Diagnosing Downburst-prone environments with machine learning

Every time that deep convection develops, the Mediterranean region faces the possible impacts of downburst-induced wind extremes that originate within, and the sub-km scale of these events is below the resolution of operative numerical weather prediction models. However, pre-convective environments often provide an opportunity to assess the possibility of thunderstorms developing downbursts, and recent studies have shown that machine learning provides skillful approaches to derive severe weather probabilities from these environments. However, pre-convective environments often provide an opportunity to assess the possibility of thunderstorms developing downbursts, and recent studies have shown that machine learning provides skillful approaches to derive severe weather probabilities from pre-convective environments.

In this study, we employ a machine learning model to derive the probability of downburst occurrence, derived from convective parameters characterizing pre-convective environments. The model has been trained and tested on 10 years of downbursts and non-severe thunderstorms over Spain that have been identified using lightning records, public reports of severe weather and automated weather stations; with the ERA5 reanalysis providing a representation of the atmospheric structure associated with these events. 

Several training procedures have been explored in order to achieve model robustness and avoid biases arising from the underrepresentation of public-source reports in rural areas. Regarding sample generation, these include spatially uniform subsamples of non-severe thunderstorms, oversampling of the minority class of downbursts or restrictions to common-points in downbursts and non-severe thunderstorms. To reduce noise and overfitting risk arising from the large number of candidate convective parameters, a forward selection retains only parameters that improve model performance. The hyperparameters of the model and resampling methods are tuned by a cross-validation based on permutating left-apart years. Finally, the model is tested on new data covering 2 years, measuring its performance with metrics and diagrams adequate for rare phenomena such as False Alarm Rates (FAR), the Critical Success Index (CSI), the Performance-Diagram and its Area Under the Curve (AUPDC) or the attributes diagram.

The most skillful configuration of the model performs better than any individual convective parameter, in terms of performance metrics. Explainability methods and convective parameter selection are coherent with the physical knowledge of downburst winds, highlighting the role of a warm unstable lower-troposphere favoring downdrafts, as well as strong mid-level winds that can be transported downward. Probability estimations improve those of a random model based on the observed climatological frequency of downburst in thunderstorms, although probability outputs close to 1 show elevated uncertainty, as indicated by bootstrap confidence intervals. Another aspect to improve concerns the elevated number of false alarms, a consequence of class imbalance (estimated in 3 downbursts per 100 thunderstorms).

The applications of this model include the coupling with km-scale models, which cannot directly resolve sub-km downbursts but can provide high-resolution depictions of the pre-convective environments that the model takes as input.