Heavy rainfall can trigger natural disasters such as floods or landslides.
Global climate models are necessary to predict the changes in the frequency of these extremes that are expected due to climate change. In a study, researchers at the Karlsruhe Institute of Technology (KIT) show for the first time a method based on artificial intelligence (AI) that can be used to increase the accuracy of the rough precipitation fields generated by global climate models.
They managed to improve the spatial resolution of precipitation fields from 32 to two kilometers and the temporal resolution from one hour to ten minutes. This increased resolution is necessary in order to be able to better predict the more frequent occurrence of local heavy precipitation and the resulting natural disasters in the future. (DOI10.1029/2023EA002906)
Many natural disasters such as floods or landslides are a direct result of extreme rainfall. Researchers expect that extreme precipitation will continue to increase as average temperatures rise. In order to adapt to a changing climate and prepare for disasters early, accurate local and global information about the current and future water cycle is essential. “Precipitation is highly variable both spatially and temporally and is therefore difficult to predict – especially at a local level,” says Dr. Christian Chwala from the Institute for Meteorology and Climate Research – Atmospheric Environmental Research (IMK-IFU), KIT’s Alpine Campus in Garmisch-Partenkirchen. “That’s why we want to increase the resolution of precipitation fields, such as those generated by global climate models, and thus, above all, improve their classification with regard to possible threats such as flood disasters.”
TrendingGreentech.LIVE Conference 23.-24.11.2023 (Fall Edition)
Finer resolution for more accurate regional climate models
Current global climate models use a grid that is not fine enough to accurately represent the variability of precipitation. High-resolution precipitation maps can only be generated using extremely computationally intensive and therefore spatially or temporally limited models. “We have therefore developed a generative neural network – called GAN – from the field of artificial intelligence and trained it with high-resolution radar precipitation fields. The GAN learns how to generate realistic precipitation fields and their temporal sequence from roughly resolved data,” explains Luca Glawion from IMK-IFU. “This means the network is able to create realistic, high-resolution radar precipitation films from the very coarse resolution maps.” These refined radar maps not only show how rain cells develop and move, but also precisely reconstruct the local rain statistics with the corresponding extreme value distribution.