SF-Mamba: A Semantic-Flow Foreground-Aware Mamba for Semantic Segmentation of Remote Sensing Images

Mamba-based methods excel in long-range modeling for remote sensing images (RSIs) semantic segmentation but lack sufficient foreground awareness. To address this limitation, we propose a Semantic-Flow Foreground-Aware Mamba (SF-Mamba) to enhance semantic segmentation performance. The encoder of SF-Mamba is a dual-branch architecture; One branch conducts local feature extraction, and the other employs a visual state space model for global context. A foreground-aware decoder is designed to learn object information. A semantic flow aggregation module (SFAM) is presented to handdle the complex scenes of RSIs. A global-guided foreground awareness module (GFAM) is designed to improve the capacity for foreground modeling through cross-layer similarity analysis between local features and global information. Experiments conducted on three datasets demonstrate that the SF-Mamba is more effective and efficient than other mainstream approaches. Visualization results and ablation experiments demonstrate the superiority in improving the semantic segmentation ability and dealing with complex RSIs, especially for small objects in RSIs. The code is available at https://github.com/TiezhuXing01/SF-Mamba.