Thunderstorm Nowcasting With Deep Learning: A Multi‐Hazard Data Fusion Model

• Published in: Geophysical research letters 2023-04, Vol.50 (8), p.n/a
• Main Authors: Leinonen, Jussi, Hamann, Ulrich, Sideris, Ioannis V., Germann, Urs
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Aviation; Data integration; Data sources; Deep learning; Digital Elevation Models; Digital imaging; Emergency response; Explainable artificial intelligence; Extreme weather; floods; Hail; Hazards; Heavy precipitation; Infrared imagery; Infrastructure; Lightning; Lightning detection; Machine learning; Meteorological radar; Modelling; Nowcasting; Numerical prediction; Numerical weather forecasting; Precipitation; Predictions; Probability theory; Radar; Radar data; Radar detection; Resolution; Satellite imagery; Storm forecasting; Temporal resolution; Thunderstorms; Weather; Weather forecasting; Weather radar
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2022GL101626

Predictions of thunderstorm‐related hazards are needed in several sectors, including first responders, infrastructure management and aviation. To address this need, we present a deep learning model that can be adapted to different hazard types. The model can utilize multiple data sources; we use data from weather radar, lightning detection, satellite visible/infrared imagery, numerical weather prediction and digital elevation models. We demonstrate the ability of the model to predict lightning, hail and heavy precipitation probabilistically on a 1 km resolution grid, with a temporal resolution of 5 min and lead times up to 60 min. Shapley values quantify the importance of the different data sources, showing that the weather radar products are the most important predictors for all three hazard types. Plain Language Summary Thunderstorms are hazardous to both people and property through various extreme weather phenomena. Predicting these hazards allows individual people, infrastructure managers and emergency services to take action in advance. To serve their needs, we use a model based on an artificial intelligence (AI) to predict the probability of the hazards occurring at a given time and place during the next 60 min. The model uses multiple sources of weather observations and predictions to construct the predictions, and can be adapted to function also when some of these sources are unavailable, increasing its reliability. We show that the model can predict the occurrence of lightning, hail and heavy precipitation, detecting and predicting the motion of thunderstorms as well as whether they are increasing or decreasing in severity. We use explainable AI methods to determine how much each of the data sources contributes to the predictions, showing that weather radar observations are the most important source of data. Key Points We present a deep learning model for nowcasting thunderstorm hazards, and demonstrate it for lightning, hail and heavy precipitation The model can provide probabilistic warnings of these hazards on a two‐dimensional grid We analyze the importance of the different data sources used in the model using explainable artificial intelligence methods