I. Introduction
Hyperspectral remote sensing images (HRSI) can merge the spatial and spectral information into data cubes from imaged scenes, allowing different land cover objects to be distinguished, and are widely used in geological exploration, environmental monitoring, vegetation investigation and other fields[1]. However, due to the restricted relationship between spectral resolution and spatial resolution, mixed pixel almost exists in every HRSI, and has been a key issue of hyperspectral image processing. Hyperspectral unmixing (HU) is the most effective techniques to address mixed pixel problems. HU aims to decompose spectral signature of the mixed pixel into a set of pure constituent spectra, called endmembers, and a collection of corresponding abundance fractions [2].