Research On Hyperspectral Imagery Unmixing Algorithms Based On Simultaneous Sparse Represent

Satellite remote sensing imaging technology can obtain the spatial information and spectral information of image at the same time, which is the important means for mankind to get both remote sensing and deep space exploration target information. Hyperspectral remote sensing image has the very high spectral resolution, with a very important application in the geological survey, environmental monitoring, land planning, target detection and tracking and many fields. Due to the conditions of the imaging spectrometer, hyperspectral image has low spatial resolution, and with a lot of mixed pixels in it, which severely restricts the wide application of hyperspectral images. Therefore, it is necessary for the research about hyperspectral unmixing algorithms.This paper discusses the traditional methods of hyperspectral unmixing first, makes a deep analysis of the existing sparse unmixing methods both at home and aboard, and improves some efficient hyperspectral unmixing methods based on the existing ones. The main research in this paper can be summarized as follows:(1) To solve the problem that the nonexisting endmembers will have a negative effect on the estimation of the abundances corresponding to the actual endmembers, this paper presents a novel algorithm termed backtracking-based simultaneous orthogonal matching pursuit(BSOMP) for hyperspectral sparse unmixing. BSOMP uses a block-processing strategy, which divides the whole image into several blocks and picks some potential endmembers from the spectral library in each block. Then, BSOMP incorporates a backtracking process, which detects the previous chosen endmemebers' reliability and deletes the unreliable ones in each iteration. Through this modification, BSOMP can identify the true endmembers set more accurately than the other silmutaneous greedy algorithms.(2) To solve the problem that the high correlation of the spectral library remains a great challenge at finding an optimal subset of endmembers for hyperspectral image, this paper presents a novel greedy algorithm termed recursive dictionary-based simultaneous orthogonal matching pursuit(RD-SOMP). At each iteration, the spectral library is projected into the orthogonal subspace and renormalized, which can reduce the correlation of the spectral library. And then RD-SOMP selects a new endmember with the maximum correlation between the current residual and the orthogonal subspace of the spectral library. RD-SOMP can recover the optimal endmembers from the spectral library, and has a better spectral unmixing accuracy.(3) This paper presents a compound regularized multiple sparse bayesian learning(CRMSBL) algorithm for sparse unmixing. On the framework of sparse Bayesian Learning model, the parameters is established with the probability, and a regularization-based multiple sparse bayesian learning model for spectral unmixing is constructed by bayesian inference, taking the non-negativity and sum-to-one property of abundances into the convex objective function. Experimental results demonstrate that the proposed method outperforms the greedy algorithms and the convex algorithms with a better spectral unmixing accuracy.(4) To improve the sparse unmixing solution, this paper discusses a multiple sparse bayesian learning with total variation regularization(MSBL-TV) algorithm. The simultaneous sparsity and spatial information between neighbour pixels are taken into the unmixing formulation, and total variation regularization is added in the optimal problem. The penalty parameter is used to adjust the degree of spatial correlative regularization. Experimental results demonstrate MSBL-TV outperforms other algorithms with a better spectral unmixing accuracy.