VEGETATION WATER AND DRY-MATTER CONTENTS ESTIMATED FROM TOP-OF-THE-ATMOSPHERE REFLECTANCE DATA - A SIMULATION STUDY

Leaf (PROSPECT), soil, canopy (SAIL), and atmosphere (6S) models were coupled and used to create a large set of simulated reflectance spectr a with corresponding water content per unit leaf area (C-w) dry matter content per unit leaf area (SLW; i.e., the specific leaf weight), lea f area index (LAI), whole canopy water content (C-w,LAI), and whole ca nopy dry matter col?rent (SLW.LAI). Multiple linear regression was use d to estimate these canopy variables from the simulated satellite refl ectance spectra within the 880-2380-nm domain. Our data set was subdiv ided into calibration and validation subsets to evaluate the predictiv e power of the relations. Canopy-level variables (C-w.LAI,SLW.LAI,LAI) were retrieved with a good accuracy, whereas leaf-level variables (C- w,SLW) were less accurately retrieved. The radiometric resolution of t he simulated sensor greatly affected the accuracy of the estimation. C onversely the spectral resolution between 10 and 20 nm was not critica l, the largest spectral resolution providing the most accurate estimat es because it smoothed the instrument noise. We used multiple linear r egression to select between five and eight wave bands for each canopy variable. Several wave bands selected were common to different canopy variables. Therefore, a set of ten wavebands centered on about 890, 10 80, 1210, 1290, 1535, 1705, 2035, 2205, 2260, and 2295 nm efficiently allowed reasonable estimates of the variables investigated with varyin g coefficients for each of the canopy variables. For each variable, ne ural networks were trained over the wave bands selected by the multipl e regression. Results showed better performances than classical multip le linear regression. Shifting the on wave bands by 10 or 20 nm when c alibrating and testing the networks slightly decreased the accuracy of the estimation. The difference was more pronounced for C-w, and SLW. Conversely, When equations were generated with the use of the wave ban ds at their optimal position and validated by using wave bands shifted by 10 or 20 nm, the accuracy of estimation for all variables except L AI was low. These results are discussed with emphasis on the design of future sensors. (C) Elsevier Science Inc, 1997.