FSMamba: A Frequency-Enhanced Spatial–Temporal Mamba Network for Local–Global Remote Sensing Image Change Detection

Remote sensing change detection (RSCD) is challenged by complex scene variations, illumination differences, misregistration errors, and the need to jointly model global semantics and fine-grained local structures. To address these issues, this paper proposes FSMamba, a Frequency-enhanced Spatio-temporal Mamba network for local–global RSCD. FSMamba integrates frequency-domain analysis, spatiotemporal interaction, and efficient long-range dependency modeling within a unified framework. Specifically, a Frequency-Domain Enhancement Module (FDEM) is designed to capture global structural discrepancies via learnable spectral filtering, effectively suppressing pseudo-changes caused by illumination variations and noise. A Spatio-Temporal Perception Module (STPM) leverages channel-wise interaction and the linear-complexity visual state space model to enhance semantic consistency between bi-temporal features. In addition, a Global–Local Collaborative Decoder (GLCD) is introduced to adaptively fuse multi-scale features, preserving large-scale structural integrity while maintaining fine-grained boundary details. Extensive experiments on five public benchmark datasets demonstrate that FSMamba consistently outperforms state-of-the-art CNN, Transformer, and Mamba-based methods in terms of accuracy, robustness, and computational efficiency, validating its effectiveness under complex RSCD scenarios. Our source code is available on https://github.com/2419924948gyt/FSMamba.