Classification using unmixing models in areas with substantial endmember variability

Abstract

Hyperspectral imagery (HSI) can significantly contribute to habitat monitoring which is essential to ecosystem management. Specifically, spectral unmixing in HSI is an important tool for habitat conservation status assessment. However, an issue found is the spectral variability of the endmembers that can be addressed by the Extended Linear Mixing Model (ELMM), which considers this spectral variability for obtaining accurate abundance maps. An analysis is performed in a mountainous ecosystem with high spectral variations. Classification maps are obtained from the abundance maps to assess the unmixing models. Results are very satisfactory, plausible abundances are estimated with accurate characterization of variability within the scene and overall accuracies over 83%are obtained. ELMM and Robust ELMM allow studying the features of each pixel, including additional information about characterization of mixed pixels, by considering spectral variability and being more robust to the absence ofpure pixels as well as to noise.