Research On Change Detection Of Remote Sensing Images Based On Local-global Feature Representatio

The research of remote sensing image change detection aims to use the multi-temporal remote sensing images and geospatial data taken in the same surface areas at different times to determine and analyze the changes of ground objects,including the changes of scope and state.It is an important way and means to realize surface observation.With the rapid development of remote sensing image observation technology,traditional change detection methods,such as change vector analysis,principal component analysis,iterative weighting and wavelet transformation,have been unable to meet the current task requirements.The objects processed by these methods are often single-channel grayscale images with simple background and clear boundaries.When faced with high-resolution RGB remote sensing images,they often produce serious false detection and missing detection.Benefiting from the strong ability of parameter learning,the deep learning algorithm can effectively solve the above problems,but the existing change detection methods based on deep learning still have some problems.(1)First of all,the existing methods can not explicitly distinguish changed areas from unchanged areas,leading to serious loss of edge details.To solve this problem,this paper proposes an Attentional Change Detection Network Based on Siamese U-shaped Structure,which uses multi-branch structure to explicitly model multiple discriminative semantic features of images,and uses attention mechanism to effectively fuse image features.(2)Secondly,the existing change detection algorithms are heavily dependent on convolutional neural networks(CNN).Limited by the size of the receptive field of the convolutional kernel,they are often difficult to capture the long-distance dependencies of image patches.To solve this problem,a parallel bi-branch fusion network of CNN and Transformer is proposed,which uses the bidirectional fusion mechanism of CNN and Transformer to simultaneously extract local and global features of images.Considering the large amount of calculation of self-attention in Transformer,this paper also proposes Axial Cross Attention to replace the original self-attention.(3)In addition,considering that the existing change detection methods based on Transformer often ignore the multi-scale problem of ground objects while modeling global features,a unified change detection network based on Transformer is proposed,which takes into account the multi-scale feature modeling problem while extracting local and global features through Transformer.Finally,comparative experiments and ablation experiments are carried out on multiple bitemporal remote sensing image change detection datasets to prove the effectiveness of the three network algorithms proposed in this paper.