The scale-aware cross-attention boundary refinement network for identifying earthquake-triggered landslides based on SAR backscatter coefficients

Abstract

Earthquake-Triggered landslides (ETLs) are among the most devastating types of post-earthquake secondary geological hazards, and their accurate and rapid automatic detection is crucial for disaster emergency response. Synthetic Aperture Radar (SAR) possesses active microwave detection capabilities and operates all-weather and all-day conditions, thereby demonstrating mapping advantages that cannot be replaced by traditional optical remote sensing. To address the challenges of ETLs identification, an intelligent network is proposed based on SAR backscatter coefficients: Scale-aware Cross-attention Boundary Refinement Network (SCAB-Net). Using pre- and post-earthquake logarithmic backscatter intensity ratio maps as the core input features, Multi-scale Spatial Channel Collaborative Attention (MSCA) module is constructed to capture landslide features at different scales. In addition, an Enhanced Boundary Refinement (EBR) module is also designed to improve segmentation boundary accuracy. In the experiment, three typical ETLs cases are used: Miling, Luding and Papua New Guinea. Compared with DeepLabV3+, RCFSNet and Vim-UNet, our proposed SCAB-Net has achieved the best results in both the average F1 score (85.42%) and average IoU (75.11%). In an independent cross-region test (Milin), the F1 score reached 84.39% and the IoU reached 72.99%, validating the method’s effectiveness and generalisation capability.