Global landslide susceptibility mapping: a 1 km resolution model derived from a 2-million-event inventory with uncertainty quantification

Landslide susceptibility mapping (LSM) at the global scale is a prerequisite for hazard risk management but has long been hindered by inventory bias and unquantified model uncertainties. Existing global products are often constrained by substantial spatial sampling biases, leading to inconsistent prediction performance across data-scarce and vegetated regions. Addressing these challenges, this study presents a robust 1-km global susceptibility model derived from a dataset of over 2 million landslide events aggregated from 24 diverse sources.

To resolve data heterogeneity, we applied an LLM-driven framework (utilizing Qwen2.5-7B) to extract and standardize attributes from unstructured descriptions across 14 languages, significantly elevating metadata completeness . Leveraging this enriched inventory and 34 environmental predictors (comprising 17 static and 17 dynamic variables), we implemented a rigorous spatial block cross-validation strategy to strictly evaluate model transferability. We evaluated nine machine learning algorithms (e.g., CatBoost, ExtraTrees) coupled with Optuna tuning. Furthermore, Monte Carlo simulations (N=50) were integrated to propagate input uncertainties, generating explicit pixel-level confidence intervals.

Our results demonstrate high predictive accuracy (spatial CV AUC > 0.99), suggesting that the density of the training data effectively bridges generalization gaps found in previous studies. Feature optimization confirms the model’s robustness even with reduced dimensions. Spatially, the model identifies high-susceptibility zones in complex tropical highlands (e.g., the Andes and Southeast Asia), aligning with independent records of fatal landslide clusters. By providing a bias-corrected and uncertainty-aware spatial baseline, this study offers a critical foundation for global hazard monitoring.