A geostationary satellite AI application for severe weather warnings (MSO project)

Development of a deep learning framework for geostationary satellite data to provide high resolution, real-time severe weather warnings

One of the responsibilities of KNMI is the development and refinement of geophysical early warning capabilities. It is KNMI’s ambition to move from national and regional severe weather warnings towards local severe weather warnings ("postcode level"). Recent scientific developments using artificial intelligence/machine learning (AI/ML) have shown that computer vision methods combining reported severe weather events and stacks of American GOES geostationary satellite imagery can provide reliable near-real-time probabilistic predictions of several severe weather types. However, similar approaches have yet to be developed for the European geostationary satellites such as the current operational Meteosat Second Generation (MSG) satellites.

At the same time Europe plans to soon switch to the new Meteosat Third Generation imagers - MTG-I1, the first one, successfully launched 13 December 2022 - containing state-of-the-art instrumentation, including the Flexible Combined Imager (FCI) and the Lightning Imager (LI). MTG-I1 will provide data with better spatiotemporal resolution and better data quality over Europe than ever before. Its observations thus should enable the development of advanced data science applications towards local severe weather monitoring and warning.

This MSO project proposes to improve upon the approach described above and develop a similar satellite-observation-based European probabilistic NRT severe weather prediction system, starting with MSG data and in preparation for the switch to MTG. It builds on a previously developed proof of concept for lightning prediction by the Chief Investigators (CIs) in collaboration with the Center for Data Science at Leiden University.

Furthermore, this MSO project will deliver not one, but a suite of AI applications, providing early warnings through nowcasting of large hail, strong wind gusts, downbursts, heavy precipitation, tornados, and intense lightning. Also, although the satellite image channels that are required to identify severe weather may depend on the type of severe weather, the tooling, data pre-processing and computational pipelines will be similar.

This MSO proposal therefore also creates a generic more widely applicable deep learning framework for satellite data, building a foundation for artificial intelligence/machine learning (AI/ML) research with MSG, MTG and other satellite data at KNMI. Once developed this framework thus will open up many new opportunities for research and operations.