Debris Flow Susceptibility Mapping Using Explainable Deep Learning in Eastern Hindukush Pakistan

ABSTRACT

Debris flows are among the most destructive geomorphological hazards in mountainous regions, causing damage to human settlements, infrastructure and ecosystems. The Eastern Hindu Kush region of Pakistan is particularly vulnerable due to steep topography, intense monsoonal precipitation, active tectonics and increasing anthropogenic disturbance. This study develops an integrated debris flow susceptibility assessment framework by combining convolutional neural networks (CNN) with SHapley Additive exPlanations (SHAP). A total of 13 conditioning factors representing topographic, hydrological, geological, environmental and anthropogenic characteristics were prepared using GIS and remote sensing datasets at a spatial resolution of 30 × 30 m. The CNN model demonstrated strong predictive performance, achieving an area under the curve (AUC) of 0.94, overall accuracy (ACC) of 0.86, precision of 0.84, recall of 0.88, an F1‐score of 0.86 and a Kappa coefficient of 0.71. The resulting susceptibility map identified approximately 26.3% of the study area as high to very high susceptibility zones, predominantly concentrated along steep river valleys and highly dissected mountainous terrain. The susceptibility map was further validated by debris flow density analysis, which showed hazard density increasing from 0 events/km ² in very low zones to 0.048 events/km ² in very high zones, with most historical debris flow events concentrated within the very high susceptibility class. SHAP analysis identified slope, NDVI, rainfall, distance to roads and distance to streams as the most influential conditioning factors controlling debris flow occurrence. The proposed CNN–SHAP framework provides a robust and interpretable approach for regional‐scale debris flow susceptibility assessment, with applicability to similar mountainous environments. These findings provide valuable information for hazard mitigation, infrastructure planning and land‐use management in data‐scarce mountain regions.