Aims: This study used machine learning models, Random Forest (RF) and Boosted Regression Tree (BRT), to predict avalanche-susceptible areas avalanche-susceptible areas in the Central Alborz watershed, Alborz Province, Iran.
Materials & Methods: We established a polygon-based framework based on three main avalanche zones: starting, track, and runout. Three training scenarios were considered for avalanche susceptibility modeling using RF and BRT: (1) the runout zone alone, (2) the starting and track zones, and (3) all three zones combined. We randomly split the avalanche inventory into 70% training and 30% validation datasets. We initially considered 21 topographic, geomorphological, and meteorological factors. We first conducted multicollinearity and Pearson correlation analyses to identify redundant predictors. We then applied Recursive Feature Elimination (RFE) with the Random Forest (RF) algorithm and out-of-bag (OOB) permutation importance to select the optimal predictor subset, retaining 13 factors for model development. We trained models under the three scenarios and evaluated them using the Receiver Operating Characteristic (ROC) curve, precision, sensitivity, F1 score, and Matthews Correlation Coefficient (MCC).
Findings: The BRT model trained with the starting and track zones showed the best overall performance across most evaluation metrics (AUC = 0.95, precision = 0.93, F1-score = 0.91, and accuracy = 0.91).
Overall, training both BRT and RF models using the starting and track zones substantially improved snow avalanche susceptibility prediction compared with the other scenarios. For the starting and track, runout, and all-zone scenarios, the RF model achieved AUC ≥ 0.90, while the corresponding BRT reached AUC ≥ 0.92 for all three scenarios.
Conclusion: The results showed that the BRT and RF models achieved their highest predictive performance under the starting and track zones scenario, with AUC values of 0.95 and 0.93, respectively. These findings indicate that incorporating the spatial characteristics of avalanche zones can improve susceptibility assessment and provide useful information for avalanche hazard management and disaster risk reduction in mountain watersheds.