Classifying convective surface wind gusts from polarimetric radar data

Thunderstorm cells generating severe straight-line winds pose significant risks to infrastructure in Switzerland, where current warning methods lack specific detection capabilities for these events.  Five high-impact thunderstorm cases (2019–2023) were analyzed to identify indicative radar signatures, revealing that descending reflectivity cores, specific differential phase (KDP) cores, and midlevel radial convergence were the most consistent radar-based indicators of downbursts. A climatology of gust-producing cells was established by integrating 5-minute thunderstorm tracking data with weather station data, crowd-sourced reports, and European severe weather database records from May–October 2019–2023. This dataset was combined with composite radar metrics to train a Random Forest model for a binary classification into gust-positive and gust-negative cells. Input features included thunderstorm tracking attributes and the identified radar features, supported by the case studies. Lightning and convergent radial shear emerged as top predictors in analyses of feature importance. The model achieved a moderately skillful probability of detection and a very low false alarm ratio, while maintaining a good critical success index, highlighting that the model struggled to classify some gust-positive cells correctly, but rarely misclassified a gust-negative cell. This approach demonstrates the potential of improving nowcasting systems in Switzerland by applying machine learning on radar data to advance severe weather detection and warnings.