HAM-CD: Hybrid Attention Mamba for Remote Sensing Change Detection

The Mamba-based change detection (CD) methods have achieved remarkable success owing to their ability to model long-range pixel dependencies with linear time complexity effectively. While effective in modeling long-range pixel relationships, these methods exhibit limitations in comprehensively learning complex image spatial structures and fully capturing local pixel dependencies critical for effective spatial-temporal representation learning and accurate change detection. To overcome this drawback, we propose a Hybrid Attention Mamba Change Detection (HAM-CD) method that achieves efficient global context modeling while simultaneously maintaining finegrained local detail representation. Our HAM-CD employs the Visual Mamba architecture as its encoder, effectively capturing both global and local spatial-temporal features from the input image. To enhance both the global and local representation ability of the Visual Mamba, we design a decoder composed of several Hybrid Attention Mamba (HAM) blocks, which integrates the Transposed Attention Block (TAB) into the Textual-aware Mamba (TAM) block. The HAM effectively captures spatial-temporal feature interactions by utilizing the TAM to strengthen local pixel dependencies and the TAB to enhance global dependencies. Additionally, to mitigate error propagation across stages and enhance more accurate change detection, we introduce an Interactive Channel-Spatial Fusion (ICSF) block that improves cross-stage feature interaction by jointly modeling spatial and channel-wise dependencies. Comprehensive experiments on three benchmark datasets demonstrate that our HAM-CD consistently outperforms existing methods, achieving state-of-the-art (SOTA) results without requiring sophisticated training schemes or specialized optimization techniques. Our source code and pretrained models are publicly available at https://github.com/guanguanboy/HAM-CD.