A Causal Analysis based on Dynamic Landslide Hazard Assessment from 1980 to 2024 in Hubei, China

Landslides rank among the most destructive geological hazards globally, with their frequency and intensity increasingly exacerbated by the dual pressures of climate change and rapid anthropogenic land modification. Traditional static landslide hazard mapping often relies on global feature importance rankings, which obscure the spatial heterogeneity of driving mechanisms. This black box nature limits the physical interpretability of hazard evolution. This study aims to establish a long-term Dynamic Landslide Hazard assessment framework to decouple the causal mechanisms of rainfall and land use in landslide evolution.

Focusing on Hubei Province (1980–2024), we integrated XGBoost for dynamic prediction with Double Machine Learning (DML) for causal attribution. To address high dimensionality, Principal Component Analysis (PCA) was employed to reconstruct comprehensive indices (cumulative variance > 90%). Central to our methodology is the proposal of a novel "Consistency-Interaction Diagnostic Framework." By coupling the global trends derived from Partial Dependence Plots (PDP) with the local heterogeneity of SHAP values, this framework constructs a 2D metric system to diagnose the physical stability and spatial interaction strength of drivers.

Application of this diagnostic framework successfully decoupled the dual physical attributes of landslide drivers, a distinction missed by traditional methods:

• The framework accurately identified land use intensity and static terrain factors as "Stable Background Stress." These factors exhibit high PDP-SHAP correlations (Consistency > 0.95) with low spatial variance, confirming their roles as domain-wide controls regardless of local micro-environments.
• In contrast, rainfall factors were diagnosed as "High-Sensitivity Pulses." For instance, antecedent summer precipitation exhibited an extremely high SHAP interaction Coefficient of Variation (CV) of 3.78. This quantitative diagnosis proves that rainfall is not a uniform stressor but a spatially selective trigger whose hazard efficiency is intensely modulated by local topography.
• Diagnostic results further reveal that the majority of environmental factors fall into the "Heterogeneous Effect" quadrant. This indicates that landslide incubation is not a linear superposition of single factors but a complex non-linear process strongly modified by local environments.

This study demonstrates that the proposed framework offers a new physical perspective for opening the black box of machine learning. By distinguishing between globally consistent factors and locally sensitive perturbations, the findings provide a scientific basis for shifting landslide risk management from homogenized meteorological warnings to fine-grained control based on spatial heterogeneity and ecological resilience.