

where Tmax is the maximum temperature and Tmin is the minimum temperature on the same day, A is the full spectrum albedo. Soil water content is usually estimated from an empirical linear equation. The equation is written as W = aAT I + b (15.5) where W is soil water content, a and b are coefficients.

410

Z. Chen et al.

15.4.2 Crop Water Stress Index The crop water stress index was proposed by Jackson et al. (1981). This index is based on the energy balance theory and calculated from the difference between canopy and air temperatures. The energy balance equation can be expressed as Rn − G = λ E + H

(15.6)

where Rn , G, λ E and H(W · m−2 ) are net radiation, soil heat flux, latent heat flux and sensible heat flux, respectively. The sensible heat flux can be expressed in terms of a temperature difference as H = ρCp

Tc − Ta γa

(15.7)

where ρ (kg · m−3 ) is the air density, Cp (J · kg−1 K−1 ) the specific heat of the air, γa (s · m−1 ) the aerodynamic resistance and Tc and Ta (K) the canopy temperature and air temperature at the reference height, respectively. The CWSI is expressed as CW SI =

(Tc − Ta ) − (Tc − Ta )ll (Tc − Ta )ul − (Tc − Ta )ll

(15.8)

where the subscripts ll and ul denote lower (well-watered crop) and upper (nontranspiring crop) limits, respectively. Substituting Eq. (15.7) into Eq. (15.6) and solving for Tc –Ta , Tc − Ta =

γa [(Rn − G) − λ E] ρCp

(15.9)

In a well-watered crop, λE is equal to the potential crop evaporation (expressed in g · m−2 · s−1 ) times the latent heat of vaporization (λ, J · g−1 ). In a completely stressed crop, λE is zero, thus Eq. (15.9) reduces to: (Tc − Ta )ul =

γa (Rn − G) ρCp

(15.10)

The CWSI has been commonly applied to the detection of water stress of plants, but difficulties in measuring canopy temperature of crops with less than 100% vegetation cover has limited its operational application.

15 Monitoring and Management of Agriculture with Remote Sensing

411

15.4.3 Transformed Normalized Difference Vegetation Index Soil water is the key factor which affects crop growth when the radiation and temperature condition show little change. The crop grows better under good soil water condition than with water stress conditions. This method calculates the long-term decadal average vegetation index from remote sensed data, then it determines the growth condition of crops and soil water condition from the difference of the decadal average vegetation index and the long-term average. Kogan (1997) proposed two useful indices VCI and TCI to monitor crop water and temperature stress when drought occurs. VCI is the Vegetation Condition Index; it reflects the growth difference of crops among different years. VCI is expressed as VCI =

(NDV I − NDV Imin ) (NDV Imax − NDV Imin )

(15.11)

TCI is the Temperature Condition Index, which reflects the different responds of crops to temperature. TCI is expressed as TCI = (BTmax − BT ) (BTmax − BTmin )

(15.12)

in Eqs. (15.11) and (15.12), NDVImin (BTmin) and NDVImax (BTmax) refer to the absolute minimum and maximum NDVI (BT) measured for a given decade (or month) over a multi-year series of image data. NDVI (BT) refers to the current year NDVI (BT) for the same decade. BT is the brightness temperatures derived from channel 4 of NOAA /AVHRR. The VCI is an indicator of the status of the vegetation cover as a function of the NDVI minima and maxima encountered for a given ecosystem over many years. It normalizes the NDVI (or any other vegetation index) and allows for a comparison of different ecosystems. It is an attempt to separate the short-term climatic signal from the long-term ecological signal and in this sense it is a better indicator of water stress conditions than the NDVI. The TCI is an equivalent indicator based on the surface skin temperature derived from NOAA AVHRR data. Both, the VCI and the TCI, are dimensionless and vary between the values of 0 and 1. Zero indicates the worst condition ever encountered over the period of available images, one indicates the best condition encountered during the same period of time. If the period covered includes dry and wet years and under the assumption that the vegetation condition is mainly related to the water availability, these indicators have a high potential for monitoring water stress.

15.4.4 Vegetation Temperature Condition Index The vegetation temperature condition index is defined as: V TCI =

LSTNDVIi. max − LSTNDVIi LSTNDVIi. max − LSTNDVIi. min

(15.13)

412

Z. Chen et al.

where LSTNDVIi. max = a + bNDVIi LSTNDVIi. min = a + b NDVIi

(15.14)

where LSTNDV Ii. max and LSTNDV Ii. min are maximum and minimum land surface temperature (LSTs) of pixels which have same NDV Ii value in a study region, respectively, and LSTNDV Ii denotes LST of one pixel whose NDVI value is NDV Ii . Coefficient a, b, a and b can be estimated from an area large enough where soil moisture at the surface layer should span from wilting point to maximum field water-holding capacity at pixel level. In general, the coefficients are estimated from the scatter plot of LST and NDVI for the study area. The shape of the scatter plot is normally triangular at a regional scale (Gillies et al., 1997; Wang et al., 2001) if the study area is large enough to provide a wide range of NDVI and surface moisture conditions. VTCI is not only related to NDVI changes in the region, but also related to LST changes of pixels with a specific NDVI value. It can be physically explained as the ratio of temperature differences among the pixels. The numerator of Eq. (15.13) is the difference between maximum LST of the pixels and LST of one pixel, while the denominator of Eq. (15.13) is the difference between maximum and minimum LSTs of the pixels. LSTmax can be regarded as the “warm edge” where there is less soil moisture availability and plants are under dry conditions; LSTmin can be regarded as the “cold edge” where there is no water restriction for plant growth. The value of VTCI ranges from 0 to 1. The lower the value of VTCI, the higher the occurrence of drought.

15.4.5 Thermal Infrared Method There is an inherent relationship between soil moisture and soil temperature. The spatial distribution can be deduced through thermal images obtained by thermal infrared remote sensing. Experiments indicated that soil moisture of bare soil can be estimated by surface temperature, and can reach 50 cm depth (Myers and Heilman, 1969). Bartholic et al. (1972) found that the maximum temperature of the bare soil surface during the day decreases according to the increase of surface soil moisture. The spatial distribution of soil moisture can be estimated by the daytime temperature of the underlying surface on a sunny and windless day. It contains two main elements for soil moisture monitoring using infrared methods, one is the measurement of land surface temperature, and the other is the establishment of the relationship between soil moisture and the land surface temperature.

15.4.6 Microwave Remote Sensing The soil dielectric constant at microwave frequencies exhibits a strong dependence on the soil water content. For example, at L band, the real part of the dielectric constant ranges from 3 for dry soil to about 25 for saturated soil. This variation can

15 Monitoring and Management of Agriculture with Remote Sensing

413

result in a change in the order of 10 dB in the magnitude of the radar-backscatter coefficient (Oh et al., 1992), and of 100 K in the magnitude of the brightness temperature. According to the operation method, there are two kinds of microwave remote sensing systems, the passive system and the active system.

15.4.6.1 Passive Systems Research on the soil moisture content with microwave radiometers has been performed since the late 1970s, and has recently been revitalized by new missions: the already in orbit AMSR-E and AMSR, and the planned SMOS and AQUARIUS (Kerr et al., 2001; Le Vine et al., 2001). Various experimental researches on soil moisture with passive systems have been carried out and different algorithms were proposed. For example, large airborne experiments, called the Southern Great Plains Hydrology Experiments, were conducted in the USA in 1997 (SGP97) and 1999 (SGP99) to address significant gaps in our knowledge, and to validate retrieval algorithms designed for the AMSR and the AMSR-E. These approaches differ primarily in the methods used to correct for the effects of soil texture, roughness, vegetation, and surface temperature. Soil moisture content retrieval approaches that have been investigated in previous studies include: (a) Single-channel retrieval with sequential corrections using ancillary data (Jackson et al., 1993; Vinnikov et al., 1999); (b) iterative forward model corrections using multi-channel brightness temperatures (Njoku and Li, 1999); (c) correction using multi-frequency polarization indices (Paloscia et al., 2001); and (d) variations or combinations of the above methods (Owe et al., 2001; Jackson and Hsu, 2001). The thickness of the soil layer through which moisture can be directly estimated by means of a microwave radiometer has been investigated by many experimental studies. Most researchers have come to the conclusion that at L band this layer is about 5–10 cm. This result matches well with the requirements of those processes such as infiltration and evapotranspiration that take place within this first layer of the soil medium. In other applications, where soil moisture profiles down to several decimeters are necessary, microwaves must be coupled to appropriate hydrological models. Galantowicz et al. (1999) pointed out the effectiveness of the Kalman filter for retrieving such a quantity by using field observations and a simulation study. The usefulness of assimilating remotely sensed measurements into land-surface models was discussed in (Houser et al., 1998; Burke et al., 2001). Burke and Simmonds (2003) explored the potential for using low-resolution passive microwaves in a twodimensional Land Data Assimilation System (LDAS) for estimating deep-soil moisture from surface-soil moisture. Houser et al. (1998) investigated four-dimensional (4D) soil-moisture assimilation using in situ and remote sensing observations.

414

Z. Chen et al.

15.4.6.2 Active Systems The possibility of monitoring soil moisture changes using SAR data has stimulated a large number of studies focused on establishing a relationship between the observed SAR response and surface soil moisture content. For a homogeneous soil with a perfectly smooth surface, the scattering of electromagnetic waves is totally forward, and depends on permittivity of the medium. For a rough surface, radiation is scattered in various directions, and also generates backscattering. Thus, two basic properties determine the backscatter response observed by the SAR system: the permittivity of the medium and the roughness characteristics of the surface. Both parameters are, in turn, related to different geophysical parameters of the soil. With the advent of the polarimetric SAR, radar remote sensing of soil moisture has attained significant prominence in the past two decades. Initially, extensive experimental studies using polarimetric scatterometers were carried out to establish a relationship between radar response and the surface roughness and soil moisture (Sarabandi et al., 1992). Extensive field experiments have also been conducted to examine retrieval algorithms ranging from simple analytical to regression/empirical models (Oh et al., 1992; Dubois et al., 1995; Kim and van Zyl, 2002). Apart from surface-roughness parameters, the existence of short vegetation on the surface makes the retrieval of soil moisture content very complicated. Vegetation cover and its temporal variations are believed to be the major stumbling blocks in monitoring soil moisture content variations using microwaves. Chiu and Sarabandi (2000) proposed a very complicated coherent-scattering model, which accounts for scattering from rough surfaces, vegetation cover, and their near-field interaction. The inverse of this model was then used to demonstrate its ability for estimating the physical parameters of a soybean field, including soil moisture from a polarimetric set of AIRSAR images. The remote sensing monitoring of soil moisture is a complicated indirect process: many factors, such as the choice of bands and characteristics of the sensor are critical for the monitoring of soil moisture; the properties of the soil, the different landforms, climate change and soil management by human being’s all affect the soil moisture to different degrees, and the characteristics of the land surface are correlative to soil moisture in some degree. So, the improved thermal inertia and CWSI (Crop water stress index) method are the mature methods of remote sensing-based soil moisture monitoring methods. The extensive application of GIS technology, the optimization of models for the remote sensing monitoring of soil moisture and the application of microwaves are the key directions for future study. Currently, theories and methods on soil moisture remote sensing monitoring are relative mature. Most approaches such as crop indices, CWSI, thermal infrared and microwaves can get good results in soil moisture retrieval. But different approaches have their own applicability and limits. So, when using a single method in soil moisture retrieval there may be large errors. For example, the thermal inertia method is used in case the soil is bare; CWSI is used on crop covered surfaces, and for partly covered soil the improved thermal inertia method is applicable. To avoid large errors in soil moisture estimation, we should select different approaches to get good results.

15 Monitoring and Management of Agriculture with Remote Sensing

415

15.5 Summary The rapid development of remote sensing technology, including novel sensors and remote sensing data sets, quantitative inversion algorithms, and so on, has greatly benefited the application of remote sensing for agricultural monitoring and management. The intrinsic characteristics of agriculture make remote sensing an ideal technique for its monitoring and management. Meanwhile, the demands arising from the applications in agricultural sectors have also enhanced the progress and innovation in remote sensing technology. For crop identification and crop mapping, traditional unsupervised and supervised classification methods are still widely used. Increasingly, nonparametric classification algorithms are being used, which make no assumptions regarding the distribution of the data to be classified, because the frequently adopted normal distribution hypothesis is usually not true for remotely sensed data sets. Auxiliary data and expert knowledge are added in the process of cropland classification, which improve the overall classification accuracy. Multi-sensor data fusion and classification of time series data are applied in cropland monitoring more and more. For crop yield monitoring and prediction, different kinds of models are applied for different purposes. Empirical relationships for crop yield monitoring and prediction are widely used for their simplicity. But its shortcomings are also obvious: the unstableness and site-specific relationship between yield and remote sensing data. Physiology-based models are mainly based on the crop physiological functions, which is its strong point. But some parameters are not consistent over a large region and/or for different crops and sometimes not easily acquired by remote sensing or in situ survey. Crop growth models have a long history and are extensively used for crop growth monitoring, yield prediction and farm management around the world. The huge work of data collection and preparation hampered its good performance at a regional scale. In recent years, crop growth models with remote sensing data assimilation have been improved greatly at regional scales for better estimation of crop parameters. Crop phenology is important in crop monitoring because it can have great impact on the monitoring accuracies of crop yield and acreage change. Accurate monitoring of crop development patterns (i.e., phenology and growth) is an important component of farm management since it enables us to assess crop growth under various regional weather conditions. Soil moisture is an important parameter for crop monitoring and management. A lot of efforts have been made for accurate monitoring soil moisture or drought. Monitoring soil moisture and drought with remote sensing is based on two basic principles. The first one is that soil water content change leads to variations of soil spectral reflectance; the other is that soil water content change can cause physiological changes of a plant, thus changes the spectral characteristics of a leaf and affects the reflectance of a canopy. There are three categories of remote sensing systems used in soil moisture monitoring: optical (including reflective near-infrared), thermal and microwave systems. Optical systems acquire data from visible wavelengths up to, and including, near infrared. They measure the amount and wavelength of the

416

Z. Chen et al.

sun’s radiant energy reflected from a target object. The inclusion of wavelengths visible to the human eye often allows an intuitive interpretation of images, especially when combined with expert knowledge of soil properties. According to different approaches, microwave systems include passive systems, i.e., radiometers; and active systems, i.e., radar systems. Currently, theories and methods on soil moisture remote sensing monitoring are relative mature. Most approaches such as crop indices, CWSI, thermal infrared and microwaves can get good results in soil moisture retrieval. But different approaches have their own applicability and limits. So, if we use a single method in soil moisture retrieval there may be large errors. For example, the thermal inertia method is used in case the soil is bare; CWSI is used on crop covered surfaces, and for partly covered soil the improved thermal inertia method is applicable. To avoid large errors in soil moisture estimation, we should select different approaches to get good results. Acknowledgements This work is funded by a National Key Technologies R&D Program of China (No. 2006BAD10A06) and a National High Technology Research and Development Program of China (863 Program No. 2006AA12Z103).

References Allen R, Hanuschak G, Craig M (2002) History of remote sensing for crop acreage in USDA’s National Agricultural statistics Service. http://www.usda.gov/nass/nassinfo/remotehistory.htm Baban SMJ, Luke C (2000) Mapping agricultural land use using retrospective ground referenced data, Satellite sensor imagery and GIS. Int. J. Remote Sens. 21:1757–1762 Badhwar GD (1984) Classification of corn and soybeans using multitemporal thematic mapper data. Remote Sens. Environ. 16:175–181 Baret F, Guyot G (1991) Potentials and limits of vegetation indices for LAI and APAR assessment. Remote Sens. Environ. 35:161–173 Barnett TL, Thompson DR (1983) Large area relation of Landsat MSS and NOAA-6 AVHRR spectral data to wheat yields. Remote Sens. Environ. 13:277–290 Bartholic JE, Namken LN, Wiegand CL (1972) Aerial thermal scanner to determine temperatures of soils and crop canopies differing in water stress. Agron. J. 64:603–608 Bastiaanssen WGM, Ali S (2003) A new crop yield forecasting model based on satellite measurements applied across the Indus Basin, Pakistan. Agric. Ecosyst. Environ. 94:321–340 Bijleveld RA (1978) A combined surface temperature, soil moisture and evaporation mapping approach. 12th International Symposium on Remote Sensing of Environment, Manila, Philippines, 20–26 Blumberg DG, Freilikher V (2001) Soil water-content and surface roughness retrieval using ERS-2 SAR data in the Negev Desert, Israel. J. Arid Environ. 49:449–464 Boogaard HL, Eerens H, Supit I, van Diepen CA, Piccard I, Kempenseers P (2002) Description of the MARS Crop Yield Forecasting System (MCYFS). METAMP-1/3. Alterra and VITO, Wageningen and Mol Bouman B, Kropff M, Tuong T, Wopereis M, Berge H, Laar H (2002) ORYZA2000: modeling lowland rice. IRRI, Manila, Philippines Broge NH, Mortensen JV (2002) Deriving green crop area index and canopy chlorophyll density of winter wheat from spectral reflectance data. Remote Sens. Environ. 81:45–57 Burke EJ, Simmonds LP (2003) Effects of sub-pixel heterogeneity on the retrieval of soil moisture from passive microwave radiometry. Int. J. Remote Sens. 24:2085–2104

15 Monitoring and Management of Agriculture with Remote Sensing

417

Burke EJ, Shuttleworth WJ, Lee K, Bastidas LA (2001) Using area-average remotely sensed surface soil moisture in multipatch land data assimilation systems. IEEE Trans. Geosci. Remote Sens. 39:2091–2100 B¨uttner G, Csillag F (1989) Comparative study of crop and soil mapping using multitemporal and multispectral SPOT and landsat thematic mapper data. Remote Sens. Environ. 29:241–249 Carlson TN, Ripley DA (1997) On the relation between NDVI, fractional vegetation cover, and leaf area index. Remote Sens. Environ. 62:241–252 Carlson TN (1986) Regional-scale estimates of surface moisture availability and thermal inertia using remote thermal measurements. Remote Sens. Environ. 1:197–247 Chakraborty M, Panigrahy S, Sharma SA (1997) Discrimination of rice crop grown under different cultural practices using temporal ERS-1 synthetic aperture radar data. ISPRS J. Photogramm. Remote Sens. 52:183–191 Chen Z, Zhou Q, Tang H (2004) The future agricultural applications’ demanding on space-born remote sensing in China. In: H Tang (ed) Research on the Agricultural Resource Utilization and Regional Sustainable Development. China Outlook Press, Beijing, pp 236–250 Chiu TC, Sarabandi K (2000) Electromagnetic scattering from short branching vegetation. IEEE Trans. Geosci. Remote Sens. 38:911–925 Choudhury BJ, Schmugge TJ, Newton RW, Chang A (1979) Effect of surface roughness on the microwave emission from soils. J. Geophys. Res. 84:5699–5706 Clevers JGPW, van Leeuwen HJC (1996) Combined use of optical and microwave remote sensing data for crop growth monitoring. Remote Sens. Environ. 56:42–51 Cohen Y, Shoshany M (2002) A national knowledge-based crop recognition in Mediterranean environment. Int. J. Appl. Earth Observ. Geoinform. 4:75–87 Dadhwal VK, Ray SS (2000) Crop assessment using remote sensing-Part II: crop condition and yield assessment. Indian J. Agric. Econ. 55:55–67 Del Frate F, Ferrazzoli P, Schiavon G (2003) Retrieving soil moisture and agricultural variables by microwave radiometry using neural networks. Remote Sens. Environ. 84:174–183 Delecolle R, Maas SJ, Guerif M, Baret F (1992) Remote sensing and crop production models: present trends. ISPRS J. Photogramm. Remote Sens. 47:145–161 Digkuhn M, Gal P (1996) Effect of drainage date on yield and dry matter partitioning in irrigated rice. Field Crops Res. 46:117–126 Doraiswamy PC, Moulin S, Cook PW, Stern A (2003) Crop yield assessment from remote sensing. Photogramm. Eng. Remote Sens. 69:665–674 Doraiswamy PC, Sinclair TR, Hollinger S, Akhmedov B, Stern A, Prueger J (2005) Application of MODIS derived parameters for regional crop yield assessment. Remote Sens. Environ. 97:192–202 Dubois PC, van Zyl J, Engman T (1995) Measuring soil moisture with imaging radars. IEEE Trans. Geosci. Remote Sens. 33:915–926 Eck TF, Dye DG (1991) Sattellite estimation of incident photosynthetically active radiation using ultraviolet reflectance. Remote Sens. Environ. 38:135–146 Field CB, Randerson JT, Malmstrom CM (1995) Global net primary production: combining ecology and remote sensing. Remote Sens. Environ. 51:74–88 Fitzgerald GJ, Pinter J, Hunsaker DJ, Clarke TR (2005) Multiple shadow fractions in spectral mixture analysis of a cotton canopy. Remote Sens. Environ. 97:526–539 Fuller DO (1998) Trends in NDVI time series and their relation to rangeland and crop production in Senegal, 1987–1993. Int. J. Remote Sens. 19:2013–2018 Galantowicz JF, Entekhabi D, Njoku EG (1999) Tests of sequential data assimilation for retrieving profile soil moisture and temperature from observed L-Band radiobrightness. IEEE Trans. Geosci. Remote Sens. 37:1860–1870 Gillies RR, Carlson TN, Cui J, Kustas WP, Humes KS (1997) A verification of the ‘triangle’ method for obtaining surface soil water content and energy fluxes from remote measurement of the Normalized Difference Vegetation Index (NDVI) and surface radiant temperature. Int. J. Remote Sens. 18:3145–3166 Gitelson AA, Kaufman YJ (1998) MODIS NDVI optimization to fit the AVHRR data series– spectral considerations. Remote Sens. Environ. 66:343–350

418

Z. Chen et al.

Glenna NF, Carr JR (2003) The use of geostatistics in relating soil moisture to RADARSAT-1 SAR data obtained over the Great Basin, Nevada, USA. Computers and Geosciences 29:577–586 Goldberg B, Klein WH (1980) A model for determination the spectral quality of daylight on a horizontal surface at any geographical location. Solar Energy 24:351–357 Groten SME (1993) NDVI – crop monitoring and early yield assessment of Burkina Faso. Int. J. Remote Sens. 14:1495–1515 Gu S, Gao H, Zhu Y, Zhao B, Lu N, Zhang W (2002) Remote sensing land surface wetness by use of TRMM/TMI microwave data. Meteorol. Atmos. Phys. 80:59–63 Gu´erif M, Duke CL (2000) Adjustment procedures of a crop model to the site specific characteristics of soil and crop using remote sensing data assimilation. Agric. Ecosyst. Environ. 81:57–69 Guerif M, de Brisis S, Seguin B (1993) Combined NOAA-AVHRR and SPOT-HRV data for assessing crop yields of semiarid environments. EARsel Adv. Remote Sens. 2:110–123 Hanan NP, Prince SD, Begue A (1995) Estimation of absorbed photosynthetically active radiation and vegetation net production efficiency using satellite data. Agric. Forest Meteorol. 76:259–276 Hochheim KP, Barber DG (1998) Spring wheat yield estimation for Western Canada using NOAA NDVI data. Can. J. Remote Sens. 24:17–27 Houser PR, Shuttleworth WJ, Famiglietti JS, Gupta HV, Syed KH, Goodrich DC (1998) Integration of soil moisture remote sensing and hydrological modeling using data assimilation. Water Resour. Res. 34:3405–3420 Hu J, Tan B, Shabanov N, Crean KA, Martonchik JV, Diner DJ, Knyazikhin Y, Myneni RB (2003) Performance of the MISR LAI and FPAR algorithm: a case study in Africa. Remote Sens. Environ. 88:324–340 Huang J, Tang S, Ousama AI, Wang R (2002) Rice yield estimation using remote sensing and simulation model. J. Zhejinag Univ. (Science Edition) 3:461–466 Jackson RD, Idso SB, Reginato RJ, Pinter PJ (1981) Canopy temperature as a crop water stress indicator. Water Resour. Res. 17:1133–1138 Jackson TJ, Hsu AY (2001) Soil moisture and TRMM microwave imager relationships in the Southern Great Plains 1999 (SGP99) Experiment. IEEE Trans. Geosci. Remote Sens. 39:1632–1642 Jackson TJ, Le Vine DM, Griffis AJ, Goodrich DC, Schmugge TJ, Swift CT, O’Neill PE (1993) Soil moisture and rainfall estimation over a semiarid environment with the ESTAR microwave radiometer. IEEE Trans. Geosci. Remote Sens. 31:836–841 Jakubauskas ME, Legates DR, Kastens JH (2002) Crop identification using harmonic analysis of time-series AVHRR NDVI data. Comput. Electron. Agric. 37:127–139 Jiao X, Yang B, Pei Z, Wang F (2005) Monitoring crop yield using NOAA/AVHRR-based vegetation indices. Trans. Chin. Soc. Agric. Eng. 21:104–108 Jonsson P, Eklundh L (2002) Seasonality extraction by function fitting to time-series of satellite sensor data. IEEE Trans. Geosci. Remote Sens. 40:1824–1832 Kalubarme MH, Mahey RK, Dhaliwal SS, Sidhu SS, Singh R, Mahajan A, Sharma PK (1995) Agromet-Spectral wheat yield modeling in Punjab. Proceedings of National Symposium on Remote Sensing of Environment with special emphasis on green revolution, pp 11–17 Kerr YH, Waldteufel P, Wigneron JP, Martinuzzi JM, Font J, Berger M (2001) Soil moisture retrieval from space: the Soil Moisture and Ocean Salinity (SMOS) mission. IEEE Trans. Geosci. Remote Sens. 39:1729–1735 Kim Y, van Zyl J (2002) On the relationship between polarimetric parameters and soil moisture. Proceedings of the IEEE 2002 International Geoscience and Remote Sensing Symposium (IGARSS 2002), Toronto, June 2002 Kimball J, McDonald K, Running S, Frolking S (2004) Satellite radar remote sensing of seasonal growing seasons for boreal and subalpine evergreen forests. Remote Sens. Environ. 90:243–258 Kogan FN (1997) Global drought watch from space. Bull. Am. Meteorol. Soc. 78:621–636 Kressler FP, Steinnocher K (1999) Detecting land cover changes from NOAA-AVHRR data using spectral mixture analysis. Int. J. Appl. Earth Observ. Geoinform. 1:21–26

15 Monitoring and Management of Agriculture with Remote Sensing

419

Le Vine D, Koblinsky C, Pellerano F (2001) The measurement of salinity from space: sensor concept, IEEE International Geoscience and Remote Sensing Symposium, Sidney, Australia, July 9–13, 2001, pp 1010–1012 Li J, Islam S (1999) On the estimation of soil moisture profile and surface fluxes partitioning from sequential assimilation of surface layer soil moisture. J. Hydrol. 220:86–103 Lobell DB, Asner GP (2004) Cropland distributions from temporal unmixing of MODIS data. Remote Sens. Environ. 93:412–422 Loveland TR, Zhu Z, Ohlen DO, Brown JF, Reed BC, Yang L (1999) An analysis of the IGBP global land-cover characterization process. Photogramm. Eng. Remote Sens. 65:1021–1032 Maas SJ (1988) Using satellite data to improve model estimates of crop yield. Agron. J. 80:655–662 McIver DK, Friedl MA (2002) Using prior probabilities in decision-tree classification of remotely sensed data. Remote Sens. Environ. 81:253–261 Mkhabela MS, Mashinini NN (2005) Early maize yield forecasting in the four agro-ecological regions of Swaziland using NDVI data derived from NOAA’s-AVHRR. Agric. Forest Meteorol. 129:1–9 Monteith JL (1972) Solar radiation and productivity in tropical ecosystems. J. Appl. Ecol. 9:747–766 Moody A, Johnson DM (2001) Land-surface phonologies from AVHRR using the discrete fourier transform. Remote Sens. Environ. 75:305–323 Moshou D, Vrindts E, De Ketelaere B, De Baerdemaeker J, Ramon H (2001) A neural network based plant classifier. Comput. Electron. Agric. 31:5–16 Myers VI, Heilman HD (1969) Thermal IR for soil temperature studies. Photogramm. Eng. Remote Sens. 35:1024–1032 Myneni RB, Knyazikhin Y, Privette JL, Glassy J (2002) Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data. Remote Sens. Environ. 83:214–231 Myneni RB, Keeling CD, Tucker CJ, Asrar G, Nemani RR (1997) Increased plant growth in the northern high latitudes from 1981–1991. Nature 386:698–702 Njoku EG, Li L (1999) Retrieval of land surface parameters using passive microwave measurements at 6–18 GHz. IEEE Trans. Geosci. Remote Sens. 37:79–93 Njoku EG, O’Neill PE (1982) Multifrequency microwave radiometer measurements of soil moisture. IEEE Trans. Geosci. Remote Sens. 20:468–475 Oh Y, Sarabandi K, Ulaby FT (1992) An empirical model and an inversion technique for radar scattering from bare soil surfaces. IEEE Trans. Geosci. Remote Sens. 30:370–381 Owe M (1988) Estimating surface soil moisture from satellite microwave measurement and a satellite derived vegetation index. Remote Sens. Environ. 24:331–345 Owe M, de Jeu R, Walker J (2001) A methodology for surface soil moisture and vegetation optical depth retrieval using the microwave polarization difference index. IEEE Trans. Geosci. Remote Sens. 39:1643–1654 Pal M, Mather PM (2003) An assessment of the effectiveness of decision tree methods for land cover classification. Remote Sens. Environ. 86:554–565 Paloscia S, Macelloni G, Santi E, Koike T (2001) A multi-frequency algorithm for the retrieval of soil moisture on a large scale using microwave data from SMMR and SSM/I satellites. IEEE Trans. Geosci. Remote Sens. 39:1655–1661 Park SJ, Hwang CS, Vlek PLG (2005) Comparison of adaptive techniques to predict crop yield response under varying soil and land management conditions. Agric. Syst. 85:59–81 Potdar MB (1993) Sorghum yield modelling based on crop growth parameters determined from visible and near-IR channel NOAA AVHRR data. Int. J. Remote Sens. 14:895–905 Pradhan S (2001) Crop area estimation using GIS, remote sensing and area frame sampling. International J. Appl. Earth Observ. Geoinform. 3:86–92 Pratt A, Ellyett CD (1979) The thermal inertia approach to mapping of soil moisture and geology. Remote Sens. Environ. 8:51–168 Price JC (1985) On the analysis of thermal infrared imagery. The limited utility of apparent thermal inertia. Remote Sens. Environ. 18:59–73

420

Z. Chen et al.

Reed BC, Brown JF, van der Zee D, Loveland TL, Merchant JW, Ohlen DO (1994) Measuring phenological variability from satellite imagery. J. Veg. Sci. 5:703–714 Sakamoto S, Yokozawa M, Toritani H, Shibayama M, Ishitsuka N, Ohno H (2005) A crop phenology detecting method using time-series MODIS data. Remote Sens. Environ. 96:366–374 Sarabandi K, Oh Y, Ulaby FT (1992) Measurement and calibration of differential mueller matrix of distributed targets. IEEE Trans. Antennas Propag. 40(12):1524–1532 Schmugge TJ, Gloersen P, Wilheit T, Geiger F (1974) Remote sensing of soil moisture with microwave radiometers. J. Geophys. Res. 79:317–323 Schmugge TJ, O’Neill PE, Wang JR (1986) Passive microwave soil moisture research. IEEE Trans. Geosci. Remote Sens. 24:12–20 Shabanov NV, Wang Y, Buermann W, Dong J, Hoffman S, Smith GR, Tian Y, Knyazikhin Y, Myneni RB (2003) Effect of foliage spatial heterogeneity in the MODIS LAI and FPAR algorithm over broadleaf forests. Remote Sens. Environ. 85:410–423 Shao Y, Fan X, Liu H, Xiao J, Ross S, Brisco B, Brown R, Staples G (2001) Rice monitoring and production estimation using multitemporal RADARSAT. Remote Sens. Environ. 76:310–325 Shimabukuro YE, Smith JA (1991) The least squares mixing models to generate fraction images derived from remote sensing multispectral data. IEEE Trans. Geosci. Remote Sens. 29:16–20 Soares JV, Renno CD, Formaggio AR, da Costa Freitas Yanasse C, Frery AC (1997) An investigation of the selection of texture features for crop discrimination using SAR imagery. Remote Sens. Environ. 59:234–247 Spanner MA, Pierce LL, Peterson DL, Running SW (1990) Remote sensing of temperate coniferous forest leaf area index. The influence of canopy closure, understory vegetation and background reflectance. Int. J. Remote Sens. 11:95–111 Sridhar VN, Dadhwal VK, Chaudhary KN, Sharma R, Bairagi GD, Sharma AK (1994) Wheat production forecasting for a predominantly unirrigated region of Madhya Pradesh (India). Int. J. Remote Sens. 15:1307–1316 Theau J, Peddle DR, Duguay CR (2005) Mapping lichen in a caribou habitat of Northern Quebec, Canada, using an enhancement of classification method and spectral mixture analysis. Remote Sens. Environ. 94:232–243 Thenkabail PS, Smith RB, De Pauw E (2000) Hyperspectral vegetation indices and their relationships with agricultural crop characteristics. Remote Sens. Environ. 71:158–182 Tompkins S, Mustard JF, Pieters CM, Forsyth DW (1997) Optimization of endmembers for spectral mixture analysis. Remote Sens. Environ. 59:472–489 Tucker CJ, Holden BN, Elgin GH Jr, Mcmurtrey E (1980) Remote sensing of total dry matter accumulation in winter wheat. Remote Sens. Environ. 11:267–277 Van Niel TG, McVicar TR (2004) Current and potential uses of optical remote sensing in rice-based irrigating systems: a review. Aus. J. Agric. Res. 55:155–185 Vina A, Gitelson AA, Rundqist DC, Keydan G, Leavit B, Schepers J (2004) Monitoring maize (Zea mays L.) phenology with remote sensing. Agron. J. 96:1139–1147 Vinnikov K, Robock A, Qiu S, Entin J, Owe M, Choudhury B, Hollinger S, Njoku E (1999) Satellite remote sensing of soil moisture in Illinois, USA. J. Geophys. Res. 104:4145–4168 Viovy N, Arino O, Belward A (1992) The Best Index Slope Extraction (BISE): a method for reducing noise in NDVI time-series. Int. J. Remote Sens. 13:1585–1590 Wang JR (1985) Effect of vegetation on soil moisture sensing observed from orbiting microwave radiometer. Remote Sens. Environ. 17:141–151 Wang N (1996) Wheat area and yield monitoring with remote sensing in China. China Science Press, Beijing Wang P, Li X, Gong J, Song C (2001) Vegetation temperature condition index and its application for drought monitoring. Proceedings of International Geoscience and Remote Sensing Symposium, July 9–14, 2001, Sydney, Australia. IEEE, Piscataway, NJ, pp 141–143 Watson K, Rowen LC, Offield TW (1971) Application of thermal modeling in the geologic interpretation of IR images. Remote Sens. Environ. 3:2017–2041 Weissteiner CJ, Braun M, Kuhbauch W (2004) Regional yield predictions of malting barley by remote sensing and ancillary data. Proceedings of SPIE – The International Society for Optical Engineering, Barcelona, Spain, vol 5232, pp 528–539

15 Monitoring and Management of Agriculture with Remote Sensing

421

Wessel KJ, De Fries RS, Dempewolf J, Anderson LO, Hansen AJ, Powell SL, Moran EF (2004) Mapping regional land cover with MODIS data for biological conservation: examples from the Greater Yellowstone Ecosystem, USA and Para State, Brazil. Remote Sens. Environ. 92:67–83 White MA, Thornton PE, Running SW (1997) A continental phenology model for monitoring vegetation responses to interannual climatic variability. Global Biogeochem. Cycles 11:217–234 Wickel AJ, Jackson TJ (2001) Multitemporal monitoring of soil moisture with RADARSAT SAR during the 1997 Southern Great Plains hydrology experiment. Int. J. Remote Sens. 22:1571–1583 Wu C (2004) Normalized spectral mixture analysis for monitoring urban composition using ETM + imagery. Remote Sens. Environ. 93:480–492 Xiao X, Boles S, Frolking S, Li C, Salas W, Berrien Moore III (2005) Mapping paddy rice agriculture in South and Southeast Asia using multi-temporal MODIS images. Remote Sens. Environ. 100:95–113 Xiao X, Boles S, Frolking S, Salas W, Moore BI, Li C (2002) Observation of flooding and rice transplanting of paddy rice fields at the site to landscape scales in China using VEGETATION sensor data. Int. J. Remote Sens. 23:3009–3022 Yang P (2005) Linking Multi-temporal Remotely Sensed Data, Field Observations and GIS-based Crop Growth Model to Estimate Winter Wheat Yield in North China Plain. Doctoral Degree Thesis of University of Tokyo, Tokyo. 104pp Zhang X, Friedl MA, Schaaf CB, Strahler AH, Hodges JCF, Gao F, Reed BC, Huete A (2003) Monitoring vegetation phenology using MODIS. Remote Sens. Environ. 84:471–475

Chapter 16

Remote Sensing of Terrestrial Primary Production and Carbon Cycle Maosheng Zhao and Steven W. Running

Abstract The objective of this chapter is to review the historical development of and the recent advances in the application of satellite remote sensing data for estimating terrestrial gross and net primary production (GPP and NPP), while also monitoring carbon cycle related ecosystem dynamics and changes. We achieve this objective by separating the topic into five sections: 1. A review of the history of using satellite data to study the carbon cycle, concentrating on using the Normalized Difference Vegetation Index (NDVI) and its derived Fraction of Photosynthetically Active Radiation (FPAR) and Leaf Area Index (LAI) for biomass and NPP estimations 2. A description of recent advances in the application of Moderate Resolution Imaging Spectroradiometer (MODIS) data to estimates of GPP and NPP, along with related findings using MODIS Land Surface Temperature (LST) and the Enhanced Vegetation Index (EVI) 3. A discussion of the application of long-term satellite data to the study of terrestrial ecosystems, including phenology monitoring, changes in regional carbon storage resulting from land use change, carbon flux changes induced by climate change, disturbance detection, and validation of ecosystem models 4. A proposed general scheme for applying satellite data to terrestrial ecosystem studies, highlighting the role of modeling 5. A summary that emphasizes the continuity of vegetation monitoring with satellites The use of remote sensing information for studying terrestrial primary production and the global carbon cycle is significant both for an increased understanding of the earth system and improved management of land and natural resources.

Maosheng Zhao and Steven W. Running Numerical Terradynamic Simulation Group, Department of Ecosystem and Conservation Science, University of Montana, Missoula, USA [email protected] S. Liang (ed.), Advances in Land Remote Sensing, 423–444. c Springer Science + Business Media B.V., 2008

423

424

M. Zhao, S.W. Running

16.1 Introduction Terrestrial Net Primary Production (NPP), the difference between Gross Primary Production (GPP) and plant autotrophic respiration (Ra), is the net carbon fixed by vegetation through photosynthesis. Quantification of NPP has socioeconomic significance, since NPP can directly measure the quantity of goods (e.g., food, fuel and fiber) provided to human beings by ecosystems (Imhoff et al. 2004). NPP provides the carbon required for maintenance of the structure and functions of an ecosystem. The current climate change caused primarily by increasing anthropogenic greenhouse gas emissions, especially CO2 , is the largest global environmental issue facing the world (IPCC, 2001), and there is now ample evidence of the ecological impacts of recent climate change (Walther et al., 2002). Climate and terrestrial ecosystems interact with and influence each other. On one hand, climate change and increasing CO2 can cause changes in NPP and carbon storage in ecosystems (Nemani et al., 2003a; Prentice et al., 2001), impacting the well-being of humans (Milesi et al., 2005). On the other hand, terrestrial ecosystems can affect climate through carbon, water, and energy exchange. For example, terrestrial ecosystems and oceans equally absorbed nearly half of CO2 emission by human activities in the 1990s (Prentice et al., 2001). Understanding the response of terrestrial NPP and the carbon cycle to climate change is therefore critical for predicting future environmental change and mitigating the impacts. Satellite remote sensing data provide us with invaluable continuous temporal and spatial information, which help us understand the processes, dynamics, and disturbances (such as land use change, wildfires, and insect outbreaks) in the biosphere, and the impacts of environmental changes on terrestrial ecosystems. Since the application of AVHRR data to the study of vegetation in the early 1980s, great progress has been made in the study of terrestrial carbon storage and fluxes, especially from the NASA Earth Observing System (EOS) program. The NASA EOS program has been planning and executing satellite-based earth monitoring for 15 years, and is the heart of global change science for the USA. The central sensor on board the Terra Satellite Platform is the Moderate Resolution Imaging Spectroradiometer (MODIS). Terra was successfully launched on December 18, 1999, and the second MODISbased satellite, Aqua, was launched on May 4, 2002. For the first time in history, we are able to obtain near-real time global vegetation growth status, including primary production, at an 8-day time interval with 1 km spatial resolution (Justice et al., 2002; Running et al., 2004). This chapter concentrates on applying satellite data to terrestrial primary production and carbon cycle studies. We (1) review the history of using satellite data to study the carbon cycle; (2) describe recent advances in the applications of MODIS data; (3) discuss monitoring terrestrial ecosystems with the long-term satellite data records; (4) propose a general scheme of applying the satellite data to terrestrial ecosystem studies, highlighting the role of modeling; and (5) summarize the chapter by emphasizing the continuity of vegetation monitoring with satellites. We focus on the optical sensors on-board polar-orbiting satellites, especially on those with medium to coarse resolution instruments. Radar and Lidar (light detection and

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

425

ranging), which can also provide valuable vegetation information from microwaves, are not covered in this chapter. In addition, land cover classification, an important aspect of carbon relevant application, is detailed elsewhere (Chapter 13).

16.2 History of the Development of Using NDVI 16.2.1 Related NDVI to Green Biomass for Vegetation Monitoring The basis for remote sensing of vegetation is the sharp contrast in reflectivity of visible (0.4–0.7 µm) and near-infrared (0.7–1.3 µm, NIR) spectra caused by the optical properties of chlorophyll and the internal structure of green leaf cells. The most widely used mathematic formula combining red and NIR reflectance (ρRED and ρNIR ) is the Normalized Difference Vegetation Index (NDVI), NDV I =

ρNIR − ρRED ρNIR + ρRED

(16.1)

Tucker (1979) evaluated several proposed vegetation indices and found that NDVI is strongly related to biomass, and in 1980, his lab experiments also revealed that NDVI is most strongly related to the percentage of green/brown herbage, suggesting the potential of using satellite data to estimate dry green biomass (Tucker, 1980). The first reported study for using NDVI from satellites (AVHRR / NOAA) to quantify grass production at the regional scale was published in 1983 by Tucker et al. (1983). They first constructed the relationship between AVHRR NDVI and clipped grass biomass on the ground for a limited number of sites, extrapolating the relationship to the study region using NDVI to estimate spatial patterns of biomass. Today, similar studies still employ this method, sometimes using relatively higher resolution satellite data (e.g., 30 m TM/ETM+, or 15 m ASTER) as an intermediate to scale up from field observations to coarse resolution satellite data to enhance the accuracy of estimations (e.g., Reeves et al., 2006). The first use of multi-temporal AVHRR NDVI to monitor the dynamics of vegetation at the continental and global scales was in 1985. Tucker et al. (1985) found the differences in temporal dynamic of vegetation reflected by NDVI are associated with variations in climate and dependent on biome types, and that the integrated NDVI over a given time interval is related to NPP. Justice et al. (1985) expanded this study to the global scale, suggesting that AVHRR NDVI is effective for monitoring phenology of global vegetation. Townshend and Justice (1986) further analyzed NDVI from different years in Africa and found inter-annual variation in NDVI can reveal the response of vegetation to climate anomalies.

426

M. Zhao, S.W. Running

16.2.2 Related NDVI to FPAR, LAI and NPP for Carbon Cycle Study Goward et al. (1985) first demonstrated the linear relationship between the growing season integrated NDVI and ground-based observations of NPP for different biomes over North America. Tucker et al. (1986) found a strong relationship between seasonal variations in atmospheric CO2 and NDVI. Both studies suggested that NDVI can be used to estimate terrestrial photosynthesis, and therefore NPP. In the following year, Fung et al. (1987) was the first to relate NDVI to annual NPP at the global scale to study atmosphere-biosphere exchange of CO2 . Fung et al.’s (1987) study simply distributed annual NPP into monthly level based on the monotonic function of monthly NDVI, without considering solar radiation or environmental stresses. However, Running and Nemani (1988) examined the seasonal relationship between photosynthesis estimated by a process-based model, FOREST-BGC, and NDVI/AVHRR for seven sites in North America representing a wide range of annual climates, and found that NDVI can not reflect the drawdown of photosynthesis induced by summer drought for water stressed sites, implying that NDVI alone could not fully represent seasonal photosynthesis. Along with the above studies to directly relate NDVI to NPP and the carbon cycle, additional studies were conducted to develop the physiological linkage between NDVI and NPP. Professor Monteith was the pioneering scientist who proposed the concept of photosynthesis efficiency logic. Monteith (1972, 1977) found that crop production under non-stressed conditions is linearly related to the amount of photosynthetically active radiation solar radiation (PAR) that is absorbed by green leaf (APAR). Kumar and Monteith (1982) decomposed the linear model into independent parameters such as incoming solar radiation, radiation absorption efficiency, and conversion efficiency of APAR, while also demonstrating how the fraction of APAR (FPAR) is related to the ratio of red reflectance to NIR. Asrar et al. (1984) found that NDVI can be simultaneously used to estimate both FPAR and leaf area index (LAI), FPAR is a linear function of NDVI, and LAI is a curvilinear function of NDVI, while Sellers (1985, 1987) also found a linear relationship between NDVI and FPAR. Running et al. (1989) first demonstrated a prototype for incorporation of the LAI derived from AVHRR NDVI into a process-based ecosystem model to simulate regional forest evapotranspiration and photosynthesis. While numerous empirical studies have found a strong relationship between NDVI and ground LAI and FPAR, the first theoretical interpretation of this was provided by Myneni et al. (1995), stating that vegetation indices are a measure of chlorophyll abundance and energy absorption. Further, Myneni et al. (1997a) proposed using a canopy radiation transfer model to derive FPAR and LAI. Models calculating regional NPP based on Monteith’s photosynthesis efficiency logic combined with FPAR estimated from remotely sensed NDVI were proposed after 1990. Prince (1991) first proposed a model of regional primary production for use with coarse resolution satellite data, accounting for the physiological costs of maintenance and growth respirations, and the environmental constraints tending to

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

427

reduce maximum light use efficiency (ε ). Prince and Goward (1995) later named the model the Global Production Efficiency model (GLO-PEM), and simulate global GPP and NPP at 8 km resolution. The first global NPP images estimated using satellite NDVI were generated with the Carnegie–Ames–Standford approach (CASA) model by Potter et al. (1993), which had a spatial resolution of one degree. Running and Hunt (1993) proposed a NDVI-based NPP model, also exploring the range and variability of ε using a mechanistic model. Ruimy et al. (1994, 1996) proposed similar models of NPP from AVHRR NDVI, but their models had no constraints on potential maximum ε resulting from environmental stresses. In addition, a forest stand growth model, 3-PG (Use of Physiological Principles in Prediction Growth) can use satellite FPAR as an input, and was developed by Landsberg and Waring (1997). Though there are some differences among the different models, Photosynthetic Efficiency Models (PEMs) can be expressed generally as, P = ∑ PAR∗ FPAR∗ εm∗ f (T )∗ F(W )

(16.2)

t

where P is GPP or NPP over a given time interval t, FPAR is derived using vegetation indices, εm is the maximum ε , and f (T ) and f (W ) are the constraints resulting from temperature and water stress. In general, water stress results from soil moisture and air vapor pressure deficit (VPD).

16.3 Advances for the MODIS Sensor 16.3.1 MODIS GPP and NPP Products MODIS may be so far the most complex instrument built and flown on a spacecraft for civilian research purposes (Guenther et al., 2002). The MODIS sensor provides higher quality data for monitoring terrestrial vegetation and other land processes than previous AVHRR, not only because of its narrower spectral bands that enhance the information derived from vegetation (Justice et al., 2002), onboard calibration to guarantee the consistent time-series reflectance (Guenther et al., 2002), and orbit and altitude satellite maneuvers to ensure sub-pixel geolocation accuracy (Wolfe et al., 2002), but also because leading scientists are working as a team to improve the accuracy of the data from low level reflectance data, to high level data, such as land cover, fire, land surface temperature, vegetation indices (NDVI and EVI; EVI is the enhanced vegetation index), FAPR / LAI and GPP / NPP (Justice et al., 2002). The other important feature of MODIS data is that all of the land products have quality flags, denoting any negative atmospheric impacts (e.g., cloudiness and aerosol) to help users screen the data for their purposes (Justice et al., 2002). Combined with the MODIS atmosphere and ocean products, MODIS vegetation data provide invaluable information for earth system study.

428

M. Zhao, S.W. Running

The algorithm for the MODIS 1 km 8-day GPP and annual NPP employs 1 km land cover and 1 km 8-day FPAR/LAI in addition to daily spatial coarse resolution meteorological data. The LAI is used to estimate the biomass of leaf, fine root and live wood for plant maintenance respiration calculations in the algorithm (Running et al., 2000; Heinsch et al., 2003; Running et al., 2004; Zhao et al., 2005). Because of the near real-time processing of MODIS GPP/NPP products and the required suite of large datasets, water stress is represented solely by air VPD rather than a full water stress from both VPD and soil moisture, thereby avoiding extremely intensive computation and the creation of the additional physical and biophysical datasets required for water balance calculation. Our study has shown that VPD alone can capture the inter-annual variability of the full water stress, although it may fail to capture the seasonality of water stresses for dry areas dominated by strong monsoons (Mu et al., 2007). The quality flags in these MODIS land data products allow us to fill the gaps in the time series of FPAR and LAI resulting from contamination by unfavorable atmospheric conditions, generating more accurate estimations of GPP and NPP (Zhao et al., 2005). For the first time, we have more than 6 years of consistent global 1 km GPP and NPP data products estimated from MODIS. Figure 16.1 shows the averaged seasonality of MODIS GPP from 6-year (2000–2005) results, in which we have aggregated 8-day values to 3 month averages (Fig. 16.1a). Spatial seasonal variations clearly

Fig. 16.1 Spatial patterns of the seasonality of MODIS GPP (a), and the mean annual cycle of GPP at an 8-day interval for four latitude bands and the globe (b). (From Zhao et al., 2006b.)

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

429

demonstrate the expected peak GPP in June, July, and August and the low values in December, January, and February over the mid- and high-latitudes of the northern hemisphere (NH). The rainforests of the Amazon Basin have higher GPP during the dry season from July to November than during the wet season, which agrees with the studies by Huete et al. (2006) and Xiao et al. (2006). Monthly precipitation in the Amazon Basin can reach approximately 100 mm in the dry season, making solar radiation, not water, the leading limiting factor in this region, limiting growth in the wet season. Figure 16.1b shows the annual cycle of total GPP for four latitudinal bands and for the entire globe. Relatively strong seasonal signals occur for the mid- and high-latitudes of the NH (i.e., north of 22.5 ◦ N). The areas south of 22.5 ◦ S have the opposite seasonal profile relative to the mid- and high-latitudes of NH, and the seasonality for the southern hemisphere is much weaker because there is significantly less land mass. For the entire tropical region (22.5 ◦ S–22.5 ◦ N), there is almost no discernible seasonality, and total GPP is always the highest among the four latitudinal bands. Therefore, at the global scale, the magnitudes of annual GPP cycle can be mostly attributed to the tropical region, while the seasonality in the global cycle is largely determined by areas north of 22.5 ◦ N. Figure 16.2 reveals the spatial pattern of the 6-year mean annual total NPP. As expected, high MODIS NPP occurs in areas covered by forests and woody savannas, especially in tropical regions. Low NPP occurs in areas dominated by harsh environments, such as high latitudes with short growing seasons constrained by low temperatures and daylength, and dry areas with limited water availability. At the global scale, from 2000 to 2005, MODIS estimated a total terrestrial annual GPP of 109.07 Pg C (std. 1.66), and an annual NPP of 52.03 Pg C (std. 1.17), ignoring

Fig. 16.2 Spatial patterns of global terrestrial MODIS NPP averaged over 6 years (2000–2005). (From Zhao et al., 2006b.)

430

M. Zhao, S.W. Running

Fig. 16.3 Interannual variations in global total C4.8 MOD17 NPP driven by NCEP and GMAO, respectively, as compared with the inverted atmospheric CO2 growth rate. A Multivariate ENSO Index (MEI) is shown in gray scale, where darker shades represent higher MEI values. (From Zhao et al., 2006b.)

barren land cover as defined by the MODIS land cover product. For vegetated areas, the mean annual GPP and NPP are 996.03 (std. 823.67) and 475.19 (std. 375.44) g C m−2 year−1 , respectively. Interannual anomalies in the MODIS global NPP record correlate well with the inverted atmospheric CO2 growth rate, corresponding to results (correlation = 0.70, P < 0.001) found by Nemani et al. (2003a) for the AVHRR period of record (1982–1999). Figure 16.3 shows the relationship between MODIS NPP anomalies and CO2 growth rates. To account for the uncertainties from inputs from different meteorological datasets, Zhao et al. (2006a) used both GMAO and NCEP meteorology to drive the MODIS GPP and NPP algorithm. For the 6-year MODIS record, the correlations are 0.85 (P < 0.016) and 0.91 (P < 0.006) using GMAO and NCEP, respectively (Zhao et al., 2006b), implying that NPP is the primary driver of the atmospheric CO2 growth rate.

16.3.2 Recent Findings Using MODIS Vegetation Products There were no standard global land surface products (temperature (LST), EVI, fire, GPP and NPP) prior to 2000 when MODIS began providing data, but information from these 1 km or sub 1 km data products help us understand the dynamics of terrestrial ecosystems and detect the underlying mechanisms at different levels ranging from local to regional scales. MODIS LST and EVI data can provide information on surface resistance and disturbance (Nemani and Running, 1989; Mildrexler et al.,

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

431

2007), and EVI has been found to be strongly related to the GPP derived from eddy covariance flux tower measurements (Xiao et al., 2004; Rahman et al., 2005; Huete et al., 2006), indicating the potentially superior utility of EVI in comparison to NDVI. However, Daniel A. Sims (personal communication on 3/2/06) has found that EVI models break down for sites subjected to summer drought as indicated by high summer VPD. This finding is consistent with the similar conclusion by Running and Nemani (1988) using NDVI, implying that neither EVI nor NDVI can reflect the total reduction in carbon uptake resulting from water stress, especially for evergreen ecosystems in dry areas. Hence, water stresses must be incorporated into global remote sensing primary production models. For Amazon rainforests, MODIS EVI has revealed enhanced vegetation growth during dry seasons (Huete et al., 2006; Xiao et al., 2006), and the MODIS GPP, incorporating such stresses, also shows this enhanced production during dry seasons (Fig. 16.1a).

16.4 Monitoring the Long-Term Dynamics of Ecosystems and Carbon Cycle The advantages of using satellites to monitor land are numerous. Not only can satellite data provide detailed spatial patterns and variations in ecosystem processes, but they also provide information on the temporal changes. With the accumulated longterm satellite data from Landsat since 1972 and AVHRR/NOAA since 1981, there are more than three decades of satellite data, enabling us to study dynamics of and changes in terrestrial ecosystems. The following subsections present several important applications of long-term satellite data.

16.4.1 Long Term Phenology Monitoring Vegetation activities at mid- and high-latitudes are largely controlled by short growing seasons resulting from temperature and daylength constraints. Recent warming (1976–2000) has been greater over the continents of the northern hemisphere (NH) than elsewhere (Folland et al., 2001), lengthening the growing season and stimulating vegetation growth. CO2 measurements have shown an advance of upto 7 days in the timing of the drawdown of CO2 in spring and early summer since the 1960s (Keeling et al., 1996). However, CO2 data alone can not depict detailed spatial information on the response of vegetation to the warming temperature. Though 3D transfer models can retrieve some spatial information on carbon sinks and sources, the spatial resolution is very coarse and the results contain relatively large uncertainties (Fan et al., 1998). Long-term AVHRR NDVI data allow us to detect vegetation responses to the warming climate spatio-temporally. Myneni et al. (1997b) utilized the AVHRR NDVI from 1981 to 1991 and found that biospheric activity increased remarkably in the northern high latitudes as a result of warming. Zhou et al. (2001)

2.5 2 1.5 1 0.5 0 −0.5 −1 −1.5 −2 −2.5

M. Zhao, S.W. Running NORTH AMERICA (40N-70N) TEMPERATURE

1.5 1 0.5 0 −0.5

GREENNESS

EURASIA (40N-70N)

2 TEMPERATURE GREENNESS

ANOMALY

ANOMALY

432

82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99

YEAR

−1 −1.5 −2 −2.5

82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99

YEAR

Fig. 16.4 Anomaly of standardized NDVI and air temperature from 1982 to 1999 for North America and Eurasia (40◦ N–70◦ N). (Redrawn from Zhou et al., 2001.)

continued the study by using a longer term NDVI record from 1981 to 1999, comparing that data to anomalies in air temperature (Fig. 16.4) for North America and Eurasia from 1982 to 1999. They found that Eurasia had larger increases in both the magnitude and duration of the seasonal cycle of NDVI than did North America. The length of active growing season increased by more than 2 weeks (18 ± 4 days) for Eurasia and nearly 2 weeks for North America (12 ± 5 days). This lengthening of the growing season has implications for many aspects of ecosystem processes, especially the carbon cycle. It may indicate an increase in NPP, but it may also enhance processes that release carbon to the atmosphere, such as decomposition, wildfires and insect outbreaks.

16.4.2 Estimation of Regional Carbon Storage Changes from Land Use Change Year-to-year variability in the growth rates of atmospheric CO2 concentrations is large, and is principally induced by inter-annual variations in terrestrial metabolism (Kindermann et al., 1996; Prentice et al., 2000; Keeling et al., 2001; McGuire et al., 2001; Nemani et al., 2003a). Over a relatively long period (e.g., decades), the mean carbon fluxes between atmosphere and land are controlled mainly by changes in biomass density, soil carbon and land use (Prentice et al., 2001; Schimel et al., 2001). Historical records and national inventories together with bookkeeping models have been used to estimate the contribution of carbon flux from land use change (e.g., Houghton et al., 1983). There are, however, large uncertainties in quantifying the role of changes in biomass density and land use in the contemporary global carbon cycle (Prentice et al., 2001; Schimel et al., 2001), largely because of varying definitions of forest cover among countries and differing time intervals (Matthews, 2001). Satellite data offer the possibility of providing spatially and temporally consistent estimates of forest cover to complement national reports. Using AVHRR data, for example, DeFries et al. (2002) estimated the changes in forest cover at the sub-pixel level over tropical regions for the 1980s and 1990s, and found that for the

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

433

1980s, the carbon emissions from land use change were 0.6 (0.3–0.8) Pg C year−1 , much less than 1.7 (0.6–2.5) Pg C year−1 , a value from IPCC report (Prentice et al., 2001), which was estimated mainly based on the statistics from United Nation Food and Agriculture Organization (FAO) and deforestation rates from the FAO Forest Resource Assessment (FRA). The measurements from satellite suggested less “missing” carbon in the global carbon budget than the previous estimates. DeFries et al. (2002) also found that, compared with the 1980s, clearing of tropical forests increased by about 10% as revealed by satellite data, as opposed to the 11% reduction reported by FRA, implying that there are increasing carbon emissions from changes in tropical land use. For forests in mid- and high-latitudes of NH, Myneni et al. (2001) simply used the fitted empirical equations between growing season cumulative NDVI and total biomass derived from inventory data of stem wood volume and long-term AVHRR NDVI records to discover an overall carbon sink for these forests, especially for forests in Eurasia and east North America. Compared with statistics and inventory data, satellite data can provide more detailed spatial information, and the results are generally consistent and accurate.

16.4.3 Estimation of Regional Carbon Fluxes Changes With a long-term AVHRR dataset and the PEMs mentioned previously, it is possible to detect the trends and interannual variability in global NPP. At the regional level, there have been several reports of increased NPP. For example, Randerson et al. (1999) used the CASA model to study the relationships among NPP, Net Ecosystem Production (NEP) and seasonal cycle of atmospheric CO2 at high latitudes and found that the enhanced photosynthesis activities in the early spring due to warming temperature are responsible for the changing trend in the seasonal cycle of CO2 . Also using the CASA model, Hicke et al. (2002a, b) reported an increase in NPP over most of North America from 1982 to 1999; Fang et al. (2003) found increased NPP from 1982 to 1999 in China. Studies have also found increased NPP at the global scale. Ichii et al. (2001) found that NPP increased during the 1980s with AVHRR NDVI. Cao et al. (2004), using the GLO-PEM model, found that NPP increased from 1981 to 2000, and the authors attribute the increased NPP to the effects of both increased CO2 fertilization effects and the favorable change in climate during the periods. Potter et al. (2003a) used CASA and AVHRR data to calculate changes in NPP and NEP during 1982–1998, determining that NEP increased at the global scale, and further, that Eurasia was a larger carbon sink than North America, contradicting the atmospheric inversion study by Fan et al. (1998). These studies are concentrated, using only one (solar radiation; Ichii et al., 2001) or two climate variables (temperature and precipitation; Potter et al., 2003a) or the combined effects of climate and CO2 (Cao et al., 2004). Nemani et al. (2003a) used a PEM developed for the global MODIS GPP and NPP products (Running et al., 2000) to study the effects of all three primary climate controls (incoming solar radiation, temperature and water) on NPP for the

434

M. Zhao, S.W. Running

Fig. 16.5 Spatial distribution of linear trends in estimated NPP (% change per year) using FPAR and LAI derived from AVHRR data for 1982–1999. (From Nemani et al., 2003a.)

period 1982–1999, and found that global NPP increased 6%, with tropical regions, especially Amazon rain forests (accounting for 42% of the global increase), being the largest contributors (Fig. 16.5). The increased NPP in the Amazon is caused by increased solar radiation resulting from reduced cloudiness during the dry season. In addition, Nemani and colleagues found that climate alone was responsible for more than 40% of the increase in the global NPP. Other factors, such as CO2 fertilization and nitrogen deposition may also played a role in enhancing NPP. Nemani et al.’s (2003a) study emphasizes the role of radiation, generally a dominant limiting climate factor in humid, highly productive rain forests covering large areas of the tropics.

16.4.4 Monitoring Disturbances Disturbances from fire, insect outbreaks, logging, etc., constitute the major components to ecosystem change and have great impact on the global carbon cycle, but many terrestrial biogeochemistry models do not account for disturbances (Canadell et al., 2000). The chief reason for omitting disturbances in these models is that there is no reliable record of the timing, distribution, spatial extent, or severity of these disturbances at the regional or the global scale. Only medium- to coarse-resolution satellites can provide the timing of major disturbances at the global scale. Potter et al. (2003b) retrieved major disturbances at the global scale with the AVHRR FPAR data for the period 1982–1999 and combined it with the NASA-CASA model to estimate the above-ground biomass carbon lost to the atmosphere during that

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

435

period. Van der Werf et al. (2004) used fire activity obtained from several satellites together with a biogeochemical model and an inverse analysis of atmospheric CO anomalies to estimate the CH4 and CO2 emissions from fire during the 1997– 2001 El Ni˜no/La Ni˜na period. They found that CO2 released from fire accounts for 66 ± 24% of the observed atmospheric CO2 annual growth rates during El Ni˜no. Randerson et al. (2005) further separated the contributions C3 and C4 vegetation to fire emissions and found that C3 vegetation is largely responsible for interannual variations in global fire emissions. MODIS provides real-time fire burned area products (Roy et al., 2005). However, other disturbances besides fires, including insect outbreaks, flood, irrigation etc., are also important contributors to the carbon cycle. With MODIS LST and EVI datasets, Mildrexler et al. (2007) have proposed a powerful Disturbance Index (DI), DI = (LSTmax /EV Imax )/(LSTmax /EV Imax )

(16.3)

to detect all disturbance, regardless of origin. In Eq. 16.3, LSTmax (EV Imax ) is the annual maximum composite MODIS LST (EVI), and LSTmax /EV Imax is the multiyear mean of the ratio of LSTmax to EV Imax . Compared with the field data, the DI can effectively detect fire scars and other disturbances at regional and global scales on an annual basis (Fig. 16.6). As mentioned previously, MODIS NPP results show that NPP anomalies can explain 72–83% of annual CO2 growth rate (Fig. 16.3), while 66 ± 24% was found by Van der Werf et al. (2004) from fire emission. Incorporating DI into MODIS NPP algorithm will help to separate the roles of NPP and disturbance in the global carbon cycle.

16.4.5 Validation of Process-Based Ecosystem Model with Remote Sensing Data Process-based ecosystem models are now playing an indispensable role in simulating the dynamics of ecosystems. These models all experience improvements through validation and advances in knowledge. However, unlike General Circulation Models (GCMs), which can be validated with observations from several thousand weather stations covering nearly all climatic zones, there is a paucity of ecological observations to validate ecosystem models. Remote sensing data, because of their spatio-temporal consistency, can be used as independent observations to validate such models. For example, Zhao et al. (2002) used AVHRR NDVI to test improvements in a biogeography model for simulation of leaf longevity (deciduous or evergreen) over China relative to the original method. Zhuang et al. (2003) compared the trend in carbon storage simulated by a process-based ecosystem model with that derived from NDVI data by Myneni et al. (2001) over the extra-tropical NH, determining that there is similar spatial pattern between the modeled and satellite data, though the modeled results had a smaller rate of change than the satellite estimates.

436

M. Zhao, S.W. Running

Fig. 16.6 Correspondence between the DI results, the MODIS active fire-detection data (black dots) and fire perimeter maps (black outlines) for (a, b) the 2003 wildfires near Missoula, Montana, and (c, d) the 2003 southern California wildfires. The southern California fires occurred in savanna and shrublands, vegetation types with the highest frequency of major disturbance at the global scale. (From Mildrexler et al., 2006.)

More importantly, results from models and satellite observations can support each other and strengthen our confidence in understanding the mechanisms behind ecosystem processes. Lucht et al. (2002) provide an excellent example of this support in their study of LAI following the Mt. Pinatubo eruption. The trend in vegetation activities derived from the long-term AVHRR dataset has long been suspect because of the need for data corrections resulting from instrument and navigational drift, intercalibration of successive instruments, and consideration of aerosol effects (Cihlar et al., 1998). In addition, there have been several different explanations for the reduced atmospheric CO2 growth rates after the effects of Mt. Pinatubo have been taken into account (Arora, 2003). Lucht et al. (2002) used a biogeochemical model of vegetation and observed climate data to discover that there is a drawdown of maximum LAI simulated by the model in the northern high-latitude growing seasons of 1992–1993, consistent with the reduced LAI in the same period derived from

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

437

AVHRR. This consistency between two independent datasets lead to the discover that the lower AVHRR-estimated LAI following the 1991 Mt. Pinatubo eruption was not caused by contamination of the satellite record from high aerosol loading, but by the volcanic-induced climate anomaly, especially low temperatures. Moreover, the model results demonstrated that the unbalanced effect of cooling on NPP and heterotrophic respiration provides a much simpler explanation for the additional high-latitude CO2 uptake than the proposed mechanism of increased NPP due to the increased diffuse sky light by Gu et al. (2003). Krakauer and Randerson (2003) also found reduced tree-ring width following Pinatubo eruption, confirming the reduced NPP due to volcanic effects.

16.5 Towards an Integrated Study of the Terrestrial Carbon Dynamics Remote sensing data are able to provide spatio-temporal dynamics of vegetation at regional and global scales. To quantify the role of terrestrial vegetation in the carbon cycle, however, requires ecosystem models. These models are validated, calibrated and refined based on the relationship between remote sensing data and the groundbased data related to carbon at local level (Heinsch et al., 2006; Turner et al., 2006; Zhao et al., 2005). Remote sensing data are then used to scale up from the local to the regional and the global scale both temporally and spatially (e.g., Zhao et al., 2005). The ecosystem models mentioned here cover a large complexity, ranging from very simple regression models such as the empirical relationship between NDVI and biomass or LAI to complex ecosystem process models. Figure 16.7 shows a scheme of quantification of the terrestrial carbon cycle using remote sensing data in an integrated system. Ground data, such as these measured at the eddy flux towers, and other observations, such as land cover, NPP, biomass, are the source for developing linkages between field-based and remote sensing data. Then the ecosystem models use remote sensing data to quantify the regional or global scale carbon exchange (Running et al., 1999). In reality, the interactions among ground data, remote sensing data, ecosystem models and carbon cycle are more complex. As depicted by Fig. 16.7, remote sensing data sometimes can also detect some mechanisms not observable with field measurements and remote sensing data may even assist in selecting appropriate sites for field observations. The results from ecosystem models may help us explain remote sensing data, and vice versa, while remote sensing data also can be used to validate or drive ecosystems models. Ecosystem models can quantify carbon exchange between atmosphere and terrestrial ecosystems, but carbon cycle data can also constrain and validate ecosystem models. Ecosystem models are the centerpiece of terrestrial carbon cycle studies, because modeling is the only method for understanding ecosystem processes in an interactive manner and predicting the impacts of biotic and abiotic factors on the carbon cycle, providing important information for policy makers and stockholders. With the improved quality and increasing spatial

438

M. Zhao, S.W. Running

Fig. 16.7 An interactive scheme for studying the terrestrial carbon cycle using remote sensing data

coverage of ground-based data, improved quality of satellite data, and the increasing knowledge gained from these data, the performance of the ecosystem models will continue to be enhanced, thereby enhancing our ability to study the earth as a system.

16.6 Summary Human beings have never before had such a great impact on the earth environment in recorded history. Our activities have changed land cover, water and nutrient cycling, the chemistry of the atmosphere, and therefore, climate systems and the structure and function of ecosystems. In turn, the anthropogenically induced environmental changes have influenced our well-being. More importantly, many current human-

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

439

induced environment issues are not restricted to a given region or country, but are having global impacts. Regular global measurements from satellites play a crucial role in monitoring the earth, which can provide advanced warning to allow for favorable environmental change (Running et al., 2006). The atmospheric CO2 concentration measurements at Mauna Loa since 1958 (the famous “Keeling curve”), has given us the advanced warning for global climate change, and led to the 1997 “Kyoto Protocol”, the international treaty on climate change, assigning mandatory targets for the reduction of greenhouse gas emissions to signatory nations. Ironically, Keeling’s measurements have been nearly halted several times due to the reluctance of the US government to provide continued support (Keeling, 1998). With the accumulated long-term time-series data, we gain more knowledge and find new discoveries (Keeling, 1998). Likewise, EOS has provided valuable scientific knowledge, and we hope that NASA continues this mission in some EOS-like project. The other lesson we learned from the use of satellite data to study the carbon cycle is that it is vital to have basic standard datasets for use in land science. We have made a lot of discoveries using only NDVI. However, the quality of NDVI is very critical for the science research, and there should be some MODIS-like land products generated continuously, with similar sensor spatial resolution, quality flags, and easy access data formats. As discussed above, for example, without standard LST and EVI datasets, it is impossible for the ecologists to effectively use satellite data (e.g., Mildrexler et al., 2007); without FPAR and LAI derived from satellite data, it is impossible for us to have a deeper understanding of the terrestrial carbon cycle following the Mt. Pinatubo eruption (e.g., Lucht et al., 2002), and the longterm trend in global NPP (e.g., Nemani et al., 2003a). With advancements in ecosystem modeling based on remotely sensing data, as well as the enhancement of computer performance and Internet technology, it is now becoming possible for land managers and policy makers to make decisions based on near real-time satellite data (Running et al., 2004), or even to take preventative action based on the information from near future forecasting using the same satellite data (Nemani et al., 2003b). Thus, the advancements in the study of terrestrial primary production and carbon cycle using satellite data are significant, not only for understanding the global carbon cycle, but also for application of satellite data to carbon-related natural resource management and land management. Acknowledgements The work is funded by the NASA/EOS Natural Resource/Education Training Center (grant NAG5-12540) and NASA MODIS Project (NNG04HZ19C). We thank Dr. Faith Ann Heinsch for comments and Dr. Liming Zhou for providing Fig. 16.4 for the chapter.

References Arora VK (2003) Decreased heterotrophic respiration reduced growth in atmospheric CO2 concentration. IGBP Global Change Newsletter 54:21–22 Asrar G, Fuchs M, Kanemasu ET, Hatfield JL (1984) Estimating absorbed photosynthetic radiation and leaf area index from spectral reflectance in wheat. Agron. J. 76:300–306

440

M. Zhao, S.W. Running

Cao M, Prince SD, Small J, Goetz SJ (2004) Remotely sensed interannual variations and trends in terrestrial net primary productivity 1981–2000. Ecosystems 7:233–242 Canadell JG, Mooney HA, Baldocchi DD, Berry JA, Ehleringer JR, Field CB, Gower ST, Hollinger DY, Hunt JE, Jackson RB, Running SW, Shaver GR, Steffen W, Trumbore SE, Valentini R, Bond BY (2000) Carbon metabolism of the terrestrial biosphere: a multi-technique approach for improved understanding. Ecosystems 3:115–130 Cihlar J, Chen JM, Li Z, Huang F, Latifovic R, Dixon R (1998) Can interannual land surface signal be discerned in composite AVHRR data? J. Geophys. Res. 103(D18):23163–23172, doi: 10.1029/98JD00050 DeFries RS, Houghton RA, Hansen M, Field CB, Skole DL, Townshend J (2002) Carbon emissions from tropical deforestation and regrowth based on satellite observations for the 1980s and 90s. Proc. Natl. Acad. Sci. USA 99(22):14256–14261 Fan S, Gloor M, Mahlman J, Pacala S, Sarmiento J, Takahashi T, Tans P (1998) A large terrestrial carbon sink in North America implied by atmospheric and oceanic carbon dioxide data and models. Science 282:442–446 Fang J, Piao S, Field CB, Pan Y, Guo Q, Zhou L, Peng C, Tao S (2003) Increasing net primary production in China from 1982 to 1999. Front. Ecol. Environ. 1:293–297 Folland CK, Karl TR, Christy JR, Clarke RA, Gruza GV, Jouzel J, Mann ME, Oerlemans J, Salinger MJ, Wang S-W (2001) Observed climate variability and change. In: JT Houghton,Y Ding, DJ Griggs, M Noguer, PJ van der Linden, X Dai, K Maskell, CA Johnson (eds), Climate Change 2001: The Scientific Basis. Contribution of Working Group I to the Third Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge, United Kingdom and New York, USA, pp 182–237 Fung I, Tucker CJ, Prentice K (1987) Application of advanced very high resolution radiometer vegetation index to study atmosphere-biosphere exchange of CO2 . J. Geophys. Res. 92(D3): 2999–3015 Goward SN, Tucker CJ, Dye DG (1985) North American vegetation patterns observed with the NOAA-7 Advanced Very High Resolution Radiometer. Vegetatio 64:3–14 Gu L, Baldocchi DD, Wofsy SC, Munger JW, Michalsky JJ, Urbanski SP, Boden TA (2003) Response of a deciduous forest to the Mount Pinatubo eruption: enhanced photosynthesis. Science 299:2035–2038 Guenther B, Xiong X, Salomonson VV, Barnes WL, Young J (2002) On-orbit performance of the earth observing system moderate resolution imaging spectroradiometer; first year of data. Remote Sens. Environ. 83:16–30 Heinsch FA, Reeves M, Votava P, et al. (2003) User’s Guide: GPP and NPP (MOD17A2/A3) Products, NASA MODIS Land Algorithm, pp 1–57 Heinsch FA, Zhao M, Running SW, et al. (2006) Evaluation of remote sensing based terrestrial productivity from MODIS using regional tower eddy flux network observations. IEEE Trans. Geosci. Remote Sens. 44(7):1908–1925 Hicke JA, Asner GP, Randerson JT, Tucker C, Los S, Birdsey R, Jenkins JC, Field CB (2002a) Trends in North American net primary productivity derived from satellite observations, 1982– 1998. Global Biogeochem. Cycles 16(2), doi:10.1029/2001GB001550 Hicke JA, Randerson J, Asner G, Randerson J, Tucker C, Los S, Birdsey R, Jenkins J, Field C, Holland E (2002b) Satellite-derived increases in net primary productivity across North America, 1982–1998, Geophys. Res. Lett. 29(10), doi:10.1029/2001GL013578 Houghton RA, Hobbie JE, Melillo JM, et al. (1983) Changes in the carbon content of terrestrial biota and soils between 1860 and 1980: a net release of CO2 to the atmosphere. Ecol. Monogr. 53:235–262 Huete AR, Didan K, Shimabukuro YE, Ratana R, Saleska SR, Hutyra LR, Yang W, Nemani RR, Myneni R (2006) Amazon rainforests green-up with sunlight in dry season. Geophys. Res. Lett. 33, L06405, doi: 10.1029/2005GL025583 Ichii K, Matsui Y, Yamaguchi Y, Ogawa K (2001) Comparison of global net primary production trends obtained from satellite-based normalized difference vegetation index and carbon cycle model. Global Biogeochem. Cycles 15(2):351–364, doi: 10.1029/2000GB001296

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

441

Imhoff ML, Bounoua L, Richetts T, Loucks C, Harriss R, Lawrence WT (2004) Global pattern in human consumption of net primary production. Nature 429:870–873 IPCC (Intergovernmental Panel on Climate Change) (2001) IPCC (Intergovernmental Panel on Climate Change). In: JT Houghton, Y Ding, DJ Griggs, M Noguer, PJ van der Linden, X Dai X (eds), Climate Change 2001: The Scientific Basis Contribution of Working Group I to the Third Assessment Report of the IPCC. Cambridge University Press, Cambridge, United Kingdom Justice CO, Townshend JRG, Holben BN, Tucker CJ (1985) Analysis of the phenology of global vegetation using meteorological satellite data. Int. J. Remote Sens. 6:1271–1318 Justice CO, Townshend JRG, Vermote EF, Masuoka E, Wolfe RE, Saleous N, Roy DP, Morisette JT (2002) An overview of MODIS land data processing and product status. Remote Sens. Environ. 83:3–15 Keeling CD, Chin JFS, Whorf TP (1996) Increased activity of northern vegetation inferred from atmospheric CO2 measurements. Nature 382:146–149 Keeling CD (1998) Rewards and penalties of monitoring the earth. Annu. Rev. Energy Environ. 23:25–82 Keeling CD, Piper SC, Bacastow RB, Wahlen M, Whorf TP, Heimann M, Meijer HA (2001) Exchange of atmospheric CO2 and 13 CO2 with the terrestrial biosphere and oceans from 1978 to 2000, I. Global aspects. SIO Ref. Ser. 01–06, Scripps Institute of Oceanography, La Jolla, CA Kindermann J, W¨urth G, Kohlmaier GH, Badeck F-W (1996) Interannual variation of carbon exchange fluxes in terrestrial ecosystems. Global Biogeochem. Cycles 10(4):737–756, doi: 10.1029/96GB02349 Krakauer NY, Randerson JT (2003) Do volcanic eruptions enhance or diminish net primary production? Evidence from tree rings. Global Biogeochem. Cycles 17(4):1118, doi:10.1029/2003GB002076 Kumar M, Monteith JL (1982) Remote sensing of crop growth. In H Smith (ed), Plants and daylight spectrum. Academic Press, London, pp 133–144. Landsberg JJ, Waring RH (1997) A generalized model of forest productivity using simplified concepts of radiation use efficiency, carbon balance and partitioning. Forest Ecol. Manag. 95: 209–228 Lucht W, Prentice IC, Myneni RB, Sitch S, Friedlingstein P, Cramer W, Bousquet P, Buermann W, Smith B (2002) Climatic control of the high-latitude vegetation greening trend and Pinatubo effect. Science 296:1687–1689 Matthews E (2001) Understanding the FRA 2000: Forest Briefing No. 1. World Resource Institute, Washington, DC, p 12 McGuire AD III, Prentice IC, Ramankutty N, Reichenau T, Schloss A, Tian H, Williams LJ, Wittenberg U (2001) Carbon balance of the terrestrial biosphere in the twentieth century: analyses of CO2 , climate and land use effects with four process-based ecosystem models. Global Biogeochem. Cycles 15(1):183–206 Mildrexler D, Zhao M, Heinsch FA, Running SW (2007) A new satellite based methodology for continental scale disturbance detection. Ecol. Appl. 17(1):235–250 Milesi C, Hashimoto H, Running SW, Nemani RR (2005) Climate variability, vegetation productivity and people at risk. Global Planet Change 47:221–231 Monteith JL (1972) Solar radiation and productivity in tropical ecosystems. J. Appl. Ecol. 9: 747–766 Monteith JL (1977) Climate and efficiency of crop production in Britain. Phil. Trans. Royal Soc. Lond. B 281:277–294 Mu Q, Zhao M, Heinsch FA, Liu M, Tian H, Running SW (2007) Evaluating water stress controls on primary production in biogeochemical and remote sensing based models. J. Geophys. Res. 112,G01012 Myneni RB, Hall FG, Sellers PJ, Marshak AL (1995) The interpretation of spectral vegetation indexes. IEEE Trans. Geosci. Remote Sens. 33:481–486 Myneni RB, Nemani RR, Running SW (1997a) Estimation of global leaf area index and absorbed par using radiative transfer models. IEEE Trans. Geosci. Remote Sens. 35:1380–1393

442

M. Zhao, S.W. Running

Myneni RB, Keeling CD, Tucker CJ, Asrar G, Nemani RR (1997b) Increase plant growth in the northern high latitudes from 1981–1991. Nature 386:698–702 Myneni RB, Dong J, Tucker CJ, Kaufmann RK, Kauppi PE, Liski J, Zhou L, Alexeyev V, Hughes MK (2001) A large carbon sink in the woody biomass of northern forests. Proc. Natl. Acad. Sci. USA 98(26):14784–14789 Nemani RR, Running SW (1989) Estimation of regional surface resistance to evapotranspiration from NDVI and thermal-IR AVHRR data. J. Appl. Meteorol. 28:276–284 Nemani RR, Keeling CD, Hashimoto H, Jolly WM, Piper SC, Tucker CJ, Myneni RB, Running SW (2003a) Climate-driven increases in global terrestrial net primary production from 1982 to 1999. Science 300:1560–1563 Nemani RR, White MA, Pierce L, Votava P, Coughlan J, Running SW (2003b) Biospheric monitoring and ecological forecasting. Earth Observ. Mag. 12(2):6–8 Potter CS, Randerson JT, Field CB, Matson PA, Vitousek PM, Mooney HA, Klooster SA (1993) Terrestrial ecosystem production: a process model based on global satellite and surface data. Global Biogeochem. Cycles 7:811–841 Potter CS, Klooster S, Myneni R, Genovese V, Tan P, Kumar V (2003a) Continental scale comparisons of terrestrial carbon sinks estimated from satellite data and ecosystem modeling 1982–98. Global Planet. Change 39:201–213 Potter CS, Tan P, Steinbach M, Klooster S, Kumar V, Myneni R, Genovese V (2003b) Major disturbance events in terrestrial ecosystems detected using global satellite data sets. Global Change Biol. 9(7):1005–1021 Prentice IC, Heimann M, Sitch S (2000) The carbon balance of the terrestrial biosphere: ecosystem models and atmospheric observations. Ecol. Appl. 10:1553–1573 Prentice IC, Farquhar GD, Fasham MJR, Goulden ML, Heimann M, Jaramillo VJ, Kheshgi HS, Le Qu´er´e C, Scholes RJ, Wallace DWR (2001) The carbon cycle and atmospheric carbon dioxide. In: JT Houghton,Y Ding, DJ Griggs, M Noguer, PJ van der Linden, X Dai, K Maskell, CA Johnson (eds), Climate Change 2001: The Scientific Basis. Contribution of Working Group I to the Third Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge, United Kingdom and New York, USA, pp 182–237 Prince SD (1991) A model of regional primary production for use with coarse resolution satellite data. Int. J. Remote Sens. 12:1313–1330 Prince SD, Goward SN (1995) Global primary production: a remote sensing approach. J. Biogeogr. 22:815–835 Rahman AF, Sims DA, Cordova VD, El-Masri BZ (2005) Potential of MODIS EVI and surface temperature for directly estimating per-pixel ecosystem C fluxes. Geophys. Res. Lett. 32, L19404, doi:10.1029/2005GL024127 Randerson JT, Field CB, Fung IY, Tans PP (1999) Increases in early season ecosystem uptake explain recent changes in the seasonal cycle of atmospheric CO2 at high northern latitudes. Geophys. Res. Lett. 26:2765–2768, doi: 10.1029/1999GL900500 Randerson JT, van der Werf GR, Collatz GJ, Giglio L, Still CJ, Kasibhatla P, Miller JB, White JWC, DeFries RS, Kasischke ES (2005) Fire emissions from C3 and C4 vegetation and their influence on interannual variability of atmospheric CO2 and d13 CO2 . Global Biogeochem. Cycles 19:GB2019 Reeves MC, Zhao M, Running SW (2006) Applying improved estimates of MODIS productivity to characterize grassland vegetation dynamics. Rangeland Ecol. Manag. 59:1–10 Roy DP, Jin Y, Lewis PE, Justice CO (2005) Prototyping a global algorithm for systematic fireaffected area mapping using MODIS time series data. Remote Sens. Environ. 97:137–162 Ruimy A, Saugier B, Dedieu G (1994) Methodology for the estimation of terrestrial net primary production from remotely sensed data. J. Geophys. Res. 99:5263–5283 Ruimy A, Dedieu G, Saugier B (1996) TURC: a diagnostic model of continental gross primary productivity and net primary productivity. Global Biogeochem. Cycles 10: 269–286, doi: 10.1029/96GB00349

16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle

443

Running SW, Nemani RR (1988) Relating seasonal patterns of the AVHRR vegetation index to simulated photosynthesis and transpiration of forests in different climates. Remote Sens. Environ. 24:347–367 Running SW, Nemani RR, Peterson DL, Band LE, Potts DF, Pierce LL, Spanner MA (1989) Mapping regional forest evapotranspiration and photosynthesis by coupling satellite data with ecosystem simulation. Ecology 70:1090–1101 Running SW, Hunt ER (1993) Generalization of a forest ecosystem process model for other biomes, BIOME-BGC, and an application for global-scale models. In JR Ehleringer, CB Field (ed), Scaling physiological processes: leaf to globe. Academic, San Diego, CA, pp 141–158 Running SW, Baldocchi DD, Turner DP, Gower ST, Bakwin PS, Hibbard KA (1999) A global terrestrial monitoring network integrating tower fluxes, flask sampling, ecosystem modeling and EOS satellite data. Remote Sens. Environ. 70:108–128 Running SW, Thornton PE, Nemani RR, Glassy JM (2000) Global terrestrial gross and net primary productivity from the earth observing system. In O Sala, R Jackson, H Mooney (eds), Methods in ecosystem science. Springer, New York, pp 44–57 Running SW, Nemani RR, Heinsch FA, Zhao M, Reeves M, Hashimoto H (2004) A continuous satellite-derived measure of global terrestrial primary productivity: future science and applications. Bioscience 54:547–560 Running SW, Nemani RR, Townshend J, Baldocchi D (2006) Next generation terrestrial carbon monitoring. American Geophysical Union Monography XX. A tribute to the career of Charles David Keeling Schimel DS, House JI, Hibbard KA, Bousquet P, Ciais P, Peylin P, Braswell BH, Apps MJ, et al. (2001) Recent patterns and mechanisms of carbon exchange by terrestrial ecosystems. Nature 414:169–172 Sellers PJ (1985) Canopy reflectance, photosynthesis and transpiration. Int. J. Remote Sens. 6:1335–1372 Sellers PJ (1987) Canopy reflectance, photosynthesis and transpiration. II. The role of biophysics in the linearity of their interdependence. Remote Sens. Environ. 21:143–183 Townshend JRG, Justice CO (1986) Analysis of the dynamics of African vegetation using the normalized difference vegetation index. Int. J. Remote Sens. 7:1435–1445 Tucker CJ (1979) Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 8:127–150 Tucker CJ (1980) A spectral method for determining the percentage of green herbage material in clipped samples. Remote Sens. Environ. 9:175–181 Tucker CJ, Vanpraet C, Boerwinkel E, Gaston A (1983) Satellite remote sensing of total dry matter production in the Senegalese Sahel. Remote Sens. Environ. 13:461–474 Tucker CJ, Townshend JRG, Goff TE (1985) African land-cover classification using satellite data. Science 227:369–375 Tucker CJ, Fung I, Keeling C, Gammon R (1986) Relationship between atmospheric CO2 variations and satellite-derived vegetation index. Nature 319:195–199 Turner DP, Ritts WD, Cohen WB, Gower ST, Running SW, Zhao M, Costa MH, Kirschbaum A, Ham J, Saleska S, Ahl D (2006) Evaluation of MODIS NPP and GPP products across multiple biomes. Remote Sens. Environ. 102:282–292 Van der Werf GR, Randersonm JT, Collatz GJ, Giglio L, Kasibhatla P, Arellano A, Olsen S, Kasischke ES (2004) Continental-scale partitioning of fire emissions during the 1997 to 2001 El Ni˜no/La Ni˜na period. Science 303:73–76 Walther G-R, Post E, Convey P, Menzel A, Parmesan C, Beebee TJC, Fromentin J-M, Guldberg OH, Bairlein F (2002) Ecological responses to recent climate change. Nature 416:389–395 Wolfe RE, Nishihama M, Fleig AJ, Kuyper JA, Roy DP, Storey JC, Patt FS (2002) Achieving subpixel geolocation accuracy in support of MODIS land science. Remote Sens. Environ. 83:31–49 Xiao X, Hollinger D, Aber J, Goltz M, Davidson EA, Zhang Q, Moore III B (2004) Satellite-based modeling of gross primary production in an evergreen needleleaf forest. Remote Sens. Environ. 89:519–534

444

M. Zhao, S.W. Running

Xiao X, Hagen S, Zhang Q, Keller M, Moore III B (2006) Detecting leaf phenology of seasonally moist tropical forests in South America with multi-temporal MODIS images. Remote Sens. Environ. 103:465–473 Zhao M, Neilson R, Yan X, Dong W (2002) Modelling the vegetation of China under changing climate. Acta Geographica Sinica 57(1):28–38 Zhao M, Heinsch FA, Nemani RR, Running SW (2005) Improvements of the MODIS terrestrial gross and net primary production global dataset. Remote Sens. Environ. 95:164–176 Zhao M, Running SW, Nemani RR (2006a) Sensitivity of Moderate Resolution Imaging Spectroradiometer (MODIS) terrestrial primary production to the accuracy of meteorological reanalyses. J. Geophys. Res. 111:G01002, doi:10.1029/2004JG000004 Zhao M, Running SW, Heinsh FA, Nemani RR (2006b) Variations of global terrestrial primary production observed by Moderate Resolution Imaging Spectroradiometer (MODIS) from 2000 to 2005. (in preparation) Zhou L, Tucker CJ, Kaufmann RK, Slayback D, Shabanov NV, Myneni RB (2001) Variations in northern vegetation activity inferred from satellite data of vegetation index during 1981 to 1999. J. Geophys. Res. 106:20069–20083 Zhuang Q, McGuire AD, Melillo JM, Clein JS, Dargaville RJ, Kicklighter DW, Myneni RB, Dong J, Romanovsky VE, Harden J, Hobbie JE (2003) Carbon cycling in extratropical terrestrial ecosystems of the Northern Hemisphere during the 20th Century: a modeling analysis of the influences of soil thermal dynamics. Tellus B 55:751–776

Chapter 17

Applications of Terrestrial Remote Sensing to Climate Modeling Robert E. Dickinson

Abstract Terrestrial processes are an important component of climate. Climate can be viewed as a nonlinear dynamical system which generates statistics to be compared with observational statistics. The surface is forced by net radiation balanced by sensible and latent fluxes, and by precipitation balanced by evapotranspiration, soil moisture storage, and runoff. These balances depend on detailed geographic descriptions of parameters required by the modeling. These details are constrained by satellite remote sensing (as demonstrated by various recent studies) with consequent substantial improvement in climate models. Some of these parameters, especially those involving vegetation, may be evolved with the climate system. When climate models characterize their radiative processes consistent with the remote sensing algorithms useful for their detection, they become physically more realistic and provide a suitable modeling framework for forward modeling data assimilation. In particular, the 3D nature of canopy radiation needs to be represented in climate models and how this connects trees and bushes to various underlying surfaces.

17.1 Introduction to the Formulation of Climate Models Climate dynamical models solve various conservation equations that describe how the climate system evolves in time, and consequently generate numerical data for its statistical characterization. These climate statistics may include, e.g., monthly mean temperature and humidity of near surface air, temperature of leaves and profiles of soil temperature and water content. Additional important statistics involve diurnal variations, annual cycles and year to year variability of such quantities. The terrestrial component of climate consists of various state variables related to the terms Robert E. Dickinson Georgia Institute of Technology [email protected] S. Liang (ed.), Advances in Land Remote Sensing, 445–463. c Springer Science + Business Media B.V., 2008

445

446

R.E. Dickinson

just mentioned, linked together by process descriptions, and constrained by appropriate observational information. Statistics of the process descriptions are also often examined as part of the study of the climatology of a model. Exactly what state variables, process descriptions, and constraining observations should be used is determined by the multiple purposes for which the terrestrial system is included in a climate model (e.g., for their contributions to weather prediction, inter-annual climate prediction, projection of long term change from greenhouse gases, study of hydrology and water resources, crop production, ecology of various natural systems, biogeochemical cycling, especially that of carbon, as an ancillary source of information for remote sensing studies, or for forward modeling of remote sensing radiances). The primary role of the terrestrial system as an interactive component of a climate model is to determine the near surface and surface variables that are used to characterize climate observationally and to control the dynamic and thermodynamic structure of the atmosphere as mentioned above. Major ingredients of this control are the surface exchanges of energy, moisture, and momentum. Land surface models have evolved from quite simple treatments of these exchanges to increasingly complex descriptions. For earlier reviews, cf. Dickinson (1983, 1984, 1989, 1992, 1995a, b), Dickinson et al. (1991), Sellers et al. (1997), Pitman (2003), Yang (2003). Land is most simply treated as a single reservoir for whatever quantities it is supposed to exchange with the atmosphere. Earliest climate models also averaged over the diurnal and annual cycles so that climate was a “steady state” system. The land component of such early models required a statement as to what fraction of the incident solar radiation it would absorb; that is, its albedo. It also required a specification of heat capacity and water holding capacity and a rule relating its evaporation to that which would come from a wet surface, depending on how much water was stored. Later models have become much more complex. Dai et al. (2003) describes one example of current models. Such models include descriptions of the global distribution of vegetation and various mechanisms changing with model time-step by which vegetation returns water to the atmosphere. They may also describe overall energy and water balances and fluxes from lakes, wetlands, glaciers; river-flows are determined as an input to ocean dynamics. Rather than describe the details of any particular such approach, we frame the problem somewhat abstractly. All appropriate approaches to modeling the land surface in a climate model, although they may look quite different, are implementations of essentially the same ideas.

17.2 What Does the Land Surface of a Climate Model Consist of? The terrestrial surface in a climate model consists of either soils covered by vegetation or bare soil. It includes wetlands and lakes and a description of land elevations at the resolution of the atmospheric model. An initial question might be: what soils and what vegetation should the model have. This question could be responded to by

17 Applications of Terrestrial Remote Sensing to Climate Modeling

447

the various descriptive names that observed such vegetation is given: e.g., a given grid-square in a climate model may be covered by an open forest with an understory of grass, overlying a soil classified as an “Alfisol”. Such names, alone, are useless for quantitative physical modeling. However, they may be used through their correlation with needed properties to infer what is needed. One application of terrestrial remote sensing has been to extrapolate over wider areas names that have been given to particular kinds of vegetation at the surface. This assignment of a geographic distribution of land cover is useful for other purposes but may be losing some of the information needed by climate modelers. If so, this is regrettable since much of the quantitative information measured by a satellite sensor, especially such as provided by reflectance imagery, is closely connected to the information directly needed by climate modelers. Initial geographic characterization of land cover was provided by Wilson (1984), Matthews (1984). More recent land cover data sets have come from AVHRR (e.g., Strugnell et al., 2001) and now MODIS (Friedl et al., 2002).

17.3 Water and Energy Balance Requirements Energy exchange at the terrestrial surface is dictated by its net absorption of solar energy. The solar flux incident at the top of the atmosphere is determined by location, time of day, and time of year. In passing through the atmosphere, some radiation is absorbed by atmospheric molecular constituents and some scattered, i.e., Raleigh scattering by molecules, and Mie scattering by aerosols and cloud droplets. Overall, on average, about 20% is absorbed and about the same lost by upward scattering. More than half the incoming solar radiation remains to reach the terrestrial surface. This radiation incident at the surface is approximately half downward scattered from the atmosphere (much less on clear days and nearly 100% on cloudy days). It can in the short term be stored, but on longer time scales must be balanced by net upward long-wave thermal radiation and turbulent fluxes carrying internal energy from atmospheric temperature (sensible fluxes), and the energy stored in evaporated water (latent fluxes). Different surfaces act in different ways to achieve a balance between their absorption of solar radiation and the thermal radiation and turbulent flux exchanges that collectively provide this balance. Because the net thermal emission and storage variation (e.g., thermal conduction into soil) are relatively regular, during the daytime these are commonly lumped for diagnostic purposes with the absorbed solar radiation to get a net forcing of the sensible and latent fluxes. As by construction, this net radiation plus storage must balance the turbulent fluxes, the primary issue is then what determines the relative amount of each. This partitioning between the turbulent flux terms is commonly characterized by the ratio of sensible to latent fluxes, referred to as “the Bowen ratio”. According to the above generalities, state variables should be connected to components of the system that store water and energy and change in time as this storage

448

R.E. Dickinson

is modified. Additional “diagnostic” variables are used to relate the state variables to fluxes. The amount of energy stored in a given material is proportional to its temperature with this proportionality factor, a heat capacity. Furthermore, sensible energy exchanges are driven by the differences in temperature between two reservoirs and moisture exchange by some difference in moisture potentials. In sum, the temperatures and moisture of various system elements must be distinguished. Models address how these elements are heated by the sun and how they lose energy to each other or to the atmosphere. The most significant contributors to these elements to be modeled are the vegetation, surface water in the form of snow or free water (e.g., lakes), and “soil”, where “soil” refers to mineral or dead organic matter stratified vertically. Vegetation generally consists of individual surfaces such as leaves or branches with spatial scales from a few mm to meters. These surfaces have an area in contact with air that is large compared to the flat area of underlying soil (thus, acting in some ways as a porous medium, but on a relatively large spatial scale). They respond to solar heating by increasing their temperatures until their heating is balanced by turbulent and convective air motions carrying thermal energy and water. Soils and snow are also porous but their spatial scales are only on a nanometer to micrometer scale so that water moves through them as a very viscous fluid and thermal energy by conduction. Soil by volume is about half mineral or organic materials, and the rest some combination of air and water. Snow consists of crystalline ice with quite a bit of air. In simple terms, the modeled atmosphere delivers solar radiation and precipitation to the modeled terrestrial surface. How this energy and water is returned to the atmosphere depends on quite a few modeling details. From the climate viewpoint, the details of the return fluxes in amount and timing can be important for temperatures and precipitation. Precipitation may either be lost at nearly the same time it falls through evaporation from canopy (referred to as interception loss) or can infiltrate the soil and be extracted weeks later as part of the transpiration fluxes (e.g., Dickinson et al., 2003). Solar heating of springtime boreal vegetation may either return directly to the atmosphere or be used to melt the snow pack. The latter is often invoked as a vegetation feedback on snow melting. More concrete descriptions of particular processes will be addressed in the following discussion of current issues that emerge in the context of remote sensing data requirements.

17.4 Role of Solar Radiation in the Climate Model Terrestrial System Climate models determine the solar radiation absorbed by the various components of the terrestrial system to determine separate energy and water balances. For example, the solar radiation incident at the surface may heat a canopy, its understory, and the underlying soil. If all these terms are lumped together into a single complex surface, the solar radiation absorbed is simply the complement of (one minus the) albedo, which is quantified by remote sensing data products such as from MODIS (e.g., Gao

17 Applications of Terrestrial Remote Sensing to Climate Modeling

449

et al., 2005). However, surfaces treated in more detail will determine total solar absorbed as the sum over the absorption by individual terms, and hence observed albedo becomes a “constraint” rather than a simple boundary condition. Climate models use leaves, green or brown, and the underlying woody stems to determine the canopy contribution to absorption of solar radiation. This absorbed solar radiation raises the vapor pressure of water inside green leaves to higher values than that of the air outside the leaves and consequently forces the leaves to transpire. Since the storage of water inside leaves is commonly small, this water loss must be compensated by extraction of water stored in the soil by plant roots. The contribution of leaves is quantified in the model in terms of the leaf-area-index (LAI). This quantity is the one-sided area of leaves (flattened if necessary, to be a spatial projection) per area of soil surface. MODIS (Myneni et al., 2002) provides remote sensing estimates of the global distribution of leaves. This term has been determined over periods of 8 days from the beginning of the year 2000 to the present. It is accompanied by a product FPAR which provides that fraction of the absorbed visible solar radiation that is taken up by the canopy. The characterization of dead leaves and stems is more problematic as not yet supported by remote sensing data. The most common approach to address sub-grid variability in vegetation is to portion the climate model grid square into various sub-grid tiles. Some fraction of the grid square must be taken as vegetated, and can be subdivided in terms of plant functional types (pft’s) (Bonan and Levis, 2002). For example, a savanna might be covered 80% by grass and 20% by trees, and these are put on separate areas and their radiative exchange determined by one-dimensional models. The estimated area of bare surfaces (i.e., whatever is not directly under a tree or grass canopy) is similarly treated as a separate tile. The area directly under canopies has been treated as bare soil. Such models assume that trees and grass are in separate clumps; consequently, if there are trees overlying grass or soil, the light environment is treated very unrealistically in terms of the effects of the 3D shading of the grass or soil by the trees. For open canopies, this treatment can underestimate the visible radiation absorbed by the canopy by at least a factor of 2. However, many canopies are closed enough, and reflections from the underlying surface sufficiently compensate, that much smaller errors are seen in comparing the modeled with observed fraction of absorbed visible radiation. Models of transpiration and photosynthesis have a nonlinear dependence on the intensity of incident light. That is, their dependence on light is linear at low light levels but asymptotes to some constant value (saturation) at high light levels. Direct sunlight either strikes a leaf and is greatly attenuated or does not hit any leaves and continues at the same intensity it started with. Thus, the appropriate statistic for the effect of direct solar radiation on photosynthesis is not average light intensity but rather what is the relative area of leaves that receive the direct sun. Diffuse light, on the other hand is idealized as coming from all sky directions with the same intensity (cf. Pinty et al., 2005, for a relaxation of this assumption.) and consequently strikes all leaf surfaces with the average intensity of the diffuse light at a given level in the canopy. Climate models commonly use an average value of the diffuse radiation, an assumption which may lead to a factor of 2 overestimation of the average shade

450

R.E. Dickinson

leaf intensity compared to that from a better scaling (e.g., Dai et al., 2004) and further serious error if light intensities are too strong to be in the light limited regime of transpiration/photosynthesis. Such may happen, e.g., for a midday summer sun, passing through a thin cloud or thick aerosol layer. Light of average intensity locally attenuates by “Beers Law” either in fractional area of sunlit leaves (direct sun) or its intensity (diffuse radiation) and thus according to dI +G·I = S (17.1) dL where I is the incident light intensity, L is path length measured in leaf area increment, S is an internal source term, accounting for scattering if needed, and G is a factor for the projection of leaves into the direction of the light (commonly assumed to be a factor of 0.5 approximating the effect of a uniform distribution of leaf orientations). For a canopy idealized as a uniform spatial distribution of leaves, Eq. (17.1) generalizes to a global expression and the path length is the vertical path length divided by a cosine projection factor. However, leaves, especially those in forest canopies, are far from uniformly distributed. The most obvious spatial heterogeneities that must be addressed are the organization of plant canopies as discrete objects, organization of leaves into “clusters” and the variety of path lengths (i.e., number of leaves) that a light ray must pass by. Although this issue of canopy heterogeneity has been recognized by climate modelers for a long time (e.g., Dickinson, 1983), practical approaches to address it are only now being developed (e.g., Pinty et al., 2006). Ideally, this part of a climate model should not be computationally much more intensive than the evaluation of few exponentials (as in analytic two-stream solutions), thus probably precluding the use of fine layering approaches with multi-scatter iteration or even the direct inversion of a large matrix as might arise out of various idealizations of the situation. The physically most realistic and complete description using Monte Carlo is slow by factors of at least thousands compared to what is needed but can be very useful for validating approaches of low computational cost (e.g., Pinty et al., 2006). Where climate model treatments of canopy radiation are weakest, they have also been most limited by lack of data. These are the situations where sparse canopies have large openings through which light can directly reach the underlying surface. Thus, however bad the treatments of radiation by climate models may appear from a conceptual viewpoint, they have done the required job within the context of the very little available information. With the archiving of several years of MODIS data, the time has arrived to evaluate and improve the treatments of radiation used in climate models. Jin et al. (2002) reported a very large range of albedos over snow covered areas depending on the masking by vegetation of the underlying snow surface. Figure 17.1, from Gao et al. (2005) compares observed albedos (visible light) for evergreen forest and grassland covers and for varying degrees of snow cover. It illustrates the increase of albedo with increasing snow but with a much greater darkening of the forested region and less sensitivity to snow. It also shows that an extremely large spatial variability of albedo especially with the largest snow covers (factor of 2

17 Applications of Terrestrial Remote Sensing to Climate Modeling

451

Fig. 17.1 Shows how albedos vary with snow cover fractions (Gao et al., 2005). Points in the figure were extracted from global high-quality retrievals during the whole year of 2001 (23 production periods). Each point in the figure represents a CMG resolution (0.05◦ ) pixel. Grassland albedos are seen to be four times as sensitive to snow fractions as that of an evergreen forest

or more). The masking of snow by forests has been included in at least some climate models since the early 1980s but not with the correct geometric considerations as discussed above. Furthermore, climate modelers have had little basis for incorporating in their global models from first principles or surface observation descriptions of how “open” canopies might be. Prior remote sensing data such as from AVHRR has been of limited value in developing climate model radiation schemes that adequately address such details. Gao et al. (2005) show that the albedos measured by MODIS have the spatial and temporal patterns appropriate to their underlying land cover classes. The seasonal variations of these albedos are consistent with the phenology of leaf cover (e.g., outside the growing season, deciduous broadleaf forests become much darker in the near-infrared and a bit brighter in the visible). The albedos over areas of sparser vegetation also show influences of underlying soil, e.g., grasslands in the band of 10–20 N and presumably mostly in North Africa, show a much higher albedo than elsewhere. Tsvetsinskaya et al. (2002) has shown in the context of North Africa how measured MODIS albedos can be applied to characterize the albedos of an arid region. They find over this region, considerable spatial variability in its surface albedos, apparently related to soil mineral composition and geographical characteristics. For example, MODIS shortwave albedos vary by a factor of about 2.5 from the darkest volcanic terrains to the brightest sand over the Sahara. Overall, the Sahara desert has much higher albedos than deserts elsewhere. Climate models have previously assumed a single albedo for desert.

452

R.E. Dickinson

17.5 Climate Model Evaluation Studies Recent studies have begun to evaluate in the context of the new satellite data the performance of the treatments of radiation in climate models and their various ancillary assumptions such as the leaf area used to produce their radiation. Two studies have evaluated the climate model described by Zeng et al. (2002). This model utilized land surface data developed from AVHRR (e.g., Zeng et al., 2000; Strugnell et al., 2001) and implemented in the CLM model as described by Dai et al. (2003). Zhou et al. (2003) found that largest discrepancies between observed and modeled albedos occur over snow covered regions and in arid regions. They emphasized a lowering of winter albedos compared to that observed by MODIS, resulting from an apparent excess of tree and bush stem areas included in the model. Tian et al. (2004a) compared the LAI assumed in the model with that observed by MODIS and the contributions of differences between model verus observed LAI to model error. They also emphasized that whereas the Zeng et al. (2002) climate model may have overstated the absorption of light by the canopy stems, that this absorption should still be included but with more appropriate optical properties and that the MODIS algorithms for LAI and FPAR have left out their effect entirely. Zhang et al. (2005a) have further demonstrated that a significant fraction of solar radiation is absorbed by non-photosynthesizing canopy surfaces. Other studies have evaluated the climate model version described by Bonan et al. (2002). It differs in its modeling of the land surface from that used by Zeng et al. (2002) in replacing some aspects of the previous version of CLM (common land model) with the canopy radiation description developed by Bonan (1996), and with land data sets developed by Bonan and Levis (2002), also from AVHRR. Although the AVHRR satellite data used by Zeng et al. (2002) may have been developed more thoroughly than that of Bonan and Levis (2002), it was expressed in terms of the IGBP classification and Bonan et al. (2002) required the land data to be formulated in terms of plant functional type tiles (pft’s) as described in Bonan and Levis (2002). In after-sight, the treatment of canopy radiation by Zeng et al. (2002) suffered some numerical flaws (e.g., as mentioned by Pinty et al., 2006), and other inadequacies so its replacement by the full two-stream computation of Bonan (1996), essentially the same as used in the Sib model (Dickinson, 1983; Sellers, 1985), was likely an improvement. To make a long story short, the CLM “took two steps forward and one step back” as a result of different individual ideas as to what was currently the best way to do canopy radiation and derive land boundary conditions from AVHRR. This CLM2 model (community land model version 2) described by Bonan et al. (2002) was assessed by Oleson et al. (2003), Wang et al. (2004), in the context of its simulation of albedos. They reported a large negative bias in the albedos over the Sahara in contrast to the large positive bias reported by Zhou et al. (2003). This negative bias was a consequence of the Sahara albedo constrained to fit data from the ERBE instrument, whereas the positive bias was a consequence of data constrained to reproduce albedos inferred from AVHRR. In sum, prior to MODIS, previous less reliable satellite estimates of the albedo of the Sahara may have provided no real information beyond that already available to climate modelers from

17 Applications of Terrestrial Remote Sensing to Climate Modeling

453

surface and aircraft observations. For the latter, soils, sands, and rock had only been distinguished by models in terms of their soil hydrological properties and albedo inferred from an original; “light”, “medium” and “dark” color classification. The ratio of near-infrared to visible albedo ratio has been fixed at 2 and any dependence on solar zenith angle neglected. Much more detailed spatial and spectral information, especially over semi-arid regions is now available from MODIS. For example, the ratio of near-infrared to visible albedos observed in MODIS over the deserts of Sahara, Taklimakan, and Australia varies from 1.6 to 2.7 (Zhou et al., 2003) and albedo in the Sahara has been shown to increase significantly with solar zenith angle (Wang et al., 2005). Zhou et al. (2005) has proposed a method using principal component analysis to economically represent a high quality soil albedo dataset over non-vegetated North Africa. An extension of their approach can be used to separate the MODIS albedos into soil and canopy contributions. Oleson et al. (2003) (as Zhou et al., 2003) reported that the land surface model, CLM2, overall reasonably simulated vegetation albedos but that the two-stream treatment of radiation appeared to have too strong a dependence on solar zenith angle so that the CLM “black-sky” albedo with sun at local noon was in better agreement with MODIS than the “white sky” diffuse radiation. Wang et al. (2004) further documented the substantial differences between model and observed dependences on solar zenith angle. Oleson et al. (2003) also emphasized an overestimate of albedos by the model over snow covered regions in contrast to the emphasis in Zhou et al. (2003) of an apparent underestimate. Evidently, it is easy for climate models to make large errors in regions of snow cover and the shading/masking effects of the vegetation can be either under or over emphasized in different approaches to the canopy radiation that are too oversimplified and unconstrained by observations to be expected to give the right answer. Tian et al. (2004b) illustrate the usefulness of new land surface datasets developed from MODIS as model boundary conditions. They related LAI, vegetation continuous field, and the land cover maps to the pft formulation of Bonan and Levis (2002), using data “collection 4” for the period of September, 2000–August, 2002. This new dataset showed large differences from the old dataset of Bonan and Levis (2002). The new LAI was larger than the older version by at least 1.5 over the Amazon, central Africa, southeastern Asia, and north Europe, and by about 0.5–1.0 over most areas beyond 60◦ N in both seasons (Fig. 17.2). These increases in LAI are likely a result of the AVHRR inversion saturating at a lower LAI than that of MODIS. The new LAI was also found to be smaller over many regions (Fig. 17.2). Their use of MODIS data decreased the amount of crops and grass by 20–40% globally, and increased the “bare” category by a large amount (Fig. 17.3). The previous version of the data assumed: “that non-tree covered land in forest, savannas, and grasslands was covered by grasses, in shrub lands by shrubs, in croplands by crops” (Bonan and Levis, 2002). In semi-arid regions, there was no information (Bonan and Levis, 2002) as to the fraction of area covered by bushes so leaves were spread uniformly. This assumption has been found to be incompatible with the existing parameterization of under canopy energy fluxes and a revised approach for the latter was implemented (Zeng et al., 2005).

454

R.E. Dickinson

Fig. 17.2 Spatial pattern of LAI difference between the old and new land surface datasets (newold as derived respectively from MODIS and AVHRR LAI products) in winter (DJF) and summer (JJA) (Tian et al., 2004c)

In the newer MODIS data for fractional tree cover (Hansen et al., 2003) fractional regions are characterized either as trees or shrubs or “bare”. Apparently, the bare category in the continuous field data refers to all understory components, but for lack of further information, Tian et al. (2004b) used literally the “bare” classification. Presumably, in semi-arid systems, understories are mostly bare soil, but by contrast, in moist forests they should mostly be dead leaves, mosses, or herbaceous small plants. Tian et al. (2004b, 2004c) used CLM2 coupled with the Community Atmospheric Model (CAM2) to investigate how the modeled surface variables such as temperature and albedo were modified by the new dataset. For snow-free regions, the increased LAI and changes in the percent cover from grass/crop to tree or shrub decreased albedo, but also decreased surface air temperatures. Increases of canopy evapotranspiration and decreases of ground evaporation over tropical regions improved the modeled surface temperature. MODIS albedo data can be used to adjust model parameters controlling absorption of solar radiation to provide a better fit to the albedo observations (e.g., Liang et al., 2005).

17 Applications of Terrestrial Remote Sensing to Climate Modeling

455

Fig. 17.3 Spatial pattern of percent cover difference (new-old) in grass/crop, tree, shrub, and bare soil at the model spatial resolution as inferred from MODIS land cover classification versus use of AVHRR (Tian et al., 2004c)

17.6 How can Terrestrial Remote Sensing Best Support Climate Models? Observational characterizations of the terrestrial surface may advance climate models more by providing information that climate models should be using rather than that required by the current formulation of the climate models. For example, we have discussed in Section 17.4, that climate models have oversimplified treatments of the interaction of solar radiation with the terrestrial surface. What observational information could support a more correct treatment of solar radiation? Observations

456

R.E. Dickinson

should characterize, to the extent possible, the 3D structure of the surface. A description of fractional tree cover (e.g., Hansen et al., 2003) is a first step in that direction. Besides the fraction of tree cover, other tree statistics needed for climate model details are their size, i.e., height and aspect ratio, and a characterization of their average distance to nearest neighbors (e.g., Widlowski et al., 2001, 2004). In addition, a climate model needs all the smaller scale parameters that contribute to the characterization of surface shading, in particular the relative areas of leaf and stems. The surfaces beneath a canopy that are affected by the canopy radiatively must be connected to the description of the canopy. Thus, the fractional areas of other surfaces besides trees should be established at an appropriate level of detail, in particular, the fraction that is radiatively connected versus unconnected and the composition of the underlying surface. The concept of radiative connectivity is somewhat vague, but it can be quantified by some simple rules: e.g., it might include all underlying surfaces within a horizontal distance 3 × H from the canopy, where H is the crown height. A description of the underlying surfaces so connected can be relatively simple. What are the characteristics of the dominant material shading the mineral soil, what fraction is photosynthesizing, and what fraction is open to the underlying soil? As discussed in the preceding sections, understory has been variously characterized as “grass” or “bare”. Either choice of surface may lead to erroneous results in absorption of radiation when, as now done, its shading by the overlying canopy is neglected except for surfaces directly under the canopy (which are currently always taken to be bare soil). Other potentially important under-canopy-cover can be, e.g., dead leaves, moss, lichen, or wetland. In general, if fc denotes the fractional cover of trees, then the sum of all understory components should be 1, not fc as currently assumed by climate modelers. The difference between surfaces shaded for overhead sun and those shaded at other times of day is not discontinuous in nature and should not be so modeled.

17.7 Terrestrial Remote Sensing as a Component of Climate Prediction In the climate context, the use of information from terrestrial remote sensing is still relatively immature. The initial MODIS products involve individual estimates of parameters at a given pixel and composited over enough scene views to obtain a variation of view angles. Views are rejected for cloud contamination and quality flags are set. What is still lacking is utilization of the expected strong space and time correlation of the data. A pixel with, e.g., a pine forest, will have neighboring pixels also classified as pine forest and for the next several years at least, they will mostly remain pine forests. An observation of neighboring pixels can be viewed as all having the same expected value plus a random noise element. There will be real differences in their albedo, LAI, etc. The random noise characterizations simply mean that these differences are too small and irrelevant to be of interest in detail.

17 Applications of Terrestrial Remote Sensing to Climate Modeling

457

Likewise, pixels should look the same from year to year except for random changes. Such changes could be as large as those caused by forest fires, which will certainly be of great interest for some issues, but may not be significant for, e.g., a climate model’s albedo over a continent. The state variables computed by a climate model are necessarily continuous in time and space whereas the pixel by pixel remote sensing products have many data holes caused by clouds, or more rarely by instrument failures. These holes must be filled before statistics can be obtained appropriate for use by climate modelers. Ideally, this filling should be done with no loss of real information. Moody et al. (2005) reports an approach to such filling for albedos measured by MODIS. What might be the optimum characterization of, e.g., pine forest albedos in July? First, what additional parameters besides the month and land cover description do we expect these albedos to depend on? Probably latitude, and fraction of tree cover need to be considered. After these correlations are quantified, there may still be significant variability that can be ascribed to characteristics of different understories. In principle, any significant variability can be inverted from its cause, provided the radiative model used is realistic enough to include this source of variability. Whatever remains is uninterpreted noise. Characterizing the amplitude of this residual noise is an important aspect of the data analysis. The state variables of vegetation change in time as a consequence of their interactions with climate. Various approaches have been developed to model this dynamical system. Ecosystems change their structure as a response to competition between different plant functional types (e.g., Lu et al., 2001; Sitch et al., 2003; Bonan et al., 2003; Woodward and Lomas, 2004; Krinner et al., 2005). These models have generated their leaves according to prescribed phenologies based on accumulated “degree days”, apparently by itself not on adequate constraint (e.g., Arora and Boer, 2005). The onset of leaves in Northern forests appears to have as much correlation with mid-summer temperature as degree days (Jenkins et al., 2002). In semi-arid or tropical systems, the phenology is largely controlled by the onset of rainy and dry seasons. Their dependences are shown in Fig. 17.4 (Zhang et al., 2005b). Detailed day by day dynamics of leaf growth can be attempted (e.g., Dickinson et al., 1992; 2002), but the underlying principles for this may be inadequately understood. Including dynamic growth of roots is another difficult modeling issue (e.g., Arora and Boer, 2003). The terrestrial state variable X moves forward in time, formally as a multivariate differential equation: dX + F(X, λ ) = Q(β ) (17.2) dt where λ are various fixed parameters, and Q is a forcing term depending on other parameters β . A satellite measures various reflectance quantities, denoted Yo , which consists of the “real” Y plus a measurement error term. The model can use its model value of X = Xm to provide an estimate of Y = Ym . For example, X can be some combination of the model soil moisture, or LAI or fc ; the latter two would be changed by dynamic vegetation versions of the model.

458

R.E. Dickinson 210

Date (DOY)

180 150 Forests Shrublands Savannas Grasslands

120 90 60 30 5

a

7

9

11 Latitude (⬚N)

13

15

17

210 Date (DOY)

180 150 2001

120

2002

90

2003

60 30 5

b

7

9

11 Latitude (⬚N)

30 25 20 15 10 5 0

c

15

25 Sahelian and subSahelian region Southern Africa

Frequency distribution (%)

Frequency distribution (%)

35

13

−80 −60 −40 −20

0

20

40

60

Sahelian and subSahelian region Southern Africa

20 15 10 5 0

80

Differnce in greenup onset and ORS (days)

17

0

d

20

40

60

80 100 120 140 160

Differnce in domancy onset and ORS (days)

Fig. 17.4 The phenology of vegetation over Africa as time of greenup from beginning of year (a, b) and (c, d) its timing relative to the onset (o) and end (e) of the rainy season (Zhang et al., 2005)

Assume Ym = g(Xm )

(17.3)

The term Ym will differ from Y either because Xm differs from the real X or from structural error in use of g. We can require that X and Y be adjusted to be as close as possible to their real values in an rms sense, i.e., find an optimum estimate Xˆ as the minimum value of 2 2 (17.4) J = X − Xˆ + Y − gXˆ .

17 Applications of Terrestrial Remote Sensing to Climate Modeling

459

ˆ is close enough where all parameters have been normalized by error estimates. If X to Xm , Eq. (17.4) can be linearized and solved by matrix computations. In this case, it reverts to linear least squares fitting, as appropriate for a statistical model with Gaussian error terms. The standard technologies of data assimilation in atmospheric models are summarized in Kalnay (2003). Detailed implementation should make use of known spatial correlations and error characterizations. Some aspects of this issue have already been addressed by hydrologists in the context of the fusion of microwave data and soil moisture modeling (e.g., Entekhabi et al., 2004; Reichle and Koster, 2005). For statistics that change a lot in time, i.e., depend on X, it has been found useful to do ensemble integrations of the system Eq. (17.2). A future target for such a data assimilation approach will be terrestrial carbon (e.g., Hese et al., 2005). Various satellites are under development to measure the variability of atmospheric CO2 with enough accuracy to infer terrestrial and oceanic sources and sinks. Such estimation can be substantially improved with the inclusion of these measurements in atmospheric data assimilation schemes along with terrestrial and oceanic process modeling. Assimilation of terrestrial reflectance imaging into models for the time evolution of canopy structure and LAI will become another major component of this activity. The implementation of such assimilation will require more advanced algorithmic approaches to canopy radiation so that the climate model simulations are realistic enough to reproduce the reflectances seen by remote sensing data.

17.8 Conclusions Climate models have become increasingly realistic in their descriptions of land surface processes. They absorb solar radiation and emit long wave radiation depending on structural details such as arrangement of leaves and soil moisture. Remote sensing instruments measure the same radiation as reproduced by climate models. Various aspects of vegetation change in time in response to variations in climate. Climate models have begun to include such changes in terms of “dynamic vegetation”. However, the climate models still use much less realistic treatments of terrestrial radiation than have been achieved by the remote sensing community. By simplification and improved efficiencies, the treatments of radiation used in remote sensing can be adopted into climate models. With such, many terrestrial remote sensing products should be derivable through forward calculations in data assimilating climate models. The logic, as used today, for atmospheric observations is that the climate model provides a first estimate, which in principle has incorporated in it all past observational information. Current observations then provide a correction to this a priori estimate. The resulting optimal estimate is essentially a weighted average of two estimates, where the weighting is determined by an understanding of correlated error statistics. Although assimilating terrestrial data should involve these general principles, the details of approaches needed will be vastly different

460

R.E. Dickinson

because of very different time and space scales of the dynamics and observations. In particular, vegetation is only spatially coupled on climate scales, has very fine spatial sampling structure, and its temporal sampling by satellite is sparse. Acknowledgements Yuhong Tian and Liming Zhou are thanked for their suggestions on the manuscript, including some content, Qing Liu for editing and Janet McGraw for her help with the manuscript. Also thanks to an anonymous reviewer. Support has been provided by the author’s NASA Grant NNG04GO61G; NASA Grant NG04GK87G, and S.L. Liang’s NASA Grant NNG04GL85G.

References Arora VK, Boer GJ (2003) A representation of variable root distribution in dynamic vegetation models. Earth Interactions 7(6): 1–19 Arora VK, Boer GJ (2005) A parameterization of leaf phenology for the terrestrial ecosystem component of climate models. Global Change Biol. 11(1): 39–59 Bonan GB, Levis S (2002) Landscapes as patches of plant functional types: an integrating concept for climate and ecosystem models. Global Biogeochem. Cycles, 16 (2), doi:10.1029/2000GB001360 Bonan GB, Oleson KW, Vertenstein M, Levis S, Zeng X, Dai Y, Dickinson RE, Yang, Z-L (2002) The land surface climatology of the community land model coupled to the NCAR community climate model. J. Climate 15: 3123–3149 Bonan GB, Levis S, Sitch S, Vertenstein M, Oleson KW (2003) A dynamic global vegetation model for use with climate models: concepts and description of simulated vegetation dynamics. Global Change Biol. 9:1543–1566 Dai Y, Zeng X, Dickinson RE, Baker I, Bonan GB, Bosilovich MG, Denning AS, Dirmeyer PA, Houser PR, Niu G-Y, Oleson KW, Schlosser CA, Yang Z-L (2003) The common land model (CLM). Bull. Amer. Meteor. Soc. 84(8):1013–1023 Dai Y, Dickinson RE, Wang Y-P (2004) A two-big-leaf model for canopy temperature, photosynthesis and stomatal conductance. J. Clim. 17(12):2281–2299 Dickinson RE (1983) Land surface processes and climate – Surface albedos and energy balance. In: B Saltzman (ed) Theory of climate, advances in geophysics. Academic Press, New York, vol 25, pp 305–353 Dickinson RE (1984) Modeling evapotranspiration for three-dimensional global climate models. Climate Processes and Climate Sensitivity. JE Hansen, K Takahashi (eds), American Geophysical Union Geophysical Monograph 29, Maurice Ewing, vol 5, pp 58–72 Dickinson RE (1989) Water and energy exchange. Chapter 6, Remote Sensing of Biosphere Functioning. R. Hobbs, H. Mooney (eds), Springer, New York, pp 105–133 Dickinson RE, Henderson-Sellers A, Rosenzweig C, Sellers PJ (1991) Evapotranspiration models with canopy resistance for use in climate models, a review, Agric. Forest Meteor. 54:373–388 Dickinson RE (1992) Land surfaces. Chapter 5, Climate Systems Modeling. K. Trenberth (ed), Cambridge University Press, Cambridge, pp 149–171 Dickinson RE (1995a) Satellite systems and Models for Future Climate Change. Chapter 7, World Survey of Climatology, 16, Future Climates of the World. A. Henderson-Sellers (ed), Elsevier, Amsterdam, pp 245–280 Dickinson RE (1995b) Land processes in climate models, International Satellite Land Surface Climatology Project special issue of Remote Sensing of Environment 51:27–38 Dickinson RE, Shaikh M, Bryant R, Graumlich L (1998) Interactive canopies for a climate model. J. Clim. 11:2823–2836

17 Applications of Terrestrial Remote Sensing to Climate Modeling

461

Dickinson RE, Berry JA, Bonan GB, Collatz GJ, Field CB, Fung IY, Goulden M, Hoffman WA, Jackson RB, Myneni R, Sellers PJ, Shaikh M (2002) Nitrogen controls on climate model evapotranspiration. J Clim. 15 (3):278–295 Dickinson RE, Wang G, Zeng X, Zeng Q-C (2003) How does the partitioning of evapotranspiration and runoff between different processes affect the variability and predictability of soil moisture and precipitation? Adv. Atmos. Sci. 20(3):475–478 Dickinson RE, Oleson KW, Bonan GB, Hoffman F, Thornton P, Vertenstein M, Yang Z-L, Zeng X (2006) The community land model and its climate statistics as a component of the community climate system model. J. Clim. 19:2302–2324 Entekhabi D, Njoku EG, Houser P, Spencer M, Doiron T, Kim Y, Smith J, Girard R, Belair S, Crow W, Jackson TJ, Kerr YH, Kimball JS, Koster R, McDonald KC, O’Neill PE, Pultz T, Running SW, Shi J, Wood E, Van Zyl J (2004) The Hydrosphere state (Hydros) satellite mission: an earth system pathfinder for global mapping of soil moisture and land freeze/Thaw. IEEE Trans. Geosci. Remote Sens. 42 (10):2184–2195 Friedl M, McIver D, Hodges J, Zhang X, Muchoney D, Strahler A, Woodcock C, Gopal S, Schneider A, Cooper A, Baccini A, Gao F, Schaaf C (2002) Global land cover mapping from MODIS: Algorithms and early results. Remote Sens. Environ. 83:287–302 Gao F, Schaaf CB, Strahler AH, Roesch A, Lucht W, Dickinson R (2005) MODIS bidirectional reflectance distribution function and albedo Climate Modeling Grid products and the variability of albedo for major global vegetation types. JGR 110 (D01104l), doi:l0.1029/2004-JD005190 Hansen MC, DeFries RS, Townsend JRG, Carroll M, Dimicelli C, Sohlberg RA (2003) Global percent tree cover at a spatial resolution of 500 metres: first results of the MODIS vegetation continuous fields algorithm. Earth Interactions 7, paper no. 10 Hese S, Lucht W, Schmullius C, Barnsley M, Dubayah R, Knorr D, Neuman K, Reidel T, Schr¨oter K (2005) Global biomass mapping for an improved understanding of the CO2 balance – the Earth observation mission Carbon-3D. Remote Sens. Environ. 94:94–104 Jenkins JP, Braswell BH, Frolking SE, Aber JD (2002) Detecting and predicting spatial and interannual patterns of temperate forest springtime phenology in the eastern US. Geophys. Res. Lett. 29(24): 54–56, doi:10.1029/2001GL014008 Jin Y, Schaaf CB, Gao F, Li X, Strahler AH, Zeng X, Dickinson RE (2002) How does snow impact the albedo of vegetated land surface as analyzed with MODIS data. Geophys. Res. Lett. 29, doi:10.1029/2001GL014132 Kalnay E (2003) Atmospheric modeling, data assimilation and predictability. Cambridge University Press, Cambridge, 341pp Krinner G, Viovy N, de Noblet-Ducoudre N, Ogee J, Polcher J, Friedlingstein P, Ciais P, Sitch S, Prentice IC (2005) A dynamic global vegetation model for studies of the coupled atmospherebiosphere system. Global Biochem. Cycles 19 (GB1015), doi:10.1029/2003GB002199 Liang X-Z, Xu M, Gao W, Kunkel K, Slusser J, Dai Y, Min Q, Houser PR, Rodell M, Schaaf CB, Gao F (2005) Development of land surface albedo parameterisation based on Moderate Resolution Imaging Spectroradiometer (MODIS) data. J. Geophys. Res. 110 (D11107), doi:10.1029/2004JD005579 Lu L, Pielke RA Sr, Liston GE, Parton WJ, Ojima D, Hartman M (2001) Implementation of a two-way interactive atmospheric and ecological model and its application to the central united states. J. Clim. 14:900–919 Matthews E (1984) Prescription of land-surface boundary conditions in GISS GCMII and Vegetation, land-use and seasonal albedo data sets: documentation of archived data tape. NASA Technical Memos 860096 and 86107, NASA, Goddard Institute for Space Studies, New York, 20pp and 9pp Moody EG, King MD, Platnick S, Schaaf CB, Gao F (2005) Spatially complete global spectral surface albedos: value-added datasets derived from Terra MODIS Land Products. IEEE Trans. Geosci. Remote Sens. 43 (l):144–158 Myneni RB, Hoffman S, Knyazikhin Y, Privette JL, Glassy J, Tian Y, Wang Y, Song X, Zhang Y, Smith GR, Lotsch A, Friedl M, Morisette JT, Votava P, Nemani RR, Running SW(2002) Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data. Remote Sens. Environ. 83:214–231

462

R.E. Dickinson

Oleson KW, Bonan GB, Schaaf C, Gao, F, Jin Y, Strahler A (2003) Assessment of global climate model land surface albedo using MODIS data. Geophys. Res. Lett. 30 (8):1443, doi:1029/2002GL016749 Pinty B, Lattanzio A, Martonchik JV, Verstraete MM, Gobron N, Taberner M, Widlowski J-L, Dickinson RE, Govaerts Y (2005) Coupling diffuse sky radiation and surface albedo. J. Atmos. Sci. 62:2580–2591 Pinty B, Lavergne T, Dickinson RE, Widlowski J-L, Gobron N, Verstraete MM (2006) Simplifying the interaction of land surfaces with radiation for relating remote sensing products to climate models. J. Geophys. Res. 111 (D02116), doi:10.1029/2005JD005952 Pitman AJ (2003) The evolution of, and revolution in, land surface schemes designed for climate models. Int. J. Climatol. 23:479–510 Reichle RH, Koster RD (2005) Global assimilation of satellite surface soil moisture retrievals into the NASA Catchment land surface model. Geophys. Res. Lett. 32 (L02404), doi:10.1029/2004GL021700 Sellers PJ (1985) Canopy reflectance, photosynthesis and transpiration. Int. J. Remote Sens. 6:1335–1372 Sellers PJ, Dickinson RE, Randall DA, Betts AK, Hall FG, Berry JA, Collatz GJ, Denning AS, Mooney HA, Nobre CA, Sato N, Field CB, Henderson-Sellers A (1997) Modeling the exchanges of energy, water, and carbon between continents and the atmosphere. Science 275:502–509 Sitch S, Smith B, Prentice IC, Arneth A, Bondeau A, Cramer W, Kaplan JO, Levis S, Lucht W, Sykes MT, Thonicke K, Venevsky S (2003): Evaluation of ecosystem dynamics, plant geography and terrestrial carbon cycling in the LPJ dynamic global vegetation model. Global Change Biol. 9:161–185 Strugnell NC, Lucht W, Schaaf C (2001) A global albedo data set derived from AVHRR data for use in climate simulations. Geophys. Res. Lett. 28(1):191–194 Tian Y, Dickinson RE, Zhou L, Zeng X, Dai Y, Myneni RB, Knyazikhin Y, Zhang X, Friedl M, Yu H, Wu W, Shaikh M (2004a) Comparison of seasonal and spatial variations of leaf area index and fraction of absorbed photo synthetically active radiation from Moderate Resolution Imaging Spectroradiometer (MODIS) and common land model. J. Geophys. Res. 109, doi:10.1029/2003JD003777 Tian Y, Dickinson RE, Zhou L, Myneni RB, Friedl M, Schaaf CB, Carroll M, Gao F (2004b) Land boundary conditions from MODIS data and consequences for the albedo of a climate model. Geophys. Res. Lett. 31(5), L05504, 10.1029/2003GL019104 Tian Y, Dickinson RE, Zhou L, Shaikh M (2004c) Impact of new land boundary conditions from Moderate Resolution Imaging Spectroradiometer (MODIS) data on the climatology of land surface variables. J. Geophys. Res. 109 (D20115), doi:10.1029/2003JD004499 Tsvetsinskaya EA, Schaaf CB, Gao F, Strahler AH, Dickinson RE, Zeng X, Lucht W (2002) Relating MODIS derived surface albedo to soil and landforms over Northern Africa and the Arabian peninsula. Geophys. Res. Lett. 29, doi:10.1029/2001GL014096 Wang Z, Zeng X, Barlage M, Dickinson RE, Gao F, Schaaf CB (2004) Using MODIS BRDF and albedo data to evaluate global model land surface albedo. J. Hydrometeorol. 5:3–14 Wang Z, Barlage M, Zeng X, Dickinson RE, Schaaf CB (2005) The solar zenith angle dependence of desert albedo. Geophys. Res. Lett. 32 (L05403), doi:10.1029/2004GL021835 Widlowski JL, Pinty B, Gobron N, Verstraete MM (2001) Detection and characterization of boreal coniferous forests from remote sensing data. J. Geophys. Res. 106 (D24): 33405–33419 Widlowski JL, Pinty B, Gobron N, Verstraete MM, Diner DJ, Davis AB (2004) Canopy structure parameters derived from multi-angular remote sensing data for terrestrial carbon studies. Climatic Change 67 (2–3):403–415 Wilson MF (1984) The construction and use of land surface information in a general circulation climate model. Unpublished Ph.D. thesis, University of Liverpool, United Kingdom, 346 pp Woodward FI, Lomas MR (2004) Vegetation dynamics – simulating responses to climatic change. Biol. Rev. 79:643–670, doi:10.1017/S1464793103006419

17 Applications of Terrestrial Remote Sensing to Climate Modeling

463

Yang Z-L (2003) Modeling land surface processes in short-term weather and climate studies. In: X Zu, X Li, M Cai, S Zhou, Y Zhu, F-F Jin, X Zou, M. Zhang (eds), Theory and modeling of atmospheric variability. World Scientific Series on Meteorology of East Asia, World Scientific, New Jersey, pp 288–313 Zeng X, Dickinson RE, Walker A, Shaikh M, DeFries RS, Qi J (2000) Derivation and evaluation of global 1-km fractional vegetation cover data for land modeling. J. Appl. Meteor. 39:826–839 Zeng X, Shaikh M, Dai Y, Dickinson RE, Myneni R (2002) Coupling of the common land model to the NCAR community climate model. J. Clim. 15:1832–1854 Zeng XB, Dickinson RE, Barlage M, et al. (2005) Treatment of undercanopy turbulence in land models. J. Clim. 18(23):5086–5094 Zhang Q, Xiao X, Braswell B, Linder E, Baret F, Moore B III (2005a) Estimating light absorption by chlorophyll, leaf and canopy in a deciduous broadleaf forest using MODIS data and a radiative transfer model. Remote Sens. Environ. 99:357–371 Zhang X, Friedl MA, Schaaf CB, Strahler AH (2005b) Monitoring the response of vegetation phenology to precipitation in Africa by coupling MODIS and TRMM instruments. J. Geophys. Res. 110 (D12103), doi:10.1029/2004JK005263 Zhou L, Dickinson RE, Tian Y, Zeng X, Dai Y, Yang Z-L, Schaaf CB, Gao F, Jin Y, Strahler A, Myneni RB, Yu H, Wu W, and Shaikh M (2003) Comparison of seasonal and spatial variations of albedos from Moderate-Resolution Imaging Spectroradiometer (MODIS) and Common Land Model. J. Geophys. Res. 108(D15):4488, doi:101029/2002JD003326 Zhou L, Dickinson RE, Tian Y (2005) Derivation of a soil albedo dataset from MODIS using principal component analysis: Northern Africa and the Arabian Peninsula. Geophys. Res. Lett. 32 (L21407), doi:10.1029/2005GL024448

Chapter 18

Improving the Utilization of Remotely Sensed Data John R. Townshend, Stephen Briggs, Roy Gibson, Michael Hales, Paul Menzel, Brent Smith, Yukio Haruyama, Chu Ishida, John Latham, Jeff Tschirley, Deren Li, Mengxue Li, Liangming Liu, and Gilles Sommeria

Abstract This chapter reports on the work of a team set up by the Committee on Earth Observation Satellites to examine the factors encouraging the increased use of space data. The main emphasis of the study was on situations where the desired outcome is a sustained capability rather than increasing the number of one-off users. Twenty-five case studies were compiled by the team. Examination of the successful case studies indicates that more of the factors favoring adoption of remote sensing products by users are within the province of the users themselves, rather than John R. Townshend University of Maryland, College Park, USA [email protected] Stephen Briggs European Space Agency Roy Gibson EUMETSAT Michael Hales, Paul Menzel and Brent Smith NOAA Yukio Haruyama and Chu Ishida JAXA John Latham and Jeff Tschirley Food and Agriculture Organization Deren Li Wuhan University Mengxue Li NRSCC Liangming Liu National Remote Sensing Center of China Gilles Sommeria WCRP

S. Liang (ed.), Advances in Land Remote Sensing, 465–483. c Springer Science + Business Media B.V., 2008

465

466

J.R. Townshend et al.

with the space agencies, but successful adoption always involves balanced cooperation between space agencies and users. In this cooperation space agencies can and should do more to ensure the greater uptake of their data and services. Critical factors from the space agency point of view are provision of user-specific products; making resources available speedily and with a minimum of bureaucracy; willingness to listen over an extended period to user needs and practical difficulties and to accommodate them to the maximum extent practicable. From the user standpoint, success appears to be very dependent on (a) the user having an understanding, and preferably first-hand experience, of the use of remote sensing data and products, and its integration into geospatial information systems; (b) a willingness to make appropriate staff available before and during the project to work with the space agency; (c) the user having a core need, such as a statutory need, for the information; (d) an ability to define rather precisely what the needs are; and (e) the user being willing and able to help develop a realistic financial scenario for an operational system. On both sides, the case studies indicate the need for innovation and flexibility in response to new opportunities for data products.

18.1 Introduction A wide range of sensors regularly supply observations of the Earth’s environment and ever improved capabilities are provided to meet increasingly stringent demands are frequently made to satisfy users (e.g., Duchossois and Sommeria, 2003). Achieving enhanced use of observations depends on many factors including the successful transition from research to operational observations (NRC, 2000, 2003a). Another major factor is data policy (Harris, 2002) which can strongly impact the availability of data through charging policies. Increasing the utilization of these data represents an improved return on the considerable investments in the space and ground segments (NRC, 2003b). There is a widely held belief that such data are often underutilized. To facilitate and encourage the increased use of space data and the products derived from them, the Committee on Earth Observation Satellites (CEOS) set up a Utilization Team to advise space agencies about their roles in achieving these goals (CEOS, 2003). The approach adopted by the Utilization Team was first to identify case studies, where adoption of remote sensing was both rapid and successful and then to examine what can be inferred from an analysis of them. The advantage of this approach was to ground the Team’s work and advice on concrete examples. Second the case studies were used to identify impediments, the necessary preconditions for adoption of remote sensing and the ways to facilitate utilization. Finally using the previous considerations specific recommendations were made to space agencies to increase utilization. In this chapter we next describe the procedures adopted, including the assembly of case studies as well as insights from interviews, which authors conducted with experienced users of satellite data. We then identify and analyse the factors, which

18 Improving the Utilization of Remotely Sensed Data

467

either facilitate or hamper the increased use of satellite data, products and services. These considerations are used to derive a set of principles, which we believe will lead to enhanced adoption of space-based data and products. Finally we summarize the recommendations which were addressed to CEOS as a body and to individual members and associates.

18.2 Approach and Procedures Basic to this study is that increased utilization of space-derived data, products and services is a good thing in its own right. While all space agencies undoubtedly wish to see the widest use of data acquired by the satellite systems for which they are responsible, it should be recognised, that some space agencies do not have a specific remit to be proactive in promoting the use of their data for users for whom the Earth observation activity was intended, nor do they necessarily have a mandate to undertake the regular re-processing necessary to ensure consistent term data sets over the longer term, crucial to many applications. We also consulted a number of experts on these issues, who are recognized in the acknowledgments. The main emphasis of the study was on situations where the desired outcome is a sustained capability rather than increasing the number of one-off users. We examined increased use of existing data, as well as increasing the potential use of future satellite missions. We considered both those situations where immediate increases in utilization are realisable as well as those where sustained collaboration between space agencies and users may be needed before operational applications can be developed. For each case study a template was completed, which provided the following information: • Characterizing the activity • Providing a chronology • Identification of the factors encouraging adoption • The principal obstacles that were overcome and how this was achieved • Related spin-off applications • The lessons learnt by the authors from this case study The templates were completed by the authors through interviews with both the users and the responsible space agencies and their partners. In Fig. 18.1 there is the completed template for one of the case studies concerned with rapid response of fires. Table 18.1 contains a listing of all of the 22 case studies compiled by the Team with the names of space agencies and partners as well as the users of the data. Discussions of potential users imply that there are always many more waiting to be found. This is not the case for all types of users: some markets may already be close to saturation. For example, although weather forecasting centres may need improved observations, the number of such centres is unlikely to grow and may even contract. Hence there may be situations where there may only be the prospect of an improved use of remote sensing, without any actual increase in the number of users.

468

J.R. Townshend et al.

CEOS ad hoc team on utilization: implementation case study template Name of activity MODIS Fire Rapid Response System Further information (URL) http://rapidresponse.umd.edu/ Name of space agency/ies and partners NASA, University of Maryland, USFS Name of adopting agency/ies USFS and the National Inter-Agency Fire Center Brief description of activity 1. Images from MODIS are down-loaded, processed and placed in a GIS/map format for use by NIFC and the USFS on a daily basis for the whole of the conterminous US plus Alaska. 2. Areas of active fires and cumulative burn scars are identified on a topographic map base showing relief, main roads the urban interface, and administrative and state boundaries. 3. The products are used by NIFC to make strategic resource allocation decisions (i.e., where to redeploy firefighters and major equipment such as aircraft as opposed to scale local tactical fire fighting decision-making) Chronological description of process of adoption 2000 USFS learnt of the availability of fire products from MODIS and examined examples prepared on an ad hoc basis. 2001 Fully implemented prototype boot-strapped system developed, using NASA, NOAA and UMD and USFS facilities to provide products on a daily basis to NIFC in time for the 2001 fire season (June 1). Algorithms simplified, main EOS-DIS system by-passed and completely automated implementation to allow sufficiently rapid delivery of products. In practice products are available 2–4 h after imaging. 2002 USFS takes over complete operations through use of own direct receiving system for western US and transfer to them of all needed software. The NASA system now acts as a back-up. NASA and UMD involvement moved to joint research in developing new products (e.g., burn intensity in addition to providing eastern US and Alaska data beyond USFS station coverage). Factors favoring adoption 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.

Adopting agency had existing experience in using remote sensing. Adopting agency had clear, strong need for products. Pre-existing scientific expertise in algorithm development for fire detection. Close working relationship developed between algorithm/system developers and adopting agency Funds quickly deployed by space agency and adopting agency allowing prototype system to be implemented. Highly competent technical environment (at NASA, USFS and UMD) allowing information system components to be built and integrated from COTS and reuse of hardware. Funds quickly deployed to allow USFS to purchase dish allowing them take over as operational agency. Products delivered in form specified by users - paper maps with information kept to the minimum required for decision-making. New products compatible with heritage formats and procedures. Those developing the system passed over operations as quickly as possible to adopting agency. Developers interacted directly with users (e.g., RR staff visited active wildfires and spoke with range of users including management, direct attack, and rehabilitation team members.

Principal obstacles that were overcome (indicate whether and how they were overcome) 1. Routine EOS-DIS data processing system for science had to be by-passed to allow sufficiently rapid delivery. 2. Need for non-operational prototype was avoided. Strictly speaking no true prototype. Other related spin-off applications 1. Rapid response system implemented for whole world with reflectance background and fire locations identified (http://rapidfire.sci.gsfc.nasa.gov/). 2. Products distributed through UN system. 3. Products made available regionally through GOFC-GOLD mechanism. Lessons learned from case study 1. 2. 3. 4. 5.

This example worked exceedingly well and quickly. Technical expertise played a major part. But so did the receptivity of the adopting agency. The quickness of the funders to back a winner when they saw one. Perhaps this example may be atypical in that everything was in its favor.

Fig. 18.1 Example of completed case study template

18 Improving the Utilization of Remotely Sensed Data

469

Table 18.1 Case studies where remote sensing was adopted successfully and rapidly Name of case study

Responsible space agency (and partners)

User agency

1

MODIS Fire Rapid Response System

NASA

2

Application of the HY-1 satellite for sea ice monitoring and forecasting (the sea ice retrieval system) Meteorological Satellite Fire Monitoring System

Chinese National Oceanic Satellite Application Center

US forest Service and National Inter-agency Fire Center National Marine Environment Forecasting Center

4

Integrated operational system of flood monitoring and assessment

5

Monitoring system of ice conditions for the Yellow River Crop monitoring system using remote sensing

Remote Sensing Technology Application Center, Ministry of Water Resources Institute of Geography, CAS National Remote Sensing Center of China and Wuhan University Institute of Remote Sensing Applications, CAS

3

6

National Satellite Meteorological Center of China

The application of CBERS-1 data in forest resources investigation Application of satellite remote sensing to fishing industry Collaborative research on sea ice observation by Earth Observation data Collaborative research on development of land utilization distribution by Earth Observation satellite data MODIS polar winds

China Center for Resources Satellite Data and Application (CRESDA) JAXA

12

Atmospheric Sounder data utilization

NOAA

13

Assimilation of satellite data in Numerical Weather Prediction

NOAA, Eumetsat

7

8

9

10

11

JAXA

JAXA

NASA, NOAA

National Satellite Meteorological Center of China, Ministry of Agriculture, China Forestry Administration Chinese Ministry of Water Resources

Information Center of Yellow River Conservancy Commission State Food Bureau, Committee for State Planning Development Guizhou Forestry and Planning Institute Japan Fisheries Information Service Center Hydrographic and Oceanographic Department, Japanese Coast Guard Kumamoto City Environmental Research Institute

European center for Medium Range Weather Forecasting US National Centers for Environmental Protection, European Center for Medium Range Weather Forecasting, UK Met. Office European Center for Medium Range Weather Forecasting (Continued)

470

J.R. Townshend et al.

Table 18.1 Continued Name of case study

Responsible space agency (and partners)

User agency

14

Regional Burned Forest Mapping

Italian Department of Civil Protection (DPC)

15

Kyoto Inventory

European Space Agency (ESA) with Telespazio, Intecs, Vitrociset, Cesia and Mestor (Italy) ESA, Intecs (IT), Planetek (IT), NEO (NL), Telespazio (IT), Dataspazio (IT), Bosdata (NL), Gamma Remote Sensing (CH), University of Trento (IT) + precursor study: VTT (FIN), European Forest Institute, Stora Enso Forest Consulting (FIN)

16

Urban expansion and monitoring (URBEX)

17

Oil Spill Monitoring Center

18

Ecosystem (environmental) monitoring

NASA, ESA, CNES

19

Ozone alert

ESA, KNMI

20

Surface albedo

21

Sea level and ocean circulation real time monitoring

EUMETSAT and Joint Research Center, Ispra CLS, CNES, EUMETSAT, JPL NASA, NOAA, US Navy

22

EnviroCast (remote sensing products for the media)

NASA, NOAA, EPA, USDA Forest Service

European Space Agency (ESA), Advance Computer Systems (ACS), Norwegian Space Centre, Kongsberg Satellite Services (KSAT)

Italian Ministry of Environment and Territory; Swiss Agency for the Environment, Forest and Landscape; Finnish Ministry of Environment/METLA; Norwegian Ministry of Agriculture/NIJOS; Dutch Ministry of Agriculture, Nature Management and Fishery World Wildlife Fund (WWF) State Pollution Authorities until 2002. Norwegian National Coastal Administration from 2003 Various NGOs including The Nature Conservancy, Conservation International, and World Resources Institute ECMWF, Met. Offices, Education Institutes, Scientific Measurements Campaign Planners, Environmental Agencies, L’Oreal EUMETSAT initially for distribution ECMWF, FNMOC, FOAM, MERCATOR, MFS, NAVOCEANO, NCEP, NOAA, NRL Private company (StormCenter Communications, Inc.) for distribution to media

18 Improving the Utilization of Remotely Sensed Data

471

It is worth noting that, compared with even 10 years ago, the costs of many elements of information systems are much lower, including costs of local ground receiving stations and computer hardware and software. These all combine to make it less expensive to integrate remote sensing products into environmental information systems, and hence potentially increases the number of users.

18.3 What Factors Favor Adoption of Remote Sensing? 18.3.1 Analysis of Case Studies Table 18.2 lists the factors identified in the case studies as favoring the adoption of remote sensing by users and the frequency of their occurrence in the case studies. Those completing the studies were asked to identify only the most important factors but were not given any fixed number of factors that had to be selected. Examination of the successful case studies indicates that more of the factors favouring adoption of remote sensing products by users are within the province of the users themselves, rather than with the space agencies, but it needs to be stressed that successful adoption always involves balanced cooperation between space agencies and users. In this co-operation space agencies can and should do more to ensure the greater uptake of their data and services. Table 18.2 Occurrence of factors important in case studies where remote sensing technology was rapidly and successfully adopted Types of factors encouraging adoption Space agency’s and partner’s factors Additional resources provided for new uses Provision of technical capability Provision of ready access to data Provision of data at affordable costs Products in user-specified forms (incl. timeliness) Inter-agency factors Strong inter-agency cooperation Rapid transfer of capability to users External linking expertise linking users and space agencies User factors Information needs strong and well quantified Additional resources allocated Resources to install ground receiving capability In-house scientific and remote sensing expertise Existing operational use of remote sensing Existing use of geospatial technologies Commitment to use information operationally Provision of field and validation information Demonstrated cost savings

Number of occurrences 7 6 3 4 10 11 2 6

19 7 1 13 6 6 6 3 2

472

J.R. Townshend et al.

Critical from the space agency standpoint is (a) provision of user-specific products; (b) making resources available speedily and with a minimum of bureaucracy; (c) willingness to listen over an extended period to user needs and practical difficulties; and (d) to accommodate these difficulties to the maximum extent practicable – going so far sometimes as to incorporate user representatives into a joint project team. Added to these factors, several case studies stressed the importance of keeping user development projects simple, and building further on reliable, successful initial projects. However, there is evidence to suggest that turning prototypes into sustained operations is frequently unsuccessful if the funding from users does not continue beyond the prototype phase. Space agencies therefore need at a relatively early stage to seek assurance that the user has sufficient interest and capacity to continue the project beyond the prototype stage. From the user standpoint, success appears to be very dependent on (a) the user having strong needs for the information and that these needs are well characterized; (b) the user having an understanding, and preferably first hand-experience, of the use of remote sensing data and products, and their integration into geospatial information systems; (c) a willingness to make appropriate staff available before and during the project to work with the space agency; and (d) the user being willing and able to help develop a realistic financial scenario for an operational system. On both sides, the case studies indicate the need for innovation and flexibility in response to new opportunities for data products. Equally important is the role that individuals can play. Sometimes, these are senior champions in either space agencies or less frequently in user organisations. In other cases it may be an individual facilitator who has experience in new remote sensing technologies and the development of algorithms and who also has close contacts with user communities; this can lead to speedy adoption by new users. Not surprisingly strong cooperation between space and user agencies was frequently recognized as being key. The development of partnerships between users, space agencies and research organizations including universities often figured prominently. In several case studies industry also played a significant role in increasing utilization.

18.3.2 Technical Factors Encouraging the Adoption of Remote Sensing Analysis of case studies showed that the principal obstacles faced, and to varying extents overcome, were overwhelmingly technical (Table 18.3) and fell into the following categories: • Need to provide adequate bandwidth for data access • Devising adequate distribution networks • Need to provide new or improved algorithms

18 Improving the Utilization of Remotely Sensed Data

473

Table 18.3 Occurrence of significant obstacles overcome in case studies, where adoption was rapid and successful Obstacles overcome Bandwidth Inadequate characterization of observations Need for improved algorithms Need for improved information system Gaining sufficient resources Improving timeliness Lack of trained personnel Lengthy prototyping Lack of understanding of potential of Remote Sensing Proving impact

Number of occurrences 5 1 10 4 3 4 1 2 2 2

• Development of faster data processing techniques (e.g., calibration techniques for SAR images) • Need to improve the basic classification systems to which the remote sensing project is designed to contribute • Need to demonstrate that continuity of space observations can be assured However, the importance of technical obstacles relative to other kinds might have been smaller had the study also considered unsuccessful case studies. Where major improvements in observational capabilities occur there is often immediate potential for improved use. For example, the availability of moderate spatial resolution MODIS (Moderate resolution Imaging Spectrometer) data with its much higher capabilities compared with the AVHRR (Advanced Very High Resolution Radiometer) has led to a number of rapid improvements in utilization. The use of MODIS data to assist fire-fighting in the United States or the use of MODIS polar winds in ECMWF (European Center for Medium Range Weather Forecasting) assimilations may be mentioned. Adoption of MODIS is encouraged by the planned follow-on of the VIIRS (Visible Infrared Imaging Radiometer Suite) sensor of NPOESS (National Polar Orbiting Environmental Satellite System) and its predecessor NPP (NPOESS Preparatory Project), with similar capabilities to MODIS. Such new sensing capabilities offer both the possibility of enhancing systems already using remote sensing and the possibility of significant new uses of the data. One notable example of a space agency positioning itself for new uses of its data belongs to EUMETSAT who has set up seven Satellite Application Facilities (SAFs) to exploit the increased capabilities of Meteosat Second Generation for non-weather and climate applications. This offers a potentially valuable model for other agencies concerned with increased exploitation of new observational capabilities. Enhancement in observational capabilities can also arise from improved acquisition strategies. The use of Landsat data increased substantially with the introduction of a much improved acquisition strategy for Landsat 7 greatly increasing coverage; increased use also stemmed from a sharp reduction in costs.

474

J.R. Townshend et al.

Where there will be major changes in observational capabilities in the foreseeable future it is normally incumbent on the responsible agencies to ensure that users are aware of these. Agencies should also consider the desirability of ensuring the provision of appropriate training and contributing to capacity building especially where it is possible to identify existing operational users who are likely to benefit from these improvements.

18.3.3 Education and Capacity Building Encouraging the Adoption of Remote Sensing Lack of trained staff and modern hardware and software were, not surprisingly, also encountered. Remote sensing can in some respects be regarded as a relatively mature technology and, as a result, in some sectors many potential users are well aware of its ability to meet their needs. However, this does not always apply in developing countries, and even users in the developed word may need additional training (as well as updated hardware and software) if they are to make best use of new data sets, which often have much higher data volumes. The availability of the technical capability to improve algorithms is clearly important in many of the case studies. To increase utilization, it appears preferable that education and capacity building is carried out in conjunction with a practical project, rather than performed generically. Where remote sensing products are being introduced it is essential at an early stage remote sensing realistically to assess the amount of education and training which will be needed in order to ensure that the project will successfully develop into an operational capacity. The understandable tendency to underestimate the cost of this element will certainly militate against success.

18.3.4 Financial Factors Affecting the Adoption of Remote Sensing There is no substitute for users being willing to spend their own time and resources in the adoption of remote sensing. Such resources will be used on staff time being allocated in the prototyping phase and in expenditures on information systems and outreach to the decision-makers and other end-users of the resultant products. Alongside space agencies and new users, third parties can also play an important role. This may be in a funding role, as in the case of the European Commission in Europe or the World Bank. Such financers rightly insist on the project being designed realistically and economically to meet user needs. In Europe, private companies often also play a major role in developing user capabilities and in linking users with space agencies.

18 Improving the Utilization of Remotely Sensed Data

475

18.3.5 Policy Factors Affecting the Adoption of Remote Sensing Two other obstacles, of a somewhat different nature, are apparent: restrictions on data access because of local data policies, and the regular complaint of the high cost of satellite data and products. It was beyond the remit of the authors to deal in depth with these two factors, but it would be unrealistic not to identify them. Perhaps the most that can be said is that space agencies should be encouraged to release data expeditiously and at the lowest cost, whenever they are needed in connection with an approved project or for a use consistent with the work which space agencies promote such as uses recognized by the World Summit on Sustainable Development and with the international environmental conventions.

18.4 Principles to Encourage the Adoption of Remote Sensing Based on the case studies we can identify three main ways in which the use of remote sensing can be enhanced. • First there are those situations where users already have information or forecast systems using remotely sensed data. Improved observations become available and these are introduced to the existing operational system. Even in these circumstances increasing utilization, may not be a trivial exercise and usually requires active cooperation between user and space agencies. • Second there are situations where users have not normally used remote sensing. This will usually require a significant effort for the needed skills and facilities to be transferred. Many of the case studies include such examples. • Third, there are situations where space agencies or contracted intermediaries are made responsible for base-lined standard products, with well-characterized properties in terms of Calibration and Validation and Quality Assessment. In these situations many users could adopt products without the need greatly to add to their technical knowledge and facilities for remote sensing.

18.4.1 Assurance of Continuity For users to adopt remote sensing there is a need for systems to have ensured continuity. Security of data and product supply is essential and operational users have to receive firm assurances to be sure that there will be long-term continuity. Operational agencies need to be assured that the significant initial costs incurred in changes in their business practices to accommodate the use of satellite data will be recovered through guaranteed long-term access to satellite data.

476

J.R. Townshend et al.

18.4.2 Need for Adequate Information Systems and Spatial Data Infrastructure In many parts of the world successful information systems may exist and the benefits of remote sensing may appear marginal compared with existing in situ sampling schemes. But in other parts of the world adequate environmental information systems often do not exist, especially in the developing world. Where aid is provided it is vital that there is not merely investment in receiving station capability (in some cases there are too many receiving stations which seem to be regarded as status symbols), but also in end-to-end capabilities from acquisition to product generation and use. Although there may be considerable potential for application of remote sensing, sustaining such capabilities in the poorest countries of the world may be difficult without long-term aid.

18.4.3 Space Agencies’ Responsibilities for Information Systems Although in some cases it is not included expressly within their mandates, space agencies have key long-term stewardship responsibilities in ensuring data delivery, data archiving and reprocessing for climate-quality data sets. Users have to be sure that these will be executed regularly and to the highest possible standards

18.4.4 Enhancing Data Delivery Systems In some situations data systems may be designed for scientific researchers and be inadequate for other applications. For example delivery of data for scientific uses may be quite acceptable even if days or weeks after acquisition. Providing a capability to access and process data almost instantaneously significantly increases the range of possible applications. This helps explain the increasing popularity of local ground receiving stations. One noteworthy example of this is the Advanced TIROS Operational Vertical Sounder (ATOVS), where direct broadcast has been maintained and a preprocessing package was created and distributed with international collaboration.

18.4.5 Ensuring that Consistent Long-Term Data Sets are Created Many uses of remotely sensed data rely on internally consistent long-term data sets especially in relation to understanding climate change and assessing its impacts on socio-economic activities. From time to time major new uses of environmental records may arise and it would be prudent for space agencies to be prepared to respond to these. Creation of such data sets usually requires very substantive efforts normally involving a considerable amount of reprocessing and analysis.

18 Improving the Utilization of Remotely Sensed Data

477

18.4.6 Access to Integrated Assets of Multiple Agencies There is a need for users to be able to use the integrated assets of multiple agencies; essential to ensure increased usage. This will often involve the merging and integration of data sets by data fusion. There is a key role for universities and research efforts outside of space agencies. We recognize that individual space agencies have to promote their own EO assets, and there is a continuing need for such promotion to be realistic in terms of capabilities. There is a need for independent mechanisms to promote integration and to carry out independent assessments of capabilities.

18.4.7 Role of Research in Stimulating Innovative Uses of Remote Sensing It is important to highlight the importance of research in general. Examples of areas where significant recent strides have been made include the development of SAR Interferometry and research in Numerical Weather Prediction. Development of applications often arises on the basis of research. Also there is a need to recognize the importance of integrated research activities between the industrial and research communities. The key roles of research institutions must be recognized in addition to the roles of space agencies and users.

18.4.8 Improved Cooperation Between Space Agencies There is a need for mechanisms to avoid potentially unhelpful duplication between space agencies. Proposals for a more integrated approach between NASA’s Pathfinder and ESA’s Explorer program are to be welcomed. Cooperation will needed as early as possible, including the definition and selection phases. The optimal situation in terms of international cooperation may often be where users pose key needs which receive a coordinated response; an exemplary case of this is of which Jason-2. Where there are international jointly funded programs such as the Global Precipitation Mission (GPM) or other missions with multiple international components, consideration should be given to joint science teams for the above activities.

18.4.9 Sustained Capacity Building With improvements in observational capabilities there will be a continuing need for capacity building in developing and developed and countries. Such outreach needs

478

J.R. Townshend et al.

to be embedded within continuing projects to secure the long-term benefits available from increased capability. This must be supported by a continuing stream of high quality technical and educational publications and outreach to users.

18.4.10 Need for Strategies to Cope with the Long Time Periods Between Problem Identification and the Arrival of an Operational System Strategies are needed to deal with very long times between problem identification and provision of comprehensive observational solutions. The timescale of response to the relatively simple problem of global stratospheric ozone, ideally suited to remote sensing techniques, is indicative of the long time constants inherent in addressing such problems. The long time period needed and gradual evolution of the system to introduce the use of operational remote sensing satellites for weather forecasting in Europe is illustrated in Fig. 18.2.

18.4.11 Recognition of New Partners and the Distinctive Needs of New Users During the last few years major new types of users are beginning to adopt remote sensing. These include NGOs who require observations for purposes such as conservation and humanitarian relief. Others include media organizations requiring very timely information. Space agencies will need to recognize and adapt to their special needs if adoption is going to be successfully achieved by these new sets of users. In developing partnerships with the private sector this should not automatically be interpreted as meaning the aerospace industry, but new partners, such as insurance companies, already providing services on a large scale through non-space means.

18.4.12 Political Ownership of Benefits There is a need to promote uses that have a high political impact. The benefits for political leaders have to be included, since political ownership of the benefits of remote sensing will encourage investment in the adoption of remote sensing. It is also important that legislators are kept aware of major advances in capabilities, which could influence the introduction of legislation, e.g., reliable monitoring of tropospheric ozone could conceivably lead to the introduction of legislation based on this capability.

18 Improving the Utilization of Remotely Sensed Data A)

479

Weather Forecasting 40 years ago in Europe Data ……………………………………...….to …………….……………………………. Information Feedback

USERS

N

Forecasts

E

Global System design

E

In situ Data

Govt Public Private

D S

Modeling,

OPS

Data Assimilation

RTD

EC-Funded Research in Universities

B)

Weather Forecasting 20 years ago in Europe Data ……………………………………...….to …………….……………………………. Information Feedback

USERS

N

Modeling,

E

Global System design

E

In situ Data

Forecasts

Data Assimilation

Govt Public Private

D

OPS

S

RTD Satellite Data ESA EC - Funded Research in Universities

C)

The Example of Weather Forecasting today in Europe Data ……………………………………...….to …………….……………………………. Information Feedback

USERS

N E E D

In situ Data

Global System design

Modeling, Data Assimilation

Satellite Data

Forecasts

Govt Public Private

EUMETSAT

OPS

S

RTD ESA, National agencies

FP6 Universities

Fig. 18.2 The evolution of the utilization of remotely sensed data in European weather forecasting. (Courtesy of D. Williams, Eumetsat.)

480

J.R. Townshend et al.

18.5 Recommendations to Space Agencies The preceding sections contain information that the authors hope will be of interest and use to at least some readers. In conclusion we emphasize a few major recommendations made to space agencies through CEOS. 1. Real-time access to data streams is leading to increasing use of remotely sensed data, especially where preprocessing packages are made available to users. The latter means that the initial responsibility to ensure the processing and delivering of products rests with the agency rather than the user. It is important that agencies make pre-processing packages available to the greatest extent possible, whether they are developed in-house or in the research or commercial communities. Real-time access includes but is not limited to direct broadcast. This means that space agencies need to make data streams available in real time and whenever possible provide any necessary preprocessing software, a benchmark data set and a focal point for questions. 2. The quality of information about instruments, made available by space agencies, varies considerably. Many space agencies need to reevaluate the way in which they provide information about past, current and future systems including the provision of high quality up-to-date information concerning instrument performance especially calibration and product quality assessments. A more standardized approach for the description of instrument performance should be adopted. 3. Best use of remotely sensed data increasingly depends on use of data from multiple sensors often from multiple agencies, but this is often significantly hindered by the lack of coordination between information systems. Space agencies need to develop information systems with more integrated catalog, search, ordering and retrieval mechanisms. 4. Increasing the utilization of remotely sensed data is dependent on developing improved partnerships with users. This is, of course, very much a matter for individual agencies, and much has already been done. Nevertheless, space agencies should commit to seeking ways of increasing and improving partnerships, including those with NGOs and universities. These could include the following but need not be limited to them: • Improve the representation of scientific and other users on their advisory bodies. • Work more intensively with the scientific community to develop new applications and services and specifically work to develop improved assimilation procedures. • Consider the possibility of selectively improving the funding of new research and other application activities. • Consider coordination between agencies in providing funding for research and other applications. • Where there are internationally coordinated remote sensing activities such as the GPM, then consider setting up international science panels in conjunction with appropriate international research organizations.

18 Improving the Utilization of Remotely Sensed Data

481

• Establish closer links with the scientific and political arms of the international environmental conventions and agreements. • Strengthen links with policy makers and international conventions for high profile data utilization in keeping with the high priority given to this by space agencies and their coordination organizations such as CEOS. Reflecting the strong emphasis of most space agencies on planning Earth observation missions for the purpose of climate and global environment studies, particular attention should be placed on improving the dialogue with relevant user communities for the planning, specification, and application of these missions. This should include establishing closer links with the scientific and political arms of the climate conventions (e.g., the UN Framework Convention on Climate Change and its Subsidiary Body for Scientific and Technical Advice) and should take account of high priority information needs as defined by bodies such as the Inter-government Panel on Climate Change. 5. Use of remotely sensed data is sometimes restricted by the fact that although products have been demonstrated in a research environment and have been produced at a regional level their value at a global level has not been demonstrated and recognized by the users. Examples include global albedo, fire products and other high priority ones identified in the Integrated Global Observation Strategy themes (IGOS, 2004) and more recently by the Group on Earth Observations

Table 18.4 GCOS Climate Monitoring Principles (http://gosic.org/GCOS/GCOS climate monitoring principles.htm) Effective monitoring systems for climate should adhere to the following principles 1. The impact of new systems or changes to existing systems should be assessed prior to implementation 2. A suitable period of overlap for new and old observing systems is required 3. The details and history of local conditions, instruments, operating procedures, data processing algorithms and other factors pertinent to interpreting data (i.e., metadata) should be documented and treated with the same care as the data themselves 4. The quality and homogeneity of data should be regularly assessed as a part of routine operations 5. Consideration of the needs for environmental and climate-monitoring products and assessments, such as IPCC assessments, should be integrated into national, regional and global observing priorities 6. Operation of historically uninterrupted stations and observing systems should be maintained 7. High priority for additional observations should be focused on data-poor regions, poorly observed parameters, regions sensitive to change, and key measurements with inadequate temporal resolution 8. Long-term requirements should be specified to network designers, operators and instrument engineers at the outset of system design and implementation 9. The conversion of research observing systems to long-term operations in a carefully planned manner should be promoted 10. Data management systems that facilitate access, use and interpretation of data and products should be included as essential elements of climate monitoring systems

482

J.R. Townshend et al.

(GEO) as having critical deficiencies. Space agencies members should provide mechanisms to allow the creation of these and other key global data sets. One possible mechanism could be for multiple agencies to make their data available and then for one or maybe two agencies to produce them globally using different software. Inter-comparisons and validation of the results would be carried out in close coordination with representative user groups. 6. A restrictive data policy can often significantly limit access to data and products and hence inhibit utilization. Commercial considerations and the need to prioritize the allocation of limited ground segment resources, for example, the planning, acquisition and delivery of high resolution data sets are legitimate reasons for restricting access to data, but, these apart, space agencies need to make data easily and cheaply available, especially when they are in response to a specific request consistent with efforts to promote sustainable development in its widest sense. 7. Capacity building, education and training remain high priorities in developing and developed countries and will remain a significant responsibility of space agencies. Experience from the case studies shows that these are often most effective in transferring capacity to local users, when performed within the context of specific projects, when the support of key decision makers is included and when working in partnership from the very beginning with the end users. 8. The case studies showed that creating long-term consistent records of the environment is of the highest importance for many key users of remotely sensed data: it is also extremely challenging for agencies responsible for their creation. As one step in this direction, it was recommended therefore that space agencies make best efforts to adopt the GCOS Climate Monitoring Principles (Table 18.4) and this was subsequently agreed to by the members of CEOS at their 17th Plenary Meeting. Acknowledgements The study was initiated by Mr. Greg Withee, when he was chair of the Committee on Earth Observation Satellites as part of an initiative to give priority to encouraging and facilitating the increased use of space date, products and services. He set up an ad hoc Team on Utilization comprising the authors of this chapter and they gratefully acknowledge his contributions and encouragement during the Team’s work. The following people gave generously of their time to the case studies and in providing more general advice: Masahiro Baba (Kumamoto City Environmental Research Institute, Japan), Philippe Bally (Spot Image), Wu Bingfang (Institute of Remote Sensing Applications, China), Anthony Busalacchi (World Climate Research Program, Climate Variability and Predictability CLIVAR), Wei Cai (China Center for Resources Satellite Data and Application), Howard Cattle (World Climate Research Program, Climate Variability and Predictability CLIVAR), John Eyre (Meteorological Office), United Kingdom, Yves Govaerts (EUMETSAT), Lujun Guo (National Satellite Meteorological Center of China), Shifeng Huang (IWHR, China), Gregg Jacobs (Naval Research Laboratory (NRL)), Anthony Janetos (The H. John Heinz III Center for Science, Economics and the Environment), Dave Jones (StormCenter Communications, Inc.), Chris Justice (Department of Geography, University of Maryland), K.R.S. Murthi (Indian Space Research Organization ISRO), Jeffrey Key (NOAA, National Environmental Satellite, Data and Information Services, NESDIS), Dieter Klaes (EUMETSAT), Pierre Yves Le Traon (Space Oceanography Division, CLS), Fing Liu (NOSAC), Guoping Liu (Institute of Remote Sensing Applications, China), Haolu Ma (Information Center of the Yellow River Conservancy Center, China, Agostino Miozzo (Department of Civil

18 Improving the Utilization of Remotely Sensed Data

483

Protection, Italy), Hiromi Nakamura (Hydrographic and Oceanographic Department, Japan Coast Guard), Marc Paganini (Science and Applications Department ESA), Bernard Pinty (Institute for Environment and Sustainability, European Commission, Joint Research Centre), Nigel Press (Nigel Press Associates), Diego Fernando Prieto (ESA), James Purdom (Italian Centre for Aerospace Research (CIRA)), Mukund Rao (ISRO), Adrian Simmons (European Centre for Medium-Range Weather Forecasts – ECMWF), Per Erik Skrovseth (Norwegian Space Centre), Rob Sohlberg (University of Maryland), Hideo Tameishi (Japan Fisheries Information Service Center – JAFIC), John Trinder, Ray Harris and other Committee Chairmen of the International Society for Photogrammetry and Remote Sensing, Espen Volden (ESA), Tomohiro Watanabe (JAXA), Bob Winokur (EarthSat Corporation) and members of the World Climate Research Program (WCRP) Satellite Working Group. Ms Judy Carrodeguas provided invaluable practical support for all aspects of the study and work of the Data Utilization Working Group.

References CEOS (Committee on Earth Observation Satellites) (2003)Utilization Team Report and Recommendations, 17th Plenary Meeting, Colorado Springs, Colorado, November 19–20, 2003, Report CEOS/17/Utilization Duchossois G, Sommeria G (2003) Update of Space Mission Requirements for WCRP, WCRP Satellite Working Group Report, WCRP, Geneva GCOS (2003) Second Report on the Adequacy of the Global Observing Systems for Climate in Support of the UNFCCC, Appendix 2, p. 57 GCOS-82; WMO TD 1143. http://www.wmo.ch/web/gcos/gcoshome.html Harris R (2002) Earth Observation Data Policy and Europe. Taylor & Francis, London IGOS (2004) Report of the IGOS International Workshop, Tokyo International Exchange, February 4–6, 2004, JAXA, Japan (http://www.igospartners.org/docsIGOS.html) NRC (National Research Council) (2000) From Research to Operations in Weather Satellites and Numerical Weather Prediction: Crossing the Valley of Death, Commission on Geosciences, Environment and Resources (CGER), National Academies Press, Washington, DC NRC (National Research Council) (2003a) Satellite Observations of the Earth’s Environment: Accelerating the Transition of Research to Operations, Committee on NASA-NOAA Transition from Research to Operations, National Academies Press, Washington, DC NRC (National Research Council) (2003b) Using Remote Sensing In State And Local Government: Information for Management and Decision Making, National Academies Press, Washington, DC

Chapter 19

Emerging Issues in Land Remote Sensing Shunlin Liang, Michael Schaepman, Thomas Jackson, David Jupp, Xiaowen Li, Jiyuan Liu, Ronggao Liu, Alan Strahler, John R. Townshend, and Diane Wickland

Abstract This chapter summarizes the key questions and issues discussed by three review panels in the 9th International Symposium on Physical Measurements and Signatures in Remote Sensing (ISPMSRS) held in October 2005 in Beijing. The panels focused on remote sensing systems and sensors, modeling and inversion, and remote sensing applications. Some emerging issues in land remote sensing are presented, including sensor networks, modeling complex landscapes, machine learning techniques for inversion, and spatial scaling.

Shunlin Liang and John Townshend Department of Geography, University of Maryland, College Park, MD 20742, USA Michael Schaepman Centre for Geo-Information, Wageningen University, Wageningen, The Netherlands Thomas Jackson USDA ARS Hydrology and Remote Sensing Lab, Beltsville, MD 20705, USA David Jupp CSIRO Marine and Atmospheric Research, Canberra ACT 2601 Australia Xiaowen Li Center for Remote Sensing and GIS of Beijing Normal University and Institute for Remote Sensing Applications, Chinese Academy of Sciences, Beijing, China Jiyuan Liu and Ronggao Liu Institute for Geographical Sciences and Natural Resource Research, Chinese Academy of Sciences, Beijing, China Alan Strahler Department of Geography and Environment and Center for Remote Sensing, Boston University, Boston, USA Diane Wickland NASA Headquarters, Washington, DC, USA S. Liang (ed.), Advances in Land Remote Sensing, 485–494. c Springer Science + Business Media B.V., 2008

485

486

S. Liang et al.

19.1 Introduction The 9th International Symposium on Physical Measurements and Signatures in Remote Sensing (ISPMSRS) in Beijing, China, from October 17 to 19, 2005, lasted for 3 days, each day ending with a panel discussion. The panels built upon and complemented each day’s papers and posters. Topics were not independent components but rather viewpoints pertinent to the science and activity of remote sensing and its applications. Panel discussions and outcomes sometimes overlapped. In this chapter, the summaries of the key questions and points discussed in the three panels and more in-depth discussions on some emerging issues in land remote sensing are provided.

19.2 Remote Sensing Sensors and Systems Panel One drew together the work addressed during the day and the overviews prepared by the presenters. It was designed to address a series of topics, such as hardware and system software, engineering physics, platforms and sensor deployment, and calibration and validation. It also addressed the modelling and integration of systems and methods in a way that takes full account of their unique features. Careful attention was paid to the framework and scales of demand that drive the work – the end-user applications. At the start, a series of questions was asked: • What have been the main areas of success and “impact” of the various systems in applications? • What are the main points of issue or limitation for their operational use in applications? • What are new or emerging opportunities in systems and sensors? • What synergies arise from combinations, including data assimilation, of systems and sensors? • Which systems and combinations hold the most promise for meaningful modelling and inversion? • Which systems and combinations will have the greatest impact on specific major issues regarding applications? After extensive discussions, it was recognized that attention to engineering issues has paid off, with significant improvements seen in many indicators, such as signalto-noise ratios, resolutions, pointing accuracies, geometric and spectro-radiometric integrity, and calibration. Many systems now incorporate multiple dimensions of observation (i.e., angular, frequencies, polarizations) at a range of scales. The panel identified a series of issues and pointed out the required actions and opportunities: • For Passive Microwave systems: L-band is needed for better soil moisture estimation; pixel sizes are still too large (e.g., 40 km) for many applications but can

19 Emerging Issues in Land Remote Sensing

• •

•

•

487

potentially be reduced by passive microwave and radar synergy; and soil moisture estimation theory in general needs further work. For Radar systems: saturation in vegetation is still a big issue; more dimensions will help with current problems; multi-dimensions can add value to radar operations; and combination with other sensors has great potentials. For Lidar systems: there is no current space-borne canopy lidar except GLAS (Geoscience Laser Altimeter System) on the Ice, Cloud and land Elevation Satellite (ICESat); airborne systems have proved the value of canopy lidar for many applications; a space sensor would be an important step; and solving issues of interpretation and modelling (rather than correlations) take priority. For Thermal systems: multiple-view-angle (MVA), multispectral, and hyperspectral data are needed; emissivities are interesting yet difficult to handle; emissivity and MVA effects have been identified and need attention (e.g., new hardware); evaporation still is not a standard product. For Hyperspectral systems: building upon the Hyperion experience in operational system design, spectral coverage of wavelengths shorter than 400–500 nm and longer than 2,000 nm is essential and must be added; true spectroscopy can add significant and valuable dimensions for land applications but demands small pixels for best results; the community is ready for a Hyperion follow-up; and the opportunities for spectroscopy of the land environment have been established.

One particular emerging issue is the sensor-web or sensor networks. Earth environments are exceptionally dynamic and interrelated through a number of systems. To understand Earth systems and dynamics, significant improvements in spatial and temporal observations are required. The consensus is that fusion of data from multiple sensor systems is currently one of the optimal solutions. However, increasingly inexpensive, yet sophisticated, chips for the computer and telecommunication industries provide a unique opportunity to develop a distributed, heterogeneous, and adaptive observing system or sensor web in the future. The sensor networks can be used for various applications (e.g., health, military, home) (Akyildiz et al., 2002). Hart and Martinez (2006) claimed that Environmental Sensor Networks (ESN) will become a standard research tool for future Earth system and environmental science. ESNs have evolved from passive logging systems (requiring manual downloading) into intelligent sensor networks that comprise a network of automatic sensor nodes and communications systems. These systems actively communicate data to a sensor network server for integration with other environmental datasets. The sensor nodes can be fixed or mobile, and range in scale appropriate to the environment being sensed. Over 50 representative examples with different scales and functions were reviewed. Lemmerman et al. (2005) conducted the cost analysis by focusing on one possible approach based on an LEO (low Earth orbiting) constellation composed of 100 spacecraft. Kung et al. (2006) described a Drought Forecast and Alert System (DFAS), which is a four-tier system framework composed of mobile users, ecology monitoring sensors, integrated service server, and intelligent drought decision system. DFAS combines wireless sensor networks, embedded multimedia communications, and neural network decision technologies to

488

S. Liang et al.

effectively achieve the forecast and alert of the drought. Paul (2006) described a new initiative in the defense sector on developing the smart sensor web - an intelligent, web-centric distribution and fusion of sensor information that provides greatly enhanced situational awareness, on demand, to fighters at lower echelons. Emphasis is on multi-sensor fusion of large arrays of local sensors, joined with other assets, to provide real-time imagery, weather, targeting information, mission planning, and simulations for military operations on land, sea, and air. Substantial progress has been made to address some of the technical issues (e.g., Akyildiz et al., 2002; Clark and Fearn, 2006). Although technological advances have facilitated these changes, it is vital to continue their use and exploration.

19.3 Modeling and Inversion The panel focused on the following questions: • What is the role of system design in making observations to invert models? Should the issue be hardware-driven or parameter-driven? • Can better use of spatial information inherent in remote sensing be made? Are assumptions that each pixel is independent of the next valid and required? If so, why? • Can the estimation of error structures in retrievals be improved? How are partial derivatives and standard errors of observations best merged? • Can error estimation lead to better validation scenarios? • Is it better to design models from the bottom up (physics first) or the top down (parameters first)? What is the role of empirical parameters in this process? • Is a probabilistic framework the best way of incorporating ancillary information? Is setting limits or bounds on parameters a valid approach? Concentrating on the above questions, the panel (1) examined primarily optical and thermal domains, with a look at microwave and optical synergy; (2) discussed two main approaches to model-based inversion and retrieval of physical parameters: (a) forward modeling to fit observed data, either iteratively or with look-up tables (LUTs); and (b) using models to teach neural nets and regression models to retrieve parameters; (3) noticed that many inversion techniques involve multiple domains: spectral, spatial, temporal, directional; and (4) looked at how to use prior information to enhance retrievals with successful examples in the optical domain. The panel identified a series of common themes: • Ill-posed nature of the inversion problem: models typically have more variables than data dimensions; correlation among model parameters is common. • Information dimensions: spectral, spatial, temporal, directional; expand the problem by adding dimensions to solve it. • Ancillary Information: not all combinations of parameters are encountered in nature or in sensing scenarios; how should the weight given to ancillary information be determined?

19 Emerging Issues in Land Remote Sensing

489

• Empirical vs physical models: empirical models or parameters require training data or calibration. • Simplicity vs complexity in models: trade-off between better description of model physics and easier to calculate, but less accurate retrieval of parameters. Several critical issues were extensively discussed in the symposium and are further elaborated here, including modeling, machine learning techniques for inversion, data fusion and assimilation, and spatial scaling. A major purpose of modeling surface radiation in remote sensing is to understand the signals observed by remote sensors, and eventually to use such models for estimating land surface variables. Some of the models are developed mainly for understanding the radiation field of the landscape and linking with other environmental modeling. Traditional numerical models are usually deterministic ones based on “first principle” equations (Clark and Gelfand, 2006). They are formulated using relevant first principles and observational data, and are usually based on solving deterministic equations (such as RT equations) and some secondary empirical components based on traditional statistical techniques like regression. Landscape complexity often demands sophisticated models, but inversion of land surface variables from remotely sensed data has to rely on simple models. For example, inversion of canopy properties often uses the plane-parallel RT models that do not describe the actual landscape well. Quantitatively, what are the impacts of the model error (uncertainty) on inversion of satellite data? There is growing demand for model outputs to be associated with some sort of uncertainty estimates. Including the stochastic components in the model is one direction to pursue. Stochasticity stands in for complexity. By making models stochastic and inferential, complexity can be represented in ways that vastly simplify modeling (Palmer et al., 2005). One solution for developing the hybrid models based on combining deterministic modeling and machine learning components may help to achieve this goal (Krasnopolsky and Fox-Rabinovitz, 2006). The approaches based on the Monte-Carlo sampling of the key parameters could be another solution to determine the model output uncertainty, although they are computationally very expensive, and the biases due to the structure deficiency also need to be characterized. Convincing arguments must be put forth in support of large-scale field measurement campaigns for calibration and validation of inversion-oriented models. Machine learning techniques, such as artificial neural network (ANN), support vector machine, and regression tree, have been widely used in many applications, however, they are often considered to be “black-box” approaches that do not reflect the physics of the modeled process in remote sensing (Krasnopolsky and Schiller, 2003; Loyola, 2006). Given the accumulated vast data and a priori knowledge, the models and inversion algorithms must rely on effective learning procedures. With the rapid development of these techniques, their potential needs further exploration. For example, neural networks have been combined into multi-neural systems with either redundant or modular elements. The combination of neural networks more effectively solves complex tasks and can furthermore result in dramatic performance improvement. It fully exploits the advantages of neural methods and expands the range of applicability. An important drawback of many ANNs is their lack of

490

S. Liang et al.

explanation capability. It is increasingly apparent that, without some form of explanation capability, the full potential of trained ANNs may not be realized. Many studies have focused on mechanisms, procedures, and algorithms designed to insert knowledge into ANNs (knowledge initialization), extract rules from trained ANNs (rule extraction), and utilize ANNs to refine existing rule bases (rule refinement) (Andrews et al., 1995; Ding and Xin, 2006; Lofstrom et al., 2004; Nunez et al., 2006; Saad and Wunsch, 2007). The land remote sensing community should pay close attentions to these developments. Given the ill-posed nature of remote sensing inversion and vast expansion of the observation data, the challenges becomes: how do we combine observations that derive from many sources, and how do we connect observations that are specific to location, time and setting with understanding that comes from a diverse body of nonspecific theory? Data fusion techniques that simply register and combine data sets together from multiple sources may not be adequate to solve our problems. Data assimilation method allows us to use all the information available to us within a time window to estimate various unknowns of land surface models (see Chapter 12). The information that can be incorporated includes observational data, existing pertinent a priori information, and, importantly, a dynamic model that describes the system of interest and encapsulates theoretical understanding. Data assimilation has been widely used in meteorology and oceanography, but more efforts are needed in the land remote sensing community to explore its potential for characterizing land surface environments. Spatial and spatiotemporal processes in the physical, environmental and biological sciences often exhibit complicated and diverse patterns across different space– time scales. Both scientific understanding and observational data vary in form and content across scales. Remote sensing data and other observations are collected at differing scales and resolutions, at different spatial locations, and in different dimensions. Many statistical issues are associated with combining such data for modeling and inference. Gotway and Young (2002) gave an overview of these issues and the approaches for integrating such disparate data, drawing on work from geography, ecology, agriculture, geology, and statistics, with the emphasis on state-of-the-art statistical solutions to this complex and important problem. Substantial progress has been made recently in combining incompatible spatial data, such as multiscale modeling using the Bayesian hierarchical framework (Wikle and Berliner, 2005), by which the combination of such information at different spatial scales can be accomplished. Such progress has not been incorporated into remote sensing modeling and inversion. Downscaling methods using unmixing algorithms have been extensively reported for estimating subpixel proportions from multi-spectral and multi-temporal remote sensing data with the traditional statistical algorithms; geostatistical methods need further exploration (Kyriakidis and Yoo, 2005; Pardo-Iguzquiza et al., 2006). The latest progress in modeling and inversion algorithms are recently reviewed (Liang, 2007).

19 Emerging Issues in Land Remote Sensing

491

19.4 Remotes Sensing Applications Panel Three identified the following factors that may lead to the disparities: data availability (e.g., data policy, costs, inadequate information systems), need for assurance of continuity of data sets, lack of familiarity of potential users with data sets, product quality insufficient for application requirements, terminological incongruities between remote sensing algorithms and application scientists, and lack of investments by some space agencies in applications. Developing successful remote sensing applications is challenging. Impediments to success include: • Lack of user involvement early in the application development process. • Lack of technical knowledge or capability in the user. • Institutional inertia – the user institution has to be interested in and willing to change or accept new approaches to decision making. Also, the initial investment can be difficult to secure. • Too great a gap between the underlying theory and concept for a new application and the first recognition/acceptance of a realistic application on the part of an end user. • Mismatches between what researchers are interested in or expect and what operational users need or can use. Successful applications are not static but evolve as new sensors, data processing and network technology emerge. Successful applications also require verification and validation of products. • The scientific credibility of the observations, measurements, derived products and model outputs must be established and maintained. • The user has the responsibility to validate the inputs to the decision-making process, i.e., to assess the input and determine that it meets the intended need. The panel presented some results of the CEOS (Committee on Earth Observation Satellites) Utilization Team research based on a series of case studies from the USA, Europe, China and Japan, and determined the critical space agency success factors and critical user success factors. The agency success factors include: • Provision of user-specific products. • Long-term continuity of data sets. • Making resources available in a timely fashion and with a minimum of bureaucracy. • Willingness of space agencies to listen extensively to user needs, their practical difficulties, and ultimately make accommodations. • Importance of keeping user development projects simple and building further on reliable, successful initial projects. • Funding by space agencies needs to be linked with user funding. Turning prototypes into sustained operations is frequently unsuccessful if the external funding from users does not continue beyond the prototype phase. • Space agencies must seek assurance that the user has sufficient interest and capacity to maintain funding.

492

S. Liang et al.

The user success factors include: • User understanding, and preferably first-hand experience of the use of remote sensing data and products and integration with geospatial information systems • Existence of key elements of spatial data infrastructure • Willingness to make appropriate staff available before and during the project to work with the space agency • Having a core need (e.g., a statutory need for the information, and an ability to define needs precisely) • Willingness and ability to help develop a realistic financial scenario for an operational system Considering both the agency and user, case studies indicate the need for innovation and flexibility in response to new opportunities for data products. In stead of conducting an extensive review of specific remote sensing applications in all areas, the application panel focused on only three areas: Land Use and Land Cover Change (LUCC) mapping, forest monitoring in the example of Canada, and regional agricultural monitoring and management. After evaluating the status of LUCC mapping from remote sensing, a series of challenging issues were identified, including: • • • •

Dynamic monitoring of LUCC with high accuracy by remote sensing Historical retrieval of LUCC by using the data with different characteristics Classifying not only the type but also the level and the structure of LUCC Standardization of the LUCC products for international exchange and global change study • Analysis of the scenarios of LUCC with coupled human–natural factors, and modeling • Analysis of the environmental impacts of LUCC, and modeling Various measures have been identified to monitor spatially and temporally the composition, structure and health of forests from remote sensing in the example of Canada, including • Forest area by type (coniferous deciduous, mixed), species, and disturbance (harvesting, fire, insect damage, blow down, etc.) • Amount of forest – timber volume and biomass • Forest health through measurements of forest chemistry Remote sensing applications to agriculture were reviewed. One of the emerging issues is the integration of remote sensing and crop growth model. Remotely sensed data substantially improves the accuracy of the regional crop growth model, but as yet, a number of critical agronomic parameters cannot be precisely inverted from remotely sensed data at large scale, such as harvest index and crop management information (e.g., applied fertilizer, irrigation, and crop variety). There are also large gaps between the promised utility and reality, for example, inaccuracy for agroparameter inversion models, ambiguity in use of various indices vs. conventional parameters (e.g., crop growth and soil moisture), and uncertainty of the availability of continuous time series data.

19 Emerging Issues in Land Remote Sensing

493

19.5 Concluding Remarks Because of considerable investment, remote sensing has advanced significantly, and the societal benefits have been tremendous. Huge amounts of data available from NASA’s EOS program and other space agencies offer promise but also present great challenges. Developing advanced sensors and systems to address application issues is constructive, but optimal integration of observations from various sensors, as well as other ancillary measurements, needs urgent attention. To realize integration, new models and algorithms, such as data fusion and data assimilation, are crucial. Developing computationally simplified surface radiation models mostly suitable for inversion of land surface variables from satellite data is also urgently required. Improved coupling of remote sensing science and applications would be most advantageous. Acknowledgements The authors are very grateful to the panel members who provided most materials presented in this chapter. Some of the ideas also come from numerous presentations by the symposium participants.

References Akyildiz IF, Su WL, Sankarasubramaniam Y, Cayirci E (2002) A survey on sensor networks. IEEE Commun. Mag. 40(8): 102–114 Andrews R, Diederich J, Tickle AB (1995) Survey and critique of techniques for extracting rules from trained artificial neural networks. Knowledge-Based Systems 8(6): 373–389 Clark JS, Gelfand AE (2006) A future for models and data in environmental science. Trends Ecol. Evol. 21(7): 375–380 Clark SD, Fearn DG (2006) The impact of advanced platform and ion propulsion technologies on small, low-cost interplanetary spacecraft. Acta Astronautica 59(8–11): 899–910 Ding XQ, Xin S (2006) Application research on extraction of rule from artificial neural networks for nonlinear regression. Dyn. Contin. Discrete Impuls. Syst. Ser. A Math. Anal. 13: 565–568 Gotway CA, Young LJ (2002) Combining incompatible spatial data. J. Am. Stat. Assoc. 97(458): 632–648 Hart JK, Martinez K (2006) Environmental Sensor Networks: A revolution in the Earth system science? Earth-Science Reviews, 78: 177–191 Krasnopolsky VM, Fox-Rabinovitz MS (2006) A new synergetic paradigm in environmental numerical modeling: hybrid models combining deterministic and machine learning components. Ecol. Model. 191(1): 5–18 Krasnopolsky VM, Schiller H (2003) Some neural network applications in environmental sciences. Part I: forward and inverse problems in geophysical remote measurements. Neural Networks 16(3–4): 321–334 Kung HY, Hua JS, Chen CT (2006) Drought forecast model and framework using wireless sensor networks. J. Inform. Sci. Eng. 22(4): 751–769 Kyriakidis PC, Yoo EH (2005) Geostatistical prediction and simulation of point values from areal data. Geogr. Anal. 37(2): 124–151 Lemmerman L, Raymond C, Shotwell R, Chase J, Bhasin K, Connerton R (2005) Advanced platform technologies for Earth science. Acta Astronautica 56(1–2): 199–208 Liang S (2007) Recent developments in estimating land surface biogeophysical variables from optical remote sensing, Prog. Phys. Geogr. 31(5): 501–516

494

S. Liang et al.

Lofstrom T, Johansson U, Niklasson L (2004) Rule extraction by seeing through the model, Neural Information Processing. Lecture Notes in Computer Science 3316, pp 555–560 Loyola RDG (2006) Applications of neural network methods to the processing of Earth observation satellite data. Neural Networks 19(2): 168–177 Nunez H, Angulo C, Catala A (2006) Rule-based learning systems for support vector machines. Neural Process. Lett. 24(1): 1–18 Palmer TN, Shutts GJ, Hagedorn R, Doblas-Reyes E, Jung T, Leutbecher M (2005) Representing model uncertainty in weather and climate prediction. Annu. Rev. Earth Planet. Sci. 33: 163–193 Pardo-Iguzquiza E, Chica-Olmo M, Atkinson PM (2006) Downscaling cokriging for image sharpening. Remote Sens. Environ. 102(1–2): 86–98 Paul JL (2006) Smart Sensor Web: tactical battlefield visualization using sensor fusion. IEEE Aerosp. Electron. Syst. Mag. 21(1): 13–20 Saad EW, Wunsch DC (2007) Neural network explanation using inversion. Neural Networks 20(1): 78–93 Wikle CK, Berliner LM (2005) Combining information across spatial scales. Technometrics 47(1): 80–91

Index

A AATSR, 66, 67, 85, 99, 100, 131, 249, 250 Advanced Very High Resolution Radiometer (AVHRR), 16, 82, 98, 123, 131, 132, 137, 174, 176, 184, 215, 221, 222, 235, 249–251, 264, 268, 269, 276, 331, 334, 353, 371, 372, 386, 398, 400, 405, 408, 409, 411, 424–427, 430–437, 447, 451–455, 473 Aerodynamic temperature, 246, 248, 253, 269–270, 273, 275 Aerosol optical depth, 78, 79, 82, 84, 229 Albedo, 6, 30–32, 40–42, 52, 83, 96–98, 102, 108, 112, 115, 116, 126, 128, 131, 155, 159, 161, 162, 184, 204, 205, 207, 220–228, 231–237, 239, 248, 264, 274, 320 Anisotropy factor, 56 ASTER, 99, 249–251, 261, 264, 266–269, 275, 276, 425 Atmospheric correction, 13, 67, 68, 70, 78–80, 82–84, 128, 219, 221, 223, 229, 234, 248, 252, 255, 261, 263–265, 334, 356 Automatic differentiation (AD), 321, 323, 324 B Bayesian network (BN), 213–214 Bidirectional reflectance factor (BRF), 55, 65, 96, 102, 113, 128, 137, 160, 177, 205, 222, 227, 228, 231 Bidirectional reflectance distribution function (BRDF), 1, 56, 71, 96, 97, 99, 102, 106, 112, 115, 118, 119, 121, 125, 131, 132, 134, 137, 183, 191, 204–207, 209, 214, 222, 224, 225, 232, 258, 376 Bihemispherical reflectance (BHR), 96, 101, 137, 222, 226–228, 232, 236

Biomass, 3, 54, 106, 123, 164, 246, 318, 320, 333, 341, 402–404, 406, 407, 425, 428, 432–434, 437, 492 Brightness temperature, 11, 15, 51, 53, 55, 57, 60, 61, 64, 66–69, 71–73, 79–82, 247, 248, 251, 252, 254–258 C Canopy openness, 112–114, 125 Carbon cycle, 54, 148, 317, 321, 331, 332, 334, 369, 370, 381, 389, 423, 424, 426, 431, 432, 434, 435, 437–439 CCD, 65, 104, 132, 137 Change detection, 26, 28, 342, 345, 346, 353, 356–359, 363 CHRIS, 98, 100, 102, 103, 113, 119, 120, 131–133, 136, 137 Classification, 10, 23, 26, 27, 29, 30, 38, 95, 110, 123, 125, 128, 134, 195, 208, 223 Clumping index, 71, 114, 115, 117, 137 Cost function, 176, 180–182, 185, 188–190, 204–206, 232, 316, 317, 321, 323, 324, 329 Crop yield, 332–324, 401–407, 415 D Data assimilation (DA), 196, 207, 210, 212, 213, 216, 273, 313, 331, 332, 379, 386, 405, 413, 415, 459, 479, 486, 490, 493 Data fusion, 14, 16, 234, 294, 297, 372, 384, 385, 388, 389, 401, 415, 477, 489, 490, 493 Decision tree classifier, 29, 374, 376, 399, 400 Dempster–Shafer theory, 377 E Emissivity, 53, 55–57, 69, 71, 74, 76, 77, 79, 80, 82, 247, 248, 251, 252, 254, 255, 257

495

496 Energy balance, 52, 59, 135, 174, 248, 270, 273, 317, 330, 331, 410, 447 Enhanced vegetation index (EVI), 376–380, 383, 386, 401, 406, 423, 427, 430, 431, 435, 439 Ensemble emissivity, 248, 252, 254, 255, 257, 260, 264, 270, 273, 275 Evapotranspiration (ET), 52, 174, 247, 330, 413, 426, 445, 454 Evidential reasoning, 377, 379–381, 384, 385, 387–389 F Filtering, 96, 123, 232, 294–296, 322, 325, 327 Foliage temperature, 53, 54, 58, 59, 69, 71, 85 fPAR (= FAPAR = FPAR), 102, 108, 110, 119, 132, 138, 148, 163–166, 177, 179, 182, 191, 196, 204, 318, 320, 374, 375, 403, 404, 423, 426–428, 434, 439, 449, 452 Fractional cover, 52, 73, 74, 85, 98, 120, 122, 456 Fuzzy set, 377, 387 G Gaussian distribution, 187, 188, 317, 324 Geometric optics, 2, 70, 163, 252, 258 Geostationary, 99, 131, 219, 220, 222, 231–235, 237, 239, 249, 251, 266, 274, 294 Gibbs distribution, 302 Goniometer, 51, 61–63, 86 H Hemispherical directional reflectance factor (HDRF), 96, 101, 138, 226, 227, 229 Hot spot, 71, 98, 104, 119, 133, 152, 156, 183, 263, 273 Hyperspectral, 97, 98, 103, 104, 107, 149, 189, 196, 211, 250, 251, 261, 265, 269, 294, 298, 487 I Ill-posed, 157, 180, 183–185, 187, 203, 216, 246, 259, 264–266, 268, 313, 320, 488, 490 Image fusion, 293–297 Instantaneous field of view (IFOV), 13, 58, 61, 62, 65, 72, 99, 113, 114, 134, 138 K Kalman filter, 321, 322, 325, 326, 413

Index L Land cover, 10, 28, 47, 53, 76, 78, 86, 95, 115, 116, 123, 125, 131, 135, 164, 207–209, 215, 220–224 Land surface temperature (LST), 10, 55, 246, 249, 253, 273, 317, 318, 320, 330, 331, 386, 411, 412, 423, 427, 430, 435, 439 Land use, 76, 208, 211, 214, 215, 260, 265, 341, 342, 344–346, 350, 352, 354–356, 358 Latent heat flux, 52, 329, 330, 410 Leaf area index (LAI), 75, 98, 102, 109, 110, 119, 120, 138, 148, 159, 165, 174, 204, 213, 276, 318, 379, 406, 423, 426, 449 Lidar, 3, 106, 107, 109, 114, 118, 135, 136, 138, 221, 320, 424, 487 Lookup table (LUT), 109, 120, 135, 157, 180, 182, 191, 259, 275, 334 M Maximum likelihood classifier, 125, 358 Meteosat, 219–221, 231, 233–237, 239 Meteosat Second Generation (MSG), 131, 138, 235, 249, 250, 276, 473 Minimization, 84, 180, 298, 302, 305, 323, 375 MISR, 98–102, 107–110, 112, 118, 121–129, 131–133, 136, 138, 182, 184 Mixture model, 52, 72, 371, 401 MODIS, 98–100, 102, 105, 115, 117, 118, 122, 124, 131, 132, 136, 138, 166, 174, 177 MODTRAN, 75, 77, 190, 191, 252, 258, 276 Monte Carlo, 39, 135, 153–155, 161, 180, 182, 252, 258, 327, 332, 345, 450, 489 N National Polar Orbiting Earth Satellite System (NPOESS), 14, 16 Net ecosystem production (NEP), 54, 332, 433 Net primary production (NPP), 106, 204, 332, 374, 375, 406, 423–430, 432–435, 437, 439, 473 Net radiation, 53, 59, 410, 445, 447 Neural network (NNT) (= ANN), 107, 109, 148, 177, 178, 204, 213, 259, 261, 262, 322, 326, 327, 341, 401, 487, 489 Normalized difference vegetation index (NDVI), 117, 123, 138, 158, 177, 186–188, 227–228, 276, 406, 409, 411, 423, 425 O Objective function (= cost function), 176, 180–182, 185, 188–190, 204–206, 232, 294, 298, 302, 307, 310, 315–317, 321–324, 329

Index Optimization, 121, 157, 180–182, 189–191, 194, 204, 259, 293, 294, 302, 306, 310, 315, 320–324, 332, 334, 414 P Phenology, 105, 226, 376, 382, 385, 386, 398, 405–407, 415, 423, 425, 431, 451, 457, 458 Photosynthetically active radiation (PAR), 97, 98, 102, 110, 119, 148, 149, 204, 227, 318, 404, 423, 426 Plant functional type (PFT), 369, 370, 374, 382, 449, 452, 457 Planck function, 73 Probability distribution function (PDF), 186–188, 205, 207, 232 Polarization, 10, 11, 20–28, 30–38, 41, 44, 46, 47, 98, 104, 221, 413, 486 POLDER, 98, 100, 104, 105, 115, 129, 131–133, 136, 138, 174, 177, 178, 184, 191, 204, 205, 221, 222 A posteriori, 189, 205, 305, 316 A priori, 3, 38, 80, 98, 180, 189, 203–208, 211, 212, 216, 223, 231, 259, 265, 266, 269, 298, 301, 307, 310, 313, 315, 334, 383, 459, 489, 490 PROSPECT, 75, 119, 120, 183, 191 p-theory, 160–162, 166 R Radiation Transfer Model Intercomparison (RAMI), 110, 111, 138, 157 Radiative transfer (RT), 11, 12, 15, 39, 51, 59, 70–72, 75, 78, 80, 83, 84, 86, 106, 109–111 Radiometric temperature, 52, 246, 248, 252, 254, 257, 259, 260, 262, 264–270, 273–275 Radiosity, 71, 119, 153, 154, 258 Recollision probability, 110, 160–162 Regularization, 185, 293, 301 S Scaling, 150, 157, 161–163, 203, 215, 216, 328, 351, 450, 489

497 Single scattering albedo, 155, 159, 184 Snow density, 20, 30, 32, 33, 35–44 Snow water equivalence (SWE), 19, 20, 23, 33–35, 39, 40, 42, 44–47 Soil moisture, 9–16, 23, 28, 29, 53, 97, 98, 130, 150, 247, 262, 274 Soil Moisture Ocean Salinity (SMOS), 11, 15, 16, 413 Soil temperature, 53–55, 58, 59, 62, 70, 71, 75, 78, 87, 272, 320, 412, 445 Soil–vegetation–atmosphere transfer (SVAT), 75, 248, 262, 277, 317, 330 Solar radiation, 57, 58, 63, 76, 81, 135, 147, 148, 150, 152, 403, 404, 426, 429, 433, 434, 446–449, 452, 454, 455, 459 Spinning Enhanced Visible and Infra-Red Imager (SEVIRI), 131, 138, 249–251, 268, 277 Split-window method, 67, 78, 80, 82 SSM/I, 10, 13, 14, 16, 408 Stochastic inversion, 154, 274 Support vector machine (SVM), 125, 489 Synthetic aperture radar (SAR), 19–35, 37–39, 41–46, 106, 107, 138, 398, 400, 414, 473, 477 T Two-stream, 152, 153, 156, 450, 452, 453 V Validation, 13, 34, 78, 86, 87, 103, 121, 150, 165, 196, 204, 207, 211, 215, 229 Vegetation continuous fields (VCF), 105, 123, 124, 138 Volumetric scattering, 117, 256 W Water balance, 428, 446, 448 Water vapor, 10, 51, 58, 59, 72, 76, 79, 81–83, 190, 234, 254, 257, 261, 262, 265, 332 Wavelet, 55, 297, 342–346, 350–352, 354, 355, 358