Probabilistic study of rainfall-induced landslides at a monthly scale in China
摘要
Objective Rainfall is a major trigger for landslides in mountainous regions of China, exhibiting distinct seasonal and spatial aggregation characteristics. To achieve quantitative identification of medium- to long-term geological disaster risks, this study aims to develop a nationwide monthly-scale probabilistic framework for rainfall-induced landslides, integrating rainfall forecasting and data-driven modeling to reveal the spatiotemporal response between rainfall variability and landslide occurrence. Methods A 1 km resolution monthly rainfall forecasting model was established using a Patch-based Long Short-Term Memory (Patch-LSTM) network trained on multi-source precipitation data spanning 1901–2023. The model’s spatial continuity and predictive performance were validated using records from independent meteorological stations. A database of rainfall-induced landslides was compiled from four typical heavy rainfall events in Yunnan Province (Dehong 2020, Daguan 2021, Gongshan 2020 and 2022), containing 8503 mapped landslides with associated topographic, geological, hydrological, and climatic factors. These datasets were used to train a probabilistic rainfall–landslide occurrence model combining logistic regression and Gradient Boosting Tree (GBT) algorithms. The model outputs were spatially coupled with monthly rainfall forecasts to generate nationwide monthly probability maps of rainfall-induced landslides. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), the precision–recall AUC (PR-AUC), and the Brier score. Results The Patch-LSTM model achieved an average absolute error (MAE) of 14.6 mm, a root mean square error (RMSE) of 35.1 mm, and a coefficient of determination (R2) of 0.51, indicating reliable capability in reproducing monthly precipitation patterns. The probabilistic landslide model showed robust predictive performance, with AUC = 0.83, PR-AUC = 0.78, and Brier score = 0.17. The spatial-temporal patter