Research On Hyperspectral Imagery Unmixing Algorithms Based On Non-convex Sparse Represent

The continuous improvement of hyperspectral remote sensing technology has promoted the development of earth observation,deep space exploration and so on.The application range of remote sensing images is extended to a large number of fields such as land resources allocation,mineral resources exploration,atmospheric composition analysis and marine resources exploration.In order to better and fully utilize the acquired hyperspectral data,it is usually necessary to identify and classify the objects in hyperspectral images effectively.If the problem of mixed pixels caused by low spatial resolution of hyperspectral images is not properly solved,the accuracy of recognition and classification will be seriously affected.The main idea of this paper is based on the sparsity of the abundance coefficient,using non-convex norms methods to establish a sparse unmixing model.In this paper,we have made the following attempts.?1?This paper presents a hyperspectral unmixing method based on the non-convex sparse and low-rank constraints,which takes the non-convex p-norm of the abundance matrix as the sparse constraint and the non-convex p-norm of the singular values of the abundance matrix as the low-rank constraint.Then the low-rank prior and the sparse prior are jointly utilized to construct a non-convex minimization model.By using the non-convex norm as the regular term and the Lagrange multiplier method to write the non-negative constraint term into the objective function as a quadratic penalty function,the original problem is transformed into a series of unconstrained optimization problems.Each single regular term is solved iteratively.Experimental evaluation carried out on synthetic and real hyperspectral data also show that the algorithm can improve the unmixing accuracy compared to greedy algorithm and convex optimization algorithm,and the algorithm can obtain better unmixing accuracy for the high SNR ratio hyperspectral data.?2?In this paper,we model the hyperspectral unmixing as a constrained L_2,q-L_2,poptimization problem,while the reconstruction error is induced by using L_2,q-norm penalty and the sparsity is induced by using L_2,p-norm penalty.To effectively solve the induced optimization problems for any q?1?q?2?and p?0?p?1?,an iteratively reweighted least squares algorithm called IRLS-L2qp is designed and the convergence of the proposed method is also demonstrated.Experimental evaluation carried out on synthetic and real hyperspectral data shows that the proposed method yields better spectral unmixing accuracy in both quantitative and qualitative evaluations than state-of-the-art unmixing algorithms.