As wildfire damage spirals out of control worldwide due to climate change, a Korean research team has developed technology that accurately predicts wildfire risk a month in advance using artificial intelligence (AI). The technology is expected to overcome the limitations of existing forecasting systems and secure a "golden time" for disaster response, while also helping reduce damage in impoverished countries that lack disaster prediction infrastructure.
A team led by Professor Im Jung-ho of the Department of Civil, Urban, Earth, and Environmental Engineering at the Ulsan National Institute of Science and Technology (UNIST) said Tuesday it has developed "FWI-Net," a global deep learning model that predicts the global Fire Weather Index (FWI) on a daily basis up to 31 days in advance.
The Fire Weather Index (FWI) is an indicator that represents the risk of wildfire spread by combining factors such as temperature, relative humidity, wind, and precipitation. Accurately predicting this index enables proactive disaster response, including the advance deployment of firefighting personnel and the issuance of evacuation orders for residents. However, the widely used numerical forecasting method of the European Centre for Medium-Range Weather Forecasts (ECMWF) had the limitation that regional prediction accuracy dropped sharply after about two weeks.
The "FWI-Net" developed by the research team broke through these shortcomings of the existing method with AI. Validation results showed that FWI-Net reduced the root mean square error (RMSE) across the entire 31-day prediction period by 6.6% compared with the existing numerical forecasting method, and succeeded in lowering the error by as much as 12.4% during the first week.
In particular, the period during which meaningful predictions are possible in extreme situations where wildfire risk is "very high" was extended by five days compared with before. When the model was applied to the recent cases of major wildfire damage—the 2023 wildfires in Canada and Chile and the 2025 Los Angeles wildfires—FWI-Net accurately captured high risk signals close to reality even in situations where existing models had underestimated the intensity of wildfire risk.
This breakthrough performance improvement was made possible because the AI simultaneously learned both "accumulated past weather" and "future weather conditions." Even when future conditions such as temperature or precipitation are the same, the AI learned on its own the pattern in which wildfire risk varies depending on how much drought and dryness had accumulated over the preceding period. The research team pre-trained the AI with the vast "ERA5" reanalysis data, which restores past weather, and then performed fine-tuning with the "SEAS5" data used in actual forecasting, complementing the drawbacks of forecasting environments where data is scarce.
Above all, this technology is drawing attention as a "warm technology" capable of resolving global disaster inequality. Prediction bias was reduced in 85% of regions with high wildfire risk exposure and socioeconomic vulnerability. Even in impoverished regions of Africa where forecasting infrastructure is severely lacking, it maintained meaningful prediction performance for an average of 22 days, proving that it can reduce the "disaster information gap" between developed and vulnerable countries.
"In a situation where wildfire damage is surging worldwide due to climate change, accurate prediction technology is a key information infrastructure directly linked to a nation's actual disaster response capability," Professor Im Jung-ho stressed. "The technology developed this time can be actively used to establish medium-term wildfire response plans and to fill the information gaps in vulnerable regions that lack their own forecasting foundations."
The study, in which Professor Kang Yu-jin of Kookmin University and UNIST researcher Lee Si-hyun participated as first authors, was conducted with support from the Ministry of Environment, the Korea Forest Service, and the National Research Foundation of Korea. It was published online in the world-renowned international journal "Communications Earth & Environment" on Nov. 28.
