Original title: Exploring the role of input data on hail nowcast skill using spatiotemporal neural networks
George Pacey, Ophélia Miralles, Ulrich Hamann, Olivia Martius · September 2026
Hail damages property and threatens public safety, so people need fast, accurate short-term forecasts called nowcasts. Scientists often use deep-learning models, which are computer systems that learn patterns from vast sets of past observations, to predict storms. Usually, researchers try to improve forecasts by redesigning these models. Here, researchers kept the model architecture unchanged and tested how altering the input data affects predictions. They changed the total training years, altered image orientation using mirroring and rotation, and varied the number of past timesteps provided. They found that training on more years of data boosted forecast skillA measure of forecast accuracy compared to a simple baseline. by up to 25 minutes at later lead times. Flipping and rotating data helped large datasets, but it hurt smaller ones. Overall, data choices greatly improve predictions.
You are learning to identify a rare bird by its call. With only one or two recordings, you often confuse it with other birds. But after listening to a hundred recordings from different seasons, you learn the subtle variations and can spot it from further away and before it even gets close. Here, the hail nowcast model gains up to 25 extra minutes of reliable prediction simply by training on more years of storm radar data.
Hail damages crops, cars, and buildings each year, and every extra minute of warning helps people and businesses take cover. This work shows that improving the dataset — not just the model — can add up to 25 minutes of reliable forecast time, at no extra computational cost. Agencies can focus on gathering more historical data rather than building fancier models.
This study only tests one model architecture and one type of data (radar). It does not explore other forms of augmentation or input data from different sensors. The experiments use a single geographic region and the results may not generalize. The reason augmentation hurts with small datasets is not explained.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Exploring the role of input data on hail nowcast skill using spatiotemporal neural networks — George Pacey, Ophélia Miralles, Ulrich Hamann et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.