糖心视频


Researchers develop AI model to generate global realistic rainfall maps

Researchers develop AI model to generate global realistic rainfall maps
Working from low-resolution global precipitation data, the spateGAN-ERA5 AI model generates high-resolution fields for the analysis of heavy rainfall events. Credit: Christian Chwala, KIT

Severe weather events, such as heavy rainfall, are on the rise worldwide. Reliable assessments of these events can save lives and protect property. Researchers at the Karlsruhe Institute of Technology (KIT) have developed a new method that uses artificial intelligence (AI) to convert low-resolution global weather data into high-resolution precipitation maps. The method is fast, efficient, and independent of location. Their findings have been in npj Climate and Atmospheric Science.

"Heavy rainfall and flooding are much more common in many regions of the world than they were just a few decades ago," said Dr. Christian Chwala, an expert on hydrometeorology and machine learning at the Institute of Meteorology and Climate Research (IMK-IFU), KIT's Campus Alpin in the German town of Garmisch-Partenkirchen. "But until now the data needed for reliable regional assessments of such extreme events was missing for many locations."

His research team addresses this problem with a new AI that can generate precise global precipitation maps from low-resolution information. The result is a unique tool for the analysis and assessment of extreme weather, even for regions with poor data coverage, such as the Global South.

For their method, the researchers use from that describe global precipitation at hourly intervals with a spatial resolution of about 24 kilometers. Not only was their generative AI model (spateGEN-ERA5) trained with this data, it also learned (from high-resolution weather radar measurements made in Germany) how precipitation patterns and extreme events correlate at different scales, from coarse to fine.

"Our AI model doesn't merely create a more sharply focused version of the input data, it generates multiple physically plausible, high-resolution maps," said Luca Glawion of IMK-IFU, who developed the model while working on his doctoral thesis in the SCENIC research project. "Details at a resolution of 2 kilometers and 10 minutes become visible. The model also provides information about the statistical uncertainty of the results, which is especially relevant when modeling regionalized events."

He also noted that validation with weather radar data from the United States and Australia showed that the method can be applied to entirely different climatic conditions.

Correctly assessing flood risks worldwide

With their method's global applicability, the researchers offer new possibilities for better assessment of regional climate risks. "It's the especially vulnerable regions that often lack the resources for detailed weather observations," said Dr. Julius Polz of IMK-IFU, who was also involved in the model's development.

"Our approach will enable us to make much more reliable assessments of where heavy rainfall and floods are likely to occur, even in such regions with poor data coverage." Not only can the new AI method contribute to disaster control in emergencies, it can also help with the implementation of more effective long-term preventive measures such as flood control.

More information: Luca Glawion et al, Global spatio-temporal ERA5 precipitation downscaling to km and sub-hourly scale using generative AI, npj Climate and Atmospheric Science (2025).

Citation: Researchers develop AI model to generate global realistic rainfall maps (2025, July 10) retrieved 11 July 2025 from /news/2025-07-ai-generate-global-realistic-rainfall.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.

Explore further

Researchers use deep learning to enhance spatial, temporal resolution of coarse precipitation maps

0 shares

Feedback to editors