A novel method based on deep learning model for national-scale landslide hazard assessment

Landslide hazard assessment is crucial for landslide monitoring and early warning. A novel method based on deep learning model for national-scale landslide hazard assessment was proposed in this study, which contains three stages: (1) landslide susceptibility analysis using three hybrid neural networks of convolutional neural network-simple recurrent unit (CNN-SRU), convolutional neural network-long short-term memory (CNN-LSTM), and convolutional neural network-gated recurrent unit (CNN-GRU); (2) landslide temporal probability prediction using the proposed spatiotemporal transformer (ST-transformer) to address the time uncertainty of rainfall threshold calculation in landslide temporal probability prediction and the matter without considering the geographical regional differences of landslide spatiotemporal probability weights; and (3) quantitative landslide hazard calculation with the preceding results using the improved landslide hazard formula. The validation of this method was conducted in conterminous United States, where the results of steps (1) and (2) demonstrated excellent performance compared with existing works. Consequently, the calculation derived from the previous two steps was effectively used for landslide hazard assessment using the improved landslide hazard formula, and its reliability was confirmed through the validation of actual landslide events. The proposed method is of practical significance for national-scale landslide hazard assessment.

Automated summary

A novel deep learning-based method for national-scale landslide hazard assessment was proposed in this study. It includes three stages: landslide susceptibility analysis, landslide temporal probability prediction, and quantitative landslide hazard calculation. The validation in the conterminous United States showed excellent performance of this method. Therefore, it has practical significance for national-scale landslide hazard assessment.