SAR Images Change Detection Based On Deep Learning

Change detection of remote sensing images is the process of identifying areas of surface change by analyzing images taken at different times in the same place.The images obtained from synthetic aperture radar(SAR)are not easy to be interfered by light,atmosphere and other conditions,and have been widely used in urban planning,land cover,disaster assessment and other fields.The deep learning technology is outstanding in the field of images because of its powerful feature extraction ability,and has attracted the attention of many researchers.Based on this,this paper has made the following research results on how to improve the accuracy of SAR images change detection:(1)A siamese discriminatory classified adversarial network is proposed to realize the task of SAR images change detection.The network consists of a siamese generative network and a discriminatory classified network,which are respectively used to get the dissimilarity probability of two time-phase samples,and to determine the class label and authenticity of the dissimilarity probability,that is,the change detection task is completed by the two networks together.Moreover,the essence of change detection is to find the differences between two time-phase samples,and the two-branch structure with the same structure and shared weights of the siamese network is conducive to obtaining the features with differences.Therefore,this idea is introduced into the design of siamese generative network,and calculate the difference features after acquiring the features of each time-phase samples,so that the network can learn the information of the changed area,reducing the false negatives,and design the difference feature extraction network based on this to obtain the dissimilarity probability between the two.In addition,the paper proposes to design the discriminatory classified network as a structure of single input and double output,that is,to add an auxiliary classifier on the basis of the original discriminator to judge the class label of the dissimilarity probability,so as to improve the accuracy of change detection.The effectiveness of this method can be verified by experiments on multiple SAR data sets.(2)A spatio-temporal fusion branch convolution stack long short term memory network is proposed to realize the task of SAR images change detection.Not only the information in the spatial dimension but also the information in the time dimension still play an importantrole in change detection in the bi-temporal images.In view of the insufficient use of time dimension information in current change detection and in order to make full use of the time dimension information and space dimension information in the bi-temporal images,the network proposed in this paper includes a convolution subnetwork and a stack long and short-term memory subnetwork,which are respectively used to extract independent and complete spatial and temporal features.In addition,considering that spatial features and temporal features have different importance to the change detection result,this paper designs a spatio-temporal feature fusion module,which sets different trainable weights for spatial and temporal features to obtain spatio-temporal features that can improve the accuracy of image change detection.Finally,a prediction classification layer can realize the classification of the spatio-temporal features.The experimental results can prove the effectiveness of the method on multiple SAR image data sets.