Machine learning tools for predicting multi-hazards caused by convective storms - the TAMIR project

<p>Convective storms and long-lasting mesoscale convective systems have the potential to cause heavy rainfall, flooding, hail, wind gusts and lightning that can result in significant damage to property and loss of lives. Accurate prediction of the location or severity of such storms (e.g. in the sub-kilometer resolution for the next hour) to assist the decision-making of civil protection authorities is beyond the capabilities of the current numerical weather prediction models. Thus, weather radar and machine learning-based methods provide an important tool to predict such events and their impacts in advance. Identifying a storm cell or system as an &#8220;object&#8221; from a radar image provides a natural way for associating different meteorological attributes of a storm with its impacts. In the TAMIR project funded by the EU Civil Protection Mechanism, we have implemented this by combining a cell tracking system with a machine learning model. The hazard levels of storms are estimated from their distance and time delay to the associated emergency reports obtained from the PRONTO database provided by the Finnish civil protection authorities. Using several meteorological attributes related to severe weather (e.g. lightning flash, hail and wind observations and indicators of convective potential), a random forest model was trained for predicting the storm hazard level. This was done by using a large sample of data during summer months between 2013-2020. The model for predicting the hazard level was verified by cross-validation. A Kalman filter-based methodology was applied for probabilistic nowcasting of future storm locations, which was combined with the model for hazard level prediction. Finally, the hazard nowcasts were combined with different exposure layers to translate them into prediction of impacts caused by convective storms. In the presentation, we demonstrate the added value of the implemented hazard and impact nowcast products with case studies. The products have also been evaluated by the Finnish civil protection authorities during the test period June-September 2021 with largely positive feedback. While the feasibility of the proposed methodology is demonstrated in Finland, discussion about its transferability to other parts of the world is also given.</p>