GlobeNet and DeepRain offer higher accuracy than existing methods.

NVIDIA announced that its GPUs are being used to train GlobeNet and DeepRain, groundbreaking typhoon and heavy rain damage prediction systems from the Korea Institute of Science and Technology Information (KISTI).

Researchers at KISTI are attracting attention for developing a system to predict damage from typhoons and heavy rain using GPU-accelerated deep learning. Dr. Minsoo Cho, Director of KISTI's HPC Research Center, stated, "While it's impossible to prevent natural disasters, the risks can be minimized with the right information."

Korea is surrounded by storms like hurricanes and tropical cyclones. Hurricanes and typhoons are known by different names depending on their geographical location: hurricanes in the Atlantic and eastern North Pacific, typhoons in the western North Pacific, and cyclones in the South Pacific and Indian Ocean. Hurricanes and typhoons are phenomena that occur worldwide.

Dr. Cho Min-soo and other KISTI researchers are working to improve the speed and accuracy of typhoon forecasts by combining deep learning technology with existing forecasting methods to create numerical weather models using GPU-accelerated supercomputers. More accurately determining a typhoon's path and intensity will help authorities issue more accurate and timely evacuation orders to residents in areas expected to be affected.

Currently, meteorologists utilize various numerical models to predict wind speed, precipitation, atmospheric pressure, and other factors that determine the path and intensity of a typhoon from its formation to its dissipation. However, KISTI researchers are using satellite and radar observation data to train two deep learning systems: GlobeNet, which predicts typhoon paths, and DeepRain, which predicts heavy rainfall. Additionally, data from numerical models are being used to train DeepTC, a tropical cyclone prediction system.

“Although these three models are still in the research phase, they have achieved significant improvements in accuracy compared to existing methods,” said Dr. Song Sa-kwang, a senior scientist in charge of developing deep learning systems at KISTI.

The KISTI research team trained these models using cuDNN with the Keras toolkit and TensorFlow deep learning framework based on NVIDIA GPUs, and deployed the trained models using GPUs in the Amazon Web Services cloud.

The system developed at KISTI has so far enabled predictions of typhoons and resulting rainfall 1-2 hours in advance. The research team plans to increase the forecasting window to six hours by next year, ultimately reaching three days, a timeframe where effective countermeasures can be implemented.

The KISTI research team's work will be utilized for flood prediction in the Seoul metropolitan area, home to approximately 30% of the nation's population, along the Imjin River, and surrounding areas. Furthermore, KISTI's system is currently undergoing technology transfer to the Air Force Meteorological Agency.

While KISTI's research was designed for domestic use in Korea, the same approach can be applied anywhere in the world. Dr. Cho Min-soo stated, "As long as sufficient satellite and radar data are available, DeepRain and GlobeNet can be applied to North America as well."