Derivation of synthetic endmembers for linear unmixing to improve parameter estimation for soil erosion modelling in agricultural ecosystems.

Abstract

Physically based erosion models have been widely accepted in soil erosion risk assessment in the past years but their application is still restricted due to a large disparity between required model input parameters and data availability. Hyperspectral imaging may offer an alternative to solve that problem providing high spatial and spectral information about earth surface materials. A major problem in image interpretation is the heterogeneous composition of each individual pixel, in the case of agricultural ecosystems a mixture of soil and vegetation signal, but also moisture conditions and illumination have a strong influence. This issue has been addressed with unmixing techniques (vii) which take individual pixel components, endmembers, into account. In this paper, we present a new algorithm which allows the derivation of synthetic spectral endmembers (SY-SEM) of soil. A systematic analysis of lab spectra from more than 100 soil samples measured at different soil moisture and illumination angles revealed stable features independent of soil moisture and illumination in SWIR I and II. Results were confirmed by an analysis of corresponding field spectra. Starting from these stable wavelength regions absolute reflectance values of all spectra were analysed adopting a linear regression analysis. Results showed coefficients of determination (R²) generally above 0.8, in the range of 1000 to 2500nm above 0.95. Based on these results, SY-SEM are derived from self-generated look-up tables containing regression parameters, and atmospherically corrected image data. In contrast to existing approaches, SY-SEM procedure is computed pixelwise over a pre-classified image data set.