COALITION-4: Probabilistic Multi-Hazard Thunderstorm Nowcasting with Deep Learning for Robust Automated Early Warnings

Thunderstorms pose serious risks through lightning, heavy rainfall, hail, and strong winds. These events develop rapidly over highly localized areas, making timely short-term forecasts essential. Deep learning has proven particularly effective for thunderstorm nowcasting by rapidly learning spatiotemporal patterns from diverse observational datasets, enabling precise multi-hazard predictions within seconds - well-suited to operational early warning systems.

COALITION-4 is a deep learning nowcasting algorithm based on an encoder-forecaster architecture with recurrent convolutional layers. It nowcasts thunderstorm-related hazards - accumulated precipitation, lightning occurrence, and hail probability - up to 60 minutes ahead at 1 km and 5 minute resolution over the full domain of Switzerland. The operational model ingests data from the Swiss dual-polarization radar network, Météorage lightning observations, and a digital elevation model. The integration of NWP forecasts and satellite imagery has also been tested at the development stage, but is not used in the first version of operational implementation. Extensive preprocessing, data augmentation, and GPU-accelerated training with an adaptive learning rate ensure robust generalization across diverse convective conditions.

We present a comprehensive validation covering six convective seasons, enabling a statistically robust assessment of model skill across a wide range of convective regimes. The evaluation framework has been extended beyond standard grid-based metrics to explicitly quantify the effective lead time of warnings prior to thunderstorm onset - a metric of direct relevance to user preparedness and protective action. Results are compared both quantitatively and qualitatively against the previous operational nowcasting system at MeteoSwiss.

Since its operational deployment, robustness and reliability have been substantially improved through continuous quality monitoring, automated fallback mechanisms for degraded or missing input data, and iterative refinements informed by forecaster feedback. Building on this foundation, an automated warning pipeline is being developed that will deliver push notifications directly to the general public via the MeteoSwiss mobile app, alongside existing support for civil protection agencies, fire brigades, and aviation ground operations. This service is planned to become operational in summer 2026, representing a key step in closing the warning value chain from nowcast to societal response.

Looking ahead, we describe a significant advancement in the ML training procedure: a substantially expanded training dataset combined with a revised sample selection strategy that better represents high-impact convective events. This update yields a marked and consistent improvement in nowcasting skill across all hazard types and lead times, and points toward further enhancements in the accuracy and reliability of automated thunderstorm warnings.