Automatic AI Severe Weather Alerting from Radar Data

Our research endeavors have recently been dedicated to advancing precipitation nowcasting through physically constrained recurrent neural networks. Our motivation stems from our vision to develop tools for protecting people everywhere from the effects of climate change and dangerous weather. One of the foundational aspects is the establishment of high-resolution, precise predictions computed in real time after new measurement becomes available. However, recognizing the necessity of delivering severe weather information to the general public lacking proficiency in interpreting weather radar data, we emphasize the need for supplementary tools to achieve our goals. Presented within this study is the culmination of our efforts - the Meteopress Severe framework, designed explicitly for the automated issuance of alerts on severe weather events. This framework is an extension of our existin radar prediction models, adapted to predict the probability of occurrence of such events. The probability is computed as a secondary output of the network, thus building on the model's existing latent representation of the weather situation. We demonstrate this ability of the network by accurately predicting the likelihood of precipitation surpassing defined intensity thresholds. The Severe framework has two integral components, the first being the automated issuance of alerts at a temporal resolution as fine as 1 minute, explicitly created for any point of interest under the radar coverage. Secondly, our framework ensures a coherent visual presentation of probability predictions to forecasters, aiding their decision-making processes. This presentation encapsulates our journey in preparing software and prediction models for safeguarding societies against the impacts of severe weather events anywhere in the world.