Change Detection For Hyperspectral Images Based On CVA And Spectral Unmixing

Remote sensing image detect change(CD)has been one of the hot topics in the research of image processing.With the development of remote sensing,hyperspectral image sequence can be obtained for the same areas,and the researches on CD for hyperspectral image are being attracted attention gradually.Nowadays,CD techniques are mainly developed for multi-spectral images.The researches on hyperspectral image CD technology are just started and the results are limited.Therefore,the study of hyperspectral image CD has certain theoretical and practical significance.Due to the limitation of spatial resolution,mixed pixels are common in hyperspectral images,which lead to the traditional CD method of hard classification comparison could not be directly used for hyper-spectral image CD.Therefore,it is necessary to solve the problem of mixed pixel decomposition.In addition,change vector analysis(CVA)is a kind of multi-class CD method,which use all band data of multi/hyper spectral images,and the detection performance of CVA is better than that of difference method and ratio method.This paper studies the hyperspectral image change detection algorithm based on the spectral unmixing and CVA.The main work of this paper includes:(1)A new method is presented to solve the problem of multiple changes detection in multi-temporal hyper-spectral images based on change vector analysis and spectral unmixing technology(CVA_SU).Firstly,the Change Vector Analysis is designed to distinguish the changes and unchanged pixels.The threshold is analyzed by using the EM algorithm in the framework of the Bayesian decision theory.Then,endmembers are extracted from the two multi-temporal hyper-spectral images respectively.The abundances are solved for each changed pixels in both two images.Taking the correlation coefficient as similarity criterion,the endmembers for the two images are matched with the threshold of similarity adaptively determined according to the fine degree of the ground classes in the images,and the category code of each endmember is assigned.Finally,the considered multiple-change detection problem is solved by comparing the classes of each pixel pairs in the changed region after assigning the class of each pixel according to the maximum abundances.(2)In order to solve the problem that the hyperspectral image pixel-level change detection algorithm could not simply determine the changes in the composition of the problem,a new method is proposed based on spectral unmixing using stacked data to detect the change composition of each object.A set of extended endmembers are constructed by all changed endmember pairs and non-changed endmember pairs,and the abundance are calculated using the stacked images.The abundance corresponding to a changed extending endmember just describe the change composition of the change type in each pixels.(3)In order to further realize the sub-pixel level change detection of hyperspectral image,we propose a sub-pixel mapping method combined with CVA method.Based on spectral unmixing and objective optimization;the minimum perimeter of image connected area as the objective function and binary particle swarm optimization algorithm.By comparing the sub-pixel location of each pixel pairs in the change region obtained by CVA,the sub-pixel level change detection result is obtained.