SMNet: A Semantic Guided Mamba Network for Remote Sensing Change Detection

In remote sensing, a significant field of research is change detection (CD) that aims at identifying semantic contrast from different images captured in identical location under various time constraints. As deep learning continues to advance, CNN and Transformer have respectively strong local modeling and global feature extraction capabilities, making great progress in the CD field of remote sensing imagery. Nonetheless, due to inherent receptive field limitations of the structure, current methods often provide limited CD performance and insufficient feature expression ability. To address the above issues, a remote sensing change detection model named SMNet was designed, effectively integrating multi-level feature representations for CD tasks. Particularly, the strengths of Mamba and RWKV in capturing far-reaching dependencies are employed in the proposed framework. Firstly, SMNet utilizes a novel Learnable Visual State Space (L-VSS) block, which dynamically adjusts the adjustment factors obtained from temporal semantics and mask semantics through the use of learnable adapters, enabling the structure to fully utilize complex spatiotemporal details to intensify the perception capability for detecting semantic variations. Furthermore, to efficiently capture global dependencies, a spatio-temporal transfer operation and a multi-directional WKV (Mi-WKV) attention mechanism are utilized in RWKV. Ultimately, the Heterogeneous Pixel Fusion module is proposed to enhance the circulation of dual temporal attributes within the pixel-based feature semantic domain. In addition, our comprehensive tests on some CD datasets that demonstrate our suggested SMNet strategy compares favorably with current leading-edge techniques.