Landslide susceptibility mapping based on Inter.iamb-Tabu algorithm considering non-landslide sampling

Abstract

Accurate landslide susceptibility mapping (LSM) remains challenging because of uncertainty in both model selection and non-landslide sample definition. To address these issues, this paper developed an LSM framework based on the Inter.iamb-Tabu algorithm while explicitly considering non-landslide sampling, using Gangu County, China, as the research area. 124 landslides were established through remote sensing interpretation and field survey. From 14 initial conditioning factors, nine were retained after correlation analysis, single-factor logistic regression and variance inflation factor analysis. Using 10 m×10 m grid cells as the mapping unit, a spatially random strategy was designed to select 26,468 non-landslide samples, and the datasets were divided into training and validation subsets at a ratio of 7: 3. Three improved Bayesian network algorithms, MMPC-Tabu, Fast.iamb-Tabu and Inter.iamb-Tabu, were then compared using Accuracy, Precision, Recall, F1-score and AUC. The results showed that Inter.iamb-Tabu achieved the best predictive performance. In addition, the proposed non-landslide sampling strategy outperformed two conventional methods, confirming its superiority in preserving spatial representativeness and reducing sample contamination. The final susceptibility map indicated that 87.1% of the documented landslides were in high and extreme susceptible areas. These areas were mainly distributed in the loess hilly region, the steep front margins of thick loess layers on the northern and southern mountains, and the secondary and tertiary gullies of the Weihe River valley. The learned DAG and CPTs revealed that slope gradient, elevation, land use and distance from fault were parent nodes of landslide and exerted direct triggering effects. Profile curvature and NDVI served as mutual feedback nodes with landslides, whereas slope aspect, distance from road and SPI mainly exerted indirect effects by interacting with other hazard factors. The proposed framework provides a robust, interpretable and transferable approach for LSM, and can support geohazard mitigation and land-use planning in regions with similar geological and geomorphological settings.