AI model predicts wildfire risk a month out with greater accuracy

UNIST professor Lim Jeong-ho's team develops deep learning model to improve fire weather index forecasting

A fire department helicopter battles a wildfire. [Korea Forest Service]

A new AI technology can predict wildfire risk up to a month in advance with greater accuracy than existing methods. The system is expected to help identify high-risk areas ahead of time, enabling preemptive responses such as deploying firefighting personnel and equipment, preventive forest management and public warning systems.

A research team led by Lim Jeong-ho, a professor in the Department of Urban and Environmental Engineering at UNIST (Ulsan National Institute of Science and Technology), announced Thursday it has developed FWI-Net, a global deep learning model that forecasts the Fire Weather Index on a daily basis up to 31 days in advance.

The Fire Weather Index is an indicator that combines temperature, relative humidity, wind and precipitation to gauge how likely a wildfire is to spread significantly once ignited. Current forecasting methods lose regional accuracy rapidly after about two weeks and tend to underestimate extreme risk conditions. Forecasting infrastructure is also often lacking in areas most vulnerable to wildfire damage, creating demand for technology that improves medium-range prediction while remaining usable in underserved regions.

FWI-Net reduced the root mean square error across the full 31-day forecast period by 6.6 percent compared with existing approaches, and cut errors by as much as 12.4 percent during the first week. The model also reduced prediction bias — both underestimation and overestimation of wildfire risk — in 85 percent of areas where wildfire exposure and socioeconomic vulnerability are both high.

Under "very high" wildfire risk conditions, the window for meaningful prediction extended by five days beyond what current methods allow. In low-income regions with limited forecasting and response infrastructure, the model maintained meaningful predictive performance for an average of 22 days — more than three weeks.

From left: Lim Jeong-ho, UNIST professor; Kang Yu-jin, Kookmin University professor; and Lee Si-hyeon, UNIST researcher. [UNIST]

The team trained FWI-Net in two stages. It first pre-trained the model on the extensive ERA5 reanalysis dataset — which reconstructs historical weather by combining actual observations with meteorological models — to maximize medium-range forecasting performance. The model was then fine-tuned on SEAS5 data, the seasonal numerical forecast product issued by the European Centre for Medium-Range Weather Forecasts that provides projections of temperature, humidity, precipitation and wind speed. Because SEAS5 is produced only once a month, it alone is insufficient to train a deep learning model adequately, so the two-stage approach compensated for that data shortage.

"In a world where wildfire damage is surging due to climate change, accurate prediction technology is an information infrastructure directly tied to a nation's disaster response capacity," Lim said. "The technology we developed can be used to build medium-range wildfire response plans and to close information gaps in regions that lack forecasting foundations."

The research was supported by the Ministry of Environment, the Korea Forest Service and the National Research Foundation of Korea, and published in the international journal Communications Earth & Environment.

nbgkoo@heraldcorp.com