"Light Rain Forecast Turns into Heavy Rain"... Improving Accuracy with Artificial Intelligence

KAIST Research Team Jointly Develops with University of Tokyo

As the peak summer season approaches, concerns about damage from heavy rain or localized downpours are rising again this year. This is due to extreme climate conditions intensifying because of global warming. In such times, accurate weather forecasting becomes crucial, but even with the Korea Meteorological Administration deploying supercomputers, it is often criticized for 'missteps' such as failing to predict heavy rainfall just a few hours in advance. In response, a domestic research team has developed a technology that significantly reduces errors in precipitation forecasts using artificial intelligence (AI) algorithms.

[Image source=Yonhap News]

KAIST announced on the 25th that an international joint research team, including Professor Kim Hyung-jun from the Moonchul Graduate School of Future Strategy at KAIST and the University of Tokyo, proposed a new machine learning method to estimate ground precipitation using observations from a microwave radiometer mounted on a satellite.

The research team succeeded in reducing the error (RMSE) in total precipitation by a minimum of 15.9% and up to 42.5% compared to existing methods. Simple data-driven models require large amounts of training data, lack physical consistency, and make causal analysis of results difficult. In this study, the team explicitly incorporated domain knowledge related to satellite precipitation estimation, enabling interdependent knowledge exchange within the learning model. Specifically, they integrated a classification model that recognizes the presence of precipitation and a regression model that estimates precipitation intensity using a deep learning technique called multitask learning, training both simultaneously.

Precipitation estimation using multi-task learning algorithms and performance comparison with existing satellite precipitation observation data. Image source: Provided by KAIST

The machine learning model proposed in this study can include various physical mechanisms beyond those incorporated this time. For example, by including classification of precipitation types such as rain, snow, or sleet, and classification of cloud types that cause precipitation such as updrafts or stratiform clouds, the accuracy of future estimations is expected to improve further.

The results of this study were published on the 16th in the international academic journal Geophysical Research Letters.

Conceptual diagram of multi-task learning and single-task learning. Image source: Provided by KAIST

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