As climate change increases the likelihood and scale of wildfires, researchers have developed artificial intelligence (AI) technology that can predict wildfire risk about a month in advance.
A team led by Professor Lim Jeong-ho in the Department of Urban and Environmental Engineering at Ulsan National Institute of Science and Technology said on the 25th it developed a Deep Learning model, "FWI-Net," that predicts the global Fire Weather Index (FWI) daily up to 31 days ahead. The study was published online in May in Communications Earth & Environment.
The Fire Weather Index is an indicator that shows how widely a wildfire could spread if it breaks out, based on temperature, relative humidity, wind, and precipitation. If this index is predicted in advance, it can be used to pre-position firefighting personnel and equipment in high-risk areas, or to prepare responses such as resident alerts and forest access controls.
Until now, a numerical forecast–based approach from the European Centre for Medium-Range Weather Forecasts (ECMWF) was mainly used, but its regional accuracy dropped quickly when the forecast period exceeded two weeks. FWI-Net reduced the root mean square error (RMSE) by 6.6% on average over the full 31-day forecast period compared with the conventional method. In particular, errors fell 12.4% during the first week of predictions.
The analysis found that FWI-Net eased the tendency to underestimate or overestimate risk in 85% of areas with both high wildfire exposure and high socioeconomic vulnerability. Under conditions of "very high" wildfire risk, the period during which meaningful forecasts were possible was extended by five days compared with the conventional approach. In poor regions lacking forecast infrastructure, it maintained significant predictive performance for an average of 22 days.
The team said performance improved because the model incorporated both past Fire Weather Index changes and future weather conditions. Even under the same temperature or precipitation, accumulated drought and dryness from earlier periods can change wildfire risk, and FWI-Net is designed to learn such temporal effects.
The researchers said, "By training the model on both past wildfire weather patterns and future forecast information, we complemented the limits of conventional numerical forecasts," and added, "It could be useful in regions where wildfire risk is high but in-house forecasting infrastructure is lacking."
Lim Jeong-ho said, "Accurate wildfire forecasts are information directly tied to disaster response planning," adding, "This technology could help with medium-range wildfire response and bridge information gaps in regions vulnerable in forecasting."
References
Communications Earth & Environment (2026), DOI: https://doi.org/10.1038/s43247-026-03692-9