Spectral-Spatial Sparse Regression For Hyperspectral Remote Sensing Imagery Unmixing

Spectral-Spatial information is the characteristic information of hyperspectral remote sensing image.Hyperspectral remote sensing technology can extract Spectral-Spatial information in a standard way,and it is widely used in ground object classification,target detection,military security and other fields.However,due to the lack of accuracy of the hyperspectral image acquisition instrument and the diversity of features in the detection area,a large number of mixed pixels exist in the acquired hyperspectral images,which will limit the information utilization rate of hyperspectral remote sensing images.As an important part of hyperspectral image information processing,solving the problem of mixed pixels is the key research direction of researchers at home and abroad.The most effective way to solve the problem of mixed pixels is unmixing to get the endmember and the abundance of each endmember.Most of the existing sparse unmixing are studied from the spectral and spatial factors of Hyperspectral images,but the sparse characterization of abundance is inadequate and the information utilization is low,which limits the understanding of the miscibility to some extent.Based on this,this paper proposes two hyperspectral image sparse miscibility models.Specifically,the research contents are as follows:（1）Spectral reweighted collaborative sparsity and total variation based hyperspectral unmixing method aimed at alleviating the lack of the sparsity of abundance in traditional methods and fully exploiting the spatial information.On the one hand,the spectral factors are utilized to estimate the weights in order to enforces the sparsity of nonzero rows,thus improving the _2,1 collaborative sparsity among all the pixels.On the other hand,the total variation based spatial regularization is employed to reinforce the smoothness within the homogenous regions,hence improving the accuracy of unmixing.The experimental results obtained from the simulated and the real datasets indicate that RCLSUn SAL-TV could significantly improve the performance of unmixing when compared to the other similar methods.（2）Spectral reweighted collaborative sparsity and graph regression for hyperspectral image unmixing.This method is presented to solve the Spatial information utillization in traditional methods.This model is based on the framework of _2,1 collaborative sparsity and graph regression.On the one hand,it introduces,the Graph Regression to obtain the high-dimensional spatial information and makes full use of the manifold structure among hyperspectral data.On the other hand,the spectral factors is used to enhance the ability to identify the endmember in the spectral library.Experimental results show that compared with other advanced sparse unmixing methods,this method canretain more spatial factorsand improve the accuracy of abundance estimation result.This thesis aimed at alleviating the lack of the sparsity of abundance in traditional methods and fully exploiting the spatial information,put forward two kinds of spares unmixing method and the reliability of the proposed algorithm has been proved by the experiment,further improve the hyperspectral remote sensing accuracy of unmixing.