Research On Hyperspectral Imagery Unmixing Methods Based On Simultaneous Greedy Algorithm

Compared with multispectral imagery and panchromatic imagery,hyperspectral imagery not only obtains the spectral data of each pixel in the imagery,but also obtains the spatial information of any spectral segment.Therefore,it can analyze and recognize the ground objects more effectively,and has broad application prospects in civil and military fields such as geological exploration,marine resources survey,battlefield environment detection and so on.However,the existence of mixed pixels is a major problem in the application of hyperspectral remote sensing.Spectral mixture analysis has attracted the attention of many scholars,and many theories and methods have been developed.Greedy unmixing algorithm searches the endmember set for mixed pixel iteratively,and then estimates the abundance coefficient by least square method.This kind of algorithms has the advantages of high reconstruction accuracy and fast reconstruction speed.The main contents of this thesis are as follows:1.In order to solve the problem that the single path search algorithm could be easy to be trapped into the local optimal solutions,a novel semi-greedy algorithm termed A* simultaneous orthogonal matching pursuit(A*SOMP)is proposed in this thesis.Firstly,the hyperspectral imagery is divided into blocks.Then for each block,in the iteration process of endmember extraction,A* search method will create a search tree whose nodes represent the endmembers in the spectral library.And then,A*search method,following the best-first search principle,uses cost function to evaluate all paths on the search tree to choose the best path and extend it.A*SOMP algorithm also incorporates pruning and path substitution techniques for a complexity-accuracy trade-off.When the cut-off condition is achieved,the algorithm stops iterating and uses the endmemers on the minimal cost path to construct endmember set.Finally,the endmembers picked in each block are associated as the endmember set of the whole hyperspectral imagery,and then the abundances are estimated by the least squares method with the obtained endmember set.Simulated and real imagery experiments are carried out for A*SOMP algorithm respectively.The performance of A*SOMP algorithm are compared with a variety of mainstream algorithms.The results of experiments show that A*SOMP algorithm has better accuracy in spectral unmixing,compared with the other unmixing algorithms..2.According to the analyses of the correlation of adjacent pixels in the hyperspectral imagery,a spatial correlation constrained simultaneous subspace pursuit(SCCSSP)algorithm is proposed.The algorithm uses a block-processing strategy to divide the whole hyperspectral imagery into severalblocks.In each block,in order to ensure that the estimated endmember set is optimal to the current hyperspectral image residuals,the algorithm first chooses the appropriate endmember to join the endmember set through preliminary test,then detects the existing endmembers in the preliminary set according to the minimum residuals criterion combined with spatial correlation constraint,and deletes the redundant endmember finally.The endmembers picked in each block are associated as the endmember set of the whole hyperspectral imagery.Finally,the abundances are estimated by the nonnegative least squares method with the obtained endmember set.Experimental results on both simulated and real imagery demonstrate that,compared with the other unmixing algorithms,the proposed algorithm selects the endmember set for the hyperspectral data more accurately,and has better spectral unmixing accuracy and capability of noise resistance.