Semi-blind source separation for the estimation of the clay content over semi-vegetated areas using VNIR/SWIR hyperspectral airborne data

Visible, near-infrared and short wave infrared (VNIR/SWIR) hyperspectral imagery has proven to be a useful technique for mapping the soil surface properties over bare soils pixels. Multivariate regression models are usually built linking a set of soil surface properties (response Y-variables) to a set of imaging reflectance spectra over bare soil pixels (predictor X-variables), and then, they are applied to all bare soil pixels to map the soil surface properties. The applicability of VNIR/SWIR hyperspectral imagery for soil properties mapping decreases when surfaces are partially covered by vegetation. The objective of this research was to develop a “Double-Extraction” approach for clay content estimation over semi-vegetated surfaces and to evaluate its performance using VNIR/SWIR HyMap airborne data acquired in a Mediterranean region over an area of 24km2. The “Double-Extraction” approach consists of 1) an extraction of a soil reflectance spectrum, s^soil, using a semi-blind source separation (SBSS) technique applied to couples of semi-vegetated spectra and 2) an extraction of clay content from the soil reflectance spectrum s^soil using a multivariate regression method. The source separation approach is semi-blind due to the use of available knowledge about expected soil and vegetation spectra. The multiplicative algorithm of Lee & Seung, belonging to the family of non-negative matrix factorization (NMF) methods, is used to solve the blind source separation (BSS) problem. The multivariate regression method used in this study is the partial least squares regression (PLSR) method. The “Double-Extraction” approach was compared to a “Direct” approach consisting of the application of the multivariate regression model built from bare soil spectra over the semi-vegetated area.Our results showed poor prediction performances for both approaches when applied to all pixels; however, a slight improvement was observed when correcting the bias prediction that occurs when using the PLSR model. Conversely, satisfactory prediction performances were obtained by restricting the prediction to the weakly vegetated area (NDVI <0.55) that covered 63% of the study area. The resulting clay map over this restricted vegetated area exhibited patterns of variations that matched the previous expertise acquired on the spatial structures of soils in this area.