Machine-Learning-Based Landslide Susceptibility Assessment Under Return Levels of Annual Rainy-Season Total Precipitation: A Case Study of Hanyuan County, China

Precipitation is often represented by long-term climatological means in landslide susceptibility assessments, leaving responses to rare seasonal precipitation and associated extrapolation uncertainty poorly understood. This study coupled hydrological frequency analysis with machine learning to assess landslide susceptibility under rainy-season precipitation scenarios in Hanyuan County, China. Daily precipitation (1985–2024) was aggregated into annual June–September totals, and peaks-over-threshold generalized Pareto models were fitted to 84 grid cells to estimate 5-, 20-, 50-, and 100-year return levels. Six machine-learning models were trained using 12 static conditioning factors, mean rainy-season precipitation (MRSP), 414 landslide samples, and 2070 non-landslide samples, and evaluated using a geographically separated 30% holdout region. XGBoost performed best (ROC-AUC = 0.917; 95% confidence interval: 0.863–0.959) and was used for scenario mapping. Median precipitation increased from 896.03 to 998.57 mm under 5-year to 100-year return levels. SHAP analysis identified elevation, MRSP, and distance to roads as the leading predictors. High-susceptibility zones were concentrated along the Dadu and Liusha river valleys. Within the common applicability domain (39.45% of the county), T100 relative to T5 reduced the very-low-susceptibility class by 3.80% points and increased the very-high class by 1.38%. Across their respective applicability domains, the mean susceptibility probability under T100 was 0.025 higher than that under MRSP. Return-levels scenarios therefore identified precipitation-sensitive areas within the model-supported domain.