Machine Learning Investigation of Downburst-Prone Environments in Canada

Abstract Thunderstorms are recognized as one of the most disastrous weather threats in Canada because of their power to cause substantial damage to human-made structures and even result in fatalities. It is therefore essential for operational forecasting to diagnose thunderstorms that generate damaging downdrafts of negatively buoyant air, known as downbursts. This study develops several machine learning models to identify environments supportive of downbursts in Canada. The models are trained and evaluated using 38 convective parameters calculated based on ERA5 reanalysis vertical profiles prior to thunderstorms with (306 cases) and without (19 132 cases) downbursts across Canada. Various resampling techniques are implemented to adjust data imbalance. An increase in the performance of the random forest (RF) model is observed when the support vector machine synthetic minority oversampling technique is utilized. The RF model outperforms other tested models, as indicated by model performance metrics and calibration. Several model interpretability methods highlight that the RF model has learned physical trends and patterns from the input variables. Moreover, the thermodynamic parameters are deemed to have higher impacts on the model outcomes compared to parcel, kinematic, and composite variables. For example, a considerable rise in the downburst probability is detected with an increase in cold pool strength. This study serves as one of the earliest attempts toward the fledgling field of machine learning applications in weather forecasting systems in Canada. The findings suggest that the developed model has the potential to enhance the effectiveness of issuing severe thunderstorm warnings in Canada, although further assessment with operational meteorologists is needed to validate its practical application. Significance Statement Severe thunderstorms demand particular attention in forecasting as their outflow can pose a serious threat to both structures and human life. This study uses machine learning techniques to predict whether or not a thunderstorm generates a damaging outflow in Canada. Atmospheric conditions that could trigger a severe thunderstorm are identified and discussed. Results show that the models have the potential to assist forecasters in better analyzing and predicting thunderstorms that generate destructive winds. Consequently, taking advantage of promising machine learning tools can yield more reliable forecasts of damaging thunderstorms, thereby mitigating the economic and societal burdens of these storms on Canadian communities.