Towards Improved Hail Detection and Size Estimation Using Convolutional Neural Networks

Hail is a significant severe weather hazard in France, capable of causing substantial damage to agriculture, vehicles, infrastructure and solar installations. For example, a severe hailstorm over Paris on 3 May 2025 resulted in insured losses exceeding €300 million. Despite the operational use of weather radar for hail detection and size estimation, current methods (such as fuzzy logic hydrometeor classification algorithms or hail proxies using vertical profiles of reflectivity) remain limited in their ability to estimate the occurrence of severe hail (>2 cm) and discriminate its size on the ground. Notably, they overlook information embedded in storm morphology.

At Météo-France, a new approach using convolutional neural networks (CNNs) has been developed to better exploit storm morphology, initially for severe hail detection. This method uses a single-timestep input comprising 19 radar-derived features, including quality-controlled polarimetric variables and storm severity diagnostics (e.g. ECHOTOP45, VIL, MESH…). These are fed into a CNN trained to predict severe hail occurrence over 30x30 km² areas. Hail cases were selected using the ESWD, while non-hail cases came from storms over densely populated areas without hail reports. Experimental results show that CNNs outperform existing hail proxies, especially when those proxies are used as input features (Forcadell et al., 2024).

This methodology has been adapted for estimating hail size, which is framed as a multi-class classification problem involving three categories: medium (20–35 mm), large (35–50 mm) and giant (≥50 mm) hail. A new radar dataset was created by extracting image sequences centred on convective cell centroids, which were tracked using a dedicated algorithm. Each sample spans six timesteps over 25 minutes, thus incorporating the storm’s temporal evolution. Several CNN architectures were tested; models using multiple prior timesteps proved more robust than those based on a single timestep. Feature importance analysis identified radar echo top as the most predictive input, followed by polarimetric variables below the freezing level.

These findings, detailed in Forcadell, V. (2024), form the basis for ongoing work. New datasets covering 2024 and 2025 hail events are currently being assembled to improve generalization and further validate the model. The methodology and updated evaluation results of this CNN-based hail size estimation algorithm will be presented.

References
Forcadell, V. (2024). PhD Thesis, Université de Toulouse. https://theses.hal.science/tel-05070872
Forcadell, V., et al. (2024). Atmos. Meas. Tech., 17(22), 6707–6734. https://doi.org/10.5194/amt-17-6707-2024