Local spatial similarity based joint-sparse regression for hyperspectral image unmixing

Abstract

Hyperspectral unmixing is indispensable for hyperspectral remote sensing technology. Exploration of spatial and spectral information helps to obtain accurate abundances. In recent years, utilizing a joint-sparse prior of the abundance matrix has earned a lot of interest. The joint-sparse-blocks structure considers the joint sparsity among column blocks (neighboring pixels) of an abundance matrix. Considering different ways of tensor expansion, the bilateral joint-sparse structure is extended to consider the joint sparsity along both horizontal and vertical directions. But it is still hard to fully utilize spectral and spatial information in hyperspectral data. In this article, we propose a local spatial similarity based joint-sparse structure by imposing the joint sparse on similar pixels in a local region, instead of only horizontally and vertically adjacent pixels. Specifically, we assume that similar pixels share identical endmembers. In this vein, we enhance the sparsity of local related pixels and have a deeper mining of the spatial information in hyperspectral data. Moreover, combined with a low-rank property, we propose a new unmixing model and derive an algorithm under the alternating direction method of multipliers framework. The validity of the proposed algorithm is verified by experiments on both synthetic and real data, compared with several state-of-the-art unmixing algorithms.