Change Detection In PolSAR Images Based On Semantic Analysis

Polarimetric Synthetic Aperture Radar(PolSAR)measures the polarization scattering characteristics of the target by transmitting and receiving polarized waves in different combinations.Therefore,compared with SAR radar,it has richer ground object information and comprehensive polarization information,so it is valued by countries all over the world.Pol SAR image change detection is one of the core content of Pol SAR image interpretation.It locates and analyzes the multi-temporal Pol SAR image data of the same geographic location to determine the change information of the area.It plays a huge role in monitoring deforestation,disaster assessment,and urban planning.Most traditional Pol SAR image change detection methods do not introduce semantic information in the change detection process,so they can only detect whether a certain area has changed,and cannot determine the target category before and after the change.But for some specific application scenarios,it is not only necessary to detect where the change has occurred,but also to identify the target type before and after the change.Therefore,this paper studies the polarization SAR change detection based on semantic analysis.The main research contents are as follows:1.In view of the traditional Pol SAR change detection method without considering semantic information,it is impossible to detect the target category before and after the change and the problem of poor localization of the change edge.A method of semantic change detection of Pol SAR images(abbreviated as CNN-MRF algorithm)based on DS evidence theory fusion is proposed.The algorithm constructs a convolutional neural network(CNN)model by defining semantic change classes,and uses it to extract semantic change features from two-phase Pol SAR image data,and obtains change evidence with semantic information through the classifier.Then,the difference map is generated by the large HLT operator on both sides,and then the difference map is modeled using Markov Random Field(MRF)theory to obtain evidence of better changes in detail preservation and noise suppression.Finally,through the DS evidence,the two change evidences are combined to complete the semantic change detection.The effectiveness and feasibility of this algorithm are proved through simulation experiments and result analysis of 4 sets of real Pol SAR data.2.Aiming at the problem of insufficient extraction of semantic change features in the process of semantic-level Pol SAR image change detection,a polarization feature and Inception structure improved residual network based Pol SAR semantic change detection method(referred to as PF-Inception-Resnet algorithm)is proposed.The algorithm uses Pauli decomposition,Cloude decomposition and Yamaguchi decomposition to decompose the polarization target of the two-phase Pol SAR image,and obtain the polarization feature map of the two-phase.These characteristics can reflect the physical characteristics of the ground object from different aspects.Then the polarization feature maps of the two-phase polarization SAR image are feature-cascaded and sent to the residual convolutional neural network improved with the Inception structure.Carrying out deep-level semantic change feature extraction,you can get richer semantic change information.Finally,the semantic change detection result is obtained through the classifier.The effectiveness and feasibility of the algorithm is proved by simulation experiments and result analysis on 4 sets of real polarized SAR data.