The objective of this study is to develop an algorithm for vegetation monitoring using neural networks. It consists in estimating canopy biophysical variables from the VEGETATION sensor reflectance data : gap fraction for diff erent view angles, leaf area index, fraction of absorbed photosynthetically active radiation, mean leaf inclination angle and integrated chlorophyll content. A synthetic top of canopy BRDF catalog of homogeneous canopies was built thanks to well known radiative transfer models. This data base was used to train neural networks to estimate the biophysical variables from the solar zenith angle and the associated idirectional reflectances collected by the sensor during its orbit cycle. We demonstrated that variables such as gap fractions or fAPAR were less sensitive to heterogeneous pixels than variables such as LAI. The validation of the algorithm on experimental data sets showed the robustness of neural network when few reflectance data are acquired (cloud occurrence or low latitudes). Neural nets thus provide gap fraction estimates within an accuracy better than 0.1. The interest of using neural network is demonstrated, especially when compared to vegetation indices based methods. Moreover, it appears that estimates of the biophysical variables depend not only on the representativity of the canopy and associated viewing conditions in the learning data base, but also on the basic assumption made within the radiative transfer model used to generate it. We additionally investigated the sensitivity of the algorithm to noisy data.
