Abstract. Radar has consistently been proven to be the most reliable source of information for the remote detection of hail within storms in real time. Currently, existing hail detection techniques have limited ability to clearly distinguish storms that produce severe hail from those that do not. This often results in a prohibitive number of false alarms that hamper real-time decision-making. This study utilises convolutional neural network (CNN) models trained on dual-polarisation radar data to detect severe-hail occurrence on the ground. The morphology of the storms is studied by leveraging the capabilities of a CNN. Two datasets of images of 60 km × 60 km containing 19 different radar-derived features are built. The first is created from severe-hail cases (≥2 cm), and the second is obtained from rain or small-hail cases (rain or hail <2 cm) selected with the help of a cell identification algorithm above densely populated areas with no hail reports. After a tuning phase on the CNN architecture and its input size, the CNN is trained to output one probability of severe hail on the ground per image of 30 km × 30 km. A test set of 1396 images between 2018 and 2023 demonstrates that the CNN method outperforms state-of-the-art methods according to various metrics. A feature importance study indicates that existing radar-based hail proxies as input features are beneficial to the CNN, particularly the maximum estimated size of hail (MESH). The study demonstrates that many of the existing hail proxies can be adjusted using a threshold value and a threshold area to achieve better performance. Finally, the output of 10 fitted CNN models in inference mode on a hail event is shown.