Algorithm Theoretical Basis Document for MERIS Top of Canopy Land Products (TOC_VEG)

This ATBD (Algorithm Theoretical Based Document) describes the proposed algorithm for Level 2 biophysical land products derived from MERIS top of canopy reflectance data with justification of the choices made. The biophysical land products considered are the following set of biophysical variables: LAI, fCover, CB ab B and LAI.CB ab B. The algorithm accepts as inputs the top of canopy reflectances, i.e. atmospherically corrected top of atmosphere data as derived from MERIS L1b images. It applies both to full and reduced resolution MERIS images. The proposed algorithm called here TOC_VEG is based on the training of neural networks over a data base simulated using radiative transfer models. The SAIL, PROSPECT models are coupled and used to simulate the The shortest wavelength bands, the oxygen and water absorption bands have not been used because they would convey significant uncertainties associated while providing only marginal information on the surface. The background optical properties are simulated using a collection of soil, water and snow typical reflectance spectra. A brightness factor is used to provide additional flexibility of the background reflectance. Finally, to account for the medium resolution of MERIS observations, mixed pixels are simulated with variable fractions of pure background and pure vegetation. The simulation of the top of atmosphere reflectance in the 11 MERIS bands requires 14 input variables. They were drawn randomly according to an experimental plan aiming at getting a more evenly populated space of canopy realization. To provide more robust performances of the network, the distributions of each input variable was close to the actual distributions and, when possible, realistic co-distributions were also used. This was achieved by considering a representative distribution of targets over the earth surface that constrains the observation geometry, as well as possible vegetation amount. A total number of 46533 cases were simulated. Half of this data set was used for training, one quarter to evaluate hyper-specialization, and the last quarter to quantify the theoretical performances. Back-propagation neural networks were trained for each variable considered. The architecture was optimized, resulting in 2 hidden layers of tangent-sigmoid neurones corresponding to a total around. The four variables were estimated concurrently with the same network to provide more consistency between the variables. The theoretical performances were evaluated over the test simulated data set. It allowed providing estimates of uncertainties. They are close to 0.06 (absolute value) for fAPAR and 0.08 for fCover,. For LAI, the rmse is close …