Bias-reduced landslide susceptibility mapping in data-scarce regions via multi-source inventory integration

Abstract

Landslide susceptibility mapping (LSM) is essential for disaster risk reduction in mountainous regions, but its reliability is often limited by spatial biases and incompleteness of landslide inventories. These reduce landslide inventory representativeness and influence the generalization of LSM models at the regional scale. To overcome this challenge, we developed a multi-source inventory integration framework that combines three typical complementary datasets: publicly available inventories, reliable field survey observations, and deformation hotspots derived from Multi-Temporal Interferometric Synthetic Aperture Radar (MT-InSAR) analysis. Four benchmark machine learning algorithms, including Random Forest, Support Vector Machine, Extreme Gradient Boosting, and Categorical Boosting, were applied to evaluate the incremental effects of different inventory integration strategies on the LSM performance. Results indicate that adding field survey data improves model reliability (AUC = 0.85), while incorporating InSAR-derived hotspots substantially enhances spatial coverage of landslides (AUC = 0.87), mitigating the geographical bias of LSM. The highest predictive accuracy (AUC = 0.89) was achieved by combining all three landslide inventories, capitalizing on the high fidelity of field data and the broad spatial reach of InSAR. This synergistic integration improves both sensitivity and precision in identifying landslide-prone slope units. The proposed framework offers a transferable approach for generating LSM in data-scarce regions.