CTST: CNN and Transformer-Based Spatio-Temporally Synchronized Network for Remote Sensing Change Detection

Abstract

Remote sensing change detection has achieved amazing results in recent years, especially the application of convolutional neural networks (CNN) and Transformer networks, which have revolutionized the field. However, the complex ground cover changes and the differences in lighting conditions caused by different times still pose challenges to the detection accuracy. In order to further extract spatial feature information and suppress irrelevant influences, we innovatively propose an edge-enhanced and time-synchronized remote sensing change detection network, called CNN and transformer-based spatio-temporally synchronized network (CTST). CTST designs a unique feature-integrated coding model with CNN and Transformer architectures, which enhances the model's understanding of the global dependencies and the extraction effect of the local features through the dynamic weight allocation method. We designed the edge salient feature enhancement module, which uses a dual operator fusion structure to combine the edge semantic information with the depth feature information, greatly enhancing the model's ability to recognize the edges of important terrain and features in remote sensing images. In addition, the spatio-temporally synchronized module is used to fuse the difference and superposition relationships between bitemporal features, and an innovative correlation mapping weighting algorithm is proposed to evaluate the similarity and difference of the fused features. Finally, the feature decoding complementary module is proposed to combine and complement features at different scales to further refine the already fused bichronological remote sensing features. The network results are optimized by the deep supervision (DS) strategy, which ensures the model's high efficiency and accuracy. CTST outperforms mainstream and state-of-the-art methods on all three datasets, with an F1 of 92.08% and an IoU of 85.33 on the LEVIR-CD dataset, an F1 of 93.25% and an IoU of 87.36% on the WHU-CD dataset, and an F1 of 93.25% and an IoU of 87.36% on the GZ-CD dataset. CD dataset the F1 is 85.95% and IoU is 75.37%, Param is 31.87 M, and FLOPS is 29.58 G.

Keywords

• Attention
• convolutional neural networks (CNN) and transformer
• high-resolution remote sensing (RS) images