Machine Learning for Seamless Thunderstorm Nowcasting from Multiple Data Sources
Webinar with Jussi Leinonen (MeteoSwiss); Moderator: Mark Higgins (EUMETSAT)
6 October 2021; 11:00 UTC;
Machine learning has recently been used for nowcasting weather phenomena in several studies and applications. One of these applications is thunderstorm nowcasting, on which several studies have already been published. However, these studies usually make use of only one data source and concentrate on a single thunderstorm hazard. In order to bring these methods to real-world applications, we are working to develop thunderstorm nowcasting methods that make use of multiple sources of data simultaneously and can be trained to create products for different hazards according to the needs of the end user.
This presentation will present the results of two related nowcasting studies. First, we discuss our analysis of the usefulness of different data sources. This analysis was carried out using the gradient boosting method, which allow the value of different predictors to be quantified in straightforward fashion. We discuss the benefits of different data sources for precipitation, hail and lightning nowcasting. The data sources considered are derived from the NEXRAD radar network, the ABI (Advanced Baseline Imager) and the GLM (Geostationary Lightning Mapper) instruments on the GOES-16 satellite, the ECMWF numerical weather prediction (NWP) forecasts, and a digital elevation model (DEM). We consider various scenarios where some of the input data might be unavailable due to data outages or geographical limitations, and analyze the impact of these situations on the quality of the nowcast.
Second, we present the current focus of our efforts, building on the results of the above-mentioned work. The goal of this work is to create a framework based on deep learning for probabilistic nowcasts of the occurrence of hazardous events. We use convolutional neural networks to analyze the spatial structure of the atmospheric fields and recurrent networks to model their time evolution. The data sources are similar to the previous study, using radar, satellite and lightning data, NWP forecasts, and static data such as the DEM. Finally, we will discuss future directions for the work and connections to other studies being carried out concurrently. The end goal is a seamless nowcasting system that can make optimal use of both near-real-time observations and NWP data and which can also be adapted for use in other applications besides thunderstorm nowcasting.