Machine Learning-Based Nowcasting of Convective Storm Impacts on a Pan-European Scale

Convective storms and the associated heavy rainfall, flooding, hail, wind gusts and lightning can result in significant damage to property and loss of lives. Thus, there is a need for accurate prediction of the future location and severity of such storms (i.e. in the sub-kilometer resolution for the next hour) to assist the decision making of civil protection authorities. The typical approach to produce such nowcasts is to identify storm cells as separate entities from radar images, which provides a natural way for associating the storm attributes with its severity. Following this approach, we have implemented a set of pan-European nowcast products in the TAMIR and EDERA projects funded by the EU Civil Protection Mechanism. This has been done by combining a cell tracking method with a random forest-based machine learning (ML) model and a Kalman filter model. In the nowcast products, storm cells are identified from the OPERA radar composites. The ML model is used for predicting the storm severity level. It is trained using a large database of meteorological features and weather hazard reports during summer months (May-September) between 2018-2022. The storm features include basic cell and track properties (i.e. area and age), lightning flash and wind observations, and also indicators of convective potential from ERA5 reanalyses. The target variable for training the ML model is the storm hazard level. The hazard level estimation is done based on the distances and time delays between the storms and the associated weather hazard reports obtained from the European Severe Weather Database (ESWD). We have trained different ML models against the hazard levels estimated for each event type (e.g. heavy rain, lightning and severe wind gusts). The above models are combined with a Kalman filter-based methodology to produce probabilistic nowcasts of future storm locations together with their severity level. The added value of such nowcasts for decision making is demonstrated with case studies and relevant verification metrics. Finally, we demonstrate how the nowcasts can be combined with different exposure layers to translate them into predictions of actual storm impacts.