Physical Scales Matter: The Role of Receptive Fields and Advection in Satellite-Based Thunderstorm Nowcasting with Convolutional Neural Networks

Machine learning has a significant impact on severe weather nowcasting, with purely data-driven 
models quickly moving into the focus of current research over traditional physically motivated 
methods. Yet, recent work on radar-based precipitation nowcasting suggests that blending physical 
insight—like advection—with deep learning can improve forecast skill. In this talk, we will present 
our attempt to test how general this idea really is by applying it to satellite-based thunderstorm 
nowcasting for the first time.
The central question we address is when and why advection should improve ML-based forecasts. 
Using a simple scale argument, we show that advection through optical flow algorithms helps 
preserve storm patterns within a model’s receptive field, which is critical for nowcasts at longer lead 
times. To test this, we trained convolutional neural networks to nowcast thunderstorms based on 
satellite imagery and lightning observations, with and without incorporating advected inputs.
When considering average performance, the impact is modest. But once we break down the results 
by lead time and wind speed, a clear pattern emerges: the benefit of including advection emerges
after about 2 hours of lead time and grows with increasing wind speed, in agreement with our 
physical reasoning.
Our findings highlight the importance of considering physical scales when designing and evaluating 
ML-based forecasting systems.