FINAL REPORT F. BARETw3.avignon.inra.fr/reseda/base/documents/reseda-report/...The ReSeDA project was originally initiated by Daniel Vidal-Madjar within the Pro-gramme National de

RESEDAASSIMILATION OF

MULTISENSOR & MULTITEMPORAL

REMOTE SENSING DATA TO MONITOR SOIL

& VEGETATION FUNCTIONINGAn EC research project

co-funded by the Environment and Climate Programme within the topic‘Space Techniques Applied to Environmental Monitoring and Research – Methodological Research’.

FINAL REPORT

F. BARET

June, 2000

Contract number ENV4CT960326

Foreword

The ReSeDA project was originally initiated by Daniel Vidal-Madjar within the Pro-gramme National de télédétection Spatiale in France. Because of the importance of the ex-periment and the large scope covered by its objectives, it was extended to the Europeanlevel to organise a collaborative effort. The start of the ReSeDA project was somewhat dif-ficult because of delays in final acceptation of the project by the Commission. We thankvery much Alan Cross for the effort spent to speed up the process. We thank also verykindly Michel Schouppe from the European commission for his interest in the project, forhis patience and for his positive comments made along the life time of the project.

After the Launch of the project, the experiment started and required intensive efforts of allthe partners. The processing of the large data set collected took more time than expectedand delayed the actual scientific processing by few months. It provides now a unique database as well as improved models and concepts that have been presented at the final meetingduring the ReSeDA special session that took place within the European Geophysical Soci-ety General Assembly in Nice. These investigations are already or will be soon published inrefereed journals.

Besides the scientific results obtained, the ReSeDA project contributed significantly to or-ganise both the French and the European scientific community concerned by application ofremote sensing to the continental biosphere. Further, it initiated a momentum around re-mote sensing data assimilation that is still growing through a range of projects ranging fromvery local applications such as precision farming up to global applications linked with hu-man impacts on climatic change.

This report is an attempt to present the main results obtained within the ReSeDA project, aswell as the limitations associated. It provides recommendations for future investigations tocarry on for setting the concepts, models, algorithms, methods and techniques up to a levelallowing operational applications. It is organised as along nine chapters:

1. Highlights: the ReSeDA objective, approach and results are presented in a syntheticway,

2. Partners: the ReSeDA partners are briefly presented,3. Rationale: the background is analysed in detail with due attention to applications to

conclude on the required investigations,4. Objectives and timeframe: The objectives of the ReSeDA project are clearly stated

and decomposed into three work-packages,5. Achievements. For each work-packages, the results are presented with enough de-

tails to understand the progress achieved within ReSeDA. Problems and limitationsof the results are also discussed,

6. Evaluation-user component. The ReSeDA results are evaluated with emphasis onactual and potential applications for users,

7. Exploitation. The results are evaluated in terms of their exploitation both for the sci-entific and user communities. Perspectives for future studies are proposed,

8. List of deliverables and publications,9. Conclusion. The conclusion reviewed the main achievements and provides clues for

future investigations.Additionally, a more detailed description of the studies and results gathered during the Re-SeDA project are presented in annexe.

The reader who wants to get a rapid flavour of the project should read chapters 1 and 9.The one who wants to evaluate the achievements will have to browse through all chapters.

Frédéric BaretHappy PI of the ReSeDA project

PLAN1. HIGHLIGHTS .............................................................................................................................................. 5

1.1 THE EXPERIMENT OVER THE ALPILLES SITE. ............................................................................................ 61.1.1 Remote sensing data........................................................................................................................ 6

1.1.1.1 airborne sensors ........................................................................................................................................... 61.1.1.2 Satellite data ................................................................................................................................................ 6

1.1.2 Ground measurements..................................................................................................................... 61.1.2.1 Soil measurements....................................................................................................................................... 61.1.2.2 Canopy characterisation .............................................................................................................................. 71.1.2.3 Microclimatic data....................................................................................................................................... 71.1.2.4 Meteorological data and atmosphere characterisation ................................................................................. 7

1.1.3 The data base .................................................................................................................................. 71.2 ESTIMATION OF CANOPY AND SOIL CHARACTERISTICS FROM REMOTE SENSING DATA ............................. 7

1.2.1 Solar domain ................................................................................................................................... 71.2.2 Thermal infrared domain ................................................................................................................ 71.2.3 µ-wave domain................................................................................................................................ 8

1.3 ASSIMILATION OF REMOTE SENSING DATA INTO PROCESS MODELS .......................................................... 81.3.1 Assimilation into SVAT models ....................................................................................................... 81.3.2 Assimilation into Canopy functioning models................................................................................. 9

2. THE PARTNERS.......................................................................................................................................... 9

2.1 LIST OF PARTNERS ................................................................................................................................... 92.2 ROLE OF PARTNERS................................................................................................................................ 10

3. RATIONALE .............................................................................................................................................. 10

3.1 MAIN ISSUES OF CONCERN ..................................................................................................................... 103.1.1 Global scale issues. ....................................................................................................................... 103.1.2 Local to regional scale issues ....................................................................................................... 11

3.2 THE NEED IN MODELLING VEGETATION AND SOIL PROCESSES................................................................ 113.2.1 Processes and models.................................................................................................................... 113.2.2 The range of spatial and temporal scales...................................................................................... 113.2.3 The limits of process models: number and accuracy of the inputs................................................ 12

3.3 THE NEED IN REMOTE SENSING OBSERVATIONS...................................................................................... 123.3.1 Role and diversity of remote sensing systems................................................................................ 123.3.2 From radiative transfer model inversion to assimilation into process models. ............................ 14

3.3.2.1 Inverse methods: exploitation of instantaneous measurements.................................................................. 143.3.2.2 Assimilation methods. ............................................................................................................................... 14

4. OBJECTIVES & TIME FRAME OF THE PROJECT........................................................................... 15

4.1 OBJECTIVES OF RESEDA ....................................................................................................................... 154.2 TIME FRAME........................................................................................................................................... 15

5. ACHIEVEMENTS...................................................................................................................................... 16

5.1 THE EXPERIMENT AND THE DATA BASE.................................................................................................. 165.1.1 Description of the Alpilles test site. ............................................................................................... 165.1.2 Satellite data ................................................................................................................................. 16

5.1.2.1 Satellite sensors and images acquired........................................................................................................ 165.1.2.2 Processing of the images ........................................................................................................................... 175.1.2.3 The data base ............................................................................................................................................. 18

5.1.3 Airborne data. ............................................................................................................................ 5-185.1.3.1 Sensors and vectors used ........................................................................................................................ 5-185.1.3.2 Flight schedule ....................................................................................................................................... 5-195.1.3.3 Processing of the data ................................................................................................................................ 205.1.3.4 The data base ............................................................................................................................................. 20

5.1.4 Ground measurements................................................................................................................... 215.1.4.1 The selection of fields................................................................................................................................ 215.1.4.2 Soil measurements..................................................................................................................................... 225.1.4.3 Canopy measurements............................................................................................................................... 225.1.4.4 Atmosphere measurements........................................................................................................................ 23

5.2 THE DATA BASE ..................................................................................................................................... 255.2.1 Structure of the data base and diffusion........................................................................................ 25

5.2.1.1 The data base ............................................................................................................................................. 255.2.1.2 The web server .......................................................................................................................................... 25

5.3 DEVELOPMENT AND EVALUATION OF INVERSION METHODS................................................................... 265.3.1 Inversion in the solar domain........................................................................................................ 27

5.3.1.1 Investigation about inversion techniques................................................................................................... 275.3.1.2 Optimal spectral and directional sampling................................................................................................. 285.3.1.3 Comparison of the inversion performances of three operational radiative transfer models ....................... 295.3.1.4 Validation of a neuronal technique for LAI and fCover estimation........................................................... 305.3.1.5 Robustness of the WDVI-LAI relationship for wheat crops...................................................................... 305.3.1.6 Using BRDF and hybrid models for canopy height estimation ................................................................. 305.3.1.7 BRDF models and albedo estimation ........................................................................................................ 315.3.1.8 Scaling issues ............................................................................................................................................ 32

5.3.2 Inversion in the Thermal Infrared domain .................................................................................... 335.3.2.1 Water content estimates from NOAA/AVHRR......................................................................................... 335.3.2.2 Estimating temperature and emissivity from the spectral variation of DAIS data. .................................... 345.3.2.3 Brightness temperature directional effects observed with the INFRAMETRICS camera. ........................ 35

5.3.3 Inversion in the µ-wave domain .................................................................................................... 355.3.3.1 Land use classification based on µ-wave................................................................................................... 365.3.3.2 Modelling the µ-wave signal ..................................................................................................................... 365.3.3.3 Biophysical parameter estimation from µ-wave data................................................................................. 39

5.4 DEVELOPMENT OF THE ASSIMILATION TECHNIQUE ................................................................................ 425.4.1 Assimilation into SVAT models. .................................................................................................... 42

5.4.1.1 Comparison and improvement of SVAT models....................................................................................... 425.4.1.2 Estimation of surface fluxes using the SEBAL algorithm ......................................................................... 435.4.1.3 Spatial heterogeneity of fluxes .................................................................................................................. 44

5.4.2 Assimilation into canopy functioning models................................................................................ 455.4.2.1 Calibration and evaluation of canopy functioning models......................................................................... 455.4.2.2 Assimilation of remote sensing data into canopy functioning models....................................................... 48

6. EVALUATION - USER COMPONENT .................................................................................................. 50

7. EXPLOITATION PLAN............................................................................................................................ 51

7.1 RADIATIVE TRANSFER MODELLING AND MODEL INVERSION .................................................................. 517.1.1 Solar domain ................................................................................................................................. 517.1.2 .............................................................................................................................................................. 527.1.2 Thermal infrared domain .............................................................................................................. 527.1.3 µ-wave domain.............................................................................................................................. 52

7.2 ASSIMILATION INTO PROCESS MODELS................................................................................................... 527.2.1 SVAT models ................................................................................................................................. 52

7.2.1.1 Using the temporal variation ..................................................................................................................... 527.2.1.2 Using the spatial variation. ........................................................................................................................ 52

7.2.2 Canopy functioning model ............................................................................................................ 52

8. LIST OF RELATED PUBLICATIONS & DELIVERABLES ............................................................... 53

8.1 PUBLICATIONS ....................................................................................................................................... 538.2 DELIVERABLES ...................................................................................................................................... 56

9. CONCLUSION............................................................................................................................................ 57

ReSeDA. Final report. June 2000. 5

1. HighlightsAwareness by the governments and the populationis increasing on Climatic changes at the global scaleand environmental issues at the local to regionalscales. In order to better understand the underlyingmechanisms, and propose solutions to mitigatethese effects mainly due to human activity, propermodels have to be developed and exploited. Theycould then be used to predict the consequences onthe environment at a range of scales under possiblescenarios of human activity and practices. Suchmodels are describing the processes occurring at thesurface of our planet, with emphasis on the dy-namics of the exchanges of mass (mainly carbonand water) and energy (heat, radiation, momentum)between the soil and the atmosphere. Obviously,the vegetation, at the interface between the soil andthe atmosphere plays a major role in the regulationof these exchanges.

Canopy state Soil canopy atmosphere exchanges

Environmental Budget Quantity & quality of the production

Canopy functioningand SVAT Models

Biophysical Variables

Remote Sensing Data

Radiative TransferModels

Inve

rsio

n

Ass

imila

tio

n

Ancillary Information

Figure 1. Flow chart diagram showing how to infervariables of interest such as canopy state, fluxes,environmental budget, production in quantity andquality from remote sensing data and ancillary in-formation. The small green arrows indicate the di-rect way the models actually run.

Beside the above environmental issues, the estima-tion of crop production in quantity and quality isalso very important, both at the local to regionalscales for optimisation of the harvest process, andat the national, continental or global scales for foodsecurity issues and the regulation of the market ofthe main crop production. Here again, canopyfunctioning models, called in this context cropmodels, can be used to understand the main drivingfactors, and to get estimates and forecasts of thequantity, quality and timeliness of the production.The soil and vegetation functioning, models, re-quire a large set of parameters and variables to run.These parameters and variables may vary stronglywith space and time, and only few of them are well

known. Therefore, remote sensing techniques, withthe ability to cover rapidly and frequently largeareas, constitute a very promising tool that providespertinent inputs and controls to the process models.The main objective of the ReSeDA (Remote Sens-ing Data Assimilation) project is to develop and testremote sensing data interpretation methods for abetter description and understanding of the soil andvegetation functioning through dedicated processmodels.The signal recorded by remote sensing sensorsaboard satellites is determined by canopy and soilcharacteristics, i.e. the biophysical variables,through physical processes of the interaction ofelectromagnetic radiation with vegetation and soil,i.e. the radiative transfer. Some of these biophysicalvariables can be used as inputs to the process mod-els, i.e. soil vegetation atmosphere transfer models(SVAT) and canopy functioning models. (Figure 1).Two main approaches could be used to exploit re-mote sensing data, depending on the degree of inte-gration of interpretation scheme.• The first one consists in deriving the canopy or

soil biophysical variables from remote sensingdata in an independent step. This step could ex-ploit our knowledge on the physical processes putin the radiative transfer models. This correspondsto radiative transfer model inversion. It couldalso exploit few ancillary information, mainly thea priori knowledge of the distribution of canopyand soil variables. The biophysical variables esti-mated from remote sensing techniques could thenbe used as inputs to the process models to derivethe variables of interest such as canopy state,fluxes, crop production, …).

• The second approach is the most integrated oneand corresponds to the assimilation of remotesensing data into radiative transfer and canopyor soil functioning coupled models. It allows toexplicitly account for the temporal dimension. Itallows also to use concurrently and in synergy thedata provided by different sensors. Further, itpermits to directly access the variables of interest,while exploiting ancillary data such as climateand soil variables. Assimilation consists in tuningsome parameters of the coupled radiative transferand canopy functioning models so that the simu-lations matches the closest possible the radiomet-ric measurements. This technique is potentiallythe most promising one because it uses the largestamount of information, both on the physical orbiological processes and on ancillary data. Wenote (Figure 1) that canopy or soil functioningmodels could provide useful information on can-opy or soil attributes that will increase the per-formances and the accuracy of radiative transfermodel simulations.

Both approaches have already been investigated forother applications such as ocean or atmosphereproblems. However it is relatively new for the con-tinental biosphere. Therefore, the ReSeDA projectwill emphasise on the development and test of suchapproaches.


The ReSeDA project was decomposed into threemain tasks:1- The experiment and the corresponding data

base2- Estimation of canopy biophysical variables

from radiative model inversion3- Assimilation of remote sensing data into proc-

ess models

1.1 The experiment over theAlpilles site.

A consistent and comprehensive data set has beenacquired during a whole growth season over theAlpilles site located in the south east of France,close to Avignon. The site is about 25km² size, flat,and corresponds to an agricultural landscape, withrelatively large fields for the region.

Figure 2. The Alpilles site. Land use map overlayed on a SPOT image.

The data correspond both to remote sensing andground measurements. All these data are includedin the data base that is available on the web(www.avignon.inra.fr/reseda).

1.1.1 Remote sensing dataRemote sensing systems have been operated in avariety of spectral ranges (optical, thermal infraredand microwave) and configurations from October1996 up to October 1997. They include

1.1.1.1 airborne sensorssuch as POLDER (4 bands in the visible and nearinfrared with directional and polarisation observa-tions), INFRAMETRICS thermal infrared camera(directional observations in a single broad band),ERASME (C and X bands, in VV and HH polarisa-tion) and RENE (S band, in HH polarisation) whichare scatterometric profilers. All these sensors wereflown several times during the growth cycle, al-lowing to get a good monitoring of the vegetationand soil dynamics. Additionally, single flights ofparticular instruments were performed. These in-struments were either imaging spectro-radiometers(DAIS visible, near infrared, short-wave infraredand thermal infrared) and IROE passive microwaveradiometer (6.8 GHz and 10 GHz). All the data

were calibrated, the atmospheric effects corrected,and registered over a single reference image.

1.1.1.2 Satellite dataThe images available during the experiment wereacquired. This includes 6 SPOT, 1 TM, 10 ERSimages and 12 Radarsat images. The data werecalibrated and corrected from the atmospheric ef-fects. Further, a geometric correction was appliedusing the single reference image. Additionally, thewhole series of NOAA/AVHRR was acquired aswell.

1.1.2 Ground measurementsThe ground measurements were performed on aselection of fields, mainly wheat, sunflower, maizeand alfalfa. They include soil, canopy and atmos-phere characteristics. Additionally, a meteorologi-cal station provided the routine climatic data.

1.1.2.1 Soil measurementsThe soil measurements included the permanentcharacteristics of the soils such as texture, chemicalproperties, thermal and hydraulic resistance andcapacity, as well as dynamic characteristics such asmoisture, water potential and roughness.


1.1.2.2 Canopy characterisationThe main structural variables such as the organ areaindex including LAI, the height and the gap fractionwere measured during the whole growth cycle. Ad-ditionally, biomass amount and partitioning intoorgans were also measured concurrently. Particularattention was paid to the water content for exploita-tion of µ-wave data.

1.1.2.3 Microclimatic dataMicro climatic measurements were implementedover a selection of fields to characterise the fluxesin the soil-vegetation-atmosphere system. Theyinclude radiation, temperature, humidity and windspeed sensors, and soil heat flux devices. The sen-sors were placed in conditions allowing to computethe fluxes from the aerodynamic and Bowen ratiomethods. These measurements span over the wholegrowth cycle. Additionally, Eddy correlation meas-urements were acquired for shorter periods. Finally,during an intensive campaign, scintillometer meas-urements and unmanned plane acquiring tempera-ture and moisture measurements were implementedto analyse the spatial variability of fluxes.

1.1.2.4 Meteorological data and atmos-phere characterisationA meteorological station was installed in the centreof the site, with all the classical variables measuredat a 20 minute time step during the whole experi-ment. Radiosoundings were performed close to thesite every day, and additionally, some balloonswere launched during intensive campaigns from thesite itself –at a higher frequency. Finally, an auto-matic CIMEL station was measuring routinely theatmospheric aerosol and water vapour opticaldepths for correction of remote sensing data.

1.1.3 The data baseA data base with all the measurements and imagesand the associated documentation files was devel-oped. It is a hierarchic data base hosted by an INRAweb server (www.avignon.inra.fr/reseda). It is ac-cessible to the whole scientific community.

1.2 Estimation of canopy andsoil characteristics from re-mote sensing data

This will be reviewed as a function of the wave-length domain.

1.2.1 Solar domainThe atmospheric correction was simply applied toimages based on sun-photometer measurements.However, the atmospheric correction based on theimage data itself is a main issue of investigationwhen using remote sensing within surface radiativetransfer models to get estimates of canopy or soilbiophysical variables. This is certainly one avenue

of research to develop. The Alpilles data base canbe used only partly for this purpose, because thisrequired more than the few bands provided bySPOT or POLDER. The DAIS data itself, becauseof their poor radiometric calibration stability are notvery pertinent to investigate this problem.The directionality of reflectance was investigatedboth for data normalisation, and for the estimationof hemispherical reflectance and thus albedo deri-vation. Models with 2 to 4 parameters (MRPV,GEN, Walthall, FLIK) allow to get a robust andaccurate estimates of the BRDF from a sub-sampling of directions available from POLDERdata. However, when estimating albedo values inthe whole solar domain (300-3000µm) from hemi-spherical reflectance observed in few bands, a sys-tematic bias appears and should be corrected for.We showed also that hemispherical reflectance andalbedo estimation is little sensitive to scaling ef-fects.The amount of information contained in the direc-tional dimension appears to be limited. However,we also showed the interest of particular configura-tions for estimation of some biophysical variablessuch as LAI and chlorophyll content. However,these results should be validated from the POLDERdata available in the ReSeDA data base. A tentativewas made for height estimation using an hybridgeometric-turbid medium model applied on thePOLDER data. However, poor performances wereobtained. This should be re-conducted over moresimple biophysical variables such as LAI and chlo-rophyll content.The spectral dimension was the one mostly ex-ploited. However, no particular investigation wasfocusing on the optimal spectral sampling problemfrom the experimental point of view, because of thelack of appropriate data.Investigation about the model inversion problemshows that:• Little differences are observed between the 3 ver-

sions of turbid medium radiative transfer modelsused.

• Techniques based on minimisation over the bio-physical variables (neural nets, vegetation indi-ces) appear performing better than those based onthe minimisation over the reflectance (Look uptables, optimisation).

• However, the, main factor to pay attention at isthe amount of information actually used in theinverse problem.

• Flux variables such as fCover and fAPAR arebetter estimated than primary variables such asLAI and chlorophyll content.

• fCover and fAPAR are little sensitive to scaling asopposed to LAI

1.2.2 Thermal infrared domainThe atmospheric correction problem was investi-gated using a range of techniques. They gave quiteconsistent results. The importance of applying at-mospheric corrections was clearly demonstrated.


The spectral dimension was exploited for emissiv-ity estimation from the DAIS data. However, thepoor calibration stability of DAIS limited the re-sults.The directional dimension was investigated thanksto the INFRAMETRICS camera. Important direc-tional effects ranging from 2° to 4° were observed.However, the variation between the acquisition ofthe directions over a single pixel due to the tempo-ral dynamics has to be carefully removed.

1.2.3 µ-wave domainThe first issue investigated is classification. Thiswas achieved using the series of ERS and Radarsatimages along the growth season. It appears criticalto filter the speckle prior to the classification whenmade on the pixel basis. The Gamma map filterappears the most efficient. The dual angle classifi-cation provided by the combination of ERS andRadarsat did not always provide improved perform-ances. However, for a single date, the potential ofthe combination of active and passive µ-waves wasclearly shown for classification purposes.

One of the main problem in retrieving soil moisturefrom µ-wave data is the confounding effect of soilroughness. Several attempts were developed tosolve this problem.• A new radiative transfer model was developed

based on a fractal description of soil roughness.Confrontation with experimental results appearsquite satisfactory.

• The IEM model is based on a description of soilroughness through the correlation length. How-ever, it appears that it does not perform satisfacto-rily when using the actual correlation length. Aneffective correlation length was adjusted over aset of experimental data. A strong correlation wasobserved between the effective correlation lengthand the actual root mean square of height. Appli-cation on the ReSeDA data base shows that thisapproach is quite robust.

• A simple linear relationship relates the soilmoisture to the backscattering coefficient. How-ever, the slope appears to be almost constant, theintercept depending on the roughness and soiltype. However, when investigating the relation-ship at a larger scale, the variation in roughness issmoothed out, and the slope and intercept aremuch better defined.

• In the passive µ-wave domain the roughnessappears to have little influence as compared tothe active domain (at least band C, 20° incidence).This results in a strong linear relationship betweenthe brightness temperature in band X, 20° inci-dence, and the first top centimetre soil surfacemoisture.

Further work is to be conducted on the comparisonof the retrieval performances of the different ap-proaches and models listed above. The ReSeDAdata base is well suited for this purpose amongstother data sets.

Concerning canopy biophysical variables retrieval,focus was mostly on LAI and canopy water con-tent. Several approaches were investigated:A new discrete model was developed that shows thedrastic differences of the response between smallleaves (wheat) and large leaves (sunflower) cano-pies. Further, the model, when inverted on actualERS data during the ReSeDA experiment appearsperforming satisfactorily.The cloud model was calibrated and then invertedover several canopies (wheat and sunflower). Itappears performing quite satisfactorily, with anuncertainty (RMSE) associated to the estimationclose to 0.5 kg.m-2 for water content and 1.0 forLAI. For the retrieval of soil and canopy character-istics, the use of simultaneous observations in twodifferent configurations appears mandatory. Opti-mal configurations were discussed based on thisdata set for wheat crops. Here again, one of themain limit is the unknown surface roughness and itspossible variation along time.

1.3 Assimilation of remotesensing data into processmodels

The assimilation of remote sensing data was inves-tigated in two separate parts corresponding to thetwo categories of process models: soil vegetationatmosphere transfer and canopy functioning mod-els.

1.3.1 Assimilation into SVAT models

• The SVAT models were first compared andevaluated. For this purpose, a strategy for com-parison was developed, based on the evaluation ofthree main criteria: (i) the accuracy of processesdescription; (ii) the portability of the models;(iii)the robustness of the calibration. In the compari-son process, the calibration and validation phaseswere as independent as possible. Results showthat despite the large range of complexityamongst the models considered, they were per-forming similarly. However, the simplest models(ISBA, MAGRET) were performing generallybetter than complex models such as SISPAT forwhich the calibration appears not very robust,particularly for the turbulent heat fluxes. This waslater investigated by a sensitivity analysis for thisSISPAT version considering a single homogene-ous soil layer. It shows that water content andland surface water and energy fluxes are very sen-sitive to soil hydraulic properties for medium tolow moisture levels. It was therefore concludedthat SISPAT has to be implemented with morethan a single layer to describe accurately surfacesoil water moisture.

• The aerodynamic resistance scheme used in thetwo-layer SVAT model, affects strongly the sur-face fluxes and temperature simulation. Six pa-rameterisation scheme implemented in the


SISPAT model have been compared. Resultsshow that the total surface fluxes were marginallyaffected. However, the main difference lies in thesimulation of the soil and vegetation tempera-ture because of the change in the partition of en-ergy between the soil and vegetation. These dif-ferences are critical for the description of the ra-diative temperature as observed by remote sens-ing in the thermal infrared.

• The SEBAL algorithm exploiting the spatialvariation in the optical domain was evaluated.The several assumptions in the algorithm wereevaluated experimentally. Then the estimates offluxes were compared to measurements and showgood results for the net radiation, but poorer re-sults for soil, latent and sensible heat fluxes. Thiswas mainly explained by the poor parameterisa-tion of the aerodynamic resistance.

• The spatial variability of fluxes was investigatedover contrasted neighbour fields using an un-manned plane. The structure of the lower atmos-phere layers described this way confirms the con-trasts at the borders, and sets the blending heightabove 30m. Scintillometer measurements set upover similar field boundaries between contrastedcrops show the importance of the aggregationscheme.

1.3.2 Assimilation into Canopy func-tioning models

• Similarly to SVAT models, the first task was di-rected to the calibration of the several modelsconsidered. This was achieved over the calibra-tion fields, or over data sets independent from theReSeDA experiment. The calibration was mainlyconsisting in tuning the parameters related to thephenology, the biomass partitioning and the waterbalance. In Addition, to use more realistic radia-tive transfer models, a more detailed canopy –structure and optical properties of elements wasimplemented for the STICS model.

• Several assimilation schemes were used.- The ROTASK canopy functioning model wasforced by estimates of LAI using vegetation indi-ces. The use of remote sensing data (SPOT andTM data) shows significant improvement of yieldestimation at the field scale, although absoluteaccuracy could be improved. The same schemewas implemented at the regional scale overMARS segments and shows that it could be op-erational.– The STICS-RT model was used to assimilatePOLDER and ERS/Radarsat data over thegrowing season. A preliminary sensitivity analy-sis was used to identify the variables to tune dur-ing the assimilation process. They mainly corre-sponds to phenology, stand density and top layersoil characterisation. The assimilation of POL-DER improved significantly estimates of totalbiomass production. Concurrent use of µ-wavedata did not contribute much to improve the drybiomass estimation, nor the water balance.

2. The partners

2.1 List of partners

The ReSeDA project is a collaborative effort of 10partners. They are listed above, with the name ofthe main investigators:INRA: CoordinatorFrédéric Baret, INRA Bioclimatologie,Site Agroparc, 84 914 cedex 9 FRANCEtel: (33) 4 90 31 60 82; fax: (33) 4 90 89 98 10Email: [email protected] Leroy,CESBIO, 18 Avenue Edouard Belin,31 055 Cedex, FRANCETel: (33) 5 61 55 85 14; fax: (33) 5 61 55 85 00Email: [email protected] Ottlé,10-12, avenue de l'Europe,78 140 Velizy, FRANCETel: (33) 1 39 25 49 18; fax: (33) 1 39 25 49 22Email: ottlé@cetp.ipsl.frBRGMChristine King,BRGM, BP 6009, Avenue de Concyr,45 060 Orléans, Cedex 2, FRANCETel: (33) 2 38 64 33 92; fax: (33) 2 38 64 33 61Email: [email protected] Clevers,Department of Landsurveying and Remote Sensing,Hesselink Van suchtelenweg 6,P.O.Box 339, 6700 AH Wageningen, THE NETH-ERLANDSTel: (31) 317 482902; fax: (31) 317 484643Email: [email protected] van Leeuween,SYNOPTICS Integrated Remote Sensing & GISApplications, P.O. Box 117,6700 AC, WageningenThe NetherlandsTel: (31) 317 42 69 36; fax: (31) 317 42 57 05Email: [email protected] Jongschaap,DLO Research Institute for Agrobiology and SoilFertility,P.O. Box 14,6700 AA, Wageningen,The NetherlandsTel: (31) 8370 75972; fax: (31) 8370 23110Email: [email protected] Pampaloni,IROE, Via Pancuatichi, 64, Firenze 50 127, ITALYTel: (39) 55 4235 205; fax: (39) 55 4235 290

mailto:[email protected]


Email: [email protected] Caselles,Departament de Termodinamica, Facultat de Fisica,Universitat de Valencia,Dr. Moliner, 50. Burjassot 46100 SPAINTel: (34) 6 386 43 50; fax: (34) 6 364 23 45Email: [email protected]

UCStephen Hobbs,Space systems and Applications Group,College of Aeronautics,Cranfield University, Bedford MK43 0AL, GBTel: (44) 1234 750111; fax: (44) 1234 752149Email: [email protected]

2.2 Role of partners

Each partner has a well identified responsibilityaccording to the following chart:

CoordinationINRA

WP1 ExperimentINRA

WP2 Model inversionCESBIO

WP3 AssimilationAB-DLO

WP11 Satellite dataBRGM INRA

WP12 Airborne dataINRA CESBIO IROE

WP13 Ground measurementsINRA CETP UV BRGM WAU

CESBIO ...

WP14 Data BaseINRA, Synoptics, BRGM

WP21 Optical DomainCESBIO, INRA, WAU

WP22 µ-wave domainCETP, INRA, IROE

WP23 Thermal Infrared domainINRA, UV, CETP

WP31 Assimilation into SVATINRA, CESBIO, AB-DLO

WP32 Assimilation into canopy functionning

AB-DLO, CESBIO, WAU

Figure 3 Chart showing the structure of the project. Each Working Package is under the responsibility of onepartner (in bold). Other partners participating actively are also indicated.

3. Rationale

3.1 Main issues of concern

Two main issues of concern are considered withinthe ReSeDA project which are related to human ac-tivity at two different scales:

3.1.1 Global scale issues.Since the last decades, evidence of significant globalclimate change due to the constantly growing humanactivity have been accumulated. This is obviously amain concern for governments because this climaticchange directly impacts human's conditions of life.Strong warning were already issued by recent inter-national initiatives organised within the United Na-tions. A large part of the scientific community isnow investigating these key questions through inter-national programmes such as WCRP (World Cli-mate Research Programme) and IGBP (InternationalGeosphere-Biosphere Programme). They need to be

mailto:[email protected]


answered to forecast climate change for the nextdecades according to scenarios of human activity,and therefore to propose recommendations for gov-ernments to minimise the impact on human beings.This was recognised during the last Internationalconferences (Kyoto, Rio) on climate change (IPCC),and actions were taken in order to mitigate the ef-fects of human activity. Vegetation constitutes theinterface between the atmosphere and the soil, andtherefore plays a key role in energy and mass trans-fers. Water and carbon cycles are obviously the mainprocesses of concern. Although the objective is tounderstand the climate at the global scale, many pro-cesses take place at the local to regional scales as wewill see later. Therefore, investigations about theglobal issues should be addressed also at the local toregional scales.

3.1.2 Local to regional scale issuesApart from the understanding of the processes ofinterest for global change issues, great attention isalso paid to the environmental problems that occurat the local to regional scales. In particular, intensiveagriculture practices induce strong constraints to thequality of the environment. The European commu-nity has taken actions to promote sustainable agri-culture practices in order to reduce the impact ofagriculture on the environment.Additionally, estimation and forecast of the produc-tion of crops, its year to year variation, as well as itsspatial distribution is a highly valuable informationfor the organisation of the harvest and the manage-ment of the market for the main crops. This issueconcerns a range of users, from the local scale suchas farmers and local traders, to the global scale forthe biggest traders, the transformation industry, aswell as governments and non governmental organi-sation for the food security problem.

3.2 The need in modellingvegetation and soil processes

3.2.1 Processes and modelsThe evolution of the global climate and its conse-quence on ecosystems has been intensively investi-gated. Global Circulation Models (GCMs) were de-veloped to simulate the climate change both at shorttime periods (weather forecast) and at longer timeperiods (global change). They require the surface tobe accurately described in order to properly modelits interaction with the atmosphere, and particularlyenergy, water, and carbon fluxes.Environmental issues that take place at the local toregional scales could be also described by vegetationand soil models. Vegetation (canopy functioning)models will describe the dynamics of the vegetationover the growing season as well as the energy, wa-ter, carbon, and sometimes nitrogen and other min-eral fluxes. Soil Vegetation Atmosphere Transfer

models (SVAT) provide also a physical descriptionof energy and mass fluxes (mainly water and car-bon) at the instantaneous time scale.All these models could be used both as a way toformalise our understanding of the processes, butalso as tools for diagnostics as well prognostics inorder to correct or remedy negative effects on theenvironment.All those processes are complex and are still lackingadequate modelling as well as inputs. They need tobe coupled together to get a comprehensive and con-sistent view of the system.

3.2.2 The range of spatial and tempo-ral scalesProcesses occur mainly at the very elementary timescale, but with a range of time constant. Therefore,in order to be able to measure significant trends, arange of typical time scales have to be considered. Itgoes from the instantaneous time scale for energyand mass fluxes within the atmosphere, up to thegrowth season for yield production, and several dec-ades for impact of climate change on vegetation dy-namics.The typical spatial scale where many processes canbe described is the local scale corresponding to fewtenths of meters. However, vegetation and soils areconstrained by the atmospheric boundary conditionsfor energy or mass fluxes at short time periods, aswell as for their evolution over longer time periods.Conversely, the atmosphere dynamics depends onthe boundary conditions determined by the soil orvegetation characteristics (Figure 4). Furthermore,hydrological processes, and atmospheric advectionconstrain the local processes at the regional scale.The processes are therefore intimately coupled at arange of scales. The high spatial and temporal vari-ability observed at the local to the regional scaleshave to be accounted for when using process mod-els. The same applies for measurements or estimatesof the state variables and the fluxes over scales withsignificant spatial heterogeneity.

LOCAL SCALE: EcosystemsWhere the elementary processes take

place

Forcingecosystems'models withGCMs, andhydrosystems

REGIONAL SCALE: Landscape, WatershedIntegration of elementary processes for hydrosystems

and ecosystems evolution

GLOBAL SCALECoupling between processes at several

scales

Boundaryconditionsfor GCMs

Figure 4. Processes are coupled over a range ofscales.


3.2.3 The limits of process models:number and accuracy of the inputsCanopy functioning models will be used as para-digm. Canopy functioning models describe the ele-mentary processes such as photosynthesis, respira-tion, biomass partitioning, water and nitrogen trans-fers. They use within a deterministic scheme, cli-mate variables and soil characteristics. These modelsallow to estimate directly variables of interest suchas the biomass production, water and nitrogen bal-ance, as well as the yield. However, they require alot of parameters and variables as inputs. Some ofthem are poorly known and should be calibrated forthe particular conditions where the canopy is sub-jected. An example of the poor performances of suchcanopy functioning models when run mainly overmeteorological and soil data was demonstrated byCramer and Field (1999). They showed a wide rangeof variation between the simulation of net primaryproduction (NPP) models in these conditions (Figure5).

Figure 5. Comparison between a range of canopyfunctioning models for the simulation of the netprimary productivity (NPP) when mainly driven bysoil and climate variables From Cramer and Field(1999).

Franks and Beven (1997) proposed a fundamentalexplanation apart from the simple differences be-tween models. Complex models use a rather largenumber of variables and parameters to describe bio-physical processes. The accuracy of estimates ofsuch variables is sometime very poor, leading tolarge uncertainties in the model simulations. Al-though information is put in the models within theknowledge about physiological processes, it is stilltoo limited to get good enough performances. There-fore, additional sources of information is requiredeither to directly estimate input variables to the pro-

cess model, or to control the behaviour of the modelby checking on few outputs of the model. Remotesensing techniques could play a very important roleat this level as we will see in the following.

3.3 The need in remote sens-ing observations

3.3.1 Role and diversity of remotesensing systemsRemote sensing techniques have been widely devel-oped and used in the past years for several applica-tion domains including the continental biosphere.However, for natural and cultivated areas, satellitedata were mostly dedicated to the mapping and theinventory of crops and natural resources. Earth ob-servation sensors (Landsat TM, SPOT), large swathsatellites (NOAA/AVHRR, METEOSAT, VEGE-TATION) and radar systems (ERS, Radarsat, JERS)were the most important sources of images. Thesesensors were mainly used in qualitative and relativeways that were nevertheless very useful. However,their potentials is under-exploited, particularly re-garding the combination with the models that can beused for a better understanding, diagnosis and prog-nosis of the processes of interest. Emphasis is cur-rently put on quantitative approaches that could pro-vide estimates of surface characteristics from remotesensing data.The signal that is recorded by remote sensing in-struments is driven by physical processes governingthe radiative transfer within the soil, canopy and theatmosphere. The interaction of radiation with cano-pies and soils depends on the optical thermal or di-electric properties of the elements as well as on theirnumber, area, orientation and position in spacewhich constitute the primary biophysical variables.Therefore remote sensing allows to derive directlyonly canopy or soil primary biophysical variables.Additionally, secondary variables that are combina-tions of the primary biophysical variables could bealso estimated generally with a quite good accuracy.The main primary and secondary variables are listedin Table 1. Each spectral domain is sensitive to par-ticular canopy or soil characteristics. The reflectiveoptical domain (400-2500nm) will provide estimatesof canopy structural variables such as LAI (leaf areaindex) and biochemical composition (pigment, wa-ter, ...). Thermal infrared domain and passive µ-wave will depend on the surface temperature andcanopy structural variables. The active microwavedomain will provide information on soil roughness,moisture, as well as canopy structure and watercontent. Therefore, the combined use of severalspectral domains is likely to be the only way to ex-tract the maximum amount of information on canopyor soil biophysical characteristics.


Spectral domains


Vis

ible

Nea

r In

frar

ed

Nea

r In

frar

edS

hort

Wav

e In

frar

ed

The

rmal

Inf

rare

d

Act

ive

µ-w

ave

(rad

ar)

Pas

sive

s µ

-wav

e

LAI +++ +++ + ++ +Leaf orientation +++ +++ + + +Leaf size and shape + + + + +Canopy height - - - ++ -

Canopystructure

Canopy water mass +++ +++Chlorophyll content +++ - - - -Water content - +++ - +++ +++

Leafcharacteristics

Temperature - - ++++ - ++Surface soil moisture - + + +++ +++Roughness + + - ++ +residues +++ ++ - -Organic matter ++ ++ - - -

Soilcharacteristics

Soil type ++ ++ +fCover ++++ ++++ ++ ++ +fAPAR ++++ ++++albedo ++++ +++

Secondaryvariables

Long wave flux - - ++++ - -

Table 1. Estimation of canopy, leaf or soil biophysical variables as a function of the spectral domain used. Thelevel of accuracy and robustness of the estimation is indicated by the “+” (“++++” accurate and robust; “-“ noestimates possible). Secondary biophysical variables are also indicated.

We note that the number of biophysical variablespotentially accessible by remote sensing techniquesis rather large. However, their estimation from theradiometric data is not always very accurate becauseof the number of biophysical variables that drive theradiative transfer and the limited information contentin the spectra variation of the radiometric signal. Theonly way to improve these estimates is to exploitother dimensions than the spectral dimension, aswell as ancillary information.Current and future Earth observation sensors aboardsatellites will almost cover the whole electromag-netic spectrum, from the optical domain to the mi-crowave (SPOT, TM, AVHRR, VEGETATION,POLDER, AATSR, RADARSAT, ERS, JERS,MERIS, MODIS, ASAR, MISR, MSG, LSPIM,SMOS ...). This situation is new and forces to inves-tigate the combined use of these various datasources. This will therefore also implies to explicitlyaccount for other dimensions than the spectral one:• Temporal dimension. All the sensors are not

borne on the same platform. They will thus pro-vide observations of the same target at differentdates or hours in the day. We need interpolationtechniques to use concurrently those multi-sensordata. The vegetation or soil are dynamic targets.Their time course brings in itself a lot of informa-tion that can be used to characterise their func-tioning. This requires to have robust methods to in-fer surface parameters from remote sensing over

the entire range of surface conditions encounteredover a whole growth cycle.

• Directional dimension. Many systems will beable to provide observations of the same target un-der different view and sun geometrical configura-tions. Algorithms dedicated to correct or normalisefrom these directional effects are required. Fur-thermore, this can be also used to get more infor-mation on the structural characteristics of thevegetation or soil. Polarisation features in the visi-ble and the microwave domains may also provideimportant information on the properties of the sur-face and the atmosphere.

• Spatial dimension. The spatial resolution is oftenclosely related to the temporal repetitivity of ob-servations. For example, large swath satellite thatare able to acquire frequent images have a lowspatial resolution (NOAA/AVHRR, METEOSAT).Conversely, high spatial resolution (SPOT, TM)have poor revisit capability. Further, to addressclimate issues by providing to global circulationmodels surface variables such as albedo, roughnessand moisture, the large scale is the one to be con-sidered. The several sensors that can be used mayhave different spatial resolution. Thus, the up-scaling and down-scaling problems are intimatelyassociated to the scientific issues addressed (envi-ronment and climate monitoring) as well as thetools available (the several satellite systems avail-able in the future).


To be able to concurrently use data provided by allthe current and future satellite systems with theirspectral/temporal/directional/spatial characteristics,vegetation and soil models provide a convenient wayto assimilate this complex information. Further, theuse of the temporal variation observed throughoutthe growth cycle is likely to provide pertinent infor-mation on the mechanisms underlying changes in thefunctioning of soil and vegetation.

3.3.2 From radiative transfer modelinversion to assimilation into processmodels.Retrieval of information on canopies or soils fromremote sensing observations can be achieved usingtwo approaches.

3.3.2.1 Inverse methods: exploitation of in-stantaneous measurements.Radiative transfer and surface reflectivity modelsdescribe the interaction between vegetation soil oratmosphere and the electromagnetic radiation. Theinversion of these models can provide estimates ofcanopy or soil biophysical characteristics from theirspectral/directional/polarisation variation acquiredinstantaneously, or at least under the assumption of asteady state for soil and vegetation. Several inver-sion methods have been developed by the geophysi-cal scientific community. However, they are gener-ally limited by the amount of information exploited,and have difficulties in solving problems such asuniqueness of the solution, ambiguities or the effectof confounding variables. Therefore, improvementof the retrieval performances could only come fromthe exploitation of additional information such asmore detailed canopy structure description, knowl-edge on uncertainties on measurements and models,and a priori information on the distribution of thevariables. Nevertheless, these methods will provideonly estimates of primary and secondary variableswhich have to be ingested into canopy and soilfunctioning models to get the full description of theprocesses of interest (Figure 6). All these limitationsof inverse methods will be largely attenuated whenusing assimilation methods as we will see in thefollowing.

3.3.2.2 Assimilation methods.Vegetation and soil process models have been de-veloped to describe our current understanding of thephysical and biophysical processes that occur be-tween the atmosphere, vegetation and soil. Thesemodels describe the energy and mass budget at theinstantaneous time scale and provide also estimatesof the time course of soil and vegetation state vari-ables over the growing season. Two types of modelsare generally distinguished at this local scale: Canopy functioning models. The biophysicalprocesses that govern canopy functioning and dy-namics are described in these models,. They con-trol plant phenological development, photosynthe-sis, respiration, water, carbon and nitrogen bal-

ance, and allocation of resources between the sev-eral parts of the canopy (leaves, stems, roots,fruits, ...). Soil Vegetation Atmosphere Transfer models.(SVAT models) describe the physical processesthat control the transfer of energy and mass(mainly water) in the soil vegetation atmospherecontinuum. They generally require information oncanopy structure and climatic variables such as theincoming radiation, air temperature and humidityand wind speed. In addition to the fluxes, they maybe used to estimate the surface temperature andmoisture distribution within the soil or the canopy.SVAT models may be coupled with canopy func-tioning models to be able to describe the energyand mass transfer along a whole growth cycle.

Canopy state, Soil canopy atmosphère exchanges Quantity & quality of the production

Environmental Budget

Canopy functioningand SVAT Models


Radiometric Data

Radiative TransferModels

Inv

ers

ion

Ass

imil

ati

on

Figure 6. Flow chart diagram showing how to inferend variables such as canopy net primary production(NPP), yield or water fluxes from assimilation ofremote sensing data into canopy and soil models.The inversion is restricted to the retrieval of canopybiophysical variables using radiative transfer modelsin the inverse mode. The green arrows indicate theway models work in the forward direction.

SVAT and canopy functioning models contain adescription of the biophysical characteristics such ascanopy structure and optical/dielectric properties ofthe soil or vegetation elements that govern the ra-diative transfer. They thus can be coupled to radia-tive transfer or surface reflectivity models to simu-late what remote sensing systems actually observe,i.e. the temporal time course of the spec-tral/directional/polarisation signature of soil andcanopies. They thus may be used to "assimilate"remote sensing data. The assimilation process willconsists in tuning SVAT and soil or canopy func-tioning model parameters so that model outputs suchas the time course of reflectance, brightness tem-perature or backscattering coefficient, agree withactual remote sensing observations. Thus the as-similation will provide deeper characterisation of thesurface processes as opposed to the state variablesprovided by the inverse methods (Figure 6). As-similation techniques can be used to set initial con-ditions to SVAT or canopy functioning models as


well as define the optimal values of parameters de-scribing particular processes. Once calibrated andinitialised, these models can be used later in a pre-dictive mode that does not require necessarily re-mote sensing information. This use of process mod-els is quite interesting in the perspective of theevaluation of median to long term effects of envi-ronmental or climatic changes.

4. Objectives & timeframe of the project

4.1 Objectives of ReSeDA

The main objective of the ReSeDA project is the useof multi-sensor and multi-temporal observationsfor monitoring soil and vegetation processes, inrelation with the atmospheric boundary layer atlocal and regional scales by assimilation of remotesensing data into canopy and soil functioning mod-els. This study aims at developing and evaluatingmethods to estimate net primary productivity, waterand energy fluxes over cultivated vegetation. It willprovide recommendations useful to: improve information retrieval from the multi-

plicity of sensors that are or will be availablewithin few years,

deliver improved methodologies for the inter-pretation of multi-temporal/multi-sensor remotesensing data,

drive the policy of technological developmentsof space observations.

The project is decomposed into three main taskscorresponding to the three work packages:

WP1-The experiment.A consistent and comprehensive data set has beenacquired during a whole growth season allowingcalibration and evaluation of the algorithms pro-posed, both at the field scale (10-100m) and the 1kmscale corresponding to large swath satellite observa-tions. Remote sensing systems will be operated tocollect data in a variety of spectral ranges and con-figurations (optical, thermal infrared and micro-wave). Measurements span over a whole year longwith emphasis on the multi-temporal aspect. Thisdata set was exploited in the later tasks for scientificpurposes. Further, it is now available to the wholescientific community via the web site:www.avignon.inra.fr/reseda.

WP2-Evaluation of inverse methods.Canopy or soil biophysical variables (structure, tem-perature, moisture, ...) have been be retrieved fromremote sensing observations using mainly the spec-tral, directional and polarisation signatures. Analyti-cal approaches using radiative transfer and surfacereflectivity models as well as semi-empirical ap-proaches have been used. Biophysical variables re-trieved through these inverse methods have beencompared to the values measured in the fields. Theconcurrent use of large scale systems providing fre-quent observations and local scale sensors have beeninvestigated and the scaling issue discussed. Thiswork package was decomposed into three sub-packages corresponding to the spectral domains con-sidered: solar, thermal infrared, and µ-wave.

WP3-Evaluation of assimilation methods.Canopy functioning models and SVAT models havebeen tuned so that the temporal, spectral, directionaland polarisation signature simulated matches theobserved remote sensing signals as close as possible.Several approaches of data assimilation have beenevaluated with emphasis on the use of the wholeremote sensing data available and knowledge of thephysical and physiological processes governing soiland vegetation functioning. They have been mainlystudied at the field scale.

4.2 Time frame

The ReSeDA project was planned for a three yearduration. According to the original schedule, thedata processing phase was largely underestimated,which resulted in delays for the two last workingpackages. These delays were mainly due to severalproblems observed in the measurements and thattook a lot of time to be sorted out and corrected for.These problems were associated mostly to the air-borne instruments which were prototype systemsthat had not the degree of operationallity expected.The delay was partly compensated by a two monthsextension.A General co-ordination meeting was organised atthe end of each year, followed by the annual report.Additionally, meetings in subgroups were also or-ganised in order to have deeper discussion around aparticular topic


Year 1 2 3Term (3 months) 1 2 3 4 5 6 7 8 9 10 11 12WP1:Data acquisitionDetails of the experimental plan and preparationDefinition of the data base format,Satellite data acquisition and processingAirborne data acquisition and processingGround level data acquisition and processingConstitution and validation of the data baseAccess to the whole scientific communityWP2: InversionSolar domain.Thermal infrared domainµ-wave domainWP3: AssimilationCalibration of SVAT and canopy functioning modelsAssimilation in SVAT modelsAssimilation in canopy functioning modelsAssimilation in coupled SVAT/canopy modelsMilestonesMain Co-ordination Meetingannual reportFinal/Mid term report

Table 2. Flowchart showing the main milestones along this 3 year project.

5. AchievementsThe achievements will be described according tothe three main work packages identified previously.

5.1 The experiment and thedata base.

5.1.1 Description of the Alpilles testsite.

The ReSeDA site is located near Avignon (SE ofFrance) in the Rhone valley (Figure 7). Its maximum dimension is approximately5km*5km. It is a very flat area. Fields are largeenough (200 m x 200 m) to extract pure pixels fromhigh spatial resolution satellites, as well as to im-plement atmospheric fluxes measurements.The main crops are wheat ( 32%), sunflower (20%),maize (9%), and grassland (16%). A detailed andexhaustive classification was achieved over the testsite (Figure 2). The measurements over the sitestarted in October 1996, and were completed inNovember 1997.

5.1.2 Satellite data

5.1.2.1 Satellite sensors and images ac-quiredThe satellite images were acquired during thewhole experiment according to the scheme pre-sented in Figure 8. They thus include SPOT, Land-sat TM, NOAA/AVHRR, ERS 2, ATSR2, and Ra-darsat and cover the whole wavelength spectrumand most of the growing season for winter andspring crops. Table 3 shows the characteristics ofthe satellite sensors used.

Figure 7. Situation of the Alpilles/ReSeDA Site.

Alpilles


Sensor Spectral domainViewangle

Swath(width)

SpatialResolution

PolarizationRevisit Fre-

quencyERS C-band 23° 100 km 20 m VV

µ-wavesensors Radarsat C-band

23°38°

150 km50 km

25 m9 m

HH

LandsatTM

0.45-0.52 µm0.52-0.60 µm0.63-0.69 µm0.76-0.90 µm1.55-1.75 µm10.4-12.5 µm2.08-2.35 µm

0° 185 km 30m

non appl. 18 days

NOAA-AVHRR

0.58-0.68 µm0.72-1.10 µm3.55-3.93 µm

10.3-11.30 µm11.5-12.50 µm

± 55° 2400 km 1.1 km

non appl. 1 daySolar &Thermalsensors

SPOTXS

0.50-0.59 µm0.61-0.69 µm0.79-0.89 µm

0° 60 km 20 mnon appl. 26 days

Table 3. Characteristics of the satellite sensors used during the ReSeDA experiment.

2/10 1/11 1/12 31/12 30/1 1/3 31/3 30/4 30/5 29/6 29/7 28/8 27/9 27/10

Date of observation

SPOT

Landsat TM

Radarsat 39°

Radarsat 23°

ERS

NOAA/AVHRR

Figure 8. The satellite images acquired during the ReSeDA experiment and available on the server.

5.1.2.2 Processing of the imagesThe geometric correction was performed for allsatellite images as well as airborne sensors thanksto a reference SPOT image. The geometric correc-tion was mainly made thanks to a network ofground control points. This reference image wasthen precisely transformed to match a raster imageof the topographic map at a scale of 1/25000. TheProjection system adopted here is the extendedLambert II.• For SPOT series in XS mode and TM image.

Prior to geo-correction, the data were correctedfrom the MTF. The data were then radiometri-cally corrected. Atmospheric correction (MOD-

TRAN) was applied, based on the ground levelmeasurements of the atmospheric characteristicscontinuously acquired during the experiment.

• Radar images: the geo-coding uses a DEM(BRGM) and the description of orbit. An impor-tant point is to be emphasised : the size of pixel inthe corrected image is the same as the DEM. Toavoid an important degradation of the radar reso-lution, a sub-sampling of the original DEM hasbeen artificially extracted: 25m for ERS and Ra-darsat mode standard, and 12,50m for Radarsat infine mode. This choice keeps relatively high spa-tial resolution of the corrected images by com-parison with the raw data, while saving computertime and minimising the size of the saved image


files. Calibration and filtering (gamma map filter)were finally applied to the data in order to get theproper physical value of the back-scattering coef-ficient σ0.

• NOAA/AVHRR raw data are simply stored onthe data base. No particular corrections or trans-formations have been applied.

5.1.2.3 The data base• Copyrights. Copyrights have been obtained so

that the data corresponding to the 5*5km² zoneare available for free to any users. The use of thewhole image must generally be made with agree-ment with ReSeDA partners. In addition, it isasked to the people who use these data to ac-knowledge the ReSeDA people who acquired andprocessed the images.

• Data format. Three export formats are archivedfor each image: ERDAS (*.img), TIFF (*.tif) for8 bit images, and binary (*.dat). In addition tothese imagettes, an EXCEL synopsis file providesthe average and standard deviation values of aselection of 43 plots within the test site. This willfacilitate the use of satellite data when the spatialarrangement is not required.

• Data base architecture. The data are available atthe ReSeDA server at INRA that is fully accessi-ble to the scientific community. The files are or-ganised in directories per sensor. Each directory isorganised in three sub-directories correspondingto the three formats used.

•

SENSOR Waveband View angle Swath Resolution Altitude Polarisation Data typeC-band 20° and 40° - 20 m 300 m VV/HH/HV profiles

ERASMEX-band 20° and 40° - 20 m 300 m VV/HH profiles

RENÉ S-band 20° and 40° - 18 m 300 m Full polar. profiles6.8GHz

µ-wavesensors

IROE10GHz

20° and 40° - 150 m H and V profiles

POLDER

443 nm550 nm670 nm865 nm910 nm

± 50° ± 3 km 30m3000m1500m

N/A images

CIMEL540 nm640 nm840 nm

0° - 20m 300m no profiles

Opticalsensors

DAIS0.45-2.45 µm

72 bands0° ± 2 km 7 m 2900 m no images

Heinman 8-14 µm 0° - 20m 300 m no profiles

Inframetrics 9-13 µm ± 40° ± 3 km60 m30 m

3000m1500 m

no imagesThermalsensors

DAIS8-12µm6 bands

nadir ± 2 km 7 m 2900 m no images

Table 3. Characteristics of the airborne sensors used during the ReSeDA experiment.

5.1.3 Airborne data.The objective here was to acquire data that will beprovided by future space systems already scheduledor that represent a very high potential with regardsto the characterisation of vegetation or soils.

5.1.3.1 Sensors and vectors usedFour types of vectors were used:• Small plane. This plane (PIPER PA28) was oper-

ated from Aix en Provence by INRA and CES-BIO. POLDER and INFRAMETRICS sensorswere mounted on it and connected to a device ac-quiring continuously the position (GPS) and theattitude of the plane. This is one of the routinemeasurements, the plane being operated at leastonce per month.

• Helicopter. It was operated by CETP and INRA.ERASME and RENE were mounted on it, with aCIMEL and a thermal infrared radiothermometer.All the instruments are profilers (except RENE).

A video camera was used to identify the fieldsactually sampled.

• ARAT. This FOKKER 27 operated by INSU(CNRS, France) was used for the passive micro-wave measurements thanks to the STAAARTEprogram. The IROE sensor was mounted on it.Additional thermal infrared data (8-14 µm) wereacquired at the same time from the same vector.This sensor was used only at two dates during theexperiment.

• DLR plane. The Dornier plane from DLR wasused for the DAIS instrument. It was used onlyonce during the campaign.The main characteris-tics of the sensors are presented in table 3. Thesensors can be split according to the frequency ofthe flights.

1. Routine sensors that provide at least monthlydata sets with two imaging instruments coveringthe visible, near infrared (POLDER) and ther-mal infrared (INFRAMETRICS) domains and aprofiler scatterometer (ERASME).

ReSeDA. Final report. June 2000. 5-19

• POLDER The Polarised and DirectionalEarth's Reflectance (POLDER) instrumentmeasures the intensity of the sun light reflectedby the Earth/atmosphere system in five spectralbands and under different viewing directions.Further, the polarisation of the light reflected ismeasured in two bands. The flight lines weredesigned in order to get a pertinent directionalsampling of the test site. (Figure 9)• INFRAMETRICS. This thermal infraredcamera is equipped with a wide angle field of

view providing directional measurements in the8-13 µm domain in the same way as POLDERdoes. The thermal camera was most of the timecoupled with POLDER. However, additionalflights were performed using the thermal cam-era only. In this case, the altitude of the flightswas 1500m rather than the regular 3000m alti-tude. This provides a better spatial resolutionalthough the directionality is poorly sampled.

1 4 5 5

Figure 9. POLDER schematic flight plan: four tracks are oriented roughly parallel to the solar direction, and onetrack perpendicular to that direction (left figure). The square represents a 5 km x 5 km area around the Alpillessite. The typical corresponding number of observation for a given pixel in 5 km x 5 km study area is given (rightfigure).

ERASME C and X band, HH and VV. These in-struments were operated concurrently with CIMELand Heinman radiometers to cover the whole spec-tral domain. The flight lines were chosen to get agood sampling of a selection of plots on which in-tensive ground measurements were made. The timeand frequency of the flights were governed by thegrowth of the vegetation, and possible variation insoil moisture and roughness conditions. For thispurpose, we defined two intensive campaigns onein April and the other in June, where we concen-trated the flights and the concurrent ground meas-urements. ERASME was mainly in its C and Xband, in VV and HH polarisation configuration.2. Additional measurements. Additional meas-

urements were acquired, depending on nationalor other European funding opportunities. Theyallow to investigate some complementary in-strument configurations. This includes• ERASME C band HV. Five flights ofERASME in cross polarisation configurationwere scheduled. They were always associated tothe routine ERASME configuration during theintensive campaigns, in order to make the com-parison possible.• RENE S polarimetric radar. RENE was set-up in S band (3.25 GHz), and HH polarisation.

It flew once in each intensive campaign. RENEprovides images unlike ERASME.• IROE sensor. Passive microwave measure-ments were acquired using the IROE sensorsaboard the ARAT thanks to the E.C.STAAARTE program. It measures the bright-ness temperature in the 6.8GHz and 10GHzbands within about 1K accuracy. The flightlines were defined to sample the fields that wereactually characterised with the other profilingsensors such as ERASME.• DAIS sensor. The DAIS flew once the sitethanks to the E.C. DAIS large scale facility pro-gram set up by DLR. The DAIS provides highspectral resolution images in the visible/near in-frared domain and six bands in the thermal in-frared (8-12µm). The flight lines were designedto get a full cover of the site.

5.1.3.2 Flight scheduleThe acquisition scheme is shown in Figure 10where the configuration for ERASME is detailed. Itshows that the objective of getting a regular moni-toring with the routine sensors (POLDER, IN-FRAMETRICS and ERASME) was achieved. Thetwo intensive campaigns in April and June werealso achieved in good conditions with complemen-tary RENE µ-wave images in S band and concur-

ReSeDA. Final report. June 2000. 5-20

rent multiple INFRAMETRICS thermal infrareddata. An other interesting period with almost all thesensors (and configuration for ERASME) available

is centred on the 8th of July with the DAIS spectro-imaging systems.

2/10 1/11 1/12 31/12 30/1 1/3 31/3 30/4 30/5 29/6 29/7 28/8 27/9 27/10

Date of observation

POLDER

ERASME C&X HH/VV

ERASME C HV/HH

RENE S

IROE C&X

DAIS

INFRAMETRICS

Figure 10. Distribution of the airborne remote sensing observations along the experiment (1996-1997)

5.1.3.3 Processing of the dataThe data have been processed in order to get "takeand play" data. This includes geometric correction,radiometric calibration and atmospheric correctionfor solar and thermal sensors. The quality of theprocessing depends on the type of sensor:• POLDER: all the images acquired at 3000 m

have been processed. The geometric correctionachieved an accuracy around ±1.5 pixels, exceptfor few images where it can reach up to 4 pixels.In these cases, better registration can be achievedlocally by correlation techniques. The absolutecalibration was performed thanks to three calibra-tion experiments performed at LOA using an in-tegrative sphere. The accuracy might be close tousual figures in these conditions, i.e. around 2%relative. The radiometric noise was very low. Theatmospheric correction was performed usingSMAC and measured aerosol optical thicknessand water vapour. The uncertainties associated tothese values are in the range of ±0.03 for aerosoloptical thickness at 550nm and ±0.2g.cm-2 forwater vapour.

• INFRAMETRICS: The 19 dates of observation at3000m, and 7 dates at 1500m have been proc-essed. The geometric correction was performedthanks to ground control points scattered on theimages. The accuracy achieved is similar to thatof POLDER. The absolute radiometric calibrationwas performed over a large number of combina-tion of black body and ambient temperature. Theaccuracy is close to ±0.3K. Atmospheric correc-tion was applied using MODTRAN code andwater vapour and temperature profile measure-ments provided by the radiosoundings. However,because of stability problems of the camera, the

data were recalibrated using the ground measure-ments over the surface temperature contempora-neously to the INFRAMETRICS flight.

• ERASME: The ERASME data have been proc-essed by CETP. The registration was achievedthanks to the video camera aboard the helicopter.The radiometric calibration was performed byexternal calibration. Validation of the results wereperformed by comparison with ERS and Radarsatdata when possible, as well as by correlation withsoil moisture. All these tests demonstrate highconsistency and accuracy of ERASME dataaround ±1dB. Both in like and cross polarisation.

• RENE: The data have been recently processedand are now available in the data base.

• DAIS: The DAIS data have been processed byDLR. The geometric registration shows quitegood performances. However, radiometric cali-bration shows stability problems both in the solarand thermal domains, and instrumental noise wasquite high. Local calibration was applied usingwell identified targets

5.1.3.4 The data base• Copyrights. No copyrights are covering these

data. Therefore, they are free for use to the wholecommunity. However, when people out of theconsortium will use these data, we would likethem to acknowledge those who spent a lot oftime and energy in acquiring and processing thedata. This information is found in the data base.

• Data base architecture. The data are available atthe ReSeDA server at INRA that is fully accessi-ble to the scientific community. The files are or-ganised in directories per sensor. In the main sen-sor directory, and information file describes thetype of data and the format associated.


5.1.4 Ground measurements

5.1.4.1 The selection of fieldsFour types of crops were investigated: wheat, al-falfa, sunflower and maize (Figure 2). This ensuresenough variability between crops to test the robust-ness of the models developed. We defined severallevels of field instrumentation. Calibration andvalidation sites were termed "permanent sites" be-cause instrumentation was installed during severalmonths. They mainly aim at the understanding andmodelling of the detailed processes. The remotesensing sites are sampled mainly during the air-borne flights and the intensive periods. They aim atthe evaluation of retrieval algorithms for canopyand soil biophysical variables through radiativetransfer models.1. Calibration fields. These fields correspond to

the highest level of instrumentation to satisfythe calibration requirements for SVAT and can-opy functioning models as well as those for thetest of remote sensing data inversion techniques.The calibration fields were equipped with afixed set of instruments measuring the surfaceenergy balance components, some climatic pa-rameters, the daily soil water balance, the sur-face temperature and the albedo. Observationsof the root system and a detailed description ofsoil physical properties were performed.

2. Validation fields. The instrumentation and thecharacterisation of the crops is lighter than thoseof the calibration fields. The measurementswere dedicated to the evaluation of SVAT andcrop models and for the test of the inversion ofremote sensing data. Beside the required datafor pure remote sensing purposes, the measure-

ments were limited to the components of the en-ergy balance, a weekly water balance and thesurface temperature. Additional characterisationof the crops and the soils were performed.

3. Remote sensing fields. A selection of fieldswas monitored to provide the surface parame-ters required to test the inversion procedures.All the soil and the vegetation measurementsmentioned above were collected concurrently toeach aircraft campaign and for some importantsatellite data acquisition.

Table 4 shows the measurements performed on thedifferent type of fields.The main emphasis of the experiment is its tempo-ral dimension. However, scaling problems will alsobe addressed. The spatial sampling strategy wasdesigned to allow a good characterisation of :• the field scale that reveals the variability be-

tween fields. This variability is to be used toevaluate the robustness of the algorithms pro-posed.

• the km scale as observed by large swath satel-lite. The individual fields sampled at the localscale can be used to estimate the correspondingkm scale characteristics.

Some problems occur in some fields due to the in-ternal variability. A specific interpretation of thedata is required in such situations. Figure 8 showsthe time distribution of the measurements. Thiscampaign was mostly organised and actually con-ducted by INRA. However, significant help wasprovided by the several partners during the twointensive campaigns.

Measurements CalibrationValidation RemoteSensing

Radiative balance (Rn, a.Rg, Trad (20°))Bowen ratio (Ta, pv, H, LE, u(2h), u(3h))Soil Heat Flux (G(0;05m))Soil Temperature profile (0.005, 0.01, 0.025, 0.075, 0.15, 0.25, 0.50, 1.00m)Moisture profile (3 neutron probe tubes)Surface moisture (2 capacitance probes at 0.025 m)Surface moisture (5 TDR at 0.025 m)Moisture profile (2 capacitance 0.025, 0.075, 0.125, 0.175, 0.25, 0.40, 0.65, 1.00m)Tensiometers (2 sets at 0.10, 0.20, 0.30, 0.50, 0.80, 1.10, 1.30 m) (2 sets at 0.20, 0.50, 0.80, 1.10, 1.30 m)Rain gaugesWet and dry mass of the organs F

Height of the crop F

Leaf area index (direct and LAI 2000) F

Table 4. Measurements performed routinely on the calibration, validation and remote-sensing fields. F meansthat these measurements are mainly performed during the flights.


0

1

2

3

4

5

6

7

8

9

10

2/10 1/11 1/12 31/12 30/1 1/3 31/3 30/4 30/5 29/6 29/7 28/8 27/9 27/10

Measurement period

Neutron Probe

Capacitance Probe Hourly

Daily

TDR Probe

Tensiometers

Soil temperature

Micrometeo & Flux

Biomass & structure

Meteorological measurements

Figure 11. Temporal distribution of the measurements on the main calibration and validation fields. Eachcolour bar corresponds to a particular field, from bottom to top for each set of measurement: 101, 120, 214,(wheat),203 (alfalfa), 102, 121, 501 (sunflower).

5.1.4.2 Soil measurements

5.1.4.2.1 Soil permanent characteristics.Soil permanent characteristics were measured overthe calibration and validation fields. This includesphysical, chemical analysis. Additionally, other soilfunctioning variables were measured and consid-ered stable during the growth period. It includesdensity, retention, hydraulic and thermal conduc-tivity curves.

5.1.4.2.2 Soil dynamics measurements.• surface moisture. Soil surface moisture was

measured at specific dates corresponding to the µ-wave instrument overpasses (either ERS, radarsat,or ERASME/RENE). They were acquired usingeither the gravimetric procedure, or using the ca-pacity gauges providing a good temporal sam-pling and TDR devices providing a good spatialsampling.

• Soil roughness. The roughness profiles weremeasured using either an automated non contactlaser profiler, 1.5 m long, with a sampling intervalof 1.7mm, or contact needles over 1-2m long witha sampling interval between 5 to 10mm.Themeasurements were carried out over 12 fields in-cluding wheat, sunflower, alfalfa and maize vali-dation and calibration plots. They were performedafter the main cultural practices (tillering, sowing,harvest) over five to eight profiles orthogonal andparallel to the row or tillage directions. The proc-essing provides the standard deviation of height,the correlation length and the calibrated digitisedprofiles.

• Soil moisture and tension profiles. Frequentmoisture profiles and tensiometer measurements

were performed over the calibration and validationfields. Figure 12 shows the good agreement be-tween the several methods used for soil moisturemeasurements.

• Temperature profiles.• Soil heat flux. Heat transducers were used.

Figure 12. Comparison of water storage in the 0-120 cm soil layer as obtained by different methodsof measurement on field 101.

5.1.4.3 Canopy measurements

5.1.4.3.1 Canopy structure and biomassThe ground data collected throughout the growingseason were processed to provide values represen-tative of the fields. They include the leaf area index(direct measurements) canopy height and the freshand dry biomass per organ. An example of the out-put is shown in figure 10. No particular problem isto be noticed, except that in some plots the hetero-geneity was important.


5.1.4.3.2 LAI2000 gap fraction and leaf areaindex. LAI2000 measurements were performed frequentlyover a selection of plots during the growing season.Ten random places were sampled using ten meas-urements to get a good representation of the fieldvalue. The data were processed to get estimates ofthe leaf area index, and the average leaf angle.Figure 13 shows the relatively good agreementbetween both types of measurements. The scattercould be explained mainly by spatial heterogeneity,and the biases by the fact that LAI2000 estimatesaccounts for the area of other components than justgreen leaves, i.e. senescent organs, stems and earsthat could represent a significant fraction of thetotal area index, particularly at the end of thegrowth cycle.

0

1

2

3

4

0 1 2 3 4

LAI planimeter

LAI 2

000

101

120

214

300

Figure 13. Comparison between LAI2000estimates of LAI and direct LAI measurements.Results over wheat fields.

5.1.4.3.3 Ground Reflectance measurementsBi-directional reflectance measurements were per-formed mainly during intensive campaign using theCROPSCAN radiometer. The instrument used hasthe following bands: 490, 550, 670, 700, 740, 780,870 and 1090 nm. A wheat field and a sunflowerfields were sampled.

5.1.4.3.4 Emissivity measurementsEmissivity measurements were performed for al-falfa and wheat crops in April, and for sunflowerand corn crops in July. The box method with one-lid or two-lids was used. For both methods (one-lidand two-lid) emissivity calculations have been cor-rected from effects due to the non-ideal nature ofthe box according to the procedure of Rubio et al.(1997). Depending on the nature of the crops,measurements were taken for pure samples (soilsand plants) and for mixed samples (plants over thenatural soil background).In addition to emissivity measurements, the coverfraction and spatial distribution of plant elementsover the plots were measured.

5.1.4.4 Atmosphere measurements

5.1.4.4.1 Meteorological stationIt measured the main climatic variables on a 20minutes frequency basis during the whole experi-mental campaign. The meteorological instrumentswere located in the centre of the test site. It wasmaintained over a natural vegetation (bare soil inwinter, and natural grass developing during the sea-son). The height of the vegetation was kept below0.20m. The measurements included:• Global, diffuse, long-wave and PAR radiation• Air temperature at 2m height• Vapour pressure at 2m height• Wind speed and direction at 2m height• Soil temperature at 0.1, 0.5 and 1.0m depth• Rain• Atmospheric pressure

Figure 14. Temporal profile of the differencebetween monthly rainfall and evapo-transpiration(P-ETP) during the ReSeDA campaign as comparedto the average values observed in Avignon.

The data from the meteorological station have beenprocessed and evaluated. This includes test of con-sistency• with the station from Avignon,• the temporal consistency• the consistency with the data available over the

calibration and validation fields.During the ReSeDA experiment, the winter wasparticularly wet, and the spring very dry (Figure14). Other periods show quite regular patterns.These climatic conditions influenced largely thegrowth of wheat crops and the installation phase ofsummer crops.

5.1.4.4.2 Atmosphere characteristicsThree sets of measurements were performed:• The CIMEL sun-photometer was permanently

installed within the meteorological site. It pro-vided the optical depth in various directions andin several wavebands. This CIMEL device wasrunning automatically. Additional measurements


with a manual sun-photometer were performed tocheck the outputs of the automatic instrument.

• The direct and diffuse global radiation of themeteorological station allows to estimate quiteaccurately the aerosol optical thickness when sunphotometer measurements were not available(Figure 15)

• The radio-sounding data from the Nîmes Cour-besac airport which is about 40 km from the ex-perimental site were systematically recorded dur-ing the flights. Additional radio-soundings wereperformed by Meteo-France during intensivecampaigns to better characterise the atmosphereboth for the correction of thermal infrared andoptical observations and for a better characterisa-tion of the planetary boundary layer used in themodelling of the fluxes between the soil, vegeta-tion and atmosphere.

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.450.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

stae

de

oso

Opt

cac

ess

Experimental data1 : 1 line 1 : 1 line

Figure 15. Comparison of Aerosols opticalthickness at 550 nm measured by solar photometerand that estimated by equivalent direct solarpyranometer measurements.

5.1.4.4.3 Energy balance and flux measure-mentsMicrometeorological measurements were per-formed over calibration and validation fields with atime step of 15 seconds and an averaging period of20 minutes. The instrumentation was installed atleast at 100 m of the upwind edge of the fields. Itconsisted in air temperature, relative humidity andwind speed measurements. All the instruments werecalibrated and inter-compared before and after theexperiment. The measurements were performed attwo levels above the canopy (at about 1.5 times and3 times the canopy height) to be able to computeheat fluxes by means of the Bowen ratio method orthe combined aerodynamic method. However, fre-quent failures in the instrumentation did not allowto use these data with a good confidence. Moreover,a misconception of the system for measuring va-pour pressure and temperature gradients did notallow to compute the fluxes with a good accuracy:at least half of the measurements were rejected.However, additional flux measurements were per-formed as a backup of the previous system:• Mono-dimensional eddy-correlation systems

were performed as often as possible and generated

a significant set of data over the three wheatfields;

• Three dimensional sonic anemometers meas-urements were also performed for some days inJune over a wheat field;

• commercial Bowen ratio systems based on dewpoint hygrometer measurements or on direct gra-dient measurements also provided data on wheatfields.

Figure 16. Comparison of sensible heat flux basedon the eddy-correlation method over three wheatfields during a particular day.

5.1.4.4.4 Spatial heterogeneity of fluxes.To investigate the effect of spatial heterogeneity ofthe fluxes, two systems were used:• The unmanned plane from Cranfield University

(Figure 17) was launched with a payload includ-ing sensors for temperature, humidity and posi-tion. Extensive data processing has been per-formed to check data quality and provide a con-sistent dataset with common time and positionreferences. Results show the expected generalfeatures of the surface layer profiles of tempera-ture and humidity and also give details of struc-ture near field boundaries.

• Scintillometer measurements were performed inJune 1997 over a transect including an importantcontrast between fields. Simultaneous eddy cor-relation flux sensors measurements were taken forthe individual fields within this transect. Figure18 shows the installation of the devices over thescintillometer transect.

Figure 17. The Cranfield University unmanedaircraft used.

Measured aerosol optical thickness

Est

imat

ed a

eros

ol o

ptic

al th

ickn

ess


Figure 18. The scintillometer measurementsperformed in the center of the Alpilles site.

5.2 The data base

The data base is one of the important output of theReSeDA project because it will provide input fortesting radiative transfer, SVAT and canopy func-tioning models, as well as the associated inversionand assimilation techniques. This will obviously beused for the ReSeDA scientific objectives them-selves, but it will also be available to the wholescientific community. For this reason, it should bevery easy to access and to use. The data base ishosted at INRA at the following address:www.avignon.inra.fr/reseda where a dedicated harddisk was installed.

5.2.1 Structure of the data base anddiffusion

5.2.1.1 The data baseThe structure described in Figure 19 was imple-mented. Directories were created for the main cate-gories of data. These directories can be divided intosubdirectories according to specific sensors ormeasurement type. In the last level sub directories,there are three types of files:• The documentation file that provides information

on the type of measurements, protocols and im-plementation of the measurements during the ex-periment. This is generally a RTF, PDF or ASCIIformat

• The update file that provides information on thelast changes. It is an ASCII file.

• The data files that contain a header with a de-scription of the data format and the data themselves. They are also ASCII files.

The consistency of the data base was checked bycombining several tests. Further, the first exploita-tion of the data that is described in the followingsections was also a good opportunity to evaluate thequality of the data. However, it is possible that inplaces, errors and uncertainties are embedded. Thatis the reason why we would like any problem en-countered to be reported in order to take the properactions for correction. The data base will thereforeevolve and improve in the next years. Therefore,the web server appears as the most convenient me-dia for distribution of the data and was preferred toCD-ROM or DVD at this stage.

When data will be used by scientists outside theconsortium, it will be asked to acknowledge thepeople and institutions who directly collected andprocessed the data as well as the funding organisa-tions, i.e. the European Commission and the FrenchProgramme National de Télédétection Spatiale.

Figure 19. General structure of the Reseda database.

5.2.1.2 The web serverThe web server originally hosted by Synoptics wastransferred to INRA Avignon. It includes threemain components:1. A general description of the ReSeDA project,the partners with proper links and the funding or-ganisations.2. A list of the papers that have been publishedfrom the project. Most of the papers will be acces-sible fully in electronic version (PDF) except whencopyrights are active.


3. A description of the data base with links to thecorresponding directories and files for download-ing.The maintenance of the web server for the nextyears will be under the responsibility of INRA Avi-gnon.

5.3 Development and evalua-tion of inversion methods

The inversion of radiative transfer models fromremote sensing data as illustrated by figure 14 canbe defined by four main aspects :1. The choice of the biophysical variables to beretrieved. These variables had to be very intimatelyrelated both to the radiative transfer process withinthe canopy and to the vegetation processes to bedescribed. We focussed on the 9 following vari-ables that were already listed in Table 1 as a func-tion of the spectral domain where it is accessible:

• albedo, the bi-hemispherical reflectance of thecanopy integrated over the 300-3000nm spectraldomain

• fCover, the cover fraction that corresponds to thegap fraction for nadir direction,

• fPAR, the daily fraction of photosynthetically ac-tive radiation absorbed by the canopy

• LAI, the leaf area index that defines the size ofthe main interface for exchange of energy andmass between the canopy and the atmosphere

• hc, the height of the canopy that is very impor-tant for the determination of the resistance totransfer used in SVAT models.

• Wc, the canopy water content• Ws, the surface soil water content• hrms, the roughness of the soil surface• Ts, the surface temperature of the canopies (soil

and/or vegetation)

2. The type of remote sensing data used. Thespectral, directional and polarisation features of theradiometric response can be used to extract canopycharacteristics. The ReSeDa has put some emphasison the directional dimension, particularly in thethermal and solar domains.3. The choice of the radiative transfer model.Radiative transfer models differ mainly by the waythey consider canopy structure and the way the ra-diative transfer equation is approximated. Themodel had to be relatively general although realis-tic. Both constraints appear however to be antago-nist because of the diversity of canopy structure.Further, for routine application, the model is de-sired to run fast, particularly when using iterativeoptimisation techniques. The choice of the radiativetransfer model is therefore a critical question tominimise uncertainties in the inversion process. Theseveral radiative transfer models used in ReSeDAare presented in Table 5. The operational modelsthat can be used for applications are distinguishedfrom reference models that attempt to describe thebest both canopy structure and the radiative trans-fer.

Model type PhysicalRadiative transfer computation Turbid medium Geometric Ray tracingRepresentation of the scatterers Continuous Discrete Geometric Realistic

Semi-empirical

PROSPECT (leaves) OperSOILSPECT (soil) Oper

CGI OperSAIL/KUUSK/IAPI/NADI OperCLAIR OperS

olar

dom

ain

BRDF (MRPV, Walthall, FLIK) Oper

SAIL-T Oper

The

rmal

Infr

ared

PrévotRef

IEM (soil) OperMIMICS RefKARAM RefACT Refµ

-wav

eac

tive

CLOUD Oper

TSA Oper

MEV Ref

µ-w

ave

pass

ive

[ω,τ] Oper

Table 5. The radiative transfer models used over the spectral domains in ReSeDA. Operational (Oper) andreference (Ref) models are distinguished.


4. The choice of inversion techniques. Two mainapproaches are classically used by the geophysicaland thus the remote sensing scientific communitiesas illustrated by Figure 20:

• Minimisation of the distance over radiance.This is generally achieved through variousoptimisation techniques where, starting froman initial set of input variables, these valuesare iteratively updated up to the point where agood match is reached between the radiancefield simulated by the models and that meas-ured by the sensors. The use of look up tablesis a special case of these iterative techniques.

• Minimisation of the distance over the vari-able of interest. This is generally achievedthrough approximation of the surface re-sponse, using mathematical methods such asneural networks that learn to approximate theinverse relationship between the radiance fieldand the variables of interest. Simpler relation-ships such as vegetation indices used to esti-mate several variables such as fAPAR havebeen investigated extensively. The split-window technique for atmospheric correctionin the thermal infrared belongs also to thatcategory of approach. The minimisation isperformed here in a first step, the learning orcalibration phase.

V

V*

R

∆V

C

M-1

M V* R*M

R

∆R

Canopy variable driven approaches:minimising on canopy variables

(vegetation indices, neural networks, …)

Reflectance driven approaches :minimising on reflectance

(iterative optimisation, look up tables)

Figure 20. The two main approaches used to estimate canopy characteristics from remote sensing data. V,R and M respectively represent the variables of interest, the reflectance and the RT model. M-1 is the inversemodel with coefficients C. V* represents the estimated value of V, and R* the corresponding simulatedreflectance. ∆ represents the distance between either R (∆R) or V (∆V).

We will review here the work achieved up to dateabout model inversion. This includes studies on:• the modelling of radiative transfer models,• comparison of radiative transfer models• comparison of inversion approaches• investigation of the optimal spectral and direc-

tional sampling• and obviously investigation on methods of canopy

biophysical variable estimation using the actualReSeDA data.

We will present the studies by wavelength domains,since the sensors and radiative transfer models arequite different. We will start from the lower wave-lengths in the solar domain, then address the ther-mal infrared domain, and finally go to the µ-wave.

5.3.1 Inversion in the solar domain

5.3.1.1 Investigation about inversion tech-niquesThree main inversion techniques have been com-pared, based on a simulation experiment:• LUT: The look up table (LUT) is the simplest

approach, although not straightforward to imple-ment. The LUT was generated by randomlydrawing the model input variables within a distri-bution law designed to better sample regionswhere reflectance is more sensitive to the vari-

able. Each variable were considered independent.A set of 280 000 cases was generated, and thecorresponding reflectance values were simulatedthanks to SAIL +PROSPECT models. The me-dian value and the standard deviation were com-puted over the 50 best solutions found in the table.

• OPT: The optimisation technique is one of themostly used at this time for radiative model inver-sion. The initial guess was derived from a simplelook up table. The quasi Newton search algorithmwas used. The cost function was designed to im-pose constraints on the variables to be estimatedand takes also into account the uncertainty associ-ated to the measurements and the models. Thedistribution of the solutions were obtained by in-version of the model over the distribution of themeasured reflectance for each case, taking intoaccount the uncertainty of each measurement.Again the median value and the standard devia-tion were used to characterise the solution.

• NNT. The neural network technique (NNT) israpidly developing for the estimation of biophysi-cal variables. In this case, a 2 layer back-propagation network was used, trained with theLevenberg-Marquart procedure. The learning dataset was made of 50000 cases taken among the280000 cases simulated for the LUT approach. Aset of 30 parallel networks for each of the fourbiophysical variables considered was used to form


the solution. The median value of the ten bestneural networks was retained as the solution.When the noise associated to the reflectance val-ues was known, it was introduced in the learningdata set.

The three inversion techniques were compared overa reference data set made of ray tracing modelsimulation applied on a 3D realistic maize archi-tecture (Figure 21). The LUT performs generallysimilarly to the OPT method although the NNT getsgenerally better results.Figure 21 shows that the fAPAR and fCover are the vari-ables the better retrieved with a relative RMSEclose to 10% which is very encouraging. Then, es-timates of the chlorophyll content Cab is retrievedwithin 15% relative accuracy. Finally, LAI is re-trieved with the worse accuracy, close to 25% rela-tive RMSE. This shows clearly the advantage ofusing secondary variables that exploits the possiblecompensations between variables and could reducethe saturation problem such as in the case of fAPAR

and fCover.

LAICab

fCoverfAPAR

NNT

LUT

OPT

0

0.1

0.2

0.3

0.4

Re

lati

ve

RM

SE

(%

)

Figure 21. Comparison of retrieval techniques forthe biophysical variables considered. Case with2.5% radiometric noise and 2% biase.

5.3.1.2 Optimal spectral and directionalsamplingThe variables considered were LAI, Cab, fAPAR andfCover. We used a simple inversion techniquebased on look-up-tables as implemented previously.The reflectance data were simulated in 9 wave-bands in the visible and near infrared domain for 32directions chosen in the principal and perpendicularplane. The sun position was fixed to 45° zenith an-gle.

5.3.1.2.1 Spectral samplingStarting from the classical combination of red andnear-infrared bands, the band providing the bestestimation performances among the 7 remainingbands was selected. The same process was repeateduntil all the nine bands were selected.Results (Figure 23) showed that, when the numberof bands used increased, the absolute RMSE valuewas decreasing until a minimum is reached. Thisdemonstrated that only a limited number of wave-bands were required for canopy biophysical vari-

able estimation. The estimation of fAPAR andfCover required less bands than for LAI and Cab .

The RMSE value difference computed over the bio-physical variables of interest between the best andthe poorest band selection demonstrated that thechoice of the bands was obviously important, par-ticularly for the first bands selected. The otherbands were generally selected in the red-edge do-main (680-740nm) where the dynamics of thevegetation spectrum is the largest.As a conclusion of this step, we decided to keep theband combination that gave the lowest RMSE foreach variable : For these optimal band sets, the op-timal directional sampling was then investigated.

Figure 22 Absolute RMSE value between actualand estimated canopy variables (LAI and Cab)as afunction of the number of bands selected for eachcanopy biophysical variable of interest. The solidline corresponds to the best band combination andthe dotted line to the worst combination.

5.3.1.2.2 Directional samplingConsidering the best set of bands for each variableas derived for nadir viewing, the same procedure asthe one used previously for the spectral samplingwas applied to select the optimal directional combi-nation.Results showed that for each of the four biophysicalvariables considered, less than six directions wererequired (Figure 23). The other directions containedredundant information and thus, their use as addi-tional bands induced noise in the retrieval process.The first selected direction was able to reduce sig-nificantly the uncertainty on the retrieval of the fourbiophysical variables considered. For this direction,


the difference between the optimal and the worstview angle was large. The first directions were al-ways located in the principal plane, either close tothe hot spot direction, or in the forward direction tohigh zenith angles.For the leaf area index estimation, only one direc-tion, the hot spot, was required to obtain the lowestRMSE value, confirming the importance of thatdirection for canopy structure estimation. Moreo-ver, the following selected angles were very closeto the hot spot, except the second one which waslocated in the forward direction with high zenithangle. The leaf chlorophyll content estimation re-quired more directions (6), mainly located in thebackward and forward directions at high zenithangles, and in the hot spot feature. One direction inthe perpendicular plane was also needed. Consid-ering fAPAR and fCover, about six directions wererequired and only the first one was actually dis-criminating. Moreover, for fCover estimation, thefirst selected direction was the nadir which wasfully consistent with the fact that it corresponds tothe gap fraction in the nadir direction. For fAPAR,the first selected direction was located in the per-pendicular plane (zenith angle of 30°) and the sec-ond one close to the hot spot direction. This wasagain consistent with the fact that fAPAR could beapproximated by the gap fraction in the solar direc-tion.

Figure 23 Absolute RMSE value between actualand estimated canopy variable (LAI and Cab) as afunction of the number of selected directions. Thesolid line and the dashed line corresponds respec-tively to the best and worse sets of directions.

5.3.1.3 Comparison of the inversion per-formances of three operational radiativetransfer modelsOperational models are based on different assump-tions and approximations on canopy architectureand the computation of the radiative transfer. It istherefore required to compare the performances ofthe models when used in the inverse mode for esti-mation of canopy biophysical variables. This wasachieved over three 1D turbid medium operationalradiative transfer models, all coupled to the PROS-PECT model for description of the optical proper-ties of the leaves:• PROSAIL, the classical 1D turbid medium model

developed by Verhoef (1984)• PROKUUSK, the model developed by Kuusk

(1995),• PROIAPI, the model developed by Iaquinta and

Pinty (1994).The main differences between these models comefrom the description of the leaf angle distribution,the hot-spot effect, and the multiple scattering com-putation. The soil background was described byMRPV BRDF model fitted over bare soils. ThePOLDER data over the 16 flights were used to ex-tract the BRDF for all the fields for which LAI wasmeasured. It includes wheat, alfalfa, sunflower andmaize. The 3 bands of POLDER were used (550,670 and 865nm) for all the directions available. Theinversion technique is based on a quasi-Newtonoptimisation algorithm implemented with a costfunction that measures the relative distance betweenmeasurements and simulations. Among the vari-ables (Cab, Cw, Cm, N, S, LAI, ALA, hot), threewere set to a priori fixed values. The inversion wastherefore concentrating on the other 5 variables:LAI, ALA, Cab, S and N.All three models allow to get a good fit of thespectral BRDF as measured by POLDER, withRMSE values in the range 0.02 – 0.04. Figure 24shows that the three models are actually estimatingLAI in a quite consistent way. The same applies forchlorophyll estimation.

Figure 24. Comparison of measured LAI and valuesestimated by inversion of three radiative transfermodels using POLDER data in three wavebandsduring the ReSeDA experiment.

It can therefore be concluded from this investiga-tion that LAI could be retrieved with a quite goodaccuracy by radiative transfer model inversion. Theslight differences between the three 1D turbid me-dium radiative transfer operational model testedhave only marginal influence on the retrieval per-formances.


5.3.1.4 Validation of a neuronal techniquefor LAI and fCover estimationThe neural network technique is based on theminimisation of the distance over the variables tobe estimated, conversely to the previous model in-version technique (OPT). A learning data set (1500cases) was generated thanks to SAIL canopy modelcoupled to PROSPECT leaf optical properties andSOILSPECT soil reflectance models. The range ofvariation and distribution of the input variableswere chosen to embrace the actual distribution overcrops. The reflectance simulated were the nadir andthe hemispherical reflectance in the three POLDERbands. The networks were trained as described ear-lier in §5.3.1.1 using 50 parallel nets from whichthe median of the output is computed.The MRPV BRDF model was adjusted on the samePOLDER data as previously. It allos to estimate thenadir and hemispherical reflectance values. As wewill see latter, the MRPV model has very good in-terpolation capacity for nadir reflectance estima-tion, and reasonable extrapolation performances forhemispherical reflectance estimation. For fCoverestimation, the input of the neural networks werethe nadir reflectance in the three POLDER bandsand the sun zenith angle. For LAI estimation, thehemispherical reflectance was required in addition.Figure 25 shows that the accuracy of the estimationis quite good.

Figure 25. Comparison between measured LAI tothe values estimated by the neural network.

Method OPT NNT NDVIRMSE 0.78 0.44 0.49

Table 6. RMSE values obtained over the Alpillesdata set by using three inversion techniques (OPT:optimisation as implemented in §5.3.1.3; NNT: thisneural network implementation; NDVI: NDVIestimation when calibrating the NDVI-LAIrelationship on the same data base as the one usedfortraining the nets).

Table 6 shows that the neuronal network outper-forms the optimisation technique. For comparison,the NDVI estimation when the relationship betweenNDVI and LAI is calibrated over the same data setas the one used for the neural nets training, appearsquite performing in these conditions.

The cover fraction (fCover) is generally better esti-mated than the LAI, with a RMSE value close to0.08, as compared to 0.13.These results are very consistent with the previoustheoretical investigations presented in §5.3.1.1. Itproves that model inversion is now very close to bean operational technique that could be used rou-tinely for applications.

5.3.1.5 Robustness of the WDVI-LAI rela-tionship for wheat crops.The WDVI (Clevers, 1991) vegetation index wascomputed from the SPOT images over the wheatfields: WDVI = ρnir - (C × ρred). The slope of the soilline (C=1.26) was calculated by using reflectancederived from the SPOT images for a bare fieldduring the first half of 1997 and for the harvestedwheat fields on August 30th, 1997.A semi-empirical model was used to relate theWDVI to LAI according to Clevers (1991):LAI = -1/α × ln(1 - WDVI/WDVI∞)Clevers (1991) obtained for α a value of 0.252 atthe vegetative phase and for WDVI∞ a value of 68.6if the WDVI is based on NIR and red reflectance.At the reproductive phase he obtained for α a valueof 0.530 and for WDVI∞ a value of 57.9. Using thisequation and the above coefficients obtained overcrops under Netherlands' conditions, Figure 26shows the good agreement and robustness of theapproach.

Figure 26. Comparison between measured LAIvalues to those estimated using WDVI applied toSPOT data over wheat and Clevers (1991) semi-empirical model.

5.3.1.6 Using BRDF and hybrid models forcanopy height estimationEstimation of canopy height could be very inter-esting because it is one the main driver of the heatexchange between the canopy and the atmosphere.A 3D radiative transfer model is thus mandatory tobe able to extract canopy height from the BRDFsignature. For this purpose, the hybrid geometric-turbid medium CGI model was implemented(Figure 27).The variables of the models were tuned for wheatcrops. Figure 28 shows that CGI allows to get agood representation of the BRDF of wheat crops asmeasured by POLDER during the growing season.The CGI model was then inverted by using a lookup table technique based on simulation of hot-spotand nadir WDVI values (WDVI= R865-α.R670 whereα is the soil line slope). The WDVI allows to ac-count for soil background reflectance variation.


Application of this technique to the ReSeDA POL-DER data over wheat fields appear to provide poorestimates of canopy height. As a matter of fact, adetailed sensitivity analysis showed that the illumi-nation conditions during the experiment (sun posi-tion) did not allow to get enough handles on theBRDF to extract accurately canopy height.

Figure 27. Typical canopy structure representationby the CGI model. The geometric protrusions areconsidered turbid medium object. This viewcorresponds to the hot-spot (no shadow).

Figure 28. Comparison between CGI simulationsand actual canopy reflectance as measured byPOLDER during the growth season over wheat.

5.3.1.7 BRDF models and albedo estima-tionBRDF models are developed to describe (ratherthan explain) the directional variation of reflec-tance. They are very useful for the normalisation ofreflectance data acquired under different viewingconditions, as well as for the derivation of thehemispherical reflectance used to compute albedo.We will first show the results obtained on the com-parison of classical BRDF models, and then addressthe problem of albedo estimation.

5.3.1.7.1 Comparison of BRDF modelsLinear kernel models are now classical to describethe BRDF as:

∑=

φθθα=φθθρk

isoiiso K

1

),,(.),,(

where ),,( φθθρ so is the bi-directional reflectance,

k is the number of kernels i used, ),,( φθθ soiK are

kernel functions, and αi the associated kernel coef-ficients. We considered 10 models:

• Walthall et al. (1985) which is mostly an empiri-cal model, with 3 parameters but without hot-spotdescription. Further it is not reciprocal

• MRPV (Engelson et al., 1996) which a semi-empirical model, with 3 parameters, reciprocal,ith a hot-spot feature. It is originally not linear, bitit could be linearised.

• Roujean et al. (1992), which is a semi-empiricalmodel with 3 parameters, reciprocal and with ahot-spot feature.

• Li-Ross (Wanner et al., 1997). Which is based onsimple geometric-optic model. It has 3 parame-ters, is reciprocal and has a hot-spot feature.

• Nadir Which is the simplest BRDF model, as-suming that the canopy is lambertian with reflec-tance equals to that at nadir. It has no parameter.

• GEN that stands for generic BRDF model. It hasa generic shape based on average MRPV func-tions. The BRDF is adjusted to the measured re-flectance by offset and gain. It has therefore 2 pa-rameters and is reciprocal and has a (generic) hot-spot feature.

• FLIK3-FLIK6 which is based on the adjustmentof empirical kernels over a BRDF data base gen-erated by radiative transfer simulations. It canhave in principle any number of kernels. In prac-tice, this number was limited to 6. Figure 29shows that the directional variation could rea-sonably be represented by a small number of ker-nels. FLIK is reciprocal and has a hot spot feature.

25

50

75

100

0 1 2 3 4 5 6

Number of kernels

frac

tion

of v

aria

nce

expl

aine

Figure 29. Fraction of the variance in the BRDFexplained, as a function of the number of kernelsused.

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

Nadir

Walt

hall

MRPV

Roujea

nRos

sGEN

FLIK3

FLIK4

FLIK5

FLIK6

RM

SE

443nm

550nm

670nm

865nm

Figure 30. Goodness of fit of the 10 BRDF modelsinvestigated on the POLDER data acquired over thewhole growing season on a selection of fields.


The results were analysed in two ways:• Fitting capacity. It measured the goodness of fit

of the models to the POLDER BRDF data. Figure30 shows that, except the lambertian assumption(nadir), and to a lesser degree GEN, all modelsare performing reasonably well, with a RMSEclose to 0.015 in the visible and 0.025 in the nearinfrared. Because of the good fitting capacity ofmost of the BRDF models, reflectance at nadirwhich is a privileged direction sampled by currentpolar orbiting sensors will always be accuratelyestimated.

• Extrapolation capacity. It describes the capacityof a BRDF model to estimate the reflectance indirections outside directions used for adjusting themodel. This criterion is very important when es-timating the hemispherical reflectance. Resultsshow (Figure 31) that most models are reasonablywell reconstructing POLDER data when adjustedin the principal plane. However, when adjusted inthe perpendicular plane, the situation is morecomplex. For the models with the larger numberof parameters, the results are very poor. The di-rectional signature in the perpendicular plane isnot sufficient to get enough constraints on theBRDF model. In this case, the generic model per-forms the best.

Figure 31. Extrapolation performances of theBRDF models. The models are adjusted either overthe data close to principal plane (left) or close to theperpendicular plane (right). They are then testedover all the POLDER data.

5.3.1.7.2 Albedo estimationAlbedo is one of the main variables governing theradiation balance and thus the energy balance. Wewill restrict the problem to the estimation of theinstantaneous albedo, i.e. the bi-hemispherical re-flectance integrated over the 300-3000nm spectraldomain, as it could be observed concurrently toremote sensing data acquisition. The estimation ofalbedo values from remote sensing data poses atleast two main problems:Directional: the estimation of hemispherical re-flectance from a limited number of directions sam-pled by the sensor. This was investigated previ-ously. We mainly selected Walthall model for itsrobustness and MRPV model for its performanceswhen sufficient directions are sampled with accu-rate reflectance data.Spectral: the estimation of albedo from a limitednumber of hemispherical reflectance. We testedthree sets of coefficients proposed by Weiss et al.(1999) and shown in Table 7.

445 560 665 865Set #1 0.000 0.000 0.570 0.460Set #2 0.000 0.680 0.080 0.350Set #3 0.060 0.690 0.001 0.350

Table 7. Spectral coefficients used to extend to thewhole spectral domain the hemisphericalreflectance observed in few wavebands.

The validation was performed over field measure-ments of albedo in the validation and calibrationfields. It shows an overestimation that shouldmainly result from the spectral coefficients thatwere derived over 400-2500nm spectral domain ascompared to the 300-3000nm spectral region con-sidered by albedo measurements. Further workshould be directed to address this issue, particularlyconsidering the fact that very little data is availableon the spectral properties of leaves and soils atthese extremities of the solar spectrum. However,an empirical correction performed over the experi-mental points (Figure 32) shows that the relativeaccuracy achieved this way is close to 8%. Possibleimprovements could be gained by considering thediffuse radiation at the time of the measurements.

Figure 32. Comparison beween measured values ofalbedo and the values estimated using POLDERdata and MRPV model with the set #3 of spectralcoefficients.

5.3.1.8 Scaling issuesThe scaling effect when observing heterogeneouslandscapes is certainly one of the main problemswhen estimating canopy biophysical variables fromremote sensing data. This is obviously linked to nonlinearity in the relationship between the radiometricdata and biophysical variables.This question was investigated both from the theo-retical and the experimental aspects.

5.3.1.8.1 Theoretical investigation.A simulation study was developed to evaluate theerror associated to the estimation of biophysicalvariables over heterogeneous pixels. Pixels with thesame value of biophysical variable such as LAI andfCover were generated under a range of spatial dis-tribution. The corresponding reflectance value was


computed. An index of heterogeneity was built thatreflects the spatial heterogeneity of the pixel. Itranges from 0.0 for homogeneous pixels, up to 1.0for the most heterogeneous pixel made of equalfraction of bare soil and vegetation cover. Results(Figure 33) show that the LAI estimation is stronglydependent on the degree of heterogeneity of thepixel. Relative errors up to 40% can be observed inthe case of heterogeneous pixels with the largestLAI values. Moreover, the LAI value is generallyunderestimated over heterogeneous pixels. Thesame was observed for chlorophyll content estima-tion. Conversely, the estimation of the cover frac-tion is little sensitive to the pixel heterogeneity. Thesame applies for fAPAR. This is mainly due to thefact that fCover and fAPAR are flux variables thatare almost linearly related to radiometric data con-versely to LAI and chlorophyll content that expressstrong curvature.

Figure 33. Relative error when estimating LAI (top)and fCover (bottom) from remote sensing data overheterogeneous pixels. Results from a simulationexperiment. The retrieval technique used here isbased on look up tables.

5.3.1.8.2 Experimental evidenceThe POLDER data were used to validate the previ-ous findings. The 16 POLDER flights were used tocompute the radiometric data over a single 1km²pixel. The neural networks trained as explained in§5.3.1.4 were fed with the 1km² averaged reflec-tance value. This estimation was compared to thatobtained when averaging over the 1km² pixel theoutput of the neural nets fed with the POLDERresolution reflectance data. Results (Figure 34) con-firms the theoretical findings, with a significantsystematic error for LAI, and little effect for fCover.These findings indicate that the estimation of bio-

physical variables such as LAI and chlorophyllcontent derived from large swath satellites with alow spatial resolution should be corrected from thea priori knowledge of the spatial heterogeneity.

Figure 34. Impact of the spatial resolution overheterogeneous pixels for the estimation of fCover(1-Po, figure on the left) and LAI (figure on theright). The comparison is made when feeding theneural net for the estimation of biophysicalvariables with (1) the reflectance values averagedover a 1 km² pixel, against (2), when using theoriginal POLDER spatial resolution and averagingthe estimated biophysical variabes over the same1km² pixel.

5.3.2 Inversion in the Thermal In-frared domainThree main issues were addressed in this domain.First, the atmospheric correction problem in orderto retrieve ground brightness temperature from thatobserved at the sensor level. This will be illustratedusing the NOAA/AVHRR data. Then, the de-coupling between emissivity and temperature,mainly using the DAIS data that provides enoughspectral features in the thermal infrared. Finally, wewill address the problem of the directionality ofcanopy temperature as observed by the INFRAM-TRICS camera.

5.3.2.1 Water content estimates fromNOAA/AVHRRThe total atmospheric water vapour content is esti-mated by a statistical analysis of the brightnesstemperatures at two wavelengths where water va-pour is the principal absorber. The split windowchannels of AVHRR instruments (11 and 12 µm)are thus good candidates. The split-window vari-ance-covariance method (Sobrino et al., 96; Ottlé etal., 97) consists in relating the atmospheric watercontent to:

∑

∑

=

=

−

−−=

N

kk

N

kkk

TT

TTTT

R

1

2110

11

1

1212110

11

)(

))((0

where the superscript and subscript refer respec-tively to the split window band and to a pixel withina window of analysis. All the AVHRR databasewas processed and the retrieved water vapour val-ues were compared to the ARPEGE model outputs.The ARPEGE model is a numeric weather predic-tion model designed by Météo-France to work atregional scale. It is based on assimilation tech-niques of radiosounding data among other input


data. Results (Figure 25) show a very good correla-tion (R²=0.91 ; RMSE=0.25cm). This method isthus quite interesting both for the estimation of theatmospheric water content, and for its use to correctthe atmospheric effects in the shorter wavelengthsfrom the water absorption.

0

0.5

1

1.5

2

2.5

3

0.6 0.7 0.8 0.9Covariance/Variance ratio

Wa

ter

Co

nte

nt

fro

m A

RP

EG

E (

cm

)

Figure 35 Atmospheric water content derived fromthe ARPEGE model compared to the Covariance-variance ratio for channel 11 and 12 ofNOAA/AVHRR. 15 days selected for clear skyconditions.

5.3.2.2 Estimating temperature and emis-sivity from the spectral variation of DAISdata.The Digital Airborne Imaging Spectrometer (DAIS7915), acquired images over the ReSeDA test siteon July 8, 1997 (Figure 36). Simultaneously, emis-sivity measurements were performed at the fieldlevel. The atmospheric correction was performedwith a radiosounding as input to the MODTRANradiative transfer code.Using the surface radiances for a view angle θ,

)(θsurfjL , the NEM algorithm (Gillespie, 1986) has

been applied. If an emissivity value, )(θε NEM , is

assumed for all channels, then the previous equa-tion can be solved for temperature for the N chan-nels of the sensor as:

)(

))(1()()(

θεθε−−θ

=NEM

skyjNEM

surfjNEM

jj

LLTB

and NEMjT is calculated through inversion of the

averaged Planck function while skyjL is obtained

from MODTRAN calculations. This provides a setof N values of temperature (one per channel) fromwhich the maximum value, is selected. The maxi-mum temperature selected maxT is now used to ob-

tain the N emissivity, through,

)()(

)()()(

max θ−

θ−θ=θε sky

jj

skyj

surfj

j LTB

LL

This provides a first estimate of the surface tem-perature and emissivity. The accuracy of these es-timates depends on the assumed )(θε NEM . To

minimise the error, we have selected )(θε NEM ac-

cording to the emissivity measurements. The NEMalgorithm has been applied to arrays of 5×5 pixelsextracted for selected fields.

Figure 36. Image of radiance at the sensor (noatmospheric correction) for DAIS channel 76,showing part of the ReSeDA test site with someselected plots indicated. North is up. Radiancesrange approximately from 0.92 to 1.51 mW/cm2 srµm (23 to 58 ºC).

0.70

0.75

0.80

0.85

0.90

0.95

1.00

8 9 10 11 12 13wavelength (µm)

em

issi

vity

DAIS-orig.; T=29.4 ºC

DAIS-recal.; T=25.0 ºC

reference

0.70

0.75

0.80

0.85

0.90

0.95

1.00

8 9 10 11 12 13wavelength (µm)

em

issi

vity

DAIS-orig.; T=52.0 ºC

DAIS-recal.; T=41.2 ºC

reference

Figure 37. Corn and soil+stubble emissivity valuesderived from the DAIS instrument before and afterre-calibration. The values are compared to groundemissivity measurements.

Results (Figure 37) show that The DAIS instrumenthas some calibration problem. This was corrected

304120

121

203201

500

126


according to ground measurements of surface tem-perature performed over two fields. The re-calibrated DAIS data were then used for estimationof emissivity over a selection of fields. The com-parison with reference emissivity derived fromground measurements (Figure 37) shows a reason-able agreement between both values. The differ-ences observed could be explained mainly by thepoor stability of the calibration of the DAIS sensorin the thermal infrared.

5.3.2.3 Brightness temperature directionaleffects observed with the INFRAMETRICScamera.Directionality of brightness temperature has re-ceived up to now little attention. However, fewstudies show the importance of these effects, bothfor normalisation of the data, but also for the ex-ploitation of this information to allow the retrievalof the temperature of canopy components, i.e. thesoil and the vegetation.

Figure 38. Schema showing the directionalsampling achieved by the INFRAMETRICScamera (top in polar representation), and thetemporal trend observed between the flight lineswhen looking at the cross points (bottom). Thepolar representation on the top shows thedirectional distribution of brightness temperatureafter temporal normalisation and smoothing.

The INFRAMETRICS camera equipped with awide field of view angle was operating the sameway as POLDER. Figure 38 shows the polar repre-sentation of its directional sampling for a particularfield. However, conversely to the solar domain, thebrightness temperature may vary significantly withtime due to climatic conditions such as wind, tur-bulence cells, and radiation. Significant differencesare therefore expected between observation of a

single pixel from different directions for which theobservations are separated by up to 20 minutes in-terval (Figure 38). Therefore, a simple procedurewas applied to reduce this artefact on the directionaleffects. It consists in normalising the several obser-vations to those acquired along the flight axis thatcrosses the four other axes close to the principalplane.

Figure 39. Directional variation of the brightnesstemperature in the principal (in-SPP) andperpendicular (V-SPP) planes for a bare soil (top)and corn (bottom) fields.

Results show a significant directional effects withan increase in the hot-spot direction correspondingto the maximum viewing of (hot) illuminated soiland vegetation. The perpendicular plane showsgenerally patterns symmetric to the nadir direction.The amplitude of variation is around 1 or 2°C inthis limited directional sampling (roughly ±30°). Inplaces it can reach up to 4°C in the hot-spot direc-tion. This investigation shows prospects for ex-ploiting this information if the hot-spot direction issampled and if larger view angles are available.

5.3.3 Inversion in the µ-wave domainTwo main issues related to µ-wave data applica-tions have been investigated. First, a qualitative onecorresponding to land use classification. Then, aquantitative one corresponding to canopy and soilbiophysical characteristics estimation. Similarly tothe solar domain, radiative transfer models havebeen used for this second issue, and therefore, im-provement of these models was also investigated.


5.3.3.1 Land use classification based on µ-wave

5.3.3.1.1 Using active µ-wave and multi-temporal data.Land use classification is quite an important appli-cation of remote sensing. It is classically performedusing the solar domain. However, because of thecost of the images and their often limited availabil-ity due to cloud occurrence, µ-wave could contrib-ute significantly to this application.The performances of land use classification basedon ERS and Radarsat multi-temporal data wereevaluated based on exhaustive ground survey.Three sets of time series of SAR data have beencompared in the perspective of operational assess-ment:• 3 dates of ERS2,• 3 dates of Radarsat,• 6 dates of bi-angular ERS and Radarsat.

The data were properly calibrated and masks wereapplied to non agricultural zones. Filtering was ap-plied to the images in order to reduce the speckle.Three filtering techniques were tested• no filter• multitemporal filter from Lopez• multitemporal + gammaMap filters• Nagao filter

Sensor Filter Whe

at(8

plo

ts)

Mai

ze(7

plot

s)

Sun

flow

er(1

1 pl

ots)

ERS

No filtermultitemporalGamma MapNagao

45.044.947.048.7

26.927.533.929.1

0.033.824.60.1

Radarsat

no filtermultitemporalGamma MapNagao

23.833.646.630.4

31.434.243.342.7

0.117.315.30.0

ERS+Radarsat

no filtermultitemporalGamma MapNagao

45.758.473.955.3

29.741.638.325.7

0.117.718.80.1

Table 8. Results of the classification for variouscombination of senors and filters for the three maincrops. Results evaluated over the validationsegment.

The results evaluated on the validation segment(Table 8), are relatively poor. This is mainly ex-plained by the limited number of plots availablewithin the 25km² Alpilles site area. However, indi-cations are given on the influence of the filteringmethod and on the combination of ERS and Radar-sat data to exploit their variation in the viewingangle. Results (Table 8) show that the gamma mapand the multi-temporal filters are performing thebest. Additional investigations based on the full

series of ERS data (10 images) confirmed that fil-tering improves classification accuracy in mostcases.The combination of ERS and Radarsat data did notalways provide improvements of the classification.Because of the small number of fields available fur-ther investigations should be conducted to confirmthese findings.

5.3.3.1.2 Combining active and passive µ-wavesSeveral combinations of active and passive µ-waves were evaluated for their discrimination per-formances. Results (Figure 40) show that combin-ing two frequencies and two incidence angles pro-vides a good way to discriminate between most ofcrops and soils over the ReSeDA site during theIROE flight.

Figure 40. Discrimination performances of passive(normalized temperature, band C, polarisation H,40° incidence) combined with active µ-wave (bandC, polarisation HH, 20° incidence). (� wheat; ✷

alfalfa; � grass; � sunflower; ✕ dry soil; � wet soil).

5.3.3.2 Modelling the µ-wave signalThe modelling activity is required to ensure that themodels to be inverted are providing a good descrip-tion of the radiative transfer within the ReSeDAcontext. It consists both in the development of newor improved models, and on the calibration andvalidation of these former models or of alreadyavailable models. The models used are preferen-tially physical. However, due to the complexity ofthe radiative transfer in the microwave, semi-empirical and empirical models will also be used.Both active and passive microwave models will beinvestigated.

5.3.3.2.1 ACF soil modelRecent studies have shown that statistical parame-ters such as the hrms (height root mean square er-ror), and the correlation length are not efficient de-scriptors of soil surfaces, since the latter are notpurely random and possess multi-scale geometry.An interesting alternative approach is to model thesurface by a continuous fractal function. Indeedfractals are suitable tools for describing mixtures ofdeterministic and random processes.


Figure 41 Simulated (dotted: Exponential ACF,dashed: Gaussian ACF, continuous: IROE’s model)and measured (black circles) backscattering in Cband as a function of incidence angle

In our research the surfaces have been characterisedby a semi-empirical spectrum with an exponentrelated to the fractal dimension. A comparison ofsimulations with experimental results shows thatthis approach gives more accurate results with re-spect to the spectra corresponding to both Gaussianand Exponential ACF (Figure 41).

5.3.3.2.2 Adaptation of the IEM soil modelThe Integral Equation Model (IEM) (Fung, 1994),is one of the most widely used model to describesoil backscattering as a function of roughness andmoisture. However, the relation between the corre-lation length used by the model and the actualroughness is not straightforward. Further, meas-urement of the correlation length is difficult andassociated to uncertainties.The aim of this study is to propose an empiricalcalibration of an effective correlation length to ob-tain a good agreement between model simulationsand observed data. Then, the effective correlationlength is related to the measured root mean squareof the surface height (hrms). This approach wasapplied to ERS and Radarsat configuration over theEuropean FLOODGEN campaign in 1998 (Pays deCaux). Figure 42 shows the very strong correlationbetween the effective correlation length and theactual hrms.

0.0 0.5 1.0 1.5 2.0 2.5 3.0rms surface height (cm)

0

5

10

15

20

25

30

35

Co

rrel

atio

n le

ng

th (

cm)

ERS-VV-23°

Lopt=exp(1.41krms - 0.95) ; R²=0.97

Lopt=0.8

(a)

Figure 42. Relationship between the effectivecorrelation length and the actual root mean squaresurface height for ERS configuration.

The reliability of this empirical relationship hasbeen validated using the ReSeDA database (tworadar configurations: ERS-23° and Radarsat-39°).A good overall agreement was observed betweenthe calibrated IEM (using the optimal correlationlength) and the measured data of the studied radarconfigurations (Figure 43).The good correlation between space-borne SARdata and the calibrated IEM indicates the feasibilityof retrieving hrms and/or soil moisture from SARobservations. These results are very promising be-cause this empirical calibration of the IEM can beused to develop an inversion technique for predict-ing soil characteristics.

-25 -20 -15 -10 -5 0Backscattering coefficient from ERS (dB)

-25

-20

-15

-10

-5

0

Bac

ksca

tter

ing

co

effi

cien

t fr

om

th

e IE

M (

dB

)

VV-23° (data set B)

rms <1cm : (radar-IEM) [dB]= -2.3±2.2rms >1cm : (radar-IEM) [dB]= +0.5±4.4all rms : (radar-IEM) [dB]= -0.8±3.7

(d): rms <1cm

: rms >1cm

-25 -20 -15 -10 -5 0Backscattering coefficient from ERS (dB)

-25

-20

-15

-10

-5

0

Bac

ksca

tter

ing

co

eff

icie

nt

fro

m t

he

IEM

(d

B)

VV-23° (data set B): validation

rms <1cm : (radar-IEM) [dB]= -0.25±2.1rms >1cm : (radar-IEM) [dB]= -0.3±1.4all rms : (radar-IEM) [dB]= -0.3±1.7

(d): rms <1cm

: rms >1cm

Figure 43. IEM performances for ERSconfiguration when using the measured correlationlength (top) compared to the use of the efectivehrsm deduced from the actual root mean square ofsurface height (bottom). Validation data set fromReSeDA.

5.3.3.2.3 A linear model for soil moistureThe relationship between soil moisture, ws, and thebackscattering coefficient as a simple linear one:σ°S = C +D ws

where C and D depend on soil roughness and tex-ture. However, we observed over the whole baresoil data set, that the slope, D, could be considered


as constant as observed earlier by (Ulaby et al.1978). The coefficients were adjusted for ERS orERASME data for band C and 20° incidence andVV polarization (Figure 44). Good performanceswere obtained for most of the fields.

Figure 44. Best fit linear relationship betweenvolumetric 0-1cm soil moisture wS (m

3.m-3) and soilbackscattering σ0 (dB).

Figure 45. Comparison between model simulation(solid line) and ReSeDA experimental data(ERS:squares, ERASME: diamonds) for C band at20° over wheat (top) and sunflower canopies.

5.3.3.2.4 A discrete canopy radiative transfermodelA model was developed, based on the radiativetransfer theory, and representing the vegetationelements as an ensemble of randomly oriented disksand almost vertical cylinders over a rough surface.The model describes the single scattering, the inter-action between the soil and vegetation (doublescattering) and the soil-vegetation-soil contribution.The simulations were compared to the ReSeDA

experimental measurements over two very con-trasted canopies: wheat and sunflower. Results(Figure 45) shows a reasonable agreement. Further,it confirms the very different behaviour of the twocanopies: wheat for which the backscattering coef-ficient decreases when LAI increases, and sun-flower that experiences an opposite behaviour.

5.3.3.2.5 Validation of the water cloud modelTo represent the influence on the backscatteringcoefficients of the surface variables such as canopywater content wc and soil water content ws, thesemi-empirical water cloud model of Attema andUlaby was evaluated. This model is a good candi-date for inversion, thanks to its simplicity and lownumber of variables and parameters. The cloudmodel writes:

osoil

o ttA σσ .)1.( 22 +−= , where 2t is the two-way

attenuation by the canopy expressed as:

)cos/..2exp(2 θ−= cwBt .

where θ is the incidence angle. The soil contribu-

tion °soilσ is expressed as a linear function of the

surface soil moisture as seen in §5.3.3.2.3:

sosoil wDC .+=σ

The parameters A, B, C and D are assumed to havethe following properties:• A, related to the scattering albedo of the canopy,

and B, related to its optical depth, are varyingonly with the frequency (C or X bands) and thepolarisation (HH or VV);

• D, the sensitivity of the soil back-scatter to soilmoisture, is assumed to be constant (Ulaby, Batli-vala et al. 1978);

• C, the intercept of the linear relationship betweenthe soil moisture and the soil back-scatter is as-sumed to be specific of the radar configuration(frequency, polarisation and incidence angle) andof the field considered, thus implicitly accountingfor the effect of the soil roughness;

all four parameters are assumed to be constant dur-ing the growing season. This last assumption mightbe difficult to be met, especially for the roughnessparameter C, but it must be kept in mind that duringthe RESEDA experiment, strong rainfalls occurredduring the fall and the beginning of winter, leadingto very smooth surface on wheat fields.

The CLOUD model was calibrated on ERS AND Ra-darsat data (C-band only) then on ERASME data (Cand X bands), using multi-angular configurations.In both cases, the leaf area index (LAI) was alsotested as the vegetation variable but the total plantwater content wc was always found to be a betterexplanatory variable than LAI on wheat canopies.Results (Table 9) shows values of the parameters ingood agreement with the literature. Further the fit-ting is quite accurate, with a rmse value close to thermse characterising the within field variability.However, we observe that Radarsat gets generallylower rmse values than ERS and Erasme (Figure46).


Sen

sor

Fre

quen

cy

Pol

aris

atio

n

Inci

denc

e (°

)A B C D

rmse

(dB

)

ER

S C VV 23 - 0.32 13.0-11.0 18.16 1.36

C VV 23 - 0.098 10.2-9.7 18.16 0.95

Rad

arsa

t

C HH 39 - 0.098 15.2-13.8 18.16 1.14

C VV 20 -18.3 0.215 12.9-11.4 17.8 1.48

C VV 40 -18.3 0.215 16.0-14.4 17.8 1.48

C HH 20 -14.6 0.000 12.7-11.3 17.8 1.52

C HH 40 -14.6 0.000 15.3-14.2 17.8 1.52

X VV 20 -14.4 0.590 10.9- 9.1 17.8 1.39

X VV 40 -14.4 0.590 16.0-12.7 17.8 1.39

X HH 20 -12.0 0.184 10.8- 9.4 17.8 1.20

Era

sme

X HH 40 -12.0 0.184 15.0-12.9 17.8 1.20

Table 9: parameters of the CLOUD model calibratedon the SAR (ERS and RADARSAT) data, along withthe rmse error (in dB). The value of the parameterC being field-specific, only its range is given foreach radar configuration.

-14 -12 -10 -8 -6 -4

measured σ° (dB)

-14

-12

-10

-8-6

-4

mo

de

led

σ°

(dB

)

ERSR23R39

Figure 46 fitting the CLOUD model on ERASME

data: modeled vs. measured σ°. Data acquired at20°: plus, at 40°: open squares. (a) C-HH, (b) C-VV, (c) X-HH, (d) X-VV. The solid line is the 1:1line.

Similarly, the cloud model was adjusted for sun-flower canopies using ERS and ERASME band C,20° incidence and HH polarisation. Instead of can-opy water content, the LAI was used to describe thecanopy status. Results (Figure 47) show quite goodagreement between the model simulations and theactual µ-wave backscattering coefficient.

0

2

4

6

8

10

12

14

0 2 4 6 8 10 12 14Measured backscattering coefficient (dB)

Sim

ula

ted

Bac

ks

catt

eri

ng

co

eff

icie

nt

(dB

)

Figure 47. Comparison between modeled andmeasured backscattering coefficient for sunflowerfields as observed by ERS and ERASME, band C,20° incidence, VV polarisation.

5.3.3.3 Biophysical parameter estimationfrom µ-wave data

5.3.3.3.1 Soil moisture retrieval

5.3.3.3.1.1 Active µ-waveERS data over the whole season have been used toevaluate their interest for soil moisture estimation.When each field is taken individually, a very nicesimilarity is observed between the back-scatteringcoefficient and the top cm soil moisture (Figure48). However, when all the fields are used to draw asingle correlation between ERS data and soilmoisture, an important scatter is observed (Figure49 top). This is mainly due to the effect of rough-ness that differs strongly from one field to an other.This effect could be minimised when consideringlarger areas where the roughness variations aresmoothed out. Figure 49, bottom, shows the im-provement of the accuracy of the correlation be-tween radar data and soil moisture when averagingover 10 fields. This shows that it could be possibleto monitor quite accurately surface soil moisture atregional scales from active µ-wave observations.

Figure 48. Time course of the backscatteringcoefficient and soil moisture in the first cm over asingle field from 30/01/97 to 30/11/97. ERS data,band C, HH, 20° incidence.


Figure 49. Relationship between back-scatteringcoefficient (Sigma-0) and soil moisture in the firstcm (SMC). All the fields (top), and average of 10fields for all the dates (bottom).

Figure 50. Normalized temperature at 10 GHz, 20°incidence as a function of soil moisture (first cm).Solid line: semi-empirical model; triangles:ReSeDA data; squares: other data.

5.3.3.3.1.2 Passive µ-waveThe X band (10GHz) data acquired from 20° inci-dence from the IROE radiometer were correlated tothe corresponding soil moisture (top cm) and com-pared to previous results. Figure 50 shows a verygood agreement between the ReSeDA data, otherdata and the semi-empirical model. This providesevidence of the interest of such observations for soilmoisture estimates at the local to the regionalscales. We observe here that the roughness appearsto have less influence of the moisture signal than

previously in band C, HH, 20° incidence with theactive µ-wave.

5.3.3.3.2 Leaf area index and biomass estima-tion from µ-wave data

5.3.3.3.2.1 Active µ-wavePrevious investigation had pointed out that the re-trieval of biomass or related variables (LAI or can-opy water content) was ambiguous when using onlyone configuration. However, for a given species,with its limited range of variation of canopy struc-ture, this appears feasible. A semi-empirical poly-nomial fit was adjusted on simulations of the dis-crete model developed previously as a function ofLAI for wheat and sunflower separately. Resultsshow a reasonable agreement (Figure 51) with Re-SeDA ERS data. However, a slight overestimationis observed for both crops.

Figure 51. Comparison between measured LAIvalues of wheat (top) and sunflower (bottom) andthose estimated from a semi-empirical polynomialmodel.

5.3.3.3.2.2 Passive µ-waveThe polarisation index (PI) was found to be quiteefficient for the estimation of canopy variables. Thefollowing equation was developed to relate PI andLAI values as a function of wavelength λ and inci-

dence angle θ: θλ⋅⋅−

⋅θ=θ cos),0(),(LAIK

ePILAIPI

Where PI(0, θ) stands for the bare soil PI value.The application of this simple model (Figure 52)shows a quite good agreement with observations ofwheat fields with the IROE radiometer in band X,20° incidence.


Figure 52. Comparison between measured LAIvalues and those estimated from the polarisationindex computed from passive µ-waves (band X, 20°incidence).

5.3.3.3.3 Concurrent estimation of soil and can-opy water contentSince the wheat canopy back-scatter depends ontwo canopy variables ws and wc, it is possible toestimate these two variables when at least two radar

configurations are available (Attema 1984, Prévotet al. 1993). This was tested based on the bi-angularRadarsat data set. Results showed that this inver-sion algorithm failed (Prévot et al. 1998) presuma-bly because on the non simultaneity of the two Ra-darsat observations. Therefore, we investigated theErasme data that present many possible simultane-ous configurations. The inversion technique usedhere is based on iterative optimisation and testedwith all the possible couples of Erasme configura-tions (28 couples). The residual errors of inversion(rmse) are given in Table 10 for ws and wc. Reason-able estimates of ws are obtained when one of thetwo radar configurations is C20HH (Table 10), thatis when the attenuation by the canopy is low. Thebest estimates of ws is achieved with C20HH &X40HH combination. whereas accurate estimates ofwc requires a dual-frequency configuration (C20VV& X20VV). The couple of radar configurationC20HH & C20VV appears as an interesting com-promise between configuration complexity (singlefrequency, single incidence angle and dual polari-sation) and quality of retrieval for both for ws andwc, as can be seen on Table 10.

ws: soil moisture wc : Canopy water contentC20VV C40HH C40VV X20HH X20VV X40HH X40VV C20VV C40HH C40VV X20HH X20VV X40HH X40VV

C20HH 0.084 0.081 0.080 0.082 0.083 0.077 0.081 0.50 0.65 0.94 0.65 0.41 1.07 1.13C20VV 0.081 0.122 0.103 0.089 0.103 0.119 0.86 1.20 1.01 0.46 1.16 1.02C40HH 0.091 0.081 0.084 0.090 0.092 0.72 0.86 0.44 0.98 1.09C40VV 0.112 0.138 0.129 0.113 1.33 0.67 0.94 0.84X20HH 0.087 0.104 0.103 0.58 1.28 1.09X20VV 0.115 0.168 0.55 0.59X40HH 0.097 0.95

Table 10: rmse on soil moisture (ws, in m3.m-3) and canopy water content (wc, in m3.m-2) estimation byinversion of the cloud model using two configurations; lower values are in bold, higher values are in italics.


5.4 Development of the as-similation technique

The assimilation of remote sensing data into proc-ess models was investigated separately for theSVAT and for the canopy functioning models. Theywill thus be described separately.

5.4.1 Assimilation into SVAT mod-els.We will first describe and evaluate the SVAT mod-els investigated, then assimilation of remote sensingdata will be presented. Additionally, the role ofheterogeneity of the surface on the distribution ofthe fluxes will be addressed.

5.4.1.1 Comparison and improvement ofSVAT models

5.4.1.1.1 Comparison of SVAT modelsAssimilation of remote sensing data will only beefficient if the process models used are representingadequately the mechanisms involved. It was there-fore critical to evaluate and compare the SVATmodels within the framework of assimilation ofremote sensing data, i.e. including an evaluation onvariables such as surface soil moisture and canopyor soil temperatures. A rigorous approach was de-veloped to quantify the performances of the modelsaccording to a range of criterions:Criterion 1: Accuracy of processes description.For this purpose, the models were run using all themeasurements available that could be used to char-acterise directly (without model tuning) some inter-nal variables or parameters.Criterion 2: Portability of the models. The modelswill be run with the values of the variables and pa-rameters extracted from the literature. Here again,no model tuning is tolerated.Criterion 3: Robustness of the calibration. Themodels are calibrated on the calibration fields andtheir performances are evaluated on the validationfields. In this case model tuning (calibration) is therule, but limited on the calibration fields.

Table 11 presents briefly the models investigated. Itcorresponds to a range of complexity, from thesimplest one, MAGRET, to the more complex andphysically based one, SISPAT.

Model Reference yearMAGRET Courault etal., 1995ISBA Noilhan and Planton, 1989TEC Chanzy and Bruckler, 1993SOIL Janssen 1998SiSPAT Braud et al., 1995

Table 11. The SVAT models investigated in theReSeDA project.

The results suggest some general comments:• model simulations were close despite the large

range of model parameterisation and complexity;• except for SISPAT, better results were obtained in

scenario 2 compared to scenario 1, which may besurprising considering scenario definition. How-ever, this was mainly due to the minimum sto-matal resistance;

• for the more complex model (SiSPAT) scenario 2did not gave better results; this was explained bythe less adequate hydraulic properties obtained byusing pedo-transfer functions; this was less sensi-tive in the case of simple models in which onlyfield capacity and wilting point were required asinput;

• on the calibration fields, scenario 3 results werebetter than for the two other scenarios; however,application to the validation fields did not alwaysprovided good results for turbulent heat fluxes,while better results were always obtained for thesoil heat flux and quite always for soil moisture;

• the use of SiSPAT considering an homogeneoussoil column generated a lot of troubles showingthat it is not pertinent to apply such type of com-plex model without taking into account soil lay-ering.

5.4.1.1.2 Improvement of SVAT models

5.4.1.1.2.1 Influence of soil hydraulic propertieson SISPAT model performances.Richards’ equation describing water movement insoil requires knowledge of the soil water retentioncurve, and the hydraulic conductivity function. Un-certainties in soil hydraulic characteristics affect theperformance of land-surface schemes used inSVAT models. A series of numerical experimentswere performed to assess the sensitivity of theSISPAT model to soil parameters and compare theresults with Alpilles-ReSeDA data set under baresoil conditions. The simulations, using the simplestversion of SISPAT with a single horizon (i.e. ho-mogeneous soil), show that water content and landsurface water and energy fluxes are very sensitiveto these parameters when soil moisture is lowerthan approximately 90% of the water content atsaturation. It is found that, for bare soil condition,latent heat flux is the most sensitive factor to soilhydraulic properties. For wetter soils the energyfluxes are totally independent from the soil watercontent. It is therefore concluded that SISPAT hasto be implemented with more than a single layer todescribe accurately surface soil water moisture.

5.4.1.1.2.2 Effect of aerodynamic resistancescheme on SISPAT performancesThe aerodynamic resistance scheme used in thetwo-layer SVAT model such as SISPAT, affectsstrongly the surface fluxes and temperatures simu-lation. A large review of existing schemes has beenconducted and 6 parameterisations based on thefirst approximation of turbulence theory (the K-


theory) have been considered. These parameterisa-tions were implemented in the SISPAT schemealready calibrated on the experimental sites of theReSeDA experiment. Results show that the differ-ent parameterisations lead to similar simulations oftotal surface fluxes but to different soil and vegeta-tion temperatures. This is due to the change in thepartition of energy between the soil and vegetationlayers considered in the model. These differencesare critical for the description of the radiative tem-perature as observed by remote sensing in the ther-mal infrared.

5.4.1.1.2.3 Estimation of the aerodynamic char-acteristics over the ReSeDA experimentAs stated previously, the description of the aerody-namic characteristics is critical for an accurate rep-resentation of the surface fluxes and the associatedremote sensing variables such as temperature andsurface soil moisture. Therefore, emphasis was puton this issue. The displacement height, roughnesslength for momentum, the stability correction termsand the scaling parameters (u∗ and T*) can be de-termined by using the measurements of wind speedand air temperature at two heights and the sensibleheat flux above the canopy through a two-concentric-loop iterative (TCLI) method. Applica-tion of this method to a wheat field shows that es-timation of d and z0 were possible only if the tem-perature gradients were measured differentially.nevertheless, consistent estimation of the scalingparameters, and therefore the Monin Obhukovlength and the stability correction terms was found.

5.4.1.2 Estimation of surface fluxes usingthe SEBAL algorithmThe SEBAL algorithms proposed by Bastiansen etal. (1998) is based on a surface energy balancemodel using the Monin-Obukhof theory for theboundary layer description. The algorithm providesinstantaneous estimates of evapotranspiration usingspatially distributed measurements of surface al-bedo, the normalised difference vegetation index(NDVI) and surface temperature. The vegetationand soil characteristics are not explicitly describedin the SEBAL algorithm. They are retrieved fromempirical relationships with the former input vari-ables.The SEBAL algorithm requires to have both com-pletely dry and wet pixels to compute the spatialdistribution of the fluxes. The wet pixels are used toestimate air temperature at the reference level. Thedry pixels are used to estimate the resistance totransfer at the canopy surface through wind speedestimates at the reference level. The selection of dryand wet pixels is based on the distribution of albedoand surface temperature (Figure 53)The SEBAL inputs were derived from the POL-DER data for albedo and NDVI, and IN-FRAMETRICS camera for surface temperature.Special procedures were used to get improved val-ues of SEBAL input:• Albedo. The method developed in §5.3.1.7.2 was

used here.

• NDVI. The NDVI values were computed fromnadir, thanks to the MRPV BRDF model(§5.3.1.7.1) applied on atmospherically correctedPOLDER data.

• Surface temperature. To minimise the directionaleffects, the selected Ts measurements were re-stricted to view zenith angles lower than 25° andoutside the hot-spot region. The pixels valueswere averaged over 2 minutes period. All the sitewas covered within approximately one hour.

• global radiation. The ReSeDA meteorologicalstation global radiation measurements were used.They were averaged over one hour.

• Atmospheric long wave radiation was derivedfrom ground measurements averaged over onehour.

albedo

Su

rfa

ce

Te

mp

era

tur

Wet pixels

Dry pixels

Figure 53. Procedure used to select the wet and drypixels using the co-distribution of albedo and sur-face temperature.

Fluxes AbsoluteRMSE (W.m-2)

RelativeRMSE (%)

Net radiation 60 10%Soil heat flux 40 30%Sensible heat flux 50 25%Latent heat flux 50 25%

Table 12. Results of the validation of the SEBALalgorithm observed over seven points located onalfalfa, wheat and sunflower crops.

The Results (Table 12) show that the SEBAL netradiation estimates were close to field measure-ments. Comparison between soil heat flux fromSEBAL simulations and field measurements werenot satisfactory although the rmse was correspond-ing to the field data accuracy. Due to inconsisten-cies in the Bowen ratio and aerodynamic methodsfor convective flux estimation, it was not possibleto get final conclusion. Only few eddy correlationdata were available and reliable. However, maps ofsurface energy fluxes (Figure 54) were generated.They show a strong spatial variability inside somefields that can reach up to 50 Wm-2 although thevariability between fields could reach up to500W.m-2.


Figure 54 Map of the evapotranspiration (latentheat flux) derived from the SEBAL algorithm usingPOLDER and INFRAMETRICS data over the Re-SeDA site, April 10th at 20m spatial resolution andaveraged over 45 min.

5.4.1.3 Spatial heterogeneity of fluxesThe atmospheric surface layer couples processes inthe atmosphere with those at the land surface.However, the effect of spatial heterogeneity is notfully understood. Therefore these investigationswere aiming at a better description of the heteroge-neity of fluxes with consequences on the scaling offluxes.

5.4.1.3.1 Boundary layer structure over hetero-geneous surfacesThe small remotely piloted aircraft flown during theReSeDA experiment measured profiles of heat andhumidity in the atmosphere's surface layer nearboundaries between different land cover types tostudy atmospheric surface layer structure over het-erogeneous terrain.

Figure 55. Mean air temperature (top) and huidity(bottom) distribution as a function of height anddistance from the boundary. Darker shadingrepresents values above the mean (mid-grey).Horizontal axis ticks are at 100 m intervals andvertical axis ticks are spaced by 10m; the verticalline shows the surface boundary position.

Figure 55 illustrates the typical structure of the at-mosphere at a boundary between two contrastedfields for neutral stability conditions. In this casethe surface to the right of the boundary was irri-

gated meadow, with dry fallow land to the left, andthe wind was blowing from the meadow to the fal-low (about 4 ms-1). The contrast near the surface oneither side of the boundary is apparent, with coolmoist air over the meadow and warm dry air overthe fallow; by 30 m above the surface the contrasthas practically disappeared. However, the sharpeststructures were observed in unstable conditionswith contrasting surface. This provides a first esti-mation of the blending height with values up toabout 30 m for these conditions, which is againconsistent with expectations.

Figure 56. SEBAL derived sensible heat fluxaround the scintillometer optical path (the blackline).

5.4.1.3.2 Heat flux scalingTwo methods for retrieving sensible heat flux over abare soil/wheat composite surface are compared(Figure 56):• The first method is based on field measurements

using large aperture scintillometers. The compari-son against reference fluxes obtained from eddycorrelation technique shows that scintillometry-derived fluxes are overestimated by about 10%(Figure 57). A numerical experiment demonstratesthat this is induced by aggregation problems result-ing both from non linearity in the relationship be-tween the structure parameter for refractive indexand the sensible heat flux, as well as from non uni-form sensitivity of the scintillometer along the path-length.

• The second method is based on the use of the SE-BAL algorithm (§5.4.1.2). A comparison againstscintillometry derived fluxes shows important dis-crepancies (Table 13). They mainly result fromlarge errors in the estimation of the roughnesslength z0 in the model. This demonstrates that theuse of an empirical relationship based on NDVIonly is inadequate for inferring this key parameterin SEBAL.


y = 1.1096x

r2 = 0.961

y = 1.159x

r2 = 0.9538

0

100

200

300

400

0 100 200 300 400

H reference 3D (Wm-2)

H s

cin

till

om

etr

y (

Wm

-2)

RL1

RL2

1:1

lin reg RL1

lin reg RL2

Figure 57. Comparison between spatially-averagedsensible heat flux derived from eddy correlationmeasurements and from scintillometry, using 2 ag-gregation schemes for roughness length.

Sensible heatflux (W.m-2)

Wheat Baresoil

Opticalpath

Eddy correlation 303 189 273Scintillometer - - 297SEBAL(initial version)

122 175 141

SEBAL(z0 prescribed)

243 206 241

Table 13. Comparison between reference eddy cor-relation, scintillometry, and SEBAL estimates ofsensible heat flux for individual fields (wheat andbare soil) and integrated over the scintillometerpath-length.

5.4.2 Assimilation into canopy func-tioning modelsSimilarly to the SVAT models, performances of theassimilation into canopy functioning models arestrongly linked to the accuracy of the canopy func-tioning models themselves. A range of canopyfunctioning models have been investigated (Table14). All these models have first to be calibratedover the particular crops considered within the Re-SeDA project. Special emphasis was put on wheatcanopies for which the largest range of groundmeasurements and remote sensing data are avail-able. We will therefore present first the perform-ances of the models after calibration, and then ad-dress the assimilation issue itself.

MODEL Reference Year Time stepV-S Cayrol et al. 1999 hourROTASK Jongschaap 2000 daySTICS Brisson et al. 1998 day

Table 14. The several canopy functioning modelsused.

5.4.2.1 Calibration and evaluation of can-opy functioning models.The results of the calibration and evaluation of thethree models considered here will be reviewedbriefly. Similarly to the SVAT evaluation proce-dure, the calibration and evaluation procedureswere kept as independent as possible to insure con-fidence in the robustness of the evaluation.

Figure 58. Comparison between the V-S modelsimulation of water storage in the 0-140cm layer tothe actual measurements over a wheat field.

Figure 59. Comparison between V-S modelsimulations and actual LAI (top) and biomass(bottom) measurements over a wheat field.

5.4.2.1.1 The V-S modelThe V-S model (for Vegetation-SVAT) was devel-oped by Cayrol et al. (1999) for simulation of en-ergy, water and carbon exchanges over the wholegrowing season. Indeed, this model couples a sim-ple SVAT model to a canopy functioning model.This model was originally developed to describesparse canopies as in the case of Sahel. The cali-bration was performed using either data from theliterature or values of the variables measured duringthe experiment. This scheme corresponds to Crite-rion 1 scenario for SVAT. The evaluation showssome discrepancy between the latent and sensibleheat fluxes and moisture in the rooting zone (Figure58). This was mainly coming from a problem in thevalue of the rooting depth. However, the canopyfunctioning model, calibrated mainly with valuesextracted from the literature was performing quitesatisfactorily. Nevertheless, some error is observedfor LAI at the end of the cycle, presumably becauseof a poor description of senescence processes. Be-cause of these deficiencies, and the large effort tospend for corrections from sparse Sahelian canopyto wheat crops, this model was not used later in theassimilation process.


5.4.2.1.2 Calibration of the ROTASK modelThe ROTASK 1.5 model is a derivation of the SU-CROS models (Spitters, 1990). It includes a tip-ping-bucket water balance without capillary rise(DRSAHE), a soil organic matter module (SOM), acrop growth module (WOFCROP), and a manage-ment module (MAN). The calibration was focus-sing on three modules:• The water balance module (DRSAHE). The four

parameters characterising the bucket storage(water holding capacity when air dry, at wiltingpoint, at field capacity, and at saturation) wereadjusted over the calibration ReSeDA wheatfields. The first layers up to 60 cm were quite wellsimulated by the original ROTASK model. How-ever, below 60cm, this was not the case (Figure60 top). The calibration of the bucket characteris-tics as well as the dynamics of the rooting systemand water uptake potential resulted in clear im-provement of the simulations (Figure 60, bottom).

• The temperature sums module (SOM°) was cali-brated to describe the developmental stage moreaccurately. Additionally, the leaf life-span pa-rameter was adjusted to better simulate the ceas-ing of the crop canopy.

• The dry matter partitioning towards roots(WOFCROP), shoots, leaves, stems and ears wascalibrated for each development stage (Figure 61).

The resulting simulations show good performancesfor LAI dynamics as seen in Figure 62, which iscritical for remote sensing assimilation purposes.Additional tests were performed to ensure the ro-bustness of the calibration.

Figure 60. Dynamics of the soil moisture in the 60-80cm layer simulated (solid line) and measured(dots). The top figure corresponds to the initialROTASK model; the bottom figure corresponds tothe model simulation after calibration of soil andcanopy parameters.

Figure 61. Dry matter partitioning coefficients usedto calibrate the ROTASK model.

Figure 62. Comparison between measured (directmeasurements) and ROTASK simulated LAI valuesafter calibration.

5.4.2.1.3 Coupling STICS and radiative trans-fer models.The STICS model is a generic model of canopyfunctioning capable to represent the behaviour of alarge range of species. It was already calibrated onwheat crops (Brisson, 1998). However, the maineffort was directed to the coupling of STICS to amulti-layer and multi-component version of theSAIL radiative transfer model. It allows to describethe structure according to the scheme presented inFigure 63. Therefore, the STICS model that origi-nally outputs only the green leaf area index, has tosimulate the dynamics of the organ are index (OAI)required for the detailed canopy structure descrip-tion (Figure 63). In the same way, the optical prop-erties have to be described according to the PROS-PECT model, i.e. by developing a module thatsimulates the chlorophyll, the water, and the drymatter of the organs, as well as a characterisation oftheir internal structure.

Soil background

Green leaves

Senescent stems

Green earsSenescent ears

Green stems

Yellow leaves

Figure 63. Schematic description of the wheatcanopy structure used in the SAIL radiative transfermodel.


Figure 64 shows the adaptations that were devel-oped from the original STICS model to simulate theOAI and the composition (chlorophyll, water, drymatter) for all the organs considered, i.e. green andyellow leaves, green and senescent stems and ears.These additional characteristics were mainly de-rived from empirical relationships with the initialoutputs of STICS, i.e. phenology, total dry matter,green LAI, nitrogen nutrition index and surface soilmoisture. These relationships were calibrated overdata sets independent from the ReSeDA experi-ment. They were then evaluated over ReSeDA.

Figure 64. Schematic description of the couplingbetween STICS and SAIL+ PROSPECT models,and the necessary adaptations that have beendeveloped for the STICS model.

Results show that generally, our parameterisationseems reasonable. However, for some variablessuch as the leaf water content, the agreement isrelatively poor (Figure 65). This will have impor-tant consequences on the success of the assimila-tion, particularly when using µ-waves or reflectancein the short-wave infrared.Figure 66 shows the resulting OAI dynamics simu-lated by our parameterisation. The agreement withactual ReSeDA experiment measurements is gener-ally good. Other adaptations were implementedconcerning the soil water balance in order to get abetter representation of the moisture of the surfaceof the soil.

Figure 65. Comparison between the simulated leafwater content using our scheme, to the actualmeasurements over ReSeDA.

Figure 66. Simulation of the OAI for a wheat fieldin ReSeDA (top). The bottom graph shows thedynamics of the green LAI as measured byplanimetry (dots) or simulated by our STICSadaptation (light green). Total green OAI is alsocompared to the LAI2000 measurements (darkgreen).

Figure 67. Comparison between our reflectancesimulations ( solid lines) based on STICS+SAIL tothe actual ReSeDA POLDER (simple dots) andSPOT data (dots with vertical bars).

Figure 68. Comparison between the STICSsimulations of the surface (0-5cm or 0-10cm) soilmoisture to actual measurements (TDR) during theReSeDA experiments.

OriginalSTICS

STICSAdaptations

SAIL orCLOUD


Finally, the simulated reflectance (Figure 67) orsurface soil moisture used in the µ-wave domain(Figure 68) were compared with actual measure-ments and show generally a good agreement. Thiswill allow to investigate the last step of the project:remote sensing data assimilation.

5.4.2.2 Assimilation of remote sensing datainto canopy functioning models

5.4.2.2.1 Yield estimation from assimilation inROTASK model at the field scaleThe ROTASK model after calibration (see§5.4.2.1.2) simulates the growth of wheat crops. Itis capable of producing each day a LAI value. How-ever, due to uncertainties in the values of somevariables or parameters, as well as model approxi-mation itself, it may deviate from the actual func-tioning. Remote sensing can be used to force themodel to follow the right course by inputting theLAI values derived from satellites. In this case, theSPOT data were used to compute the WDVI and toestimate LAI according to results presented in§5.3.1.5. Three methods of remote sensing assimi-lation for yield estimates were compared:1. Using the LAI simulated by ROTASK without

exploiting remote sensing data2. Resetting the LAI value at each date of image

acquisition, and in the interval the ROTASKmodel provides the updated values.

3. Using ROTASK LAI up to the first image, andthen using interpolated LAI values drivingROTASK.

Results show a small advantage to method 3, i.e.ROTASK forced with remote sensing estimates ofLAI. However, the field variability and the fact thatROTASK was calibrated on part of the fields limitsthe validity of the results. Additional tests shouldtherefore be conducted.

Figure 69. Comparison of the LAI time coursecorresponding to the three methods used to drivethe ROTASK model over field 120 (method 1:dotted line; method 2: bold solid line; method 3:light solid line).

CropGrowth&

waterbalance

Init, AEZ,Act.inputCrop growthmodel

Crop stagesDevelopmentstage TimePotentialgrowth

Crop stagesDevelopment

stageH2O, N limited

MeteoGrowthdelay

RemoteSensing

location wheatin region

Regional yieldgap

signatureó(t) &

Plant water

regionalsignature

Upscaling

wheatfields

Calibration

Figure 70. Combination methodology for theestimation of yield at the regional level usingremote sensing and crop growth simulation.AEZ=Agro Ecological Zone, N=nitrogen.

5.4.2.2.2 Use of Remote sensing and the RO-TASK model for regional yield estimationThis study aimed at extending wheat productionestimates from field to regional scales concerning.Regional information available from the MARSproject (Soil map, DEM, crop statistics, ancillarydata, etc.), and field segment information acquiredby the group of BRGM, were used to generate aland use map using optical satellite imagery. Thelocally calibrated Cloud-model (§5.3.3.2.4) wascoupled to the ROTASK model to assimilate ERStime-series (Figure 70). ROTASK was initializedfor the local growth conditions and fed with theactual meteorological data (temperature, rainfall,etc.) to predict wheat grain production.The flowering stage corresponds to the transitionbetween the vegetative and the generative growthcycles. This was used to adjust the crop growthmodel. As a matter of fact, this stage can be clearlyrecognized in the average time-series of the ERSdata (Figure 71).

m ean DN value of w interwheat fie lds

200

250

300

350

400

450

500

550

19/12/96 23/01/97 27/02/97 03/04/97 08/05/97 12/06/97 17/07/97 21/08/97 25/09/97

Tim e

DN

val

ue

Figure 71. Typical evolution of ERS data overwheat fields, with the flowering date indicated.

A comparison was achieved between initial RO-TASK production estimates and those using ERSdata for adjustment of the flowering date. Table 15shows that the use of ERS data changes slightly theyield estimation, and therefore the production. Al-though small, the difference ranges from 1% to 6%which is very important for large scale productionestimation. Table 15 shows also the great impor-tance of management practices for which only little

Flowering


information is available at the regional scale. Re-mote sensing could therefore offer one way to getbetter estimates of the yield by implicitly account-ing for these effects through the impact on phenol-ogy and LAI as demonstrated previously(§5.4.2.2.1).

Sce

nar

io

Soi

l typ

e

Acr

eage

(ha)

Yie

ldw

ith

out

ER

S

Yie

ldw

ith

ER

S

Rel

ativ

ed

iffe

ren

ce

Potential - 3000 11.4 11.5 -1%1 2700 6.1 5.8 5%H2O limited2 300 5.3 5.0 6%1 2700 4.7 4.5 4%H2O+

N limited 2 300 4.2 4.1 2%

Table 15. Comparison of yield estimation (t.ha-1)from ROTASK with or without ERS data for fordetermination of the flowering date.

5.4.2.2.3 Assimilation of solar and µ-wave datainto STICS-RS modelThe more advanced technique of remote sensingdata assimilation consists in using coupled canopyfunctioning and radiative transfer models to be ableto simulate the dynamics of radiometric data withthe physical and physiological information embed-ded in the models, and ancillary data such as cli-mate and soil variables. The STICS-RT model pre-sented in §5.4.2.2.3 was used to assimilate concur-rently the solar (POLDER data) and the µ-wave(ERS and Radarsat) data. The corresponding radia-tive transfer models used are theSAIL+PROSPECT model for the solar domain, andthe CLOUD model for the µ-wave domain.The assimilation scheme consists here in tuning aselection of parameters and variables of the canopyfunctioning model in order to minimise the distancebetween the simulated and measured radiometricdata (Figure 72).

Radiometric data(simulated)

RTmodels

Radiometric data(measurements)

adjustement

LAI, Cabwater content

heightstructure

soil moisturetemperature

canopy and soilvariables

Canopy state

production

SVA fluxes

Environ. budgetSVATmodel

Canopy &soilforcing

Canopy func-tioning model

Figure 72. Schematic descrition of the assimilationof remote sensing data into coupled canopyfunctioning and radiative transfer models (STICS-RT).

Therefore, a preliminary sensitivity analysis wasconducted to select the parameters and variablesthat could be tuned to change the dynamics of theSTICS-RT model simulation of canopy reflectance(Table 16) and backscattering coefficient (Table17). Results showed that the main variables selectedwere those driving the phenology of the canopy,plant density and some water characteristics of theupper soil layer.

param. waveband

LEV AMF LAX SEN DRP MAT

adens REDPIRMIR

---

++++++

+++++++++

+++++++++

+++++++++

---

stlevamf REDPIRMIR

---

+++

-++

--+

-++

---

iplt REDPIRMIR

---

++++

-+-

---

---

++++++

hcc1 REDPIRMIR

+++++++++

---

---

---

---

+++

Table 16: Summary of the sensitivity of thereflectance in the red (RED), near infrared (NIR)and short wave middle infrared (MIR) to theparameters of STICS for the 6 main phenologicalstages (LEV: emergence to MAT: maturity). Thesensitivity increases with the number of +. Theparameters are: adens: compensation between stemnumber and plant density, stlevamf: sum ofdevelopment units between the LEV and AMFstages, (iplt: sowing date, density: number of plantsper m² at emergence, extin: extinction coefficient ofPAR by the canopy, hcc1 and hcc3: volumetric soilmoisture at the field capacity for the first and thirdhorizons, h1: initial moisture of the first horizon,n1: initial nitrogen content of the first horizon.

Para-meter

DM LAI σ°σ°σ°σ°ERS

σ° (23°)σ° (23°)σ° (23°)σ° (23°)Radarsat

σ° (39°)σ° (39°)σ° (39°)σ° (39°)Radarsat

iplt ++++ +++ + + +density ++ ++ + + +extin +++ +hcc1 ++ + ++++ ++++ ++++hcc3 ++ + +++ +++ +++h1 ++ + +++ +++ +++n1 ++ + ++ ++ ++

Table 17: Same as Table 16 but for thebackscattering coefficient as observed by ERS andRadarsat.

These variables and parameters were finally ad-justed in the assimilation process using the POL-DER and ERS and Radarsat data over the wholegrowth season.Figure 73 illustrates some results acquired over awheat validation field. It clearly shows that the useof remote sensing observations allows to correct thevalues of the canopy functioning model variablesand parameters identified in the sensitivity analysisstep, leading to a better description of the actualprocesses and of the resulting variables of interest.The LAI dynamics is mainly corrected by shiftingthe time course by almost 20 days earlier (Figure73). This induces an increase of the biomass pro-


duction (DM) as compared to the initial simulationsand makes the final result closer to the actualground measurements.

Figure 73. Comparison between simulations of theSTICS-RT model before ("direct") and afterassimilation of POLDER and ERS and Radarsatdata (Assim) over field 300. Four variables arerepresented here, from top to bottom: LAI, totalabove ground dry matter (DM).

Figure 74. Comparison of measured top layer soilmoisture values to those estimated by assimilationof POLDER and ERS and Radarsat data (left) andPOLDER only (right) over field 300.

However, the situation is not so clear for the soilmoisture in the top layer that is mainly determiningµ-wave data (Figure 74). This reflects the difficultyin describing with a simple scheme as used inSTICS the top surface water content, particularlybecause of the daily time step that may not be fineenough to account for the sometime large diurnalvariation. Additional investigation shows that the µ-wave data did not contribute much to the improve-ment of canopy functioning description (Figure 74).It is mainly governed by the information in the so-lar domain.

6. Evaluation - usercomponentThe ReSeDA project was mainly and directly ori-ented towards methodological objectives. There-fore, it is not obvious to identify direct and classicalend users. However, the application objectives pur-sued within the project were addressing the fol-lowing issues:• Agriculture policy• landscape management• precision farming• hydrology• global vegetation monitoring

Therefore, three main categories of users could beidentified:1. The scientific community who is looking for

issues such as those addressed by ReSeDA, i.e.• Canopy and soil functioning;• Estimation of canopy characteristics from

remote sensing observations;• Scaling issued;• Assimilation of data into process models;• synergy between different spectral domains;• exploitation of the directional variation of

reflectance and brightness temperature2. Remote sensing data providers such as space

agencies and industrial companies. The key is-sues to be addressed are:• The definition of the optimal configuration

of sensors;• The prototyping and evaluation of algorithm

for the generation of biophysical products.• The synergy between sensors. One issue ap-

pears obvious: there will be in the comingyears a range of sensors, and it is necessaryto exploit in synergy these sensors to im-prove the accuracy of the information on soiland vegetation surfaces. This project con-tributes significantly to the development ofstrategies of remote sensing data ingestionby several sensors using the algorithms thatwere proposed.

3. "true" remote sensing users such as servicecompanies, agriculture bodies, national orEuropean governmental organisation dealingwith agriculture and environment, as well asthe global change community.

Although ReSeDA was not close to the communityof end users, except the two first categories listed,the project has initiated a momentum for develop-ing studies exploiting the data base or the algo-rithms developed. Here is a list of the main studiesinitiated by the ReSeDA project and sponsored bythird parties, proving their interest in the project.• VEGETATION: part of the data base was used to

evaluate algorithms dedicated to derive biophysi-cal products from VEGETATION sensor. The


scaling problem was also investigated for this ap-plication.

• HIBISCUS, the future generation of globalmonitoring instrument planned by CNES. The di-rectionality of observations and scaling issues arecritical for this mission.

• VALERI: a project aiming at evaluating the bio-physical products of large swath satellites. WithinVALERI, an issue is devoted to the developmentof algorithms, and to the scaling effects. It istherefore largely inspired by the methods devel-oped within ReSeDA.

• precision-farming: studies are on going withCNES, ASTRIUM, AVENTIS, SPOT image andSCOT on the definition of products for precisionfarming. Here again, ReSeDA has served as abaseline for the assimilation part of this project.

• CROMA, a European project on methodologicalaspects for precision farming

• ADAM, a project funded by CNES in collabora-tion with Romania and aiming at the exploitationof multi-temporal data for agriculture applicationsincluding precision farming.

• Synergy between SAR and LSPIM missions, astudy sponsored by ESA for supporting LSPIMand its following, with important outputs fromReSeDA on data assimilation which is the mainway to ingest data coming from different sensors.

• WATERMED, a European project dealing withhydrology and remote sensing.

• EC-MARS: evaluation of the Crop GrowthMonitoring System (CGMS) developed by JRCIspra within the MARS project and its future.

• Météo-France who is looking for improving thesurface scheme description for weather predictionmodels. The ISBA model from météo-France wastherefore included in the series of SVAT modelsinvestigated, for comparison to others and im-provements.

7. Exploitation planThe exploitation opportunity is mainly driven bythe definition of the sensor specification and theactual use of current satellites. This data set offersthe unique opportunity to investigate concurrently• The spatial scaling problem• The synergy between observations obtained in

different spectral domains• The temporal revisit frequency• The functioning of soils and canopies for envi-

ronmental issuesSpace agencies as well as space industry companiescould be potentially (and have been) interested inexploiting this data base and the results obtained.National or European governments could be also(and have been) potentially interested in developingtools for the monitoring of crops and the evaluationof the environment

The exploitation plan will first consists in writingarticles corresponding to the material presented atthe EGS in Nice in April 2000, for publication ininternational journals and presentation in confer-ences. In parallel, the investigation will be pursued,including further use of the data base, as well asvalidation and improvements of the results overindependent data sets. This will be described here-after according to the same hierarchy as previously.We should note that what is listed here are the po-tential investigation to be undertaken. From thisinitial list, some investigations are already under-taken, while some others will start depending onexternal funding or availability of researcherswithin the consortium, or outside, but in collabora-tion with the ReSeDA partners. Most of the investi-gations that are already started are listed in theTechnology Implementation Plan.

7.1 Radiative transfer model-ling and model inversion

7.1.1 Solar domain

• Experimental validation of theoretical resultsobtained on model inversion. The investigationconducted on model inversion was mainly basedon reference model simulations. Some approacheshave been validated (neural network, optimisa-tion, vegetation indices), but a more completevalidation is required, where the contribution ofthe spectral and directional information could beanalysed. Further, the importance of the a prioriinformation put in the system, derived by theknowledge of the species should be also evaluatedfrom this experiment.

• Improvement of spectral integration for albedoestimates. The validation of albedo estimationfrom PODLER data shows that a bias affects theretrieved values. This was identified as a spectralintegration problem. This could be quite easilycorrected for by developing an improved set ofspectral coefficients to extend the few POLDERbands hemispherical reflectance to the whole so-lar spectral domain.

• Non linearity, spatial heterogeneity and scaling.In case of non linear processes relating a givensurface variable to radiances observed by remotesensing sensors over heterogeneous scenes, thevalues of a the variable will depend on the scaleof observation. This has been demonstrated forLAI over the ReSeDA site for VEGETATIONspatial resolution. However, It should be possibleto limit the problem by inputting a simple estima-tion of the spatial heterogeneity. This issue will bepartly investigated within the VALERI project.


7.1.2 Thermal infrared domain

• Modelling the directional distribution of thebrightness temperature. The First results ob-tained on the directionality of brightness tem-perature demonstrated the range of variation ofthe effects, and also the difficulty encountereddue to the temporal variation during directionalsampling. Investigation on the modelling of thedirectional distribution of brightness temperaturecould now be initiated. Presumably, the similitudewith the general BRDF patterns invites to test theclassical BRDF models used in the solar domain.

• Estimation of the component temperature fromcombined solar and thermal domains. The de-termination of the component temperature (i.e.soil, vegetation, eventually illuminated and shad-owed fractions) could be very pertinent for SVATmodelling and emissivity-temperature de-coupling. The potential of the combination ofthermal and solar domains is very strong for theseissues. The ReSeDA experiment offers a uniqueopportunity for their evaluation.

7.1.3 µ-wave domain

• Improved land use mapping by per field dis-crimination. One of the main difficulty in ex-ploiting µ-wave for land cover mapping, is theimportant scattering associated to the data. Fil-tering is therefore required, and results indicatethat some filters are more efficient than others.However, for agriculture landscapes, the organi-sation of space into field units that are latter asso-ciated to a particular crop, should be exploitedand will therefore lead to improved classificationresults because the scatter in the data will besmoothed out over the field.

• Evaluation of alternative inversion techniques.In the present work presented, the inversion tech-niques used are either based on empirical relation-ships, or on optimisation algorithms. However,very little a priori information is used, and the un-certainties in the data and the model are not ac-counted for. Therefore, additional investigationshould be directed to implement advanced inver-sion techniques that could exploit concurrentlyseveral configurations as demonstrated in§5.3.3.3.3.

• Solving the soil roughness problem. The effecton soil roughness on the estimation of moisture orcanopy characteristics from the backscattering co-efficient is difficult to account for and was identi-fied as a main source of inaccuracies of inversealgorithms. Passive µ-wave could be a solution tothe problem. However, on an operational basis, itis associated to low spatial resolution (more than10 kilometres). The use of several incidence, po-larisation or frequencies should be investigated tominimise the roughness effect. For this purpose,the work conducted on the IEM model and the ef-fective correlation length has to be exploited to

simply eliminate the roughness effect when usingconcurrently several configurations.

• Choice of canopy biophysical variables depend-ing on canopy structure. In the µ-wave domain,the backscattering coefficient is generally not asimple function of a given canopy characteristicssuch as LAI or water content. However, for agiven canopy that follows a genetically deter-mined structure dynamics, the strong covariancebetween several structure characteristics couldlead to choose the pertinent canopy characteristicsas a function of the development stage and spe-cies. We should note that this could be a problemfor classical operational applications where a sin-gle well identified variable is to be used in deci-sion making. The assimilation is therefore quitepromising for exploiting the µ-wave radiometricinformation.

7.2 Assimilation into processmodels

7.2.1 SVAT modelsTwo ways of exploiting remote sensing data havebeen investigated:

7.2.1.1 Using the temporal variationOnly part of the work was achieved during the Re-SeDA project life time. This was mainly explainedby the important underestimation of the time re-quired for data processing, inducing significantdelay in the data delivery and thus processing.Therefore, the SVAT assimilation, based on thetemporal variation was mostly concentrating on theevaluation, comparison and improvement of SVATmodels. Assimilation into SVATs based on optical,µ-wave and thermal data is thus still to be com-pleted. This is to be accomplished for a selection ofmodels within the WATERMED European project,with comparison of results acquired over othersites.

7.2.1.2 Using the spatial variation.Assimilation with the SEBAL algorithm leads tointeresting results and possible ways of improve-ment. These improvements, particularly regardingthe estimation of the roughness, as well as the betterretrieval of some inputs such as albedo will be alsoimplemented within the WATERMED project, withcomparison to other sites. This will be also testedusing the ASTER images that provides both solarand thermal infrared data at a relatively fine spatialresolution.

7.2.2 Canopy functioning modelCanopy functioning models will be mainly ex-ploited for agriculture and environmental applica-tions. The coupling scheme implemented for wheatcrops will be evaluated under different conditionswithin the ADAM project funded by CNES and


taking place in Romania. It will focus on the ex-ploitation of concurrent use of the three SPOT sat-ellites to get high temporal repetitivity. ERS andRadarsat data will be also acquired during the sameperiod. Intensive ground truth is organised in orderto be able to evaluate and improve the assimilationscheme developed within ReSeDA. The protocolsand experience gained within ReSeDA was highlyexploited. This project will allow the definition ofthe optimal revisit frequency.The combination of µ-wave and solar observationswill be investigated within a study initiated by theEuropean Space Agency to further the Land Sur-face Processes Mission. ReSeDA data base anddevelopments will serve as the main basis for thisstudy.The work achieved over wheat will be extendedover maize and sunflower canopies of the ReSeDAexperiment. This will be part of the CROMA Euro-pean project. This will be also used within theCHRIS-ALIDE (CHRIS Assimilation of Land In-formation Data Experiment) project for which theCHRIS imaging spectrometer aboard the PROBAplatform will be launched next spring. The ReSeDAsite will serve as the experimental field.

8. List of relatedpublications & de-liverables

8.1 Publications

Bacour, C., Jacquemoud, S., Leroy, M., Haute-coeur, O., Prévot, L., Weiss, M., Bruguier, N.,and Chauki, H. 2000. Estimation of vegetationcharacteristics by inversion of three canopy re-flectance models on airborne POLDER data.Physics and chemistry of the Earth v. submitted.

Baghdadi, N., King, C., Chanzy, A., and Wigneron,A. 2000. An empirical adaptation of the IEMmodel for retrieving surface parameters fromSAR observations. Physics and chemistry of theEarth v. submitted.

Baret, F. 1998. ReSeDA: Assimilation of Multisen-sor & Multitemporal remote sensing data tomonitor vegetation and soil functioning,. INRA,Avignon (France).

Baret, F. 1999. ReSeDA: Assimilation of Multisen-sor & Multitemporal remote sensing data tomonitor vegetation and soil functioning,. INRA,Avignon (France).

Baret, F., Knyazikhin, J., Weiss, M., Myneni, R.,and Pragnère, A. 1999. Overview of canopybiophysical variables retrieval techniques. InALPS'99, Meribel (France).

Baret, F., Weiss, M., and Pragnère, A. 1999.Evaluation of model inversion techniques. Ap-plication to the retrieval of key biophysical vari-ables.,. INRA/ESA, Avignon (France).

Baret, F., Weiss, M., Troufleau, D., Prévot, L., andCombal, B. 2000. Maximum information ex-ploitation for canopy characterisation by remotesensing. In Remote sensing in agriculture, Vol.60, pp. 71-82. Association of Appied Biologists.

Braud, I., Chanzy, A., F., B., Calvet, J. C., E., G.-s.,C., K., Prevot, L., Olioso, A., Ottle, C., Taconet,O., Thony, J. L., and Wigneron, J. P. 1998. As-similation of remote sensing data in SVATmodels: description of the alpilles experimentand first results. In European Geophysical So-ciety, XXIII General Assembly,, Nice 20-24April 1998.

Cantelaube, P., Baret, F., Biard, F., and Clastre, P.1999. Provision of thematic studies on agrome-teorological Crop Growth Modelling. Calibra-tion of CGMS physiological crop coefficients,Rep. No. report for contract n° 14165-1998-07.INRA-GEOSYS-SAI, Avignon, France.

Chanzy, A., King, C., Prévot, L., Remond, A.,Wigneron, J. P., Calcagna, P., Mehrez, Z., andDesprats, J. F. 1998. Comparison of ERS andmulti-angle RADARSAT measurements on baresoils : first results of the ReSeDA experiment.In 2nd International Workshop on Retrieval ofBio- and Geo-Physical parameters from SARData for land Applications, Noordwijk (Neth-erland).

Clevers, J. G. P. W., Vonder, O. W., Jongschaap, R.E. E., Desprats, J. F., King, C., Prévot, L., andBruguier, N. 2000. Monitoring wheat growth bycalibrating a crop growth model using opticalsatellite data. Physics and chemistry of the Earthv. submitted.

Coll, C., Caselles, V., Rubio, E., Sospedra, E., andValor, E. 1999. Temperature and emissivity in-version using DAIS and ground measurements.Application to the ReSeDA test site. InALPS99 Symposium, Meribel (France).

Coll, Caselles, Rubio, Sospedra and Valor, 2000,Analysis of thermal infrared data from theDigital Airborne Imaging Spectrometer, Int. J.Remote Sensing. accepted

Coll, C., Caselles, V., Rubio, E., Valor, E., and So-spedra, F. 2000. Field emissivity measurementsduring the ReSeDa experiment. Physics andchemistry of the Earth v. submitted.

Coll, C., Caselles, V., Rubio, E., Valor, E., So-spedra, F., and Baret, F. 2000. Estimating tem-perature and emissivities from the DAIS in-strument. Physics and chemistry of the Earth v.submitted.

Coll, C., Schmugge, T. J., and Hook, S. J. 1998.Atmospheric effects on the temperature emis-sivity separation algorithm. In InternationalSymposium on Remote Sensing, SPIE-EUROPTO, Barcelona, Spain.

Demarty, J., Ottlé, C., Braud, I., Mangeney, A., andFrangi, J. P. 2000. Effect of aerodynamic con-ductances modelisation on SISPAT-RS simu-


lated surface fluxes. Physics and chemistry ofthe Earth v. submitted.

Desprats, J. F., Borne, F., Bégué, A., King, C.,Prévot, L., Baghdadi, N., and Bourguignon, A.2000. ERS and Radarsat images for land useclassification at a regional scale : possibilitiesfor future A-SAR configuratio. Physics andchemistry of the Earth v. submitted.

Francois, C., Ottlé, C., and Olioso, A. 2000. Cor-rection of silicon sensors albedo measurementsusing a canopy radiative transfer model. Physicsand chemistry of the Earth v. submitted.

François, C., Ottlé, C., and Taconnet, O. 1998.Coupling SVAT model with multispcetral ra-diative transfer models into the canopy. First re-sults with the Alpilles experiment. In EGSXXIII general assembly, Nice (France).

Gu, X. F., Jacob, F., Hanocq, J. F., Prévot, L., Yu,T., Liu, Q. H., Tian, G. L., Li, X. W., andStrahler, A. H. 2000. Measuring and analysingof thermal infrared emission directionality overcrop canopies with an airborne wide-anglethermal infrared camera. Physics and chemistryof the Earth v. submitted.

Hobbs, S., Courault, D., Olioso, A., Lagouarde, J.P., Kerr, Y., Mc Anneney, J., and Bonnefond, J.M. 2000. Surface layer profiles of micrometeo-rological variables as measured from unmannedaircraft. Physics and chemistry of the Earth v.submitted.

Jacob, F. 2000. Utilisation de la télédétection cour-tes longueurs d'onde et infrarouge thermique àhaute résolution spatiale pour l'estimation desflux d'énergie à l'échelle de la parcelle agricole.PhD, Université de Toulouse III, Toulouse(France).

Jacob, F., Olioso, A., Gu, X. F., Leroy, M., Haute-coeur, O., and Hanocq, J. F. 2000. Mapping sur-face fluxes using visible, near infrared andthermal infrared data and the SEBAL algorithm.Physics and chemistry of the Earth v. submitted.

Jacob, F., Olioso, A., Weiss, M., Francois, C.,Leroy, M., Hautecoeur, O., and Ottlé, C. 2000.Albedo estimation from Polder data. Physicsand chemistry of the Earth v. submitted.

Jongschaap, R. E. E. 2000. Calibration and valida-tion of ROTASK 1.5 simulation model on fielddata of the ReSeDA project in southern Francewith special reference to winter wheat.,. AB-DLO, Wageningen (NL).

Lagouarde, J. P., Jacob, F., Gu, X. F., Olioso, A.,Bonnefond, J. M., Kerr, Y., Mc Anneney, J.,and Irvine, M. 2000. Spatialisation of sensibleheat flux over heterogeneous landscapes. Phys-ics and chemistry of the Earth v. submitted.

Lagouarde J.P., Chehbouni A., Kerr Y.H., McAne-ney J., Bonnefond J.M. and Irvine M.,1999 : Use of the scintillation method for theevaluation of surface fluxes at regional scale. InALPS'99, Meribel (France), 4p.

Lagouarde J.P., Bonnefond J.M., Kerr Y.H., McA-neney J. and Irvine M., 2000 : Estimation of thesensible heat flux integrated over a two surfacecomposite landscape by scintillometry. Bound-ary layer Meteorol. submitted.

Leroy, M., Hautecoeur, O., Berthelot, B., and Gu,X. F. 2000. The airborne Polder data during theReSeDA experiment. Physics and chemistry ofthe Earth v. submitted.

Macelloni G., Paloscia S., Pampaloni P., Ruisi R.and Susini S., Monitoring crop biomass and soilmoisture with airborne microwave radiometersProc. Int. Geosci. Remote Sensing Symp.(IGARSS=99), Hamburg, vol IV, 1999, pp.2327- 2329.

Macelloni G., Paloscia S., Pampaloni P., Ruisi R.,Dechambre M., Valentin R., Chanzy A., Wign-eron J.P., Passive and active microwave datafor soils and crops characterization,” Proc. Int.Geosci. Remote Sensing Symp. (IGARSS2000), Honululu, vol, 2000, pp. 2522-2524

Macelloni G., Paloscia S., Pampaloni P., MarlianiF., Gai M., “The relationship between the back-scattering coefficient and the biomass of narrowand broad leaf crops,” To be published in IEEETrans. Geosci. Remote Sensing.

Macelloni G.., Paloscia S., Pampaloni P., Ruisi R.,Susini C. and Wigneron J.P. - AAirborne pas-sive microwave measurements on agriculturalfields,@ in Microwave Radiometry and RemoteSensing of the Earth' s Surface - P. Pampaloniand S. Paloscia Eds., VSP Press, The Nether-lands, 2000, pp. 59-70.

Macelloni, G., Paloscia, S., Pampaloni, P., Ruisi,R., Dechambre, M., Valentin, R., Chanzy, A.,and Wigneron, J. P. 2000. Active and passivemicrowave for soil and crops chazracterization.Physics and chemistry of the Earth v. submitted.

Macelloni, G., Paloscia, S., Pampaloni, P., Susini,C., and Ruisi, R. 1998. Airborne micro-waveradiometer measurements on an agriculturalsite: the IROE-STAAARTE mission. In EGSXXIII general assembly, Nice.

Macelloni, G., Pampaloni, P., Paloscia, S., andRuisi, R. 1998. Microwave Emission Featuresof Crops with Vertical Stems. IEEE Transac-tion on Geoscience and Remote Sensing v. Ge-36, 332-337.

Maceloni, G., Paloscia, S., Pampaloni, P., Ruisi, R.,Dechambre, M., Valentin, R., Chanzy, A.,Prévot, L., and Bruguier, N. 2000. Modellingradar backscatter from crops during the growthcycle. Physics and chemistry of the Earth v.submitted.

Marliani F., Paloscia S., Pampaloni P.and. Kong J.A, “A coherent scattering model for interpretingInSAR data from vegetation,” Proc. Int. Geosci.Remote Sensing Symp. (IGARSS 2000), Ho-nululu, vol, 2000, pp.911-913

Marliani F., Paloscia S. and Pampaloni P., “A co-herent scattering model for canopy coveredsoils”, Proceedings AP2000, Davos, Switzer-land, April 2000


Mangeney, A., Demarty, J., Ottlé, C., and Braud, I.2000. Influence of soil hydraulic properties onsurface fluxes, temperature and humidity in theone dimensional SISPAT model. Physics andchemistry of the Earth v. submitted.

Moulin, S., Kergoat, L., Cayrol, P., Dedieu, G.,Prévot, L., Bruguier, N., Desprats, J. F., andKing, C. 2000. Using coupled canopy function-ing and SVAT model in the ReSeDA experi-ment. Assimilation of SPOT/HRV observationsinto the model. Physics and chemistry of theEarth v. submitted.

Moulin, S., Kergoat, L., and Dedieu, G. 1999. Cou-pling of a vegetation model and a SVAT modelin the RESEDA experiment. In ALPS'99, Mé-ribel (France).

Olioso, A., Autret, H., Bethenot, O., Bonnefond, J.M., Braud, I., Calvet, J. C., Chanzy, A.,Courault, D., Demarty, J., Ducros, Y., Gaudu, J.C., Gonzales-Sosa, E., Gouget, R., Jongschaap,R., Kerr, Y., Lagouarde, J. P., Laurent, J. P.,Lewan, E., Marloie, O., Mc Anneney, J.,Moulin, S., Ottlé, C., Prévot, L., Thony, J. L.,Wigneron, J. P., and Zhao, W. 2000. Compari-son of SVAT models over the Alpilles-ReSeDAexperiment. I Description of the framework andthe data. Physics and chemistry of the Earth v.submitted.

Olioso, A., Autret, H., Bethenot, O., Bonnefond, J.M., Braud, I., Calvet, J. C., Chanzy, A.,Courault, D., Demarty, J., Ducros, Y., Gaudu, J.C., Gonzales-Sosa, E., Gouget, R., Jongschaap,R., Kerr, Y., Lagouarde, J. P., Laurent, J. P.,Lewan, E., Marloie, O., Mc Anneney, J.,Moulin, S., Ottlé, C., Prévot, L., Thony, J. L.,Wigneron, J. P., and Zhao, W. 2000. Compari-son of SVAT models over the Alpilles-ReSeDAexperiment. II Models and Results. Physics andchemistry of the Earth v. submitted.

Olioso, A., Prévot, L., Baret, F., Chanzy, A., Au-tret, H., Baudin, F., Bessemoulin, P., Bethenod,O., Blavoux, B., Bonnefond, J. M., Boubkraoui,S., Bouman, B. A. M., Braud, I., Bruguier N.(1),C. J. C., Caselles V.(7), Chauki H.(1), CleversJ.G.P.W.(8), Coll C.(7), Company A.(9),Courault, D., Dedieu, G., Delécolle, R., Denis,H., Desprats, J. F., Ducros, Y., Dyer, D., Fies, J.C., Fischer, A., François, C., Gaudu, J. C., Gon-zalez, E., Gouget, R., Gu, X. F., Guérif, M., Ha-nocq, J. F., Hautecoeur, O., Haverkamp, R.,Hobbs, S., Jacob, F., Jeansoulin, R., Jong-schaap, R. E. E., Kerr, Y., King, C., Laborie, P.,Lagouarde, J. P., Laques, A. E., Larcena, D.,Laurent, G., Laurent , J. P., Leroy, M., McAne-ney, J., Macelloni, G., Moulin, S., Noilhan, J.,Ottle, C., Paloscia, S., Pampaloni, P., Podvin,T., Quaracino, F., Roujean, J. L., Rozier, C.,Ruisi, R., Susini, C., Taconet, O., Tallet, N.,Thony, J. L., Travi, Y., Van Leewen, H., Vau-clin, M., Vidal-Madjar, D., Vonder, O. W., andWigneron, J. P. 1998. Spatial Aspects in the Al-pilles-ReSeDA Pro ject. In Intermational wor-shop: Scaling and modelling in forest. Interestof remote sensing and GIS, Montreal (Canada).

Paloscia S., Macelloni G. and Pampaloni P., 1998,A The relations between backscattering coeffi-cient and biomass of narrow and wide leafcrops, Proc. Int. Geosci. Remote SensingSymp. (IGARSS=98), Seattle, USA pp 100-102.

Paloscia S., Macelloni G., Pampaloni P. and Sigis-mondi S., @The potential of C- and L- bandSAR in estimating vegetation biomass: theERS-1 and JERS-1 experiments@, IEEE TransGeosci Remote Sensing, Vol 37, N. 4, pp 2107-2110,1999

Prévot, L., Baret, F., Chanzy, A., Olioso, A., Wign-eron, J. P., Autret, H., Baudin, F., Bessemoulin,P., Bethenod, O., Blamont, D., Blavoux, B.,Bonnefond, J. M., Boubkraoui, S., Bouman, B.A. M., Braud, I., Bruguier, N., Calvet, J. C.,Caselles, V., Chauki, H., Clevers, J. G. P. W.,Coll, C., Company, A., Courault, D., Dedieu,G., Degenne, P., Delécolle, R., Denis, H., De-sprats, J. F., Ducros, Y., Dyer, D., Fies, J. C.,Fischer, A., François, C., Gaudu, J. C., Gon-zalez, E., Goujet, R., Gu, X. F., Guérif, M., Ha-nocq, J. F., Hautecoeur, O., Haverkamp, R.,Hobbs, S., Jacob, F., Jeansoulin, R., Jong-schaap, R. E. E., Kerr, Y., King, C., Laborie, P.,Lagouarde, J. P., Laques, A. E., Larcena, D.,Laurent, G., Laurent, J. P., Leroy, M., McAne-ney, J., Macelloni, G., Moulin, S., Noilhan, J.,Ottlé, C., Paloscia, S., Pampaloni, P., Podvin,T., Quaracino, F., Roujean, J. L., Rozier, C.,Ruisi, R., Susini, C., Taconet, O., Tallet, N.,Thony, J. L., Travi, Y., Van Leewen, H., Vau-clin, M., Vidal-Madjar, D., Vonder, O. W., andWeiss, M. 1998. Assimilation of Multi-Sensorand Multi-Temporal Remote Sensing Data toMonitor Vegetation and Soil: the Alpilles-ReSeDA project. In IGARS'98, Seatle (WA,USA).

Prévot, L., Chauki, H., Rémond, A., King, C.,Wigneron, J. P., Chanzy, A., Calcagno, P., andDesprats, J. F. 1998. Comparison of ERS andmulti-angular RadarSat measurements over ag-ricultural canopies: first resukts of the AlpillesReSeDA campaign. In 2nd International Work-shop on "Retrieval of Bio- and Geophysical Pa-rameters from SAR Data for Land Applica-tions". ESA Publications, Noordwijk (TheNetherlands).

Prévot, L., Chauki, H., Troufleau, D., Weiss, M.,and Baret, F. 2000. Assimilating optical and ra-dar data into the STICS model for wheat crops :Preliminary results. Physics and chemistry ofthe Earth v. submitted.

Prévot, L., Chauki, H., Troufleau, D., Weiss, M.,Baret, F., and Brisson, N. 2000. Coupling theSTICS canopy functioning model to radiativetransfer models for assimilating remote sensingdata in the solar and micro-wave domains.Physics and chemistry of the Earth v. submitted.

Prévot, L., Wigneron, J. P., Chanzy, A., Baghdadi,N., King, C., and Dechambre, M. 2000. Moni-toring wheat development from radar observa-tions during the ReSeDA experiment. Physicsand chemistry of the Earth v. submitted.


Paloscia, S., Macelloni, G., Pampaloni, P., andSigismondi, S. 1999. The potential of C- and L-band SAR in estimating vegetation biomass: theERS-1 and JERS-1 experiments. IEEE Transac-tion on Geoscience and Remote Sensing v. ac-cepted for publication.

Schouten, L., van Leeuwen, H., Desprats, J. F.,King, C., Baghdadi, N., Prévot, L., Dechambre,M., and Valentin, R. 2000. Land use classifica-tion based on time series of microwave data.Physics and chemistry of the Earth v. submitted.

van Leeuwen, H., Scouten, L., Jongschaap, R., andKing, C. 2000. Predicting wheat production atthe regional scale by the assimilation of remotesensing data with the crop growth simulationmodel ROTASK v1.5. Physics and chemistry ofthe Earth v. submitted.

Vonder, O., Clevers, J., Desprats, J. F., King, C.,Prévot, L., and Bruguier, N. 2000. Estimation ofbiophysical variables using multiple view an-gles. Physics and chemistry of the Earth v. sub-mitted.

Weiss, M. 1998. Développement d'un algorithme desuivi dela végétation à large échelle. PhD thesis,Ecole des Mines de Paris, Avignon.

Weiss, M., Baret , F., Leroy, M., Hautecoeur, O.,Prévot, L., and Bruguier, N. 2000. Validation ofneural net techniques to estimate canopy bio-physical variables. Physics and chemistry of theEarth v. submitted.

Weiss, M., and Baret, F. 1998. What canopy bio-physical variables can be estimated from largeswath satellite data. In EGS XXIII general as-sembly, Nice.

Weiss, M., and Baret, F. 2000. Evaluation of can-opy biophysical variables retrieval perform-ances from the accumulation of large swath sat-ellite data. Remote Sensing of Environment v.in press.

Weiss, M., Baret, F., Leroy, M., Begué, A., Haute-coeur, O., and Santer, R. 1999. Hemisphericalreflectance and albedo estimates from the ac-cumulation of across track sun synchroneoussatellite data. Journal of Geophysical Researchv. 104, 22,221-22,232.

Weiss, M., Baret, F., Leroy, M., Hautecoeur, O.,Prévot, L., and Bruguier, N. 2000. Validation ofneural network techniques for the estimation ofcanopy biophysical variables from VEGETA-

TION data. In VEGETATION preparatory pro-gramme. Final meeting. G. Saint, Belgirate (It-aly).

Weiss, M., Baret, F., Myneni, R., Pragnère, A., andKnyazikhin, Y. 2000. Investigation of a modelinversion technique for the estimation of cropcharcteristics from spectral and directional re-flectance data. Agronomie v. 20, 3-22.

Weiss, M., Jacob, F., Baret, F., Pragnère, A., Leroy,M., Hautecoeur, O., Prévot, L., and Bruguier, N.2000. Evaluation of BRDF models on POLDERdata. Physics and chemistry of the Earth v.submitted.

Weiss, M., Troufleau, D., Baret, F., Chauki, H.,Prévot, L., Olioso, A., Bruguier, N., and Bris-son, N. 2001. Coupling canopy functioning andcanopy radiative transfer models for remotesensing data assimilation. The case of wheat.Agricultural and Forest Meteorology v. submit-ted.

Wigneron, J. P., Prévot, L., Chanzy, A., Baghdadi,N., and King, C. 2000. Monitoring summer cropdevelopment from radar observations during theReSeDA experiment. Physics and chemistry ofthe Earth v. submitted.

Zhao, W., Olioso, A., Lagouarde, J. P., Kerr, Y.,Mc Anneney, J., Bonnefond, J. M., Bethenod,O., and Gouget, R. 2000. Estimation of aerody-namic parameters from energy balance model-ling and measurements. Physics and chemistryof the Earth v. submitted.

8.2 Deliverables

The main deliverables consist in• the intermediate reports that are listed within

the publications,• the technology implementation plan• The core report• The articles presented during the EGS and that

will be published in a special issue of Physicsand Chemistry of the Earth

• The other articles already published or in thereview process.

• the data base itself that is accessible by the webserver: www.avignon.inra.fr/reseda

http://www.avignon.inra.fr/reseda


9. ConclusionThe ReSeDA project was a collaborative effort of10 partners aiming at the development and evalua-tion of remote sensing data assimilation into proc-ess models. The project was also supported by theFrench Programme National de Télédétection Spa-tiale, and the Programme National de Recherche enHydrologie .

The project was based on the development of anextensive experiment where the spectral, direc-tional, temporal and spatial dimensions have beensampled. It is therefore a very unique effort wherethe radiometric signal was monitored over a5*5km² area, along the growth cycle using sensorsobserving under different directions and coveringthe whole spectral domain, from the with solar,thermal infrared up to the µ-wave. Ground meas-urements have been concurrently acquired to char-acterise the growth, energy and mass exchangesbetween the soil, the vegetation and the atmos-phere.

A hierarchic data base was generated to make theobservations available to the whole scientific com-munity. It is under the form of a web server(www.avignon.inra.fr/reseda) containing the datafiles along the necessary documentation files. Thegeneration of the data base captured significantlymore effort than expected, particularly for the proc-essing of the airborne sensors that all were proto-type sensors. This resulted in delays in the deliveryof calibrated and validated data. Nevertheless, thisdata base constitutes now a unique source of con-current ground and remote sensing observationsover the growth cycle of crops such as wheat, sun-flower and maize.

Canopy biophysical variable estimation from re-mote sensing data was the preliminary step investi-gated before tackling assimilation. Intensive effortshave been put on this issue, leading to significantadvances in the improvement of radiative transfermodels and the development of model inversiontechniques. The identification of the main biophysi-cal variables accessible directly from remote sens-ing data is now quite clear. It includes albedo, LAI,fCover, fAPAR, leaf chlorophyll and water contentfor the solar domain, temperature and emissivity forthe thermal infrared, LAI or canopy water contentfor the µ-wave. For the soil, emphasis was put onsurface soil moisture in the µ-wave domain. Resultsshow thatIn the solar domain, quite accurate estimates ofalbedo, LAI, fCover and fAPAR could be retrievedfrom the spectral and directional variation of thesignal. Because of the lack of sensor sampling suf-ficiently the spectral dimension, no attempt wasdirected to the estimation of chlorophyll or water

content. For the same reason, atmospheric correc-tion were applied based on atmospheric character-istics measured from the ground level. This issueneeds further effort to be investigated and wouldconstitute a major step in remote sensing data ex-ploitation.In the thermal infrared red domain, the main em-phasis was put on the directional variation of thesignal that was presenting a significant range ofvariation. However, the observation of the direc-tional variation could not be instantaneous, and thetemporal and directional effects have to be sortedout. However further work is needed for the ex-ploitation of this directional variation of the tem-perature, to extract the component temperature thatis required for more accurate energy balance esti-mation. This could be probably improved by con-current use of the directional variation in the solardomain that will provide estimates of the fraction ofsoil and vegetation cover, with possible decompo-sition into the illuminated and shadowed parts. Thespectral variation provided by the DAIS sensor wasexploited to extract the emissivity. However thepoor calibration and stability of the sensor limitedthe conclusions. Further attempts of emissivity es-timation should be conducted with the ASTER sen-sor that is now on orbit.In the µ-wave domain, classification using ERSand Radarsat images along the growth season pro-vided fair results if the speckle effect is minimisedwith a Gamma map filter. However, for a singledate, the potential of the combination of active andpassive µ-waves was clearly demonstrated althoughthe spatial resolution of passive µ-wave sensorsaboard satellite will probably constitute a severelimitation to the application of this principle. Forthe estimation of canopy and soil characteristics,one of the main problem identified was the con-founding effect of soil roughness. Several attemptswere developed to solve this problem that showspossible solutions, both by a better modelling andby the use of multiple configurations. The passiveµ-wave, by its little sensitivity to soil roughnesscould be a solution for low spatial resolution appli-cations. Physical and semi-empirical backscatteringmodels were tested and showed good performancesin the inversion mode. The use of simultaneousobservations in two different configurations appearsmandatory. Optimal configurations were discussedbased on this data set for wheat crops. However,because of the large sensitivity of µ-waves to can-opy structure, these results should be evaluated overother crops. This could be a following study basedon ReSeDA data base for sunflower and maizecrops. However, one key issue still pending and thatmay be has no definite answer outside the assimila-tion approach, is the choice of the canopy biophysi-cal variable to be retrieved.The scaling issue was investigated. It shows thatestimates of biophysical variables such as LAIcould be seriously affected in case of heterogeneouslandscapes. Flux variables such as albedo, fCoveror fAPAR, because of their linear relationship withradiometric data, are very little sensitive to hetero-

http://www.avignon.inra.fr/reseda


geneity, thus to scaling. These issues have to beinvestigated for µ-wave domain, although strongeffects are expected because of the non linear andhighly variable effect of canopy structure, at leastfor the lower frequencies.Investigation about the model inversion problemshows that techniques based on minimisation overthe biophysical variables (neural nets, vegetationindices, …) appear performing better than thosebased on the minimisation over the radiometricsignal (Look up tables, optimisation). However, thisshould be confirmed in the case of larger number ofvariables to be retrieved simultaneously, with dueattention to the computer resources required foroperational applications. The minimisation overbiophysical variables techniques have also to betested in the µ-wave domain. However, the, mainfactor to pay attention at is the amount of informa-tion actually used in the inverse problem. There-fore, better radiative transfer models that describemore accurately the physical processes, and overallthe canopy architecture and optical or dielectricproperties of the elements are required. Some ad-vanced modelling have been proposed, mainly inthe µ-wave domain. Additionally, a priori informa-tion on canopy variables and uncertainties both onmeasurements, models and variables and parame-ters have to be included in the inverse process.Ways to include this information in the range ofinverse techniques have therefore to be developed.Nevertheless, inverting radiative transfer models toretrieve canopy biophysical variables is sometimesa strong limitation in remote sensing data exploita-tion. As a matter of fact, some variables that areaccessible by remote sensing observations could notbe used directly for applications, because they arenot directly involved in the processes of interest.This is the case of the canopy and leaf water con-tent. The only way to get rid of this main limitation,is to assimilate directly remote sensing data intocoupled radiative transfer and process models.

Assimilation of remote sensing data into processmodels was therefore the main objective of the Re-SeDA project. Assimilation presents several ad-vantages over radiative transfer model inversionthat could be resumed hereafter:Assimilation allows:• to exploit our knowledge on the physical and

physiological processes;• to exploit remote sensing data provided by differ-

ent sensors in a range of spectral domains;• to exploit the multi-temporal information;• to exploit ancillary information such as climate

and soil variables;• to get a more detailed description of canopy ar-

chitecture and leaf properties that are derivedfrom the dynamic canopy functioning and SVATmodels;

• to retrieve higher level canopy and soil character-istics such as carbon (biomass), water, energy andnitrogen fluxes;

Assimilation is obviously based on process modelsand these process models have to be robust andaccurate. The first task of the assimilation part wasto focus on the evaluation and coupling of processmodels with radiative transfer models. It showsthat complex SVAT models such as SISPAT wasnot very robust when applied to the validation sites,while it performed the best on the calibration sites.Further, the required detailed description of soilcharacteristics might be too difficult to implementon an operational basis. More simple models suchas MAGRET and ISBA were generally more robustand therefore performed better. They should bepreferred for use in the assimilation process. How-ever, one difficulty that has still to be solved in thecoupling between SVAT models and radiativetransfer models is the simulation of the surfacetemperature as a function of canopy structure andthe aerodynamic temperature. The use of direc-tional observations of surface temperature mighthelp solving this problem, by providing access tothe component temperature. The mapping of en-ergy and water fluxes was investigated using theSEBAL algorithm. It shows that it provides a firstapproximation of evapotranspiration rates. How-ever, improvements are foreseen by using the abetter parameterisation of the aerodynamic resis-tance. Nevertheless it demonstrated that largevariation between fields, and even within fields areobserved. This heterogeneity of fluxes was high-lighted by the use of scintillometer measurementsas well as the 3D description of air temperature andhumidity thanks to the unmanned plane.

A range of canopy functioning models were cali-brated and evaluated over wheat crops. Canopyfunctioning models were adapted to provide a de-tailed description of canopy structure used for im-proving the accuracy of radiative transfer models. Asensitivity analysis allowed to select the variablesto adjust during the assimilation process. The vari-ables governing phenology, LAI dynamics, and topsoil characteristics were the most important ones.Concurrent assimilation of solar (Polder) and µ-wave (ERS and Radarsat) data appears to improvesignificantly estimates of biomass production, butmarginally the water balance. Closer inspectionshows that the main limitation is the capacity ofdaily time step models to accurately model proc-esses that may have significant diurnal variationsuch as top layer surface moisture. This also ex-plains the marginal contribution of µ-wave data tothe improvement of canopy functioning modelsimulations. A simpler version of assimilationscheme of ERS data was applied at the regionalscale based on MARS project segments. It clearlyshows the feasibility of the assimilation process atthis scale, although results were not very conclusivein terms of yield estimation improvement.


The ReSeDA project demonstrated that the assimi-lation of remote sensing data was possible both atthe field and regional scales. It showed some limitsof the models and indicated ways of improvement.This is certainly one of the main advances of thisproject, right in line with its main objectives. How-ever, a lot of work is still to be accomplished, bothat the extension of the technique to other crops, butalso in the refinement and improvements of theprocess models and their coupling with radiativetransfer models. This could be partly achieved overthe large and unique data set acquired within thisproject. However, this has also to be investigatedunder different conditions. Projects have alreadystarted from the momentum initiated by ReSeDA

(CROMA, WATERMED, ADAM, CHRIS-ALIDE,…). They aim both at a better exploitation of actualremote sensing data, but also at the preparation ofthe next generation of sensors, with improved capa-bility in terms of spectral, spatial, temporal anddirectional sampling. The methodological compo-nent is still very important, showing that the devel-opments are not yet enough mature to be exploiteddirectly into operational applications. However, thedistance to this application goal decreases signifi-cantly, and the user community starts to be highlyinterested by these techniques and shows it by sig-nificantly contributing to the funding of such proj-ects.

Documents

FINAL REPORT F. BARETw3.avignon.inra.fr/reseda/base/documents/reseda-report/...The ReSeDA project was originally initiated by Daniel Vidal-Madjar within the Pro-gramme National de