Spatial-spectral Hyperspectral Sparse Unmixing Method Based On Spectral Library

Recently, the development of hyperspectral remote sensing technology is remarkable. Due to the high spectral resolution, it has been widely used in mineral identification, environmental monitoring, vegetation surveys, and other fields.However, because of the low spatial resolution of remote sensor, and the complexity and diversity of the natural surface features, mixed pixels usually exist in the scenes,which significantly affect the accuracy of target identification and classification.Therefore, how to solve the problem of mixed pixels has become one of the key points to hyperspectral image analysis and quantitative applications.In this paper, based on the linear mixed model and having a huge spectral library in advance, we mainly focus on studying the correlation between neighboring pixels and sparsity of abundance coefficients to further improve the accuracy of hyperspectral unmixing. This paper mainly researches the following two aspects:(1) A novel linear hyperspectral unmixing method based on collaborative sparsity and total variation (TV) regularization is proposed. This method is a sparse linear mixed model based on spectral library, and the total variation is introduced to impose a constraint on the correlation between neighboring pixels. Then, the collaborative sparsity is also imposed to depict the row-sparse characteristic of the fractional abundances. At last, the proposed model is solved by the alternating direction method of multipliers (ADMM).(2) A novel linear hyperspectral unmixing method based on the, l1-L2(L1 norm minus L2 norm) sparsity and total variation (TV) regularization is proposed. This method utilizes the l1-l2 norm to promote sparser results than the l1 norm, which is to further weaken the influence of high mutual coherence of the spectral library.Moreover, total variation is minimized to enforce the spatial smoothness by considering the spatial correlation between neighboring pixels. This method combines l1-l2 sparsity and TV regularization simultaneously, and could be easily solved by the alternating direction method of multipliers (ADMM).Experimental results on simulated hyperspectral data show that the performance of proposed methods in this paper are better than other state-of-the-art algorithms,and the experimental results on real hyperspectral data also verify the effectiveness of our algorithms.