A threshold updating model for rainfall-induced landslide

Rainfall threshold is an effective way for landslide early warning (LEW). Many threshold calculation models based on statistical principles have been proposed, or have been applied in national or regional early warning of geological hazard. However, due to global warming, frequent extreme rainfall and other factors, the triggering conditions of geological hazards have changed. And a fixed single rainfall threshold may no longer be applicable. In addition, the traditional threshold model requires a large number of high-quality landslide records in the region, and it is easy to ignore the heterogeneity of geological environment. Therefore, taking the slope unit as the object, we explore the dynamic updating model of rainfall threshold based on machine learning. 
Based on 170 high-risk slope units in Lin 'an District of Zhejiang Province and 65625 warning data from 2021 to 2023, we collected landslide records, rainfall station information and hourly rainfall data simultaneously. According to E-D threshold model and effective antecedent rainfall model, the general law of rainfall-induced landslide in Lin 'an District is derived, which is used as prior knowledge for model training. Then, through the warning data recorded in each slope unit, Decision Tree (DT), Bayesian Ridge (BR), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP) are selected as base learners, and ensemble strategies such as Bagging, Boosting, and Stacking are considered to dynamic updating of rainfall thresholds. 
Taking R2, Mean Square Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) as evaluation metrics, the Stacking model shows the best prediction performance. In addition, two new landslides occurred in 170 slope units in 2024 were used to verify, and it was found that the updated threshold reduced the redundant workload and gave an accurate early warning of landslides. However, the volume of warning data and its distribution on different rainfall indicators are important factors affecting the threshold update. The accuracy of updating threshold needs to be tested with more experience and practice. 
The warning data reflects the response of slope to rainfall under different rainfall conditions, which is of great significance to the threshold update of slope unit. The dynamic updating of rainfall thresholds using machine learning meets the application requirements of current climate change and provides new ideas for disaster prevention and mitigation in the new era.