A GCN-Mamba-based model for remote sensing change detection using local-global feature aggregation and dual-perspective adaptive fusion

High-resolution remote-sensing image change detection has become a crucial technical approach in urban ecological monitoring and disaster assessment. However, its practical application still faces several challenges: object-scale variation, blurred boundaries in change regions, the difficulty of capturing subtle changes, and temporal discrepancies across large time spans. These factors increase the complexity of change detection, making it difficult for existing models to achieve optimal feature learning and detection accuracy in real-world scenarios. To address these issues, we propose a GCN-Mamba-based model for high-resolution remote sensing change detection. A local and global feature aggregation (LGFA) module is designed to address the omissions and representational inconsistencies caused by object-scale variations in high-resolution imagery, fusing local edge details with global semantic structures extracted by Mamba to preserve multi-scale features. Then, a sequential enhancement module (SEM) is introduced, which improves temporal-spatial continuity and enhances the detection of subtle changes and boundary integrity. A dual-perspective adaptive gated multi-feature fusion (DPAGMF) module is proposed to reduce interference from long-term temporal discrepancies. This module utilizes the acquired multi-scale boundary-enhanced features to capture synergistic relationships via dual-perspective modeling, improving robustness under complex spatiotemporal conditions. The proposed method achieves F1 gains of 1.20% and 0.53% on WHU-CD and LEVIR-CD, along with a 0.27% increase on WCD, demonstrating strong generalization capability. Overall, the proposed method demonstrates excellent performance in short-term change detection scenarios, but faces challenges in long-term cases influenced by significant environmental differences. Future work will focus on improving the collaborative modeling of bi-temporal features across extended time spans.