TOC-VEG Algorithm

Algorithm Specification

The proposed algorithm called here TOC-VEG is based on the training of neural network over a data base simulated using radiative transfer models (RTM, Figure 1). The SAIL and PROSPECT models are coupled and used to simulate the reflectance in the 11 MERIS bands considered (490 nm, 510 nm, 560 nm, 620 nm, 665 nm, 681.25 nm, 708.75 nm, 753.75 nm, 778.75 nm, 865 nm, 885 nm). 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 two MERIS first bands are also eliminated since they correspond to strong Rayleigh and aerosol scattering and to low canopy reflectances, this induces that top of canopy reflectance in this region are very sensitive to errors in the atmospheric correction. 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 46 533 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 network was trained for estimating the four variables considered. The architecture was optimized, resulting in 2 hidden layers of tangent-sigmoid neurones corresponding to a total around 300 coefficients to adjust, and providing a good ratio (50-100) with the size of the training data base.


Figure 1: Overview of the TOC-VEG algorithm


The theoretical performances were evaluated over the test simulated data set. They are close to 0.05 (absolute value) for fAPAR and fCover, and close to 25% (relative value) for LAI and LAI.Cab that shows some loss of sensitivity for the larger values of LAI and LAI.Cab due to saturation effects.