Application of Machine Learning to Severe Weather Prediction from Storm Top Indicators

We present a few case studies demonstrating the application of a neural network (NN) model for detecting overshooting tops (OTs) using high-resolution visible (HRV) channel from the SEVIRI instrument on MSG satellites. The model was trained on manually labeled data using OT shadows (database created by Ján Kaňák) and the product of this model provides per-pixel probabilities of OT presence.

To assess the model's relevance for severe weather forecasting, we compared the detected OTs with reports from the European Severe Weather Database (ESWD). The comparison showed varying levels of agreement - several OTs corresponded well with hail or another type of events, while in other cases, strong convection was detected without reported impacts, and vice versa. This may be due to multiple factors on both sides - limitations in our model’s predictions as well as in the event database (e.g., missed reports).

This variability highlights both the usefulness and the limitations of OT detection as a proxy for severe weather. The model performs well in identifying deep convective features but should be interpreted alongside other data sources for operational use.

Our results suggest that ML-based OT detection from HRV imagery can contribute to nowcasting applications, especially when integrated with additional observational and model data.