SCPDNet: A Network for Landslide Detection Combining Deformable Attention and Multiscale Fusion in Remote Sensing Imagery

Accurate segmentation of landslides from remote sensing imagery is crucial for timely disaster monitoring and effective emergency response. However, current deep learning methods struggle to achieve high segmentation accuracy in complex terrains, especially for small-scale landslides, due to inadequate feature representation and insufficient preservation of fine-grained spatial details. To address these issues, we introduce a novel SAM-based Cross-Pyramid Deformable Network (SCPDNet) based on a pretrained Segment Anything Model (SAM) encoder, leveraging its robust and generalizable features. Specifically, we propose a Cross-Pyramid Deformable Attention (CPDA) module, effectively integrating the geometric adaptability of deformable convolutions with global context modeling from multi-head attention, significantly enhancing boundary detection. Furthermore, we propose a Multi-Scale Attention Fusion (MSAF) mechanism that hierarchically integrates deep semantic information with shallow spatial features, substantially improving segmentation of small-scale landslides. Extensive experiments on Bijie and Landslide4sense datasets show that SCPDNet consistently outperforms state-of-the-art methods in segmentation accuracy, while maintaining moderate trainable parameters. Ablation studies further confirm the CPDA module's critical role in performance gains. The proposed SCPDNet effectively bridges global context reasoning and local deformation modeling, offering a powerful and practical solution for accurate landslide detection and reliable post-disaster assessment using remote sensing imagery.