Machine learning (ML) is currently revolutionising weather prediction at the short- to medium-range timescale. Nowcasting using ML has attracted less attention in comparison, especially for convective hazards such as lightning, hail and extreme precipitation. Conventional nowcasting approaches for convective hazards are typically based on Lagrangian extrapolation by advecting radar or satellite observational fields. Limitations of this approach include difficulties representing the intensification and decay of convective systems as well as identifying convective initiation during the forecast window. ML presents a potential avenue to make significant improvements to existing approaches by leveraging the large amount of historical data available from different sources (e.g. dual-pole radar, Meteosat satellites, lightning networks). At MeteoSwiss, a probabilistic deep learning nowcast model for lightning, hail and precipitation is currently being operationalised for use in Switzerland. The model outperforms Lagrangian benchmarks when validating on one convective season. Furthermore, a seamless (nowcasting through to short- and medium-range) ML-based prediction system is envisaged for the coming years.Here, we extend the current framework and explore where further gains may be possible by investigating advanced network architectures, novel input features and diversifying the training data. We focus on the Swiss radar domain, which presents unique challenges due to complex alpine topography. Nowcasts are generated at a 1 km and 5 minute spatial and temporal resolution, respectively, up to a lead time of one hour. Our work contributes towards the continued improvement of ML-based nowcast models, providing vital guidance and damage mitigation for sectors including emergency services, aviation and the public.
Pacey et al. (2025) studied this question.