I. Introduction
Remotely sensed spectral images show a great potential for environmental monitoring, target detection, military surveillance, especially in a large scale. However, limited spatial resolution leads to a single pixel sometimes composed of several different materials, producing a mixed pixel. Mixed pixels greatly hamper the applications of spectral images. Spectral unmixing [1], decomposing the mixed pixel into a set of pure spectral signatures (called endmembers) and their corresponding abundance fractions, is a useful tool to address this mixed pixel problem.