Research on landslide hazard spatial prediction models based on deep neural networks: a case study of northwest Sichuan, China

The Wenchuan earthquake in 2008 induced thousands of geological hazards. Among them, earthquake-induced landslides are extremely disastrous, causing considerable social and economic losses and damage to the ecology and environment. Therefore, it is of great significance to accurately predict the spatial distribution of earthquake-induced landslides. The research in this paper focuses on ten extremely earthquake-stricken areas of the Wenchuan earthquake; landslide hazard data are collected from the study areas, and the terrain information entropy, distance to rivers, distance to faults, distance to roads, lithology, normalized difference vegetation index (NDVI), peak ground acceleration (PGA) and other landslide conditioning factors are extracted. Combined with a deep neural network (DNN), a landslide hazard spatial prediction model is constructed. Through Dataset 1 and Dataset 2, which are generated randomly, the training dataset (70%) and the validation dataset (30%) are divided to verify the robustness and accuracy of the model, and the landslide susceptibility map of the study area is obtained. Moreover, considering the area under curve (AUC), the impacts of different conditioning factors on the landslide hazard prediction model are analyzed. The research results show that the DNN-based landslide hazard spatial prediction model (AUC Mean = 93.66%, Recall Mean = 85.70%) has a better prediction performance in the training step. From the validation step, the influences of various factors on the spatial prediction model of landslide hazards established in this paper (from high to low) are as follows: distance to faults, lithology, distance to rivers, PGA, terrain information entropy, NDVI, and distance to roads. The landslide hazard point data and landslide susceptibility map are highly consistent, indicating that the application of a deep learning algorithm in hazard susceptibility assessment is effective and can provide a scientific basis for landslide hazard early warning and disaster prevention and mitigation in mountainous areas prone to earthquake disasters.