Debris flow susceptibility and propagation modeling: a deep learning and flow-R framework

Debris flows are among the most destructive natural hazards, characterized by rapid initiation and movement, posing significant challenges for accurate prediction and mitigation. In the Gilgit District of Pakistan, particularly along the Karakoram Highway, North Pakistan, debris flows frequently disrupt transportation routes, damage infrastructure, and hinder economic development. This study developed a comprehensive inventory of 64 debris flow events from Jaglot to Gilgit City. Fourteen causative factors were generated from remote sensing and topographic data. They were evaluated for their relative importance using a Random Forest (RF) classifier, Variance Inflation Factor (VIF), Tolerance, and SHAP (SHapley Additive exPlanations) identifying rainfall, elevation, and lithology as the most influential predictors. Three DL models, recurrent neural networks (RNN), artificial neural networks (ANN), and convolutional neural networks (CNN), were trained to generate debris flow susceptibility mapping (DFSM), where the ANN model achieved the highest predictive accuracy (AUC = 0.924). Furthermore, the susceptibility output was coupled with Flow-R modeling to evaluate spatial runout behavior, simulating debris flow propagation at a regional scale using the ANN-derived “very high” susceptibility zones as the initiation source. The results indicated that approximately 13.15% of the area falls under very high propagation susceptibility, 22.94% under high, and 63.91% under low, emphasizing areas at significant risk of runout impact. The results strongly corresponded between simulated flow paths and field observations, validating the approach. The resulting propagation patterns demonstrate significant spatial alignment with mapped debris flow paths, enhancing the practical applicability of the integrated approach. In areas where frequent debris flows occur, this combined framework provides a robust basis for identifying initiation and potential impact zones, facilitating more effective hazard mitigation and land-use planning.