Cross-domain coseismic landslide segmentation: local boundary enhancement & global pixel contrastive learning

Abstract

Accurate and automated landslide mapping is essential for disaster response, yet existing deep learning methods struggle with blurred boundaries and limited feature discriminability. To address these challenges, we propose MBENet, a segmentation framework based on a ‘local enhancement–global optimization’ philosophy, integrating two synergistic modules: a proactive Boundary Feature Enhancement Module (BFEM) for mid-level supervision to preserve detail, and a Pixel-level Contrastive Learning Module (PCLM) to strengthen semantic representations. We conducted bidirectional cross-domain validation on two large-scale public datasets focused on coseismic landslides, GDCLD and CAS. Compared to eight SOTA models like SegFormer and SegNeXt, MBENet achieved the highest performance, with a peak mIoU of 71.99%—a 1.22% absolute improvement over the runner-up. Ablation studies confirmed a strong super-additive effect, with the full model boosting the baseline mIoU by up to 5.74%. Our findings also indicate that precise boundary delineation is strongly coupled with semantic discriminability, as evidenced by our model’s 90.28% precision in the CAS (Formula presented.) GDCLD task, offering valuable insights for future model design. While MBENet has higher theoretical complexity (100.73 M parameters, 393.85 GFLOPs), its competitive 171 ms inference time on an NVIDIA RTX 3090 GPU shows a favorable trade-off between accuracy and practical efficiency for this specific task.