A Staged Approach With Structural Sparsity for Hyperspectral Unmixing

Accurate identification of pure substance and mapping the corresponding distribution are challenging because of the existence of complex distribution mixed pixels in hyperspectral image (HSI). There are numerous different methods to decompose the HSI into pure substance spectral signatures (endmembers) and corresponding fractions (abundances). However, most of them only consider local or global representations and do not fully exploit the intrinsic feature information among and inside endmember candidate domains (preliminary abundance maps), therefrom leading them to be ineffective when similar substances appear in close proximity or under noisy occlusion. In this article, we advocate a robust staged approach with structural sparsity. The proposed approach not only thoroughly exploits the intrinsic relationships among endmember distribution regions and local spatial similarities, but also maintains the spatial fidelity structure among the local regions inside each endmember distribution region. In particular, the representations among endmembers and local similarities are learned jointly. Moreover, perturbations are tackled when searching for the best endmember signature distribution region. Furthermore, by means of imposing the structural sparsity regularization on the framework of nonnegative matrix tri-factorization (NMTF), the merits of most existing sparsity constraints are accommodated and surpassed. An extensive number of qualitative and quantitative experiments are conducted on simulated and benchmark real hyperspectral data sets. The results demonstrate that the proposed method performs better than several recent classical unmixing methods.