Coseismic Landslide Mapping Based on Trans-UNet and Transfer Learning

Coseismic landslides often cause significant property damage and loss of life, necessitating timely and accurate detection for emergency response and hazard mitigation. Optical remote sensing imagery plays a critical role in landslide identification; however, sufficient high-resolution images might not be accessible after an earthquake. Traditional detection methods struggle with the coarse spatial details of low-resolution images, limiting their reliability in such scenarios. This study introduces a novel landslide mapping method leveraging the Trans-UNet model and a transfer learning strategy to overcome these challenges. The Trans-UNet model integrates a UNet-like encoder–decoder structure with a Transformer module to enhance global context extraction, a U-shaped full-scale feature extraction module to preserve multiscale spatial details, and a convolutional decoder for effective feature fusion and resolution restoration. During model training, a transfer learning strategy is employed to enhance the performance of the model on low-resolution images by utilizing rich feature information obtained from high-resolution images. In addition, this study constructs an open dataset of coseismic landslides triggered by the 2022 M_S 6.8 Luding earthquake, using the Gaofen-2 images with a 1-m resolution and the Sentinel-2 images with a 10-m resolution. The effectiveness and superiority of the proposed method are validated through its application to the Luding Landslide Dataset and comparisons with other state-of-the-art models. This method provides a reliable solution for rapid post-earthquake assessment in resource-constrained environments.