Simulation of a Blind Hyperspectral-Unmixing Algorithm Incorporating Spatial Correlation and Spectral Similarity

For hyperspectral unmixing, a multi-scale spatial regularization method based on a modified image segmentation algorithm to generate super-pixels is proposed in which the super-pixels are used to extract contextual information from spatial correlations and spectral similarity in hyperspectral images (HSIs). The unmixing problem is decomposed into two simple unmixing subproblems regarding the approximate super-pixels and the original pixels. The unmixing results of these two subproblems have spatial-correlation constraints. Introducing a novel regularization term to constrain the abundance matrix to promote the homogeneous abundances helps in making effective use of the spatial correlations and spectral similarity of the abundances from HSIs. Experimental results obtained from synthetic data demonstrate that the proposed algorithm yields an accuracy greater than other conventional methods.