Localized Forest Fire Risk Prediction: A Department-Aware Approach for Operational Decision Support

Benjamin Aynes; Christophe Guyeux; Hassan Noura; Nicolas Caron

arxiv: 2506.04254 · v2 · submitted 2025-06-01 · 💻 cs.LG · cs.AI

Localized Forest Fire Risk Prediction: A Department-Aware Approach for Operational Decision Support

Nicolas Caron , Christophe Guyeux , Hassan Noura , Benjamin Aynes This is my paper

Pith reviewed 2026-05-19 11:00 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords forest fire predictiondepartment-aware modelingAI benchmarkFranceoperational decision supportrisk assessmentmachine learningclimate change impact

0 comments

The pith

Department-specific models make forest fire risk predictions more actionable for French firefighting units by reflecting local terrain and history.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that forest fire risk should be modeled separately for each French department instead of with one national model. Firefighting units operate at the department level and face different terrain, climate, and historical fire patterns, so uniform predictions give them less useful guidance for daily operations. The authors build department-aware AI models and run the first national-scale benchmark on metropolitan France data to test this idea. They treat the task as producing region-specific risk levels rather than simple binary ignition labels. This setup aims to deliver outputs that align directly with how operational decisions are made on the ground.

Core claim

Forest fire risk prediction improves when models are built separately for each department to capture its unique terrain, climate conditions, and historical fire events. Traditional approaches treat the problem as uniform binary classification across the whole country, but this does not match the departmental organization of French firefighting services. The authors therefore train and evaluate state-of-the-art models on per-department data slices, establishing the first national-scale AI benchmark for metropolitan France on this dataset and demonstrating more locally relevant risk outputs for operational use.

What carries the argument

Department-aware modeling, which partitions the national fire dataset by administrative department and trains independent predictors using local environmental and historical features to produce region-specific risk assessments.

If this is right

Risk maps and alerts can be issued at the exact scale used by each firefighting unit without extra translation steps.
Model performance can be tracked and improved independently for departments that experience different fire regimes.
The national benchmark provides a shared reference point for testing future AI methods on French fire data.
Operational planning gains a quantitative basis for allocating resources according to department-level probabilities rather than broad averages.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same localization idea could be tested for other administratively divided countries facing similar environmental risks such as floods or storms.
Departments with very few historical fires might require data-sharing rules or transfer techniques before independent models become reliable.
Real-time sensor feeds could be routed per department to update the localized models more quickly than a central system allows.

Load-bearing premise

Local conditions and data per department are distinct enough from other departments that separate models produce meaningfully better actionable predictions than a single national model.

What would settle it

A head-to-head test that trains the same models once as one national predictor and once as a collection of department-specific predictors, then compares accuracy on held-out data plus reported usefulness in operational scenarios for each version.

Figures

Figures reproduced from arXiv: 2506.04254 by Benjamin Aynes, Christophe Guyeux, Hassan Noura, Nicolas Caron.

**Figure 2.** Figure 2: Database construction process applied in this study. Apart from the target-related process, for which we show [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Distribution of fire occurrence (a) and burned [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Training and evaluation process and models architecture used in this article. Evaluation part show a comparison [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Area-normalized F1 comparison for the GRU model: multi-class versus binary formulations as the number of departments varies. the top 15 features for fire occurrence across different forecast horizons (0, 7, 15, and 31 days). A corresponding figure for burned area prediction is provided in the supple7 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 7.** Figure 7: Final score achieved by each model. shorter-term forecasts, the model relies almost exclusively on historical, static, or slowly evolving features. Vegetation and NDVI gain importance, and socio-cultural variables also start to contribute. 7 Future work In this section, we outline several directions for future work aimed at improving current model performance. - Clustering Current risk classes are based so… view at source ↗

**Figure 6.** Figure 6: Catboost prediction summary at 0 days horizons for Bouches du Rhone in multi-classification (a) and binary (b), and Charente in multi-classification (c) and binary (d). mentary materials. At 0 day, fire prediction is primarily driven by history and short-term signals. Encoded temporal variables, spatial clustering, and immediate weather conditions dominate the model’s decision-making. Fire danger indices,… view at source ↗

**Figure 8.** Figure 8: Top 15 features computed on multi-classification [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗

read the original abstract

Forest fire prediction involves estimating the likelihood of fire ignition or related risk levels in a specific area over a defined time period. With climate change intensifying fire behavior and frequency, accurate prediction has become one of the most pressing challenges in Artificial Intelligence (AI). Traditionally, fire ignition is approached as a binary classification task in the literature. However, this formulation oversimplifies the problem, especially from the perspective of end-users such as firefighters. In general, as is the case in France, firefighting units are organized by department, each with its terrain, climate conditions, and historical experience with fire events. Consequently, fire risk should be modeled in a way that is sensitive to local conditions and does not assume uniform risk across all regions. This paper proposes a new approach that tailors fire risk assessment to departmental contexts, offering more actionable and region-specific predictions for operational use. With this, we present the first national-scale AI benchmark for metropolitan France using state-of-the-art AI models on a relatively unexplored dataset. Finally, we offer a summary of important future works that should be taken into account. Supplementary materials are available on GitHub.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Department-aware fire risk prediction for France is a practical idea, but the abstract leaves the performance improvements unproven.

read the letter

The one or two things to know about this paper are that it proposes tailoring fire risk predictions to individual French departments for better operational support and presents what it calls the first national-scale AI benchmark for metropolitan France. On the positive side, the work does a good job explaining why a uniform national model might not serve end users well. Firefighting is organized by department, each with distinct terrain, climate, and fire history, so modeling at that level could lead to more relevant predictions. Applying current AI techniques to a relatively new dataset for this purpose sets up a useful reference point for future studies in the area. The paper also mentions supplementary materials on GitHub, which is helpful for anyone wanting to build on it. The discussion of future works shows awareness that this is not the end of the line. Where things get soft is in the lack of concrete evidence. The abstract does not include any specific results, such as accuracy scores, comparisons to non-department-aware baselines, or details on how they handle varying data sizes across departments. This makes it difficult to assess whether the department-aware models truly outperform simpler approaches or provide the claimed actionability. The assumption that per-department modeling will yield better outputs relies on local factors being dominant, but without the numbers, that remains to be seen. This paper would appeal to researchers focused on applied AI for climate and disaster risks, or to those developing decision support tools for emergency services. A reader interested in adapting ML models to administrative boundaries could find the setup informative. Given that it engages honestly with the practical context and literature on fire prediction, it deserves a serious referee to evaluate the full experiments and data aspects. My recommendation is to send it for peer review, with notes to strengthen the validation section if needed.

Referee Report

2 major / 2 minor

Summary. The paper proposes a department-aware modeling strategy for forest fire risk prediction in metropolitan France. It tailors AI models to individual departments to account for local terrain, climate, and historical fire data, aiming to produce more actionable and region-specific outputs for operational decision support by firefighters. The work claims to deliver the first national-scale AI benchmark using state-of-the-art models on a relatively unexplored dataset and outlines future research directions.

Significance. If the empirical results establish that per-department models measurably outperform pooled baselines on operational metrics such as precision for high-risk ignition events or decision-support utility, this would represent a meaningful advance in localized environmental risk modeling. The national-scale benchmark on French data could serve as a reproducible reference point for similar applications in other countries facing climate-driven fire risks.

major comments (2)

[Results] Results section: The central claim that department-aware models yield more actionable predictions rests on the assumption that tailoring improves performance over uniform models, yet no quantitative comparisons (e.g., AUC, F1, or operational utility scores), error analysis, or statistical significance tests versus pooled baselines are presented. Without these, the improvement cannot be verified.
[Data and Methods] Data description and experimental setup: The paper does not report per-department sample sizes or data-volume analysis. If some departments have very few historical fire events, the tailored models risk unstable estimation, directly undermining the load-bearing assumption that local data suffice for reliable department-specific modeling.

minor comments (2)

[Abstract] Abstract: The motivation is clear, but the abstract would benefit from one sentence summarizing the key empirical outcome (e.g., average improvement over baseline) to allow readers to gauge the contribution immediately.
Notation and terminology: Ensure consistent use of 'department-aware' versus 'localized' throughout; the current phrasing occasionally blurs whether the approach is strictly per-department or merely regionally stratified.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive review. We address each major comment below and describe the revisions we will incorporate to strengthen the manuscript.

read point-by-point responses

Referee: [Results] Results section: The central claim that department-aware models yield more actionable predictions rests on the assumption that tailoring improves performance over uniform models, yet no quantitative comparisons (e.g., AUC, F1, or operational utility scores), error analysis, or statistical significance tests versus pooled baselines are presented. Without these, the improvement cannot be verified.

Authors: We agree that explicit quantitative comparisons are necessary to substantiate the claim that department-aware models produce more actionable predictions. The initial submission focused on presenting the first national-scale benchmark but omitted direct head-to-head evaluations against pooled baselines. In the revised manuscript we will add a new subsection in Results that reports AUC, F1, and operational utility metrics (e.g., precision at high-risk thresholds) for both per-department and pooled models. We will also include error analysis and statistical significance tests (paired Wilcoxon signed-rank tests across departments) to quantify and verify the observed improvements. revision: yes
Referee: [Data and Methods] Data description and experimental setup: The paper does not report per-department sample sizes or data-volume analysis. If some departments have very few historical fire events, the tailored models risk unstable estimation, directly undermining the load-bearing assumption that local data suffice for reliable department-specific modeling.

Authors: We acknowledge that reporting per-department sample sizes is essential for assessing model reliability. The current version does not provide this breakdown. In the revision we will add a data-volume analysis, including a table of historical fire events and total samples per department, together with a discussion of low-data departments and the regularization or transfer-learning strategies employed to stabilize estimation in those cases. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper frames its contribution as an empirical application of existing state-of-the-art AI models to a new department-aware fire-risk benchmark on external French national data. No derivation chain, equations, or self-citations are presented that reduce a claimed prediction or uniqueness result to a fitted parameter or prior self-referential definition. The central claim rests on comparative performance of localized versus pooled models against operational metrics, which is falsifiable on held-out data and does not rely on self-definitional or ansatz-smuggled steps.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on standard machine learning assumptions plus one key domain assumption about the value of departmental partitioning; no new physical entities are postulated and hyperparameters are typical of any ML pipeline.

free parameters (1)

Model hyperparameters (e.g., learning rate, architecture depth)
Standard in any neural network or ML training pipeline; chosen or tuned to fit the fire-risk dataset.

axioms (1)

domain assumption Departmental boundaries capture the dominant local variations in terrain, climate, and fire history relevant to risk
Invoked when the paper states that fire risk should be modeled sensitive to local conditions and does not assume uniform risk across regions.

pith-pipeline@v0.9.0 · 5734 in / 1432 out tokens · 51636 ms · 2026-05-19T11:00:45.074477+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We transformed the classical binary and regression problem into a multi-class classification problem. For each department, we created an ordinal 5-class signal for both occurrence and burned area using the K-Means algorithm.
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

CatBoost ensemble emerges as the top performer... global versus local evaluation reveals that models may appear accurate in aggregate while underperforming in sparsely burned departments

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

[1]

G. Boeing. Osmnx: New methods for acquiring, con- structing, analyzing, and visualizing complex street networks. Computers, Environment and Urban Sys- tems, 65:126–139, 2017. ISSN 0198-9715. doi: https: //doi.org/10.1016/j.compenvurbsys.2017.05.004

work page doi:10.1016/j.compenvurbsys.2017.05.004 2017
[2]

W. E. Cascio. Wildland fire smoke and human health. Science of the Total Environment, 624:586–595, May

work page
[3]

doi: 10.1016/j.scitotenv

ISSN 1879-1026. doi: 10.1016/j.scitotenv. 2017.12.086

work page doi:10.1016/j.scitotenv 2017
[4]

H. Chen, J. M. Samet, P. A. Bromberg, and H. Tong. Cardiovascular health impacts of wildfire smoke ex- posure. Particle and Fibre Toxicology, 18(1):2, Jan

work page
[5]

doi: 10.1186/s12989-020-00394-8

work page doi:10.1186/s12989-020-00394-8
[6]

J. Chen, X. Wang, Y . Yu, X. Yuan, X. Quan, and H. Huang. Improved prediction of forest fire risk in central and northern china by a time-decaying precip- itation model. 2022. URL https://doi.org/ 10.3390/f13030480

work page doi:10.3390/f13030480 2022
[7]

J. Chen, Y . Yang, L. Peng, L. Chen, and X. Ge. Knowledge graph representation learning-based for- est fire prediction. Remote Sensing, 14(17), 2022. ISSN 2072-4292. doi: 10.3390/rs14174391

work page doi:10.3390/rs14174391 2022
[8]

Cisneros, J

D. Cisneros, J. Richards, A. Dahal, L. Lombardo, and R. Huser. Deep graphical regression for jointly 9 moderate and extreme australian wildfires. Spatial Statistics, 59:100811, 2024. ISSN 2211-6753

work page 2024
[9]

Dupire, T

S. Dupire, T. Curt, and S. Bigot. Spatio-temporal trends in fire weather in the french alps. Science of The Total Environment, 595:801–817, 10 2017. doi: 10.1016/j.scitotenv.2017.04.027

work page doi:10.1016/j.scitotenv.2017.04.027 2017
[10]

Dupire, S

T. Dupire, S. Curt, T. Bigot, and T. Fr´ejaville. Vul- nerability of forest ecosystems to fire in the french alps. European Journal of Forest Research, 138(5): 813–830, 2019. doi: 10.1007/s10342-019-01206-1. hal-02163692

work page doi:10.1007/s10342-019-01206-1 2019
[11]

S. E. Finlay, A. Moffat, R. Gazzard, D. Baker, and V . Murray. Health impacts of wildfires. PLoS Cur- rents, 4:e4f959951cce2c, Nov 2012. doi: 10.1371/ 4f959951cce2c

work page 2012
[12]

K. Hardy. Comparative analysis of forest fires in france, spain, portugal, greece, and italy in 2022. 2024

work page 2022
[13]

R. Lam, A. Sanchez-Gonzalez, M. Willson, P. Wirns- berger, M. Fortunato, F. Alet, S. Ravuri, T. Ewalds, Z. Eaton-Rosen, W. Hu, A. Merose, S. Hoyer, G. Hol- land, O. Vinyals, J. Stott, A. Pritzel, S. Mohamed, and P. Battaglia. Graphcast: Learning skillful medium- range global weather forecasting, 2023

work page 2023
[14]

Michail, C

D. Michail, C. Davalas, L.-I. Panagiotou, I. Prapas, S. Kondylatos, N. I. Bountos, and I. Papoutsis. Fire- castnet: Earth-as-a-graph for seasonal fire prediction, 2025

work page 2025
[15]

M. M. Muller. Dynamic Time Warping, pages 69–84. Springer Berlin Heidelberg, Berlin, Heidelberg, 2007. ISBN 978-3-540-74048-3

work page 2007
[16]

Mutakabbir, C.-H

A. Mutakabbir, C.-H. Lung, K. Naik, M. Zaman, S. A. Ajila, T. Ravichandran, R. Purcell, and S. Sampalli. Spatio-temporal agnostic sampling for imbalanced multivariate seasonal time series data: A study on forest fires. Sensors, 25(3), 2025. ISSN 1424-8220. doi: 10.3390/s25030792

work page doi:10.3390/s25030792 2025
[17]

Pimont, H

F. Pimont, H. Fargeon, T. Opitz, J. Ruffault, R. Bar- bero, N. Martin-StPaul, E. Rigolot, M. Rivi´Ere, and J.-L. Dupuy. Prediction of regional wildfire activity in the probabilistic bayesian framework of firelihood. Ecological Applications, 31(5):e02316, 2021. doi: https://doi.org/10.1002/eap.2316

work page doi:10.1002/eap.2316 2021
[18]

Prokhorenkova, G

L. Prokhorenkova, G. Gusev, A. V orobev, A. V . Doro- gush, and A. Gulin. Catboost: unbiased boosting with categorical features. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018

work page 2018
[19]

X. Wu, G. Zhang, Z. Yang, S. Tan, Y . Yang, and Z. Pang. Machine learning for predicting forest fire occurrence in changsha: An innovative investigation into the introduction of a forest fuel factor. Remote Sensing, 15(17), 2023. ISSN 2072-4292. doi: 10. 3390/rs15174208. 10

work page 2023

[1] [1]

G. Boeing. Osmnx: New methods for acquiring, con- structing, analyzing, and visualizing complex street networks. Computers, Environment and Urban Sys- tems, 65:126–139, 2017. ISSN 0198-9715. doi: https: //doi.org/10.1016/j.compenvurbsys.2017.05.004

work page doi:10.1016/j.compenvurbsys.2017.05.004 2017

[2] [2]

W. E. Cascio. Wildland fire smoke and human health. Science of the Total Environment, 624:586–595, May

work page

[3] [3]

doi: 10.1016/j.scitotenv

ISSN 1879-1026. doi: 10.1016/j.scitotenv. 2017.12.086

work page doi:10.1016/j.scitotenv 2017

[4] [4]

H. Chen, J. M. Samet, P. A. Bromberg, and H. Tong. Cardiovascular health impacts of wildfire smoke ex- posure. Particle and Fibre Toxicology, 18(1):2, Jan

work page

[5] [5]

doi: 10.1186/s12989-020-00394-8

work page doi:10.1186/s12989-020-00394-8

[6] [6]

J. Chen, X. Wang, Y . Yu, X. Yuan, X. Quan, and H. Huang. Improved prediction of forest fire risk in central and northern china by a time-decaying precip- itation model. 2022. URL https://doi.org/ 10.3390/f13030480

work page doi:10.3390/f13030480 2022

[7] [7]

J. Chen, Y . Yang, L. Peng, L. Chen, and X. Ge. Knowledge graph representation learning-based for- est fire prediction. Remote Sensing, 14(17), 2022. ISSN 2072-4292. doi: 10.3390/rs14174391

work page doi:10.3390/rs14174391 2022

[8] [8]

Cisneros, J

D. Cisneros, J. Richards, A. Dahal, L. Lombardo, and R. Huser. Deep graphical regression for jointly 9 moderate and extreme australian wildfires. Spatial Statistics, 59:100811, 2024. ISSN 2211-6753

work page 2024

[9] [9]

Dupire, T

S. Dupire, T. Curt, and S. Bigot. Spatio-temporal trends in fire weather in the french alps. Science of The Total Environment, 595:801–817, 10 2017. doi: 10.1016/j.scitotenv.2017.04.027

work page doi:10.1016/j.scitotenv.2017.04.027 2017

[10] [10]

Dupire, S

T. Dupire, S. Curt, T. Bigot, and T. Fr´ejaville. Vul- nerability of forest ecosystems to fire in the french alps. European Journal of Forest Research, 138(5): 813–830, 2019. doi: 10.1007/s10342-019-01206-1. hal-02163692

work page doi:10.1007/s10342-019-01206-1 2019

[11] [11]

S. E. Finlay, A. Moffat, R. Gazzard, D. Baker, and V . Murray. Health impacts of wildfires. PLoS Cur- rents, 4:e4f959951cce2c, Nov 2012. doi: 10.1371/ 4f959951cce2c

work page 2012

[12] [12]

K. Hardy. Comparative analysis of forest fires in france, spain, portugal, greece, and italy in 2022. 2024

work page 2022

[13] [13]

R. Lam, A. Sanchez-Gonzalez, M. Willson, P. Wirns- berger, M. Fortunato, F. Alet, S. Ravuri, T. Ewalds, Z. Eaton-Rosen, W. Hu, A. Merose, S. Hoyer, G. Hol- land, O. Vinyals, J. Stott, A. Pritzel, S. Mohamed, and P. Battaglia. Graphcast: Learning skillful medium- range global weather forecasting, 2023

work page 2023

[14] [14]

Michail, C

D. Michail, C. Davalas, L.-I. Panagiotou, I. Prapas, S. Kondylatos, N. I. Bountos, and I. Papoutsis. Fire- castnet: Earth-as-a-graph for seasonal fire prediction, 2025

work page 2025

[15] [15]

M. M. Muller. Dynamic Time Warping, pages 69–84. Springer Berlin Heidelberg, Berlin, Heidelberg, 2007. ISBN 978-3-540-74048-3

work page 2007

[16] [16]

Mutakabbir, C.-H

A. Mutakabbir, C.-H. Lung, K. Naik, M. Zaman, S. A. Ajila, T. Ravichandran, R. Purcell, and S. Sampalli. Spatio-temporal agnostic sampling for imbalanced multivariate seasonal time series data: A study on forest fires. Sensors, 25(3), 2025. ISSN 1424-8220. doi: 10.3390/s25030792

work page doi:10.3390/s25030792 2025

[17] [17]

Pimont, H

F. Pimont, H. Fargeon, T. Opitz, J. Ruffault, R. Bar- bero, N. Martin-StPaul, E. Rigolot, M. Rivi´Ere, and J.-L. Dupuy. Prediction of regional wildfire activity in the probabilistic bayesian framework of firelihood. Ecological Applications, 31(5):e02316, 2021. doi: https://doi.org/10.1002/eap.2316

work page doi:10.1002/eap.2316 2021

[18] [18]

Prokhorenkova, G

L. Prokhorenkova, G. Gusev, A. V orobev, A. V . Doro- gush, and A. Gulin. Catboost: unbiased boosting with categorical features. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018

work page 2018

[19] [19]

X. Wu, G. Zhang, Z. Yang, S. Tan, Y . Yang, and Z. Pang. Machine learning for predicting forest fire occurrence in changsha: An innovative investigation into the introduction of a forest fuel factor. Remote Sensing, 15(17), 2023. ISSN 2072-4292. doi: 10. 3390/rs15174208. 10

work page 2023