A geological disaster risk prediction method using susceptibility and ensemble learning algorithms: direct economic loss prediction

Geological disasters have led to significant economic losses and casualties in South China. However, existing indicators for predicting geological disaster risk are inconsistent and generally lack physical meanings. This study aims to develop an intuitive and reliable geological disaster prediction method by using direct economic loss as the output indicator to establish a geological disaster economic loss prediction (GELP) model. First, multi-source data, including exposure, geological disaster susceptibility, and disaster-causing factors, are integrated to create a GELP database using data related to six types of geological disasters, i.e., landslides, rockfalls, debris flows, ground fissures, land subsidence, and ground collapse. Next, multiple GELP models are developed using three ensemble learning algorithms, including categorical boosting (CatBoost), Stacked Generalization (Stacking) and Extreme Gradient Boosting (XGBoost), and two single machine learning algorithms. The performance is evaluated using R-squared (R2), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE), whereas their reliability is validated through geological disaster cases. The results show that integrating multi-source data with ensemble machine learning models can produce a more reliable geological disaster economic loss prediction model, as the CatBoost model performs the best overall, achieving an R2 of 0.67. Additionally, the CatBoost model demonstrates high accuracy at a county scale. When the losses are smaller, the models’ predictions are closer to the actual values. Furthermore, integrating historical data from multiple geological disasters not only enhances the models’ applicability but also enlarges the scale of training datasets, allowing for more thorough training, which significantly improves performance. These findings offer valuable insights for developing reliable GELP models.