Spectral-spatial endmember extraction by singular value decomposition for AVIRIS data

Spectral Mixture Analysis (SMA) has been widely utilized for hyperspectral remote sensing image analysis and quantification to address the mixed pixel problem, in which Endmember Extraction (EE) plays an extremely important role. Distinct from the traditional EE algorithms which are only based on spectral information, a novel EE algorithm integrating spectral characteristics and spatial distribution is proposed in this paper. Purity of pixels presenting in a spatial neighborhood (SN) is examined by the Singular Value Decomposition (SVD) based on not only spectral characteristic but also spatial distribution, which effectively addresses the spectral deviation problem. Spectral deviation inside an SN is eliminated by selecting the average of the pixels in pure SNs as endmember candidates, while spectral deviation among different areas in an image is eliminated by clustering these endmember candidates. In addition, a graph theory based spatial refinement algorithm is proposed to reduce the number of endmember candidates, which can save a lot computation in the subsequent clustering step. Experimental results on AVIRIS hyperspectral data demonstrate that the proposed Spectral-spatial EE algorithm outperforms the other three popular EE algorithms, N-finder algorithm (N-FINDR), unsupervised fully constrained least squares (UFCLS) algorithm, and the automated morphological endmember extraction (AMEE) algorithm.