Hyperspectral Unmixing Via Approximate Sparsity And Total Variation Regularization

Mixed pixels are often encountered in hyperspectral remote sensing image due to the resolution limitation of the sensors and the variability of the ground surface. The existence of mixed pixels affect object recognition and classification accuracy, and increase inversion error of surface parameters. The spectral unmixing technique is the most efficient way to solve the mixed pixel, has been the research hotspot of hyperspectral remote sensing. Sparse representation, can capture data with a small amount of coefficient of main information and the intrinsic geometric structure, is applied to hyperspectral unmixing and obtained good results. In this thesis, we taking into account the sparsity of abundance and spatial information, a approximate sparsity and total variation spatial regularization sparse unmixing method has been proposed for spectral unmixing. The main research contents are:Using the known spectral library to replace the original hyperspectral data as input, and approximate LO norm are utilized to manifest the sparseness, then we design a mixed pixel decomposition algorithm by establishing a cost function for the sparse mixed pixel decomposition problem based on approximate LO norm. By using variable splitting and the generalized Lagrange algorithm, a numerical method for solving the approximate LO norm sparse unmixing model of hyperspectral remote sensing image solution, namely APSSU algorithm. The proposed algorithm have stronger sparsity than the SUnSAL algorithm based on the L1 norm. At the same time, Experimental results on both simulated and real hyperspectral data show that the APSSU algorithm can better estimate the endmember abundance values, and obtain better results.The total variation model is introduced into the sparsity regression of unmixing problem, a novel approximation LO norm and total variation regularization sparse unmixing method(APSSU-TV) has been proposed, exploiting the spatial-contextual information present in the hyperspectral image, and improve the unmixing accuracy. The approximate sparse as objective function of sparsity, and TV regularization as the smoothness of the objective function. Experimental results on both simulated and real hyperspectral data show that the proposed APSSU-TV algorithm than SUnSAL algorithm and SUnSAL-TV algorithm for the decomposition of mixed pixels.