SAR Image Change Detection Method Based On Semantic Enhancement

Remote sensing image change detection is a technology that quantitatively analyzes and determines the process of surface change characteristics from multi-temporal remote sensing images of the same geographic location.It is widely used in many fields,such as land cover monitoring,agricultural resource exploration,military target dynamic monitoring and so on.It is of great significance to study the interaction between humans and the natural environment.Synthetic aperture radar(SAR)technology has the advantages of high resolution and all-weather operation.Therefore,SAR image has become an important data source in the field of change detection.The traditional SAR image change detection uses the difference image to locate the initial change information.However,the difference operation will cause serious loss of remote sensing semantic information,and the change boundary is difficult to define.For the purpose of remote sensing semantic enhancement,this paper introduces UNet and conditional random field(CRF)model,designs a Neural-CRF sequential iterative structure based on global feature association,and designs a DR-UNet-CRF fusion iterative structure based on dynamic feature changes.The method proposed in this paper effectively compensates the semantic information of remote sensing changes and improves the boundary localization ability of the model.The work done in this paper is as follows:1.A SAR image change detection method based on Neural-CRF sequential iterative structure is proposed.Firstly,TR-UNet model is designed by using the multi-head attention mechanism in Transformer to provide unary potential function for CRF.The TR-Attention module is embedded between the encoder and decoder of UNet to capture global feature associations from spatial and channel dimensions respectively.The semantic enhancement is achieved by fusing the features of two dimensions,which improves the semantic perception ability of UNet.Secondly,the CRF model with sequential iterative structure is designed.The co-iteration of unary and binary potential functions expands the search range of the energy function and improves the pixel-level label refinement capability of the CRF model.2.A SAR image change detection method based on DR-UNet-CRF fusion iterative structure is proposed.Firstly,DR-UNet model is designed by using dynamic region-aware convolution(DRConv)to provide unary potential function for CRF.Among them,the multiscale fusion guided mask generation method based on feature pyramid network(FPN)structure improves the ability of DRConv to process the variable distribution of spatial semantics,and further improves the generalization ability of UNet to the modal perception of difference image.Secondly,the CRF model with fusion iterative structure is designed.By extracting the boundary priori of the multi-modal difference image,a multi-modal fusion binary potential function based on local boundary entropy is constructed.This not only compensates the semantic loss of the difference image with information fusion,but also improves the boundary awareness of the CRF model.The simulation results show that the change detection results generated by the method in this paper have complete change areas and clear boundaries.Compared with other change detection methods,it can obtain better detection performance and has certain application value.