Landslide hazard zoning under return-period rainfall scenarios: Integrating MaxEnt susceptibility with Pearson Type III frequency analysis

This study develops scenario-based landslide hazard maps for 10-, 20-, 50-, and 100-year rainfall conditions in Longhui County, Hunan Province, China. The study area covers 2867.71 km 2 , and the dataset comprises 188 landslide and 188 non-landslide samples, with 132 samples from each class used for model development and 56 for validation. Landslide susceptibility was modelled using Random Forest, XGBoost, and MaxEnt based on 11 conditioning factors, and model performance was evaluated using AUC, Precision, Recall, and F1-score. Design rainfall estimates derived from Pearson Type III (P-III) frequency analysis of the 2001–2022 annual maximum monthly rainfall series were combined with the normalized susceptibility indices to produce hazard maps for the four scenarios. Under the adopted non-spatial held-out split, MaxEnt showed the best comparative performance, with an AUC of 0.901, a Precision of 0.892, a Recall of 0.876, and an F1-score of 0.884. The estimated design monthly rainfall amounts for the four return periods were 341.410, 368.476, 403.475, and 428.675 mm, respectively. Bootstrap confidence intervals widened with return period, and the 95% interval for the 100-year estimate was 321.67–495.78 mm. From the 10-year to the 100-year scenario, the mapped area classified as high and very high hazard increased by 94.7% for XGBoost, 140.7% for RF, and 242.7% for MaxEnt. High and very high hazard areas were concentrated in the western, northwestern, and northern mountainous sectors. The resulting maps provide scenario-specific spatial information for prioritising county scale field investigations and rainy season monitoring.