Single synthetic aperture radar image for landslide mapping using attention U-Net

Abstract

Rapid landslide mapping is critical for effective emergency response. Optical imagery is frequently hindered by continuous cloud cover during severe storm events. Synthetic Aperture Radar (SAR) offers stable observations in all weathers, but automated mapping using a single post-event image remains a challenge. Standard dual polarization SAR data (e.g. Sentinel-1) suffers from speckle noise and struggles to distinguish landslide debris from surrounding vegetation. This study developed a SAR analysis framework using single post-event Sentinel-1 image in central Taiwan based on an Attention U-Net model to overcome these challenges. Radar Vegetation Index (RVI) was incorporated as a third input channel to capture structural changes in vegetation and enhance the contrast of bare soil. The Attention U-Net model trained solely on a descending Sentinel-1A image outperformed a Bayesian classifier and a standard U-Net. It achieved an average 33% improvement over the standard U-Net across all metrics. A case study in Wanrong Township further demonstrated that incorporating RVI consistently enhances landslide delineation. Detection performance nearly tripled in post-landslide scenarios compared to the baseline model with two channels. These results highlight the potential of combining physical polarimetric indices with deep learning to provide timely and accurate information for rapid post-event response.