The technique of remote sensing image change detection,aiming at the earth’s surface changes identifying from bi-temporal images,plays an important role in those fields such as disaster assessment,environmental monitoring,and land supervision.The rapid development of remote sensing imaging and deep learning techniques provides new means for remote sensing image change detection.Focusing on the extraction and processing of change information in bi-temporal images,using depth learning methods to conduct change detection research on high-resolution remote sensing images has theoretical significance and application value.The research and related findings of this thesis are as follows:(1)An attention-guided multi-level feature fusion network for high-resolution remote sensing image change detection is investigated.Aiming at the problem that existing methods extract image features only through a single early or late fusion structure,which is prone to local information loss or insufficient spatiotemporal context information,early fusion and late fusion structures are simultaneously used to extract context features and local correlation features of bi-temporal images.In addition,in order to fully fuse the extracted bi-temporal features,the local correlation of the difference features is enhanced by extracting correlation features,local features and difference features and manipulating them with an improved attention module.Finally,high-resolution features are used as auxiliary features for decoding operations.The experimental results on the high-resolution remote sensing image change detection dataset LEVIR-CD demonstrate that the method can extract both contextual features and local correlation features of bi-temporal images,and can compensate for the loss of spatial information caused by direct upsampling.(2)A hierarchical cross-scale global feature fusion network for highresolution remote sensing image change detection is investigated.Aiming at the problem that the same semantic information has different semantic features due to the difference of imaging conditions,and CNN is lack of modeling long-term dependency,both CNN and Transformer are introduced to extract local and global contextual features of bi-temporal images.In addition,for the problem of noise interference such as unchanged information in the bi-temporal paired features,the correlation features are enhanced first before extracting the difference features.Meanwhile,for adjacent features,the attention module is used to extract the difference information from paired features after fusion.Finally,the extracted difference features are fused and sent to Transformer for decoding operation.Experimental results on the LEVIR-CD dataset show that the method is able to simultaneously improve the local correlation between differential features,and the ability to characterize contextual information,as well as the ability to model global dependencies. |