Predictive models for earthquake-induced landslides: machine learning based on real case histories

Abstract

In this study, the deformation of earth slopes under earthquakes was evaluated using machine learning techniques. While traditional empirical models have been widely used to estimate seismic slope deformations, they often suffer from limited accuracy and generalizability due to their reliance on simplified assumptions and region-specific datasets. To address this gap, extensive real case histories on seismic deformations of earth slopes during earthquakes in different regions across the world were gathered and examined. Most important factors affecting earthquake-induced deformations of the slopes were characterized. Five models were then developed for prediction of seismic deformation of earth slopes (D) using extreme learning machine (ELM), random forest (RF), genetic programming (GP), support vector regression (SVR), and hybrid whale optimization algorithm (WOA)-SVR. Subsequently, the accuracy of developed models was measured. The results indicated that WOA-SVR model (R² = 0.821, RMSE = 0.819) has higher accuracy than SVR (R² = 0.780, RMSE = 0.852), GP (R² = 0.763, RMSE = 0.972), RF (R² = 0.634, RMSE = 1.133), and ELM (R² = 0.533, RMSE = 1.214) models. Finally, the performance of developed models was investigated through comparing with the previous relationships for calculation of earthquake-induced earth slope deformations. The results indicated that the developed machine learning-based predictive models can provide more precise forecasts in comparison to the available recommendation.