Comparisons of statistical, machine learning, and numerical models for debris flow susceptibility mapping in the Northwestern Himalayas

Abstract: Debris flows triggered by climatic variability pose major threats to life and infrastructure in the northwestern Himalayas. This study maps debris-flow susceptibility and propagation along the Karakoram Highway in northern Pakistan. A field-validated inventory of 111 debris-flow locations was compiled, and 70 catchments were analyzed morphometrically, to evaluate ten conditioning factors. Then, a coupled modelling framework integrating Weighted Overlay, Random Forest, XGBoost, and Flow-R were applied. XGBoost achieved the highest predictive accuracy (AUC = 0.834), followed by Flow-R (0.822), Random Forest (0.756), and Weighted Overlay (0.710). Flow-R simulation of propagation from high and very high susceptibility zones provided realistic runout extents. The results demonstrate that machine-learning models, particularly XGBoost, combined with process-based Flow-R, offer superior performance for debris-flow hazard assessment and mitigation planning in high-mountain terrains.