Fast Mapping of Large-Scale Landslides in Sentinel-1 SAR Images Using SPAUNet

Abstract

Rapid and accurate mapping of large-scale landslides is crucial for postdisaster prevention and risk analysis. This research introduces a new method named synthetic aperture radar (SAR) pixel-attention UNet (SPAUNet) to address the challenges of noise and information redundancy when utilizing SAR amplitude data for landslide detection. This process is based on constructing a UNet model and incorporating a pixel-attention mechanism. This study uses pre- and postdisaster SAR amplitude images as input data, generates seven pairs of dual-band combinations through registration, and inputs them into the model in sequence to automatically explore the best polarization amplitude combination, achieving rapid per-pixel mapping of large-scale landslides. Empirical analysis of landslide events in Milan and Papua New Guinea shows that SPAUNet outperforms the baseline models, improving the F1 score by 17% in Milan and 19% in Papua New Guinea. Moreover, this study emphasizes the importance of choosing the appropriate polarization combination for the region. The results indicate that SPAUNet, along with the appropriate polarization amplitude combination, improves the accuracy and reliability of landslide mapping, aiding disaster assessment and recovery work. This improvement holds significant implications for landslide disaster assessment and postdisaster recovery, providing a valuable direction for further enhancing landslide monitoring capabilities.