Rainfall-Induced Landslide Hazard Assessment Considering Multi-Temporal-Scale Rainfall Factors in Jiangwan Town, Shaoguan, Guangdong Province, China

Abstract

Shallow landslides triggered by extreme rainfall often occur abruptly and show pronounced spatial clustering. Accurate representation of rainfall conditions is therefore important for event-based landslide hazard assessment. This study investigated shallow landslides triggered by the April 2024 extreme rainfall event in Jiangwan Town, Guangdong Province, China. An event-based landslide inventory, half-hourly GPM IMERG precipitation data, and multiple environmental factors were used for hazard modelling. Three rainfall descriptors were considered: event cumulative rainfall, 21-day cumulative rainfall, and maximum 1 h rainfall. Controlled incremental experiments using Random Forest (RF) were conducted to evaluate the additional predictive information provided by these rainfall descriptors. RF, XGBoost, and an Artificial Neural Network (ANN) were then compared in terms of predictive performance. Incorporating all three rainfall descriptors increased the RF test-set AUC from 0.907 to 0.945. Among the three predefined rainfall descriptors, 21-day cumulative rainfall produced the largest incremental improvement. XGBoost achieved the highest overall predictive performance, with a test-set AUC of 0.971, and was selected for final hazard mapping. The high- and very-high-hazard zones predicted by XGBoost covered 33.2% of the study area and contained 89.7% of the observed landslide samples. These zones were concentrated mainly in the central and northeastern parts of Jiangwan Town. SHAP analysis identified elevation as the most influential predictor, while all three rainfall descriptors ranked among the important predictors. Overall, the results indicate that rainfall information at multiple temporal scales provides complementary predictive information and can improve event-conditioned shallow-landslide hazard assessment.