Abstract
Accurate delineation of landslides from high-resolution optical imagery is important for hazard assessment and infrastructure safety in mountainous regions. However, automatic segmentation remains challenging because landslides vary greatly in scale, morphology, surface appearance, and boundary clarity. These difficulties are particularly evident for old landslides, whose surface characteristics may become less distinct after long-term erosion, vegetation recovery, and surface modification. To establish a consistent baseline, five representative semantic segmentation models were evaluated on the Landslide Recognition along the Sichuan–Tibet Transportation Corridor (LRSTTC) dataset under a unified experimental protocol. DeepLabV3+ showed the strongest overall baseline performance and was selected as the base architecture. We then developed a Boundary-Aware Attention-Enhanced DeepLabV3+ (BAA-DeepLabV3+) framework, in which a Convolutional Block Attention Module (CBAM) is applied after Atrous Spatial Pyramid Pooling (ASPP) for high-level feature refinement, while an auxiliary boundary branch provides additional spatial supervision during training. In the controlled ablation experiment with seed 42, BAA-DeepLabV3+ improved the Intersection over Union (IoU) from 0.3644 to 0.4229 and the Dice coefficient from 0.5341 to 0.5944, with only a negligible increase in model parameters. Across three random seeds, the mean IoU increased from 0.3599 ± 0.0275 to 0.3876 ± 0.0315. External validation on the independent Bijie landslide dataset further improved the mean IoU from 0.7295 to 0.7444, with BAA-DeepLabV3+ outperforming the baseline under all three random seeds. These results indicate that attention-based feature refinement and boundary-aware supervision can work complementarily to improve landslide segmentation while maintaining a favorable balance between segmentation performance and computational efficiency.