Estimating rainfall using Meteosat infrared data: generative models are better at predicting extremes, whilst deterministic models are better at predicting the average

Using fifteen years of Météo-France radar data, a team is comparing U-Net, Transformers, GANs and diffusion models to produce detailed precipitation maps. Deterministic models consistently underestimate heavy rainfall.

Geostationary satellites continuously observe the infrared brightness temperature of clouds, but this information is mainly based on the cloud tops, not on the rain that falls on the ground. Drawing a detailed estimate of precipitation, which is useful for hydrology and agriculture, therefore remains tricky. Matthieu Meignin, Cécile Mallet and Nicolas Viltard have systematically compared the main families of deep learning on this task, over metropolitan France.

A compromise between precision and realism

Their dataset covers 2008 to 2023 and combines the infrared channels of Meteosat Second Generation with the radar mosaics of Météo-France, taken as a reference, with sampling adapted to the intermittent and highly asymmetric nature of the rain. U-Net and transformer-based architectures give the best pixel-by-pixel scores and robust averages, but recurrently underestimate intense episodes. Conditional GANs and diffusion models better reproduce the entire distribution, including rare and violent showers, and produce more credible spatial structures, at the cost of less local accuracy.

For the authors, generative approaches are therefore to be preferred for detecting extreme events and for probabilistic uses. They propose a reproducible framework for evaluating these methods.