CS-Mamba: A Novel Mamba-Based Framework for Landslide Detection With Attention Guided Cross-Scale Context Semantic Integration

Landslides are a frequent natural hazard, and the precise detection of affected areas using high-resolution optical imagery is of great significance for geological disaster emergency response and risk assessment. In optical images, landslides often exhibit substantial scale variations, complex and diverse morphologies, and blurred boundaries, and are susceptible to environmental influences. Existing deep learning approaches have not fully exploited cross-scale dependencies among pixels or ensured the consistency between feature representations and semantic embeddings, leading to insufficient model robustness. To address these limitations, we propose a novel Mamba-based framework for landslides detection with attention-guided cross-scale context semantic integration (CS-Mamba). In the encoder, parallel multi-scale convolutional branches are introduced to enhance the network's representation capability. At the core, a multi-scale-aware state-space module in the neck adopts bidirectional information flows and a gated fusion mechanism to integrate global and local features, thereby improving semantic consistency in boundary segmentation. In the decoder, an attention-guided strategy is employed to fuse multi-level features, achieving precise preservation of boundary details and strong robustness to cross-domain data. The proposed method is evaluated on multiple representative landslide datasets, including the LMHLD, Bijie, and GLVM datasets. Experimental results demonstrate that CS-Mamba achieves superior performance on nearly all metrics and exhibits strong generalization ability in domain generalization analysis.