XGBoost-Based Avalanche Activity Prediction with Integrated Rescue Optimization for the Western Himalayas

Avalanches in the Western Himalayas kill dozens of people each year, with fatalities concentrated on a handful of dangerous days along transport corridors and high-altitude posts. Operational forecasting in this setting is difficult for three reasons. Avalanche days are rare compared with quiet ones, so a model can appear highly accurate simply by never predicting an event. Physical snowpack simulations demand dense inputs and heavy computation, and lose skill where weather stations are sparse. And forecasters will not commit rescue teams on the strength of a model they cannot inspect. The costs of error are also lopsided: a missed avalanche can be fatal, whereas a false alarm typically results in a road closure. A useful system must therefore lean toward detection, run quickly on existing weather feeds, and explain its reasoning. This work presents an end-to-end, leakage-controlled framework that pairs XGBoost with fold-wise SMOTE for the rare positive class, SHAP for auditable explanations, probability calibration, and a rescue-allocation layer. Because open operational Himalayan records do not exist, the framework is developed and validated on the open EnviDat Swiss wet-snow benchmark (3,653 station-days, 66 stations, 2001–2020, with 11.1% avalanche days), the closest openly licensed analog. Imputation and SMOTE are fit within time-ordered cross-validation folds, so no future or synthetic information leaks into training, and the most recent 20% of the record is held out as an untouched test set. Rather than being fixed at 0.5, the decision threshold is swept on out-of-fold data to maximize the Kuipers skill score. On the held-out season, XGBoost with SMOTE achieves the best balance of the two costliest errors, reaching a precision of 0.63, recall of 0.85, F1 of 0.72, a Kuipers skill score of 0.71, and a precision–recall AUC of 0.89. Its SHAP ranking recovers the physics of wet-snow release, led by air temperature and followed by solar radiation, slope aspect, and liquid-water fraction. One caveat is retained honestly: a class-balanced logistic regression is a strong baseline on this feature set, which suggests that the SNOWPACK-derived features are already highly informative rather than indicating that the tree model should be discarded. Once trained, the model scores a day in milliseconds with no simulation in the loop, allowing a regional forecast to refresh as new weather arrives. Validation currently relies on Swiss data; transferring the framework to the Western Himalayas through regional partnerships and cross-range recalibration remains the next step.