Domain shift adaptation methodologies for statistical modeling of hailstorm occurrence in Europe

Hailstorms are among the costliest extreme weather events, particularly for sectors like agriculture. However, physically representing hailstorms in models is challenging, as they depend on high-resolution, sub-grid deep convective processes that are both difficult and expensive to simulate. To address this, some approaches based on artificial intelligence, and especially Machine Learning (ML), have recently emerged to bypass some of these limitations of physical models. These approaches usually rely on hailstorm occurrence data from reports as the model’s target. This data is often scarce and exhibits spatial and temporal biases. Although there is a relatively large and consistent database of hail reports in the US, such data is much scarcer in Europe. The European Severe Weather Database contains tens of thousands of reports, but the hail reporting rate has only increased in recent years and remains low in some countries. Since these ML methodologies require large and consistent datasets, it is challenging to construct statistical model based only on European data. In this study, we explore how to build ML classifiers for hailstorm occurrence in data-scarce regions like Europe, using datasets from different regions and applying domain shift adaptation techniques, alongside local data.

We first describe our baseline ML model, which uses meteorological variables from ERA5 reanalysis associated with deep convection, and hailstorm occurrence data from the US. The model is trained to learn the relationship between these variables and hailstorm occurrence, and is then used to infer the probability of such events in locations and days not seen by the model.

Next, we explore how to build a similar hailstorm occurrence classifier for Europe. We compare several approaches: applying the baseline US classifier in Europe, training a separate classifier on Europe’s limited data, and using domain shift adaptation techniques to combine both datasets. We tested direct dataset mixing and a more advanced approach where the model is pre-trained on US data—where it is more abundant—and fine-tuned on Europe’s limited dataset.

We benchmark these different techniques in Europe and the US, offering insights into how to build generalizable ML models for hailstorm occurrence. We also present a calibration method to ensure model output accuracy and show how the model can be used to construct hailstorm hazard maps for use by various stakeholders.