Research On Change Detection Of Remote Sensing Images Based On Weighted Nonnegative Matrix Factorization

Change detection of remote sensing(RS)images aim to obtain the change information of target region by comparing and analyzing bitemporal or multitemporal RS images,acquired over the same geographical area,but at different times.With the gradual improvement of the RS images quality and development of images processing technology,change detection has become an important research direction in the field of RS.Many nonnegative matrix factorization(NMF)based algorithms have been successfully applied to change detection of the RS images due to its simple implementation,small storage space,and strong interpretability of the factorized results.In this thesis,the weighted NMF(WNMF)algorithm is applied to change detection of the RS images.The specific research is stated as follows.(1)We first discuss the feasibility of the WNMF algorithm for change detection of the RS images.In order to effectively explore the change information between two RS images,WNMF is applied to extract features from difference image.A new weight matrix is designed by using difference values of the corresponding pixels based on radial basis function(RBF),for which the obvious changed pixels can possess the larger weights,and vice versa,thus enlarging the separability of the changed pixels and the unchanged pixels.In addition,sparsity constraint based on L1/2 regularization is introduced into WNMF to learn more effective sparse features.Experimental results on four pairs of RS images show that the proposed algorithm performs better.(2)In order to further learn the discriminative features,we propose a change detection algorithm based on discriminative weighted NMF(DWNMF).In DWNMF,by combining the learning feature and classification,the classification error is embedded into the defined objective function,and the labeled samples generated by the pre-classification algorithm are used to learn the discriminative features,thus further improving the detection performance.Finally,the experimental results on four pairs of RS images show that the proposed algorithm can obtain better detection performance than the WNMF algorithm.