Integration of spatial-spectral information for endmember extraction in hyperspectral images

Endmember extraction for spectral mixture analysis is a key step when endmember information is unknown. Under the linear mixture model assumption, the observation pixels form a simplex whose vertices correspond to the endmembers. However, simplex-based endmember extraction methods identify endmembers by accounting for the spectral similarity of pixels only, which are susceptible to noise and outliers. In this paper, an improved method is proposed to integrate both spatial context a n d spectral similarity for endmember extraction. The proposed method adopts a sequential manner to find endmember one by one. At first, the initial endmember candidates are identified by integrating spectral similarity constraint and spatial constraint into the orthogonal subspace projection process. Then, the best representative endmember is determined in terms of simplex volume maximization criterion. Experimental results, which were obtained using both synthetic and real hyperspectral data sets, indicate that the proposed method can obtain more realistic endmembers and can be used to accurately model the original hyperspectral scene using a linear mixture model.