A Generalized Deep Learning-Based Method for Rapid Co-Seismic Landslide Mapping

Abstract

The rapid mapping of co-seismic landslides is essential for emergency management and loss assessment. Deep learning algorithms generally follow a supervised learning workflow, where the trained model is used to predict landslides in surrounding areas, achieving landslide mapping with high accuracy. For a new study area landslide extraction task, the performance of the model trained on a specific dataset will be greatly reduced due to the varying data distribution of co-seismic landslides. Considering the urgent need for large-scale co-seismic landslide mapping, we developed a generalized deep learning-based landslide identification method. First, a new model—ResU-SENet is developed to generate semantic segmentation maps of landslides. The proposed model adaptively emphasizes the channel-wise weights of the input data. Three multidomain models are then designed by combining annotated landslide samples from two different domains to improve the model generalization ability. Finally, the trained models are applied directly to completely unknown domains to test model generalizability. Experiments in Iburi and Jiuzhaigou showed that the proposed model yielded the recall values of 5.93% and 7.51% higher than ResU-Net. The adoption of multidomain models effectively reduced the number of new training samples required by 50% and maintained a similar identification performance as if trained entirely with new samples. Applying the models trained by Jiuzhaigou and Iburi samples directly to Palu, the F1-score under the ResU-SENet model reached 0.6875. Moreover, the connections between model generalization and data distribution was demonstrated. This work could provide a fast response for future large-scale co-seismic landslide mapping.

Keywords

• Change detection
• co-seismic landslide inventory mapping
• generalization
• landslide identification
• medium-resolution imagery
• multidomain trained model