Extreme precipitations are widely recognized as the principal trigger of catastrophic landslides in mountainous terrain. However, purely data-driven statistical models rely on high-precision rainfall records, constraining their applicability to regions with scarce historical data and failing to account for landslide heterogeneity or underlying physical processes. Herein, we present a dynamic evaluation framework that integrates spatial and temporal probabilities to map rainfall-induced landslide hazard in the Wanzhou District of the Three Gorges Reservoir Area. The proposed approach synthesizes data-driven susceptibility models with physics-driven rainfall threshold analysis. First, a detailed field investigation was conducted to delineate accumulation and rock areas according to landslide types, and advanced ensemble learning models were developed to map the susceptibility of different landslides independently. Then, a grey wolf optimizer-enhanced bidirectional long short-term memory model (GWO-BiLSTM) was trained on extensive field-survey data to predict spatial variations in soil thickness. The probabilistic approach was proposed to quantify the rainfall thresholds for each type of landslide, considering the uncertainty of geotechnical properties, thereby providing temporal probabilities for landslide hazard zoning (LHZ). The results indicate that the landslide-specific ensemble learning strategy substantially enhances prediction accuracy and reliability. Accumulation areas failure exhibits greater sensitivity to extreme precipitation and occurs at lower rainfall intensities than rock areas. In addition, field validation during an actual heavy-rain scenario confirms that all observed landslides are correctly classified within the very high hazard zone. This integrated data-mechanics framework offers a promising reference for the global mitigation of rainfall-induced landslide hazards. • Differential mechanisms and triggers exist between soil and rock landslides. • Development of Stacking-DT-ANN model to promote landslide susceptibility zonation. • GWO-BiLSTM enhances the accuracy of soil thickness prediction. • Probabilistic approach quantifies rainfall thresholds for landslide considering geotechnical uncertainty. • Landslide dynamic hazard mapping is refined based on combined susceptibility and rainfall scenarios.
Zhao et al. (Sun,) studied this question.