RM-MambaSeg: Region-Aware Modulation for U-shaped Mamba in SITS Segmentation

Satellite Image Time Series (SITS) provide dense observations for agricultural monitoring, yet efficient and accurate segmentation remains challenging due to long-range temporal dependencies and complex spatial structures. Existing Mamba-based methods often suffer from semantic blindness caused by fixed geometric scanning, and lack efficient temporal aggregation. We propose RM-MambaSeg, a lightweight U-shaped framework where Mamba serves as the core operator for both spatial and temporal modeling, delivering linear complexity and global receptive fields. To overcome semantic blindness, we introduce a region-aware modulation (RM) module that guides spatial feature extraction at early encoder stages: it generates pixel-adaptive soft masks to preserve intra-region coherence and sharpen boundaries, overcoming the limitations of fixed scanning patterns. Temporal dynamics are efficiently captured by a Mamba encoder followed by mean pooling for robust aggregation. Experiments on the MTLCC benchmark demonstrate that RM-MambaSeg achieves competitive performance with high efficiency, validating the effectiveness of region-aware modulation for SITS segmentation.