Landslide mapping based on a hybrid CNN-transformer network and deep transfer learning using remote sensing images with topographic and spectral features

Abstract

Landslides frequently cause serious property damage and casualties. Therefore, it is crucial to have rapid and accurate landslide mapping (LM) to support post-earthquake landslide damage assessment and emergency rescue efforts. Many studies have been conducted in recent years on the application of automatic LM methods using remote sensing images (RSIs). However, existing methods face challenges in accurately distinguishing landslides due to the problems of large differences in features and scales among landslides, as well as similarities among different ground objects in optical RSIs. Here, we propose a semantic segmentation model called SCDUNet++, which combines the advantages of convolutional neural network (CNN) and transformer to enhance the discrimination and extraction of landslide features. Then, we constructed a multi-channel landslide dataset in the Luding and Jiuzhaigou earthquake areas using Sentinel-2 and NASADEM data. We evaluated the performance of SCDUNet++ on this dataset. The results showed that SCDUNet++ can extract and fuse spectral and topographic information more effectively. Compared with other state-of-the-art models, SCDUNet++ achieved the highest IoU and F1 score in all four test areas. In addition, the models achieved significant improvements in mapping the landslides of the Jiuzhaigou area after knowledge transfer and fine-tuning. Compared with direct prediction, eight models, namely DeepLabv3+, Segformer, TransUNet, SwinUNet, STUNet, UNet, UNet++, and SCDUNet++, demonstrated improvements in IoU ranging from 8.33% to 27.5% and F1 from 6.58% to 23.67% after implementing deep transfer learning (DTL). This finding highlights the significant practicality of using DTL for cross-domain LM in data-poor areas.