Research on interpretable machine learning modeling and spatial prediction of cross-landform landslide disaster mechanism: A case study of Shaanxi Province

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: CatBoost AUC: 0.8307 (northern Shaanxi), 0.8252 (southern Shaanxi)

• Primary Outcome: Landslide susceptibility prediction accuracy

• Nnt Or Sensitivity: Population density and topographic moisture index contribute 37.2% in northern Shaanxi; elevation and lithology dominate in southern Shaanxi

• Confidence Interval: Not reported

Clinical Application

Methodology could be adapted for Australian landslide-prone regions with appropriate local data Relevant for disaster preparedness in regions like Blue Mountains, Victorian highlands, and Queensland coastal ranges. Could inform Emergency Management Australia strategies and state-based disaster response planning Limited direct clinical application - primarily relevant for disaster medicine, emergency preparedness, and public health planning

Abstract

The differential understanding of the landslidetriggering mechanisms across various geomorphic units is vital for enhancing regional disaster prevention. This study investigates the Loess Plateau area in northern Shaanxi and the Qinba Mountain Area in southern Shaanxi as the study areas, and constructs a 15dimensional evaluation factor system covering topography, geology, vegetation, and human activities. After eliminating collinearity factors via double tests of the Pearson correlation coefficient and variance inflation factor, Bayesian optimization is used to optimize hyperparameters for CatBoost, Random Forest, LightGBM, and XGBoost, and the SHAP framework is combined to perform global attribution, single-factor dependency analysis, and interaction-effect quantification. The results demonstrate that CatBoost performs best across both geomorphic areas, achieving AUCs of 0.8307 and 0.8252 in the test sets. In northern Shaanxi, a discrete patch pattern driven by "human-water" coupling is observed, with population density and the topographic moisture index contributing 37.2%. In contrast, Southern Shaanxi exhibits a continuous, band-shaped distribution controlled by "structure-topography" is observed, and elevation and lithology dominate the model's decision-making. The SHAP-dependent map identified cross-geomorphic differences in the population density threshold of 99.83 people/km2, the slope threshold of 2.21° in northern Shaanxi, and the elevation threshold of 1194.73m in southern Shaanxi. Furthermore, the interaction network revealed an antagonistic effect of DEM and POP in northern Shaanxi and of DEM and LULC in southern Shaanxi. The framework proposed a quantitative, explainable basis for differentiated disaster prevention strategies in this research.

References

1. 1.Han, M., & Yang, Y. (2026). Research on interpretable machine learning modeling and spatial prediction of cross-landform landslide disaster mechanism: A case study of Shaanxi Province. PLoS One. https://doi.org/10.1371/journal.pone.0347996