A Graph Neural Network for Global Daily Fire Radiative Power Prediction at Medium-Range Lead Times

arXiv:2610.11022 · physics.ao-ph, cs.LG · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "A Graph Neural Network for Global Daily Fire Radiative Power Prediction at Medium-Range Lead Times".

Tom: The gist: A data-driven Graph Neural Network predicts global daily fire radiative power one to seven days ahead from recent observations,

Jane: First, who's behind it and why it matters.

Paper summary: Jane: So, looking at the paper "A Graph Neural Network for Global Daily Fire Radiative Power Prediction at Medium-Range Lead Times," the core idea is that this machine learning approach can forecast daily global fire radiative power out to seven days.

Lu: The authors adapted a spatiotemporal graph neural network using reanalysis meteorology, land-cover and vegetation information, recent fire history and GBBEPx FRP as their training target <ref:2610.11022#pg3>.

Meng: This is a big step for using AI in this domain because it moves us past just relying on persistence for fire input when we're trying to model things like aerosols.

Tom: The paper identifies intensity calibration and small-fire placement as the main remaining challenges before predicted FRP can support operational aerosol forecasts <ref:2610.11022#pg3>.

Jane: This means that while the model is powerful, we still need to nail down how strong those big fires actually are when we want to use it for real-time aerosol predictions.

Lu: The paper demonstrates that this machine learning model can instead forecast daily global fire radiative power out to seven days, reducing error relative to that assumption by a third or more <ref:2610.11022#pg3>.

Meng: So the practical implication is that we can use this model to get better fire input for our systems later in a seven-day window than we would with current persistence methods.

Tom: It’s about getting better fire input for the right time, which is exactly when the operational assumption of no fire evolution is weakest <ref:2610.11022#pg3>.

Jane: This work shows that this approach can give us a better handle on fire activity for aerosol forecasts than just sticking to persistence in certain parts of the forecast window.

Lu: The authors have made this model available at https://doi.org/ten point five two eight one/zenodo.twenty-three million one hundred ninety-six thousand four hundred fourteen <ref:2610.11022#pg3>.

Meng: So the paper is providing a tool for improving how we predict fire power, even if it still needs calibration work to be fully operational <ref:2610.11022#pg3>.

Tom: The main thing to remember is that intensity calibration and small-fire placement are the main remaining challenges before predicted FRP can support operational aerosol forecasts <ref:2610.11022#pg3>.

Conclusion: Tom: So, to wrap up this look at the paper "A Graph Neural Network for Global Daily Fire Radiative Power Prediction at Medium-Range Lead Times," we're talking about how they built this system to predict fire power further out than usual, and what that really means for us.

Jane: Basically, the researchers developed a sophisticated AI model that takes all sorts of climate and fire data and tries to guess how much heat is being released by global fires up to seven days in advance. <ref:2610.11022#pg3>

Lu: It’s built using a graph neural network structure, which is really good at looking at connections across space and time, using everything from land cover to past fire history as its inputs. <ref:2610.11022#pg3>

Meng: And the big win here is how they use the AI to predict this power field one day at a time, rather than just sticking to what we saw yesterday—that's a big change for forecasting systems. <ref:2610.11022#pg3>

Tom: The authors show that this machine learning approach cuts the error compared to just using the persistence method by more than a third, and that gain is biggest in those later days when persistence is weakest. <ref:2610.11022#pg3>

Jane: It means we can get a much better handle on the fire activity for aerosol predictions during those longer forecast windows, which is where things get really tricky. <ref:2610.11022#pg3>

Lu: The authors highlight that the main hurdles left are getting that intensity calibration right; they're underestimating the strongest fires by about three to four times. <ref:2610.11022#pg3>

Meng: So, while the prediction gets better at telling us *where* a fire is, we still have a big job ahead for accurately guessing *how strong* those fires actually are when we try to use this for real-world aerosol models. <ref:2610.11022#pg3>

Tom: That’s the core challenge they’ve laid out, and it sets the stage for what comes next in improving these fire forecasts.

Li Zhang, Jun Wang, Isidora Jankov, Yongxin Liu, Gonzalo A. Ferrada, Ravan Ahmadov, Ligia Bernardet, Haonan Chen

CIRES, University of Colorado Boulder, Boulder, CO, US · Global Systems Laboratory, OAR/NOAA · NOAA/NWS/OMD · CIRA, Colorado State University · NOAA/NESDIS Center for Satellite Applications and Research

physics.ao-ph, cs.LG

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: Submitted to Artificial Intelligence for the Earth Systems

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 69/100

The gist: The gist: A data-driven Graph Neural Network predicts global daily fire radiative power one to seven days ahead from recent observations, substantially outperforming persistence by reducing mean

Key concepts

Graph Neural Network (GNN)
A type of machine learning model designed to process data structured as a graph. In this context, the land surface is treated as a network where each grid cell is a node and its neighbors are connected by edges. The GNN learns complex spatial relationships between these cells to make predictions.
Fire Radiative Power (FRP)
A measure of the energy radiated by fires on Earth, expressed in megawatts (MW). This paper focuses on predicting this value daily across the globe. FRP is crucial for understanding fire intensity and its impact on the atmosphere, making it a key variable for environmental modeling.
Hurdle Formulation
A two-part loss function used during training to handle the extreme sparsity and heavy-tailed distribution of fire data. It separates the prediction task into two parts: classifying whether a pixel is burning or not, and then regressing its intensity if it is burning. This helps the model manage noisy data effectively.
Domain Training
The model is trained separately for five distinct global fire regimes (e.g., North America, EMEA). This approach acknowledges that fire behavior varies significantly by region, allowing the model to learn more accurate and context-specific patterns for different parts of the world.

Summary

The gist: A data-driven Graph Neural Network predicts global daily fire radiative power one to seven days ahead from recent observations, substantially outperforming persistence by reducing mean squared error by a third or more at one-day lead and 40% at seven days at 0.1° resolution

Model Architecture and Data

The prediction model is a spatiotemporal graph neural network adapted from the GraphCast framework, referred to here as FRP-GraphCast-Lite. At forecast initialization, the model ingests a 21-day history of predictor fields over a region and produces, for each land grid cell, a prediction of FRP at each of the seven subsequent daily lead times. Inference is performed on overlapping 128 × 128-cell spatial patches (stride 96 cells in each direction), which are reassembled into the full regional field by averaging over the overlap regions; this tiling keeps the graph at a tractable size while providing halo overlap so that patch boundaries are not visible in the reassembled field. The model architecture comprises three components: an input projection, two residual graph-convolution layers, and a temporal encoder.

Training Strategy and Loss Function

The model is trained on a continuous three-year record from 1 January 2020 to 31 December 2022. The training uses two complementary predictors for fire history, including the observed daily mean FRP field (MeanFRP t) and a duration-of-fire predictor (dof). To address the extreme sparsity and heavy-tailed distribution of FRP, the prediction head adopts a hurdle (two-part) formulation. The total loss is a weighted sum of five terms: L = 2.0·Lcls + 10.0·Lreg + 50.0·Lbg + 1.0·Lvol + 0.5·Lspec. The regression term, Lreg, is evaluated in physical units (MW) using a per-pixel smooth-L1 (Huber) error scaled by an intensity weight and an asymmetry factor.

Domain Training and Constraints

Global fire regimes are sufficiently distinct over regions that the land surface is partitioned into five fire-regime domains: North America, South America, EMEA, Southeast Asia, and the Australia–Pacific sector. A separate model is trained for each domain. The model introduces a physically motivated, climatology-constrained perpixel intensity cap at inference time. This cap limits predicted FRP to a region-specific multiple of that cell's historical dayof-year 90th-percentile FRP climatology, subject to an absolute floor.

Verification and Key Findings

The model is verified against two complete calendar years, 2023 and 2024, both strictly independent of the training period. The model reproduces the global seasonal cycle of daily FRP at both resolutions with median absolute daily global-mean bias below 0.05 MW per 0.1° pixel at the shortest lead in both years. The clearest result is the asymmetry between detection and intensity skill, where ETS declines by only a few hundredths across the seven-day forecast window, whereas KGE declines by an order of magnitude more in every month of both years. The model retains its ability to identify which pixels are burning across the forecast window; what degrades is its ability to reproduce how strongly they burn. The principal remaining limitation is intensity calibration, as the largest-fire stratum is underestimated by a factor of roughly three to four.

Operational Relevance

Relative to persistence of the most recent observed field—the reference implied by current operational practice—it reduces mean squared error by roughly a quarter to a third at one-day lead and by around 40% at seven days at 0.1° resolution. The advantage is largest precisely where the operational assumption is weakest — in the later days of the forecast window, which are also the days for which a predictive capability is most needed. The model's effective resolution is close to the scale at which the receiving model can represent the emitted plume, even though the emission fields are supplied on a finer grid.

The paper demonstrates that a machine-learning model can instead forecast daily global fire radiative power out to seven days, reducing error relative to that assumption by a third or more, with the largest gains at the end of the forecast period where the assumption is weakest. This work identifies intensity calibration and small-fire placement as the main remaining challenges before predicted FRP can support operational aerosol forecasts.

The research was supported in part by the Bipartisan Infrastructure Law (BIL) NA23OAR4050201I, NOAA Weather Program Office (https://ror.org/01q70bz83) under award NA23OAR4590387, and NOAA cooperative agreement NA22OAR4320151, for the Cooperative Institute for Earth System Research and Data Science (CIESRDS). The GBBEPx v4 fire radiative power data used for model training and evaluation are produced by NOAA/NESDIS and are available from https://www.ospo.noaa.gov/pub/Blended/GBBEPx/GBBEPx-all01GRID (last accessed: 22 September 2026). The ERA5 reanalysis dataset (0.25° grid), from which the daily meteorological predictors were derived, was obtained from the NSF NCAR GDEX repository (https://doi.org/10.5065/BH6N-5N20, last accessed: 22 September 2026). Vegetation Health Index data are available from NOAA/NESDIS (https://www.star.nesdis.noaa.gov/smcd/emb/vci/VH/index.php; last accessed: 22 September 2026). Land-cover data are available from NASA’s EarthData system (https://doi.org/10.5067/MODIS/MCD12C1.061). The model code is available at https://doi.org/10.5281/zenodo.23196414. The statements, findings, conclusions, and recommendations are those of the author(s) and do not necessarily reflect the views of NOAA or the U.S. Department of Commerce. The GBBEPx v4 fire radiative power data used for model training and evaluation are produced by NOAA/NESDIS and are available from https://www.ospo.noaa.gov/pub/Blended/GBBEPx/GBBEPx-all01GRID (last accessed: 22 September 2026). The ERA5 reanalysis dataset (0.25° grid), from which the daily meteorological predictors were derived, was obtained from the NSF NCAR GDEX repository (https://doi.org/10.5065/BH6N-5N20, last accessed: 22 September 2026). Vegetation Health Index data are available from NOAA/NESDIS (https://www.star.nesdis.noaa.gov/smcd/emb/vci/VH/index.php; last accessed: 22 September 2026). Land-cover data are available from NASA’s EarthData system (https://doi.org/10.5067/MODIS/MCD12C1.061). The model code is available at https://doi.org/10.5281/zenodo.23196414.

--- Page 37 ---

References

Abatzoglou, J. T., and A. P. Williams, 2016: Impact of anthropogenic climate change on wildfire across western US forests. Proc. Natl. Acad. Sci. USA, 113, 11770–11775, https://doi.org/10.1073/pnas.1607171128. Ahmadov, R., and Coauthors, 2024: Using VIIRS fire radiative power data to simulate biomass burning emissions, plume rise and smoke transport in a real-time air quality modeling system. 2024 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Fort Worth, TX, IEEE, 2806–2808, https://doi.org/10.1109/IGARSS.2024. Andela, N., and Coauthors, 2017: A human-driven decline in global burned area.

Improvements for AI systems

  1. Bold header: Hurdle-type prediction head with composite loss function

This formulation addresses extreme sparsity and heavy-tailed distribution of FRP by separating fire occurrence from intensity, allowing the model to locate large fires reliably but under-predict their intensity.

  1. Bold header: Five-domain training strategy

Training a separate model for each fire regime—North America, South America, EMEA, Southeast Asia, and Australia–Pacific—accommodates strong contrasts among global fire regimes, leading to domain-specific tuning of parameters.

  1. Bold header: Climatology-constrained intensity cap

The model applies a climatologyconstrained per-pixel intensity cap at inference time to enforce physical plausibility, limiting predicted FRP based on the historical day-of-year 90th percentile, which prevents implausible intensity excursions.

  1. Bold header: Lead-time degradation analysis

By quantifying skill relative to persistence—1 − MSE model / MSE persistence—the system demonstrates that it reduces mean squared error by roughly a quarter to a third at one-day lead and by around 40% at seven days, showing the advantage increases as the operational assumption of persistence becomes weaker.

  1. Bold header: Effective resolution identification

The analysis shows that the effective resolution of current AI weather prediction models is several times coarser than their nominal grid spacing, with the model's effective resolution being close to where receiving models operate, allowing for a more accurate assessment of operational utility at 25 km grid scales.

Related papers