Applying blind source separation on hyperspectral data for clay content estimation over partially vegetated surfaces

Equipe Organisation Spatiale et Fonctionnement des Paysages Cultivés

International audience

Английский. Hyperspectral imagery has proven to be a useful technique for mapping soil surface properties. However vegetation cover has a significant influence on spectral reflectance and the applicability of hyperspectral images for soil property estimations decreases when surfaces are partially covered by vegetation. To maximize information extraction from hyperspectral data, we apply a "double-extraction" technique: 1) extraction of a soil reflectance spectrum s, using blind source separation (BSS) techniques from mixed hyperspectral spectra without any information about the proportion of the components in the mixture nor the original spectra that composed the mixed spectra and 2) extraction of soil property contents from the soil reflectance spectrum s by classical chemometric methods. The Infomax algorithm is used as the BSS algorithm for this approach, and the chemometric method is the partial least squares regression (PLSR). The estimated soil property after soil signals extraction is the clay content, and the hyperspectral datasets are from Hymap airborne data. First, experiments were performed using simulated linear spectral mixtures of one soil spectrum and one vegetation spectrum (vineyards). Second, the "double-extraction" method was applied to grids of 3 x 3 Hymap mixed spectra, which were centered on surfaces partially covered by vineyards. Our simulated experiments and applications to Hymap data show that the BSS concept provides accurate soil reflectance spectra for clay content estimation. The clay content estimations are accurate compared to physico-chemical values (the mean error of estimation is always inferior to 50 g/kg in simulated experiments and predominantly inferior to 90 g/kg in Hymap mixed pixels treatments). We conclude that the "double-extraction" method, which requires no a priori information is a promising method for soil property prediction using hyperspectral imagery over partially vegetated surfaces.