Integrating data-driven and physically based landslide susceptibility methods using matrix models to predict reservoir landslides

Abstract

One of the most hazardous types of landslide disasters is reservoir landslides. Spatially predicting the probability of these events when the hydroelectric power station is not impounded can provide a foundation for their prevention and management. A reservoir landslide prediction study was conducted using the reservoir of the Baihetan Hydropower Station in Southwest China as the research area. The study applied both the widely used data-driven landslide susceptibility assessment method and the physics-based landslide stability evaluation approach. Additionally, a matrix prediction model based on a combination of the aforementioned techniques was suggested for reservoir landslide prediction. The applicability of the three approaches was assessed based on the findings of a field investigation of reservoir landslides during the initial impoundment period. The results demonstrated that, in the absence of historical reservoir landslide data, the prediction success rate of the susceptibility mapping method was incredibly poor, at only 30.19%. While the stability prediction approach had an 88.68% success rate, its frequency ratio was only 5.22. The approach significantly overpredicted reservoir landslides, according to a study of the prediction results for regions. On the other hand, the frequency ratio of the matrix prediction model, which combines the two approaches, was 2.24 times that of the stability prediction approach. Thus, the matrix prediction model was found to be more reliable than other methods and capable of predicting the vast majority of reservoir landslides with fewer extremely high probability reservoir landslide areas. To further increase the accuracy of reservoir landslide prediction, it is recommended to use the matrix prediction model.