Predictive hazards from convective systems with deep learning

Convection is a complex spatiotemporal process, which has made it a particularly attractive application for deep learning, which excels at both spatial and temporal reasoning. We have developed deep learning models for predicting the occurrence of hazards caused by convective storms, so that this information may be used by forecasters, emergency services and infrastructure managers to respond to the threats caused by these hazards.

Our network is based on a recurrent-convolutional architecture that can process input data at multiple resolutions. It issues probabilistic predictions of hazard occurrence, currently up to 1 hour to the future. As inputs, we use data from weather radars, geostationary satellites, ground-based lightning detections, numerical weather predictions and digital elevation models. We have studied the importance of each data source to the quality of the predictions, finding that radar-based inputs contribute most to the prediction quality; however, some hazards can be well predicted also without radar, indicating that it is plausible to create warning systems for these hazards in areas where radar networks are not available.

In this presentation, we will describe the model architecture and case studies, as well as our experiences so far in bringing the model to real-time use by forecasters and automated warning systems at MeteoSwiss. We will also discuss future directions of this research.