Rainfall Threshold Calculation for Landslides Based on Machine Learning Methods

Guangdong Province is one of the most geologically disaster-prone regions in China, with rainfall being a primary triggering factor. Statistics show that from 2003 to 2022, a total of 5 597 slope-related geological disasters occurred in the province, 4 456 (93%) of which took place during the flood and rainy seasons. This study constructs a sample database using historical landslide data from Guangdong and develops a rainfall threshold model for landslide prediction. With the advancement of artificial intelligence, the integration of statistical methods and machine learning techniques offers new avenues for improving model accuracy and applicability. While landslide susceptibility assessment provides a static evaluation of spatial risk variability, rainfall threshold calculation serves as a dynamic early warning approach. This study proposes an integrated dynamic-static landslide warning framework, aiming to achieve refined and real-time risk management.Firstly, rainfall events are divided into positive and negative samples, determines the antecedent effective rainfall coefficient, and calculates the antecedent effective rainfall. Then, machine learning methods are employed to predict the slope of the rainfall threshold line, deriving graded threshold lines for various probability levels. A landslide susceptibility assessment model is constructed based on the Random Forest algorithm, and a susceptibility map is generated for Guangdong Province. The main conclusions are as follows: ① The landslide rainfall threshold model based on the Rd-Rp threshold line and machine learning demonstrates good predictive performance. The antecedent effective rainfall coefficient α=0.9 was determined using Pearson bivariate correlation, with the slope of the threshold line set at 0.89. Based on the numerical values and frequency distribution of antecedent effective rainfall and daily triggering rainfall for each landslide event, critical threshold line equations for four probability levels (10%, 50%, 70%, and 90%) were established. Using the yellow warning level as the benchmark for issuing alerts, 157 landslide events were classified at or above this level, achieving an early warning success rate of 67.1%. ② The landslide susceptibility assessment conducted using the Random Forest model achieved an accuracy of 0.736 3 and an AUC of 0.81. Taking Guangdong Province as an example, the study identifies Meizhou City, northeastern Chaozhou City, central-western Heyuan City, southwestern Shaoguan City, northwestern Qingyuan City, and northern Zhaoqing City as landslide-prone areas. This study shows promising performance in landslide early warning and susceptibility assessment, contributing to enhanced risk management capabilities for regional landslide disasters. In the future, large language models could be leveraged to further improve the efficiency and accuracy of landslide risk warnings.