DFSMCG-Net: A Siamese change detection network based on Differential Feature Selection and Multi-Scale Guidance Strategies

Change detection technology effectively identifies surface changes but encounters significant challenges, including class imbalance between foreground and background and interference from pseudo-changes caused by factors such as illumination variations and geometric distortions. We propose a location-sensitive Differential Feature Selection and Multi-Scale Change Feature Guidance Network (DFSMCG-Net) to address these issues. The DFSMCG-Net introduces a Differential Feature Selection Module (DFSM) that leverages the spatial location information of bi-temporal features. This module captures spatiotemporal differential features at the exact location along the X-axis and Y-axis and integrates these features through cross-fusion to establish long-range pixel dependencies. The resulting multi-level differential features provide the network with a detailed temporal context for detecting changes. We develop a Multi-Scale Change Feature Guidance Module (MCFGM) based on a multi-head self-attention mechanism to further enhance the fusion of multi-level differential features and suppress interference from non-differential features. This module assigns each attention head a distinct non-overlapping window, dynamically adjusting window sizes according to the feature map dimensions. This approach facilitates the integration of multi-scale differential features, improving the network’s capacity to represent change-related features. Experimental results demonstrate that the proposed DFSMCG-Net performs significantly better than state-of-the-art methods on benchmark datasets, including LEVIR-CD, CDD, SYSU-CD and S2Looking. The model is particularly effective in mitigating pseudo-change phenomena under conditions of extreme class imbalance.