I. Introduction
Hyperspectral imagery measures reflected light (i.e., radiance) at different wavelengths for each pixel of the image. The radiance is converted into a reflectance spectrum in order to remove the effects of atmospheric gases or water absorption [1]. The reflectance spectrum can be modelled by a linear combination of reference spectra representing the surface materials [2]. A linear mixing model (LMM) or a nonlinear mixing model (NLMM) has been developed to model the relationship between reference spectra (endmembers) and observed reflectance spectra [3]. LMM enables spatial distributions of materials to be quantitatively estimated. While many mixing models have been used for a variety of applications reviewed by [3], there are still two major problems.