Landslide hazard zonation using heuristic and machine learning approaches for steep terrains

Abstract

Landslides are a frequent and destructive hazard in the Nilgiris District of Tamil Nadu, where steep slopes, fragile geological formations, intense rainfall, and increasing human activity combine to create highly unstable conditions. This study developed a comprehensive landslide hazard zonation framework by integrating traditional heuristic and machine learning approaches. A total of 509 validated landslide locations and 509 non-landslide points were used for model development and validation. Twelve landslide conditioning factors, namely slope, aspect, elevation, lithology, geomorphology, soil, land use/land cover, rainfall, stream density, lineament density, topographic position index, and distance from roads, were derived from multiple geospatial datasets and incorporated into the analysis. Three modeling approaches, weighted index overlay analysis, gradient boosting machine (GBM), and Extreme gradient boosting (XGBoost), were implemented to generate landslide susceptibility maps. Model validation using independent datasets and confusion matrix metrics demonstrated that XGBoost outperformed the other approaches, achieving an accuracy of 0.78, recall of 0.97, F1-score of 0.82, and an AUC of 0.86. GBM also exhibited strong predictive capability, whereas WIOA produced comparatively lower accuracy (0.67) and lacked probabilistic discrimination. Shapley additive explanations (SHAP) analysis identified slope, rainfall, lithology, distance from roads, and lineament density as the dominant factors influencing landslide occurrence, emphasizing the combined role of terrain characteristics, hydrological conditions, geological controls, and anthropogenic disturbances in slope instability. Spatial analysis revealed that very high hazard zones were concentrated in Kundah Taluk, Coonoor, Kotagiri, Manjoor, and the upper Gudalur basin, where steep slopes, weak lithological formations, dense drainage networks, and intensive human interventions coincide. The results demonstrate the superiority of ensemble machine learning approaches over conventional heuristic methods for landslide hazard assessment and provide a robust framework for disaster risk reduction and sustainable land-use planning in mountainous regions.