A meta-stacking ensemble framework for landslide susceptibility mapping using LightGBM, Histogram Gradient Boosting and Decision Tree

Abstract. Landslides are a major natural hazard in mountainous regions, causing substantial socio-economic losses and posing persistent threats to infrastructure and human safety. This study introduces a Meta-Stacking Ensemble model for advanced landslide susceptibility mapping in the Darjeeling Himalayas, India. The proposed framework integrates LightGBM, Histogram Gradient Boosting, and Decision Tree algorithms through a stacking approach that maintains the original geospatial features while enhancing ensemble diversity. In this way, fourteen key conditioning factors (i.e., topographic, geological, hydrological and anthropogenic variables) were analyzed. Validation using 1830 landslide polygons demonstrated the model’s superior predictive performance, achieving an AUC of 0.93, overall accuracy of 87%, Recall of 0.84 and F1-score of 0.72, outperforming all individual base models. Spatial analysis indicated that 22% of the area falls within the "Very High" risk zone, with slope gradient (28% importance) and proximity to tectonic faults (30%) identified as dominant controlling factors. The developed framework produces GIS-compatible susceptibility maps with quantified uncertainty metrics, providing valuable insights to support disaster risk management and mitigation planning.