From typhoon rainfall to slope failure: optimizing susceptibility models and dynamic thresholds for landslide warnings in Zixing City, China

Typhoon-specific rainfall-induced landslides pose critical hazards in mountainous regions, yet existing warning systems inadequately capture the distinct rainfall dynamics of these extreme events. To address this limitation, we propose an integrated framework combining optimized susceptibility predictions with dynamic rainfall thresholds tailored to typhoon patterns. The approach enhances machine learning accuracy through buffer-based negative sampling and variable weighting. It also introduces a spatiotemporal rainfall analysis to distinguish between short-term intense downpours and cumulative soil saturation. Tested in Zixing City, Hunan Province, China, where over 700 landslides were triggered by Typhoon Gaemi, the framework proved effective. The support vector machine (SVM) model achieved the best performance using frequency ratio (FR) inputs with a 0.5 km buffer (F1-score: 0.859, AUC: 0.914), correctly classifying 86.4 % of landslides as high or very high susceptibility. The rainfall analysis identified 24 h intensity combined with 7 d antecedent rainfall as the optimal trigger, effectively capturing both immediate and cumulative moisture effects. Spatially, rhyolite and granite slopes and areas near roads emerged as hotspots for failure (distance