Biomass fires: a graph neural network predicts fire intensity up to seven days in advance

NOAA’s aerosol forecasts capture the fires observed 1.5 days earlier. A graphical model predicts their global evolution over 1–7 days and reduces the error by between 24 and 43 per cent.

To forecast air quality, NOAA's GEFS-Aerosols system relies on satellite measurements of radiative fire power (FRP) from the GBBEPx product. There are two weaknesses in these forecasts: the data arrive about a day and a half late, and then they remain frozen during the five days of operational forecasting (seven in the experimental version), as if the fires were no longer evolving.

Fires that are no longer supposed to stand still

Researchers, including several from NOAA's GSL laboratory, propose a graph-based spatiotemporal neural network that predicts FRP on a global scale from one to seven days in advance. It combines reanalysis weather, land use, vegetation and recent fire history. Trained on 2020-2022 and then tested on 2023 and 2024, it reconstructs the global seasonal cycle and clearly beats the persistence hypothesis: at 0.1° resolution, the root mean square error drops by 32% at one day and 43% at seven days in 2023, and by 24% and 40% in 2024.

Fire detection is resilient to the extension of the time frame, but the estimation of their intensity deteriorates more quickly, and very large fires are systematically underestimated. Calibrating the intensity and better positioning small fires remain the two areas of focus before operational use. The study is submitted to the journal Artificial Intelligence for the Earth Systems.