PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2610.11022v1physics.ao-phcs.LG

TL;DR

Because operational fire inputs are delayed and persisted, the paper develops a GNN framework for predicting fire occurrence and intensity. It reproduces global seasonal FRP patterns and improves on persistence, though intensity calibration remains a limitation.

  • Problem

    Operational fire inputs arrive about 1.5 days late and are then persisted, so they cannot reflect evolving fire activity as forecast lead time increases.

    Source: §1 · pp. 3–4

  • Method

    The paper develops a GNN framework that predicts fire occurrence and intensity using a hurdle-type head and a composite loss for sparse, heavy-tailed FRP.

    Source: §1 · p. 5

  • Results

    The model reproduces the global seasonal FRP cycle and, versus persistence, reduces error by roughly a quarter to a third at one-day lead and around 40% at seven days at 0.1° resolution.

    Source: §6 · p. 31

  • Takeaways & Limitations

    The results demonstrate useful predictability of fire activity several days ahead beyond a persistence forecast.

    Source: §6 · p. 31

    Intensity calibration remains the principal limitation, with the largest-fire stratum underestimated to a third of observed intensity.

    Source: §6 · pp. 31–32

Abstract

from arXiv · show

Skillful prediction of biomass-burning activity several days in advance is important for air-quality forecasting and aerosol prediction. Two operational constraints motivate this work. First, the GBBEPx satellite fire radiative power (FRP) product used to initialize NOAA's GEFS-Aerosols is available with about a 1.5-day latency, so each forecast cycle relies on the most recently available, but already outdated, fire observations. Second, these fire inputs are then held fixed throughout the subsequent 5-day operational forecast, or 7 days in the GSL experimental system, effectively assuming no evolution in fire activity. We develop a data-driven model that predicts global FRP one to seven days ahead from the most recent available observations. The model adapts a spatiotemporal graph neural network using reanalysis meteorology, land-cover and vegetation information, recent fire history, and GBBEPx FRP as the training target. It is trained on 2020-2022 data and evaluated for 2023-2024. The model reproduces the global seasonal cycle and substantially outperforms persistence. At 0.1$^\circ$ resolution, mean squared error is reduced by 32% at one-day lead and 43% at seven days in 2023, and by 24% and 40% in 2024. At 1$^\circ$ resolution, the critical success index ranges from 0.32 to 0.60. Detection skill declines only modestly with lead time, whereas intensity skill degrades more rapidly. Large fires are detected reliably, but their radiative power is systematically underestimated. These results demonstrate useful predictability of fire activity several days ahead and identify intensity calibration and small-fire placement as the main remaining challenges before predicted FRP can support operational aerosol forecasts.

1. Introduction

Operational fire-emissions forecasts rely on delayed observations held fixed, while existing approaches do not provide global gridded FRP directly. This study develops a GNN framework for global FRP prediction.

1. Introduction

  • FRP is a globally observable, physically based proxy for fire intensity and combustion rate, providing a direct basis for predictive fire emissions.Source: §1 · p. 3
  • GBBEPx’s 1.5-day latency makes operational fire inputs outdated, and fixed persistence cannot anticipate ignitions, rapid growth, intensity changes, spread, or extinction.Persistence becomes increasingly problematic with lead time.Source: §1 · pp. 3–4
  • Existing approaches largely target regional or short-range fire behavior, occurrence, spread, or danger rather than a global gridded FRP field usable as an emissions boundary condition.Source: §1 · p. 4
  • Fire activity depends on evolving environmental conditions, while sparse, heavy-tailed FRP makes mean-squared-error training prone to near-zero predictions and missed large fires.This motivates separating occurrence from intensity and emphasizing large fires.Source: §1 · pp. 4–5
  • The proposed GNN predicts fire occurrence and intensity with a hurdle head and composite loss, and is verified at one-to-seven-day leads against withheld 2023–2024 data.Source: §1 · p. 5
  • The study introduces five-domain training settings to accommodate contrasts among global fire regimes.Source: §1 · p. 5
  • A climatology-constrained per-pixel intensity cap anchors predictions to the observed historical FRP distribution at inference time.Source: §1 · p. 5
  • The evaluation directly compares forecasts with persistence and quantifies the margin at the 0.1° resolution used for emission fields.Source: §1 · p. 5

2. Data

The study pairs daily GBBEPx v4 FRP targets with meteorological, land-surface, and fire-history inputs on a common 0.1° grid, training on 2020–2022 and evaluating independently in 2023–2024. The verification sample covers two complete annual cycles for lead-time- and month-resolved assessment, with 2024 outside the climatological record used for predictors and the intensity constraint.

2.1. Fire Radiative Power Target

  • Daily mean FRP from GBBEPx v4 is the target, a gridded, gap-filled global product at 0.1° × 0.1° resolution.Its spatial continuity, resolution, and operational use motivate its selection for training and verification.Source: §2.1 · pp. 5–6

2.2. Meteorological Predictors

  • ERA5 supplies 27 meteorological predictors aggregated to daily statistics and interpolated to the target grid, including daily extrema for 700-hPa humidity and lapse rate.Precipitation is represented by daily accumulation and exponentially weighted antecedent precipitation; wind gust uses the daily 90th percentile and 500-hPa vertical velocity the daily minimum.Source: §2.2 · p. 6

2.3. Land-Surface and Fire-History Predictors

  • Predictors combine land-cover information, a 21-day history of observed FRP, consecutive active-fire duration, and pixel-specific day-of-year FRP climatological statistics as a local prior.Source: §2.3 · pp. 8–9
  • The encoding sensitivity test found detection skill essentially unchanged (CSI within ±0.005), with small, mixed-sign changes in intensity and MSE skill scores.The main results are therefore insensitive to encoding choice, although one-hot encoding is planned for future versions.Source: §2.3 · p. 8

2.4. Training and Evaluation Periods

  • Training uses 2020–2022 data, while evaluation covers the independent calendar years 2023 and 2024; seven-day forecasts are verified at lead times of one to seven days.A 21-day input window yields 1069 initialization days for training and was chosen to capture the evolution of fires that can burn for more than a month.Source: §2.4 · p. 10
  • The verification sample is temporally dense and spans two complete annual cycles.Source: §2.4 · p. 10
  • The sample supports lead-time-resolved and month-resolved assessments in Section 5.Source: §2.4 · p. 10
  • 2024 lies outside the 2012–2023 climatological record used for climatological predictors and the intensity constraint.Source: §2.4 · p. 10

3. Methods

The methods adapt a spatiotemporal graph neural network to predict daily FRP using separate occurrence and intensity heads, a composite loss, and independently trained regional models. They describe safeguards and climatological caps applied during tiled-patch inference, and verify FRP-model skill using reanalysis meteorology, making the reported skill an upper bound for use with forecast meteorology.

3.1. Model Overview

  • A 21-day history feeds FRP-GraphCast-Lite, a GraphCast-adapted spatiotemporal GNN that predicts seven daily leads using separate fire-occurrence and conditional-intensity heads.The forecast is produced for each land grid cell at each lead time.Source: §3.1 · pp. 10–11
  • A four-connected native-grid lattice supports local message passing at 0.1° resolution and allows transfer to UFS FV3 and MPAS meshes.The lattice reflects the short-range nature of fire spread rather than requiring long-range spherical connectivity.Source: §3.1 · p. 11
  • A hurdle head separates occurrence from log-space intensity and multiplies predicted occurrence by intensity, zeroing estimates below 0.15 MW to limit sparsity effects.This separation allows the occurrence and intensity tasks to use objectives suited to each.Source: §3.1 · p. 11

3.2. Loss Function

  • The composite objective weights occurrence, intensity, background, volume, and spectral losses 2.0, 10.0, 50.0, 1.0, and 0.5, respectively; focal loss downweights easy background cells.The intensity loss penalizes under-prediction ten times more strongly than over-prediction.Source: §3.2 · p. 12

3.3. Region-based Training Strategy

  • Five fire-regime domains receive independently trained models because regional seasonality, fuels, and intensity differ; global fields are formed by mosaicking their predictions.The models share an architecture, predictor set, and loss formulation but do not share learned weights.Source: §3.3 · p. 14
  • Regional tuning clips FRP inputs and targets more tightly in Southeast Asia, while FRP and fire-duration inputs are jointly zeroed in 5% of training samples to reduce persistence reliance.Southeast Asia uses a 1500 MW clip, compared with 3000 MW in Australia–Pacific and 5000 MW elsewhere.Source: §3.3 · pp. 14–15

3.4. Climatology-Constrained Intensity Cap

  • A data-driven model lacks a physical constraint against implausible intensity excursions for rare predictor combinations, a concern for heavy-tailed FRP dominated by a few extremes.Source: §3.4 · p. 15
  • At each forecast lead, the cap limits each pixel to the greater of its day-of-year 90th-percentile FRP climatology times a regional factor and a 5000 MW floor.Factors are 5 for North America, South America, and Australia–Pacific; 4 for EMEA; and 3 for Southeast Asia.Source: §3.4 · p. 15
  • The 2012–2023 climatology used by the cap includes 2023, one evaluation year, which the authors note as a limitation.A pixel with one fire day in 4627 valid observations lacked a usable day-of-year climatology and was masked to zero in prediction and verification.Source: §3.4 · pp. 15–16
  • The climatology cap was active on 1.6% of region-days outside Southeast Asia, removing at most 6.8% of predicted radiative power; in Southeast Asia it acted almost entirely on masked non-burning cells.With a 5000 MW floor, the cap rarely engages.Source: §3.4 · p. 16

3.5. Operational Configuration and Inference

  • The model is intended to use forecast meteorology operationally, but verification uses reanalysis inputs to isolate FRP-model skill and therefore represents an upper bound; the reduction is unquantified.Global fields are formed by applying regional models to tiled patches and merging regions before post-processing.Source: §3.5 · p. 16

4. Verification Framework

Forecasts are verified against independent GBBEPx observations at native 0.1° and aggregated 1° resolution, separately for leads of one to seven days and pooled across leads. The framework assesses both fire detection and FRP intensity, including skill relative to persisted initialization-time FRP.

4. Verification Framework

  • Verification spans native 0.1° and aggregated 1° resolution, with skill reported for one- to seven-day leads and pooled across leads.The two resolutions distinguish exact spatial placement from a scale less sensitive to small displacement errors.Source: §4 · p. 17
  • Categorical detection uses a 0.5 MW per 0.1° pixel fire threshold and reports POD, FAR, CSI, HSS, ETS, and FBI.Categorical scores use unweighted contingency counts.Source: §4 · p. 17
  • Continuous intensity skill over fire-affected cells is assessed with RMSE, MAE, MBE, Pearson correlation, KGE, and NSE.Continuous metrics are cosine-of-latitude weighted, and persistently non-burning cells are excluded.Source: §4 · p. 17
  • Persistence holds initialization-time observed FRP fixed throughout the forecast, matching the operational assumption used as the benchmark.Relative skill is 1 − MSE_model / MSE_persistence; positive values indicate lower error than persistence.Source: §4 · p. 17
  • 0.006% of 2023 nonzero grid cells were below the 0.5 MW threshold, so thresholding had negligible effect on observations.That fraction was 214 of 3.62 million nonzero grid cells.Source: §4 · p. 17

5. Results and Discussion

Results assess ML-FIRE’s FRP forecast performance across lead times and spatial scales, including seasonal and broad spatial patterns and comparison with persistence. They also characterize detection and intensity skill, placement and magnitude errors, and the limitations of the examined day-of-year climatology.

5.1. Overall Skill, Seasonal Cycle and Spatial Structure

  • 0.048 MW per 0.1° pixel in 2023 and 0.044 in 2024 were the median absolute daily biases; annual predicted power was within 2% of observed in 2023 and 12% above it in 2024.83% and 77% of days, respectively, were within ±0.1 MW.Source: §5.1 · p. 18
  • At 1°, ETS spans 0.30–0.55 in 2023 and 0.34–0.59 in 2024; KGE spans −0.12 to 0.46 and 0.00 to 0.62, with positive KGE in 77 of 84 2023 month–lead combinations.At 0.1°, ETS spans 0.10–0.23 in 2023 and 0.13–0.35 in 2024; KGE is negative in all 84 combinations in 2023 and 72 of 84 in 2024.Source: §5.1 · p. 21
  • The forecasts capture the contrasting fire seasons of 2023 and 2024, retaining the seasonal cycle at D+7 while damping sharp daily peaks.Predicted daily peaks at short lead can lag and exceed observed peaks.Source: §5.1 · pp. 20–21
  • At 1°, CSI spans 0.40–0.55 in 2023 and 0.41–0.60 in 2024, versus 0.16–0.25 and 0.17–0.37 at 0.1°; the gap reflects more false alarms from imprecise placement.POD is comparable across resolutions, while FAR is two to three times higher at 0.1°.Source: §5.1 · p. 21, Figure 3 · p. 22
  • The model reproduces the location and seasonal migration of major burning regions, but native-resolution difference fields show displacement errors; aggregation to 1° removes most of their cancellation-prone structure.July predictions omit scattered high-intensity fires observed in eastern Siberia, although comparable North American boreal fires are captured.Source: §5.1 · pp. 24–25

5.2. Lead-Time Degradation: Detection versus Intensity

  • At 0.1°, MAE stays near 23 MW per 0.1° pixel in 2023 and 14 MW in 2024 across lead times, while mean bias remains negative and grows with lead.At 1°, MAE rises modestly, by about 8% over the seven-day window.Source: Figure 6 · p. 25, §5.2 · p. 26
  • From D+1 to D+7 at 0.1°, median ETS declined 0.032 in 2023 and 0.054 in 2024, while KGE fell 0.33 and 0.36 in every month of both years.The predicted-field standard deviation fell from 6.7 to 4.0 MW between D+1 and D+7, against an observed 10.2 MW.Source: §5.2 · p. 27
  • At one-day lead, FSS rose from 0.33 at 0.1° to 0.50 at 0.3° in 2023 and from 0.42 to 0.63 in 2024; thresholds (0.506, 0.511) occurred at approximately 0.34° and 0.19°, respectively.Source: §5.2 · pp. 28–29

5.3. Skill Relative to Operational Practice

  • At 0.1°, MSE fell 32%, 40%, and 43% at 1-, 3-, and 7-day leads in 2023, versus 24%, 31%, and 40% in 2024 relative to persistence.At 1°, the corresponding reductions were 9%, 21%, and 34% in 2023, and 1%, 14%, and 30% in 2024.Source: Figure 9 · p. 29, §5.3 · p. 30
  • At 0.1°, the model’s advantage over persistence widens with lead time; by D+7, it reduces error by a third or more against the week-old persisted field.The model’s skill declines, but persistence degrades faster; at D+1 the recent observation remains a strong predictor, so the model adds relatively little at the aggregated scale.Source: §5.3 · p. 30
  • The day-of-year climatology was unreported: it was undefined for 99.8% of pixel–day combinations and, after gap-filling, represented only about a fifth of observed variance.Its low mean squared error reflected the variance term of the error decomposition rather than forecast accuracy.Source: §5.3 · p. 30
  • Persistence is the appropriate comparison because its variance matches observations by construction and it is the reference used operationally.Source: §5.3 · pp. 30–31

6. Summary and Discussion

The model predicts daily FRP across one- to seven-day leads and outperforms persistence while reproducing the global seasonal cycle. Detection skill holds up better than intensity skill, with intensity calibration, spatial displacement, and regional generalization remaining challenges.

6. Summary and Discussion

  • The approach uses a hurdle prediction head, a fire-tailored composite loss, separate training over five fire-regime domains, and a per-pixel cap anchored to historical FRP climatology.Source: §6 · p. 31
  • The model reproduces the daily global FRP seasonal cycle 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.Source: §6 · p. 31
  • At 0.1° resolution, mean squared error falls by roughly a quarter to a third at one-day lead and around 40% at seven days versus persistence.The advantage over the persisted field increases with lead time.Source: §6 · p. 31
  • Across the seven-day window, detection ETS declines by only a few hundredths, while intensity Kling–Gupta efficiency declines by an order of magnitude more; detection probability rises with observed fire intensity.Large fires are located more reliably than small fires.Source: §6 · p. 31
  • Skill varies regionally: Siberian boreal-season detection is poorer than North America's, which the authors attribute primarily to heterogeneity in the Southeast Asia training domain.The authors also note that verification uses reanalysis meteorology, so reported skill is an upper bound for operational use; converting FRP to emissions adds separate uncertainty.Source: §6 · p. 32
  • The largest-fire stratum is underestimated by roughly a factor of three to four, and spatial aggregation does not recover the missing radiative power.The authors identify intensity calibration as the principal remaining limitation.Source: §6 · p. 32

Disclaimer

The disclaimer attributes the statements, findings, conclusions, and recommendations to the authors and says they do not necessarily reflect NOAA or U.S. Department of Commerce views.

Disclaimer

  • The authors’ statements, findings, conclusions, and recommendations do not necessarily reflect NOAA or U.S. Department of Commerce views.Source: Disclaimer · p. 34

APPENDIX

The appendix compares monthly mean FRP predictions with GBBEPx observations for representative months in 2023 and 2024, at two lead times and two spatial resolutions.

Spatial Evaluation of Predicted Fire Radiative Power

Loading 2610.11022v1…