Sparse unmixing analysis for hyperspectral imagery of space objects

Spectral unmixing analysis for hyperspectral images aims at estimating the pure constituent materials (called endmembers) in each mixed pixel and their corresponding fractional abundances. In this article, we use a semi-supervised approach based on a large spectral database. It aims at finding the optimal subset of spectral signatures in a large spectral library that can best model each mixed pixel in the scene and computes the fractional abundance which every spectral signal corresponds to. We use l(2)-l(1) sparse regression technical which has the advantage of being convex. Then we adopt split Bregman iteration algorithm to solve the problem. It converges quickly and the value of regularization parameter could remain constant during iterations. Our experiments use simulated pure and mixed pixel hyperspectral images of Hubble Space Telescope. The endmembers selected in the solution are the real materials' spectrums in the simulated data and the approximations of their corresponding fractional abundances are close to the true situation. The results indicate the algorithm works well.