AI Predicts Downpours Six Hours Ahead... Seoul National University Develops Ultra-Short-Term Precipitation Forecasting Technology [Reading Science]

Accurate Forecasting of Heavy Downpours Using Only Radar Images
Selected for Presentation at ICLR 2026

Domestic researchers have developed an ultra-short-term precipitation prediction system that utilizes artificial intelligence (AI) to precisely forecast various weather conditions, including extreme precipitation, using only radar imagery. This technology is expected to enhance weather warnings and disaster response systems by detecting sudden weather disasters, such as heavy rainfall, in advance.

The National Research Foundation of Korea announced on March 17 that the research team led by Professor Youngjun Hong of the Department of Mathematical Sciences at Seoul National University has developed an AI-based ultra-short-term precipitation prediction model combining a local spatiotemporal attention mechanism and a novel upsampling structure. The research results have been accepted by ICLR 2026 (International Conference on Learning Representations), one of the world's top academic conferences in the field of artificial intelligence, and will be presented in April.

Research Model Architecture and Data Distribution Histogram. (Top) The model structure that receives past radar precipitation images as input and extracts spatiotemporal information through a 3D Swin Transformer with an encoder-decoder architecture, predicting future precipitation images. (Middle) The decoder layers apply a triple 3D upsampling technique and a temporal extraction layer to restore local patterns. (Bottom left) Precipitation distribution map of the KMA dataset. It shows an imbalance with mostly clear or normal precipitation and a very small number of extreme heavy rain events. (Bottom right) Comparison of results between the existing upsampling method and the proposed method. Image and description: Professor Youngjun Hong, Seoul National University.

AI Detects 'Extreme Precipitation Patterns'

As climate change leads to an increase in extreme weather events such as local heavy rainfall, the need for technologies capable of predicting sudden disasters that occur within short periods has grown. However, existing numerical weather prediction (NWP) models are limited by vast computational requirements, resulting in restricted spatiotemporal resolution, and it has been difficult to accurately predict Korea's precipitation patterns due to its complex topography.

To overcome these limitations, the research team introduced a "local attention mechanism" that focuses computational resources on rapidly changing precipitation patterns within radar images. This structure enables accurate reconstruction of fine-grained information that causes ultra-short-term heavy rainfall while reducing unnecessary computations, thereby improving prediction efficiency.

The model developed by the research team can predict precipitation distribution for up to six hours based on one hour of past radar imagery. Validation using meteorological observation data from the Korea Meteorological Administration (KMA), as well as the United States and France, demonstrated higher accuracy than existing state-of-the-art global models, with particularly significant improvements in prediction performance for Korea's extreme precipitation events.

In terms of computational efficiency, the new model also showed significant improvement over conventional models. According to the research team, computational efficiency for extreme precipitation prediction improved by more than 20 times, confirming its potential for application in real-time disaster response systems. In experiments applying the model to domestic heavy rainfall cases in 2023, it was also found to be capable of detecting risk possibilities in advance.

'Limit Challenge R&D' Achievements... Expected Use in Disaster Response

Professor Youngjun Hong of Seoul National University's Department of Mathematical Sciences stated, "We have implemented a prediction model that reliably operates even under extreme weather conditions while reducing the AI 'black box' issue based on mathematical rigor. We hope that, in the future, this can contribute to protecting public safety by being linked with disaster response systems."

Choi Wonchun, Lead Program Manager of the Limit Challenge Strategy Center at the National Research Foundation of Korea, said, "This is a representative achievement resulting from the bold challenges of researchers combined with a flexible research management system led by the responsible program manager. We will continue to support the development of public-interest technologies that can contribute to national disaster response."

This research was carried out with support from the 'Limit Challenge R&D Project' promoted by the Ministry of Science and ICT and the National Research Foundation of Korea. In this program, the responsible program manager identifies innovative research topics and supports high-impact, high-difficulty research through flexible management methods such as adjusting research directions.

This content was produced with the assistance of AI translation services.