Change Detection In SAR Images Based On Spatial Constraint Semi-nonnegative Matrix Factorization

With the rapid development of remote sensing and computer technology,synthetic aperture radar(SAR)image change detection has become a new direction after intersection of the two disciplines.It refers to analyze bitemporal or multitemporal SAR images acquired at different time of the same region,and obtain change information of the objects,scenes or targets of interest.However,with inherent speckle noise and strong correlation between pixels come a challenge to current algorithms.Therefore,it is the most important issue at present that how to detect change information quickly and accurately from SAR images.Therefore,research works in this thesis focus on SAR image change detection based on spatial constraint semi-nonnegative matrix factorization(semi-NMF)and deep learning.The proposed methods in this thesis fully consider the characteristics of SAR images and provide good performance for change detection.It is stated as following two aspects:1.Motivated by researches of semi-NMF in SAR image change detection,in order to consider spatial contextual information as well as handle speckle noise and outliers,correntropy induced metric(CIM)and total variation(TV)are employed to build a robust semi-NMF with TV model.Firstly,the difference image is produced by bitemporal or multitemporal SAR images,acquired at different instances in the same region.Secondly,the feature matrix of difference image is obtained by using principal component analysis(PCA).Then base matrix and coefficient matrix are alternatively updated to learn the proposed model of the robust semi-NMF with TV.Finally,the coefficient matrix is classified into two classes by maximum criterion.In this thesis,experiments of three real SAR data sets are assessed feasibility and effectiveness of the proposed method.2.In order to extract potential hierarchical features of SAR images,another convolutional auto-encoder(CAE)and TV algorithm is proposed,which is combined with deep learning and taken into account spatial information.After generation of the difference image,CAE is utilized to obtain hierarchical feature matrix,and it is factorized into base and coefficient matrices by using semi-NMF with TV model.Finally,maximum criterion classifies coefficient matrix into changed class and unchanged class.Compared with the experimental results of first model,it's illustrated that the deep feature extracted from CAE can effectively improve accuracy of the change detection with three same SAR data sets,at the same time the feasibility and effectiveness of the algorithm are verified.