Deep Learning vs. Traditional Satellite-Based Thunderstorm Nowcasting: Outline of a Model Benchmark Study

The ability to produce reliable short-range forecasts for thunderstorms is crucial for issuing timely emergency warnings to the general population, protecting vital infrastructure and alerting first responders in advance. Regularly affected by thunderstorms and associated hazards are remote areas, mountainous regions and air traffic, necessitating broadly available nowcasting solutions. Geostationary satellites provide an ideal data source for thunderstorm nowcasting in these cases.

Traditional approaches employ deterministic algorithms to predict the occurrence and evolution of thunderstorms essentially by solving the optical flow problem. Recently, machine learning (ML) has emerged as an alternative and already shown promising results. While U-Net-based architectures have proven to be a very robust baseline, ML-based methods currently represent an extremely active area of research and optimal approaches have yet to crystallise. Particularly for satellite-based thunderstorm nowcasting, the superiority of either traditional methods or a specific ML architecture has yet to be established.

It is our goal to narrow this knowledge gap by performing an extensive benchmark. In this talk we present a preliminary study featuring a set of nowcasting tools continuously improved over the past 20 years at the German Aerospace Centre (DLR) as well as a newly developed ML-model. The presented evaluation pipeline is developed jointly with the German Meteorological Service (DWD) and the regional Italian meteorological service Arpae Emilia-Romagna as part of the Italia-Deutschland Science-4-Service network. It serves as a baseline for a broader follow-up study including satellite-based thunderstorm nowcasting operationally used at DWD and Arpae. All models are evaluated over a period of multiple years for a fixed region in central Europe, using only satellite imagery from the SEVIRI instrument onboard MSG as input, while lightning recorded by the LINET network serves as ground truth. The accuracy of each model is measured for a set of lead times up to 180 minutes according to established metrics in a homogeneous setting.

Our work aims to guide future research and development efforts, ultimately paving the way for improved satellite-based thunderstorm nowcasting.