PaperlayerView exact revision on arXiv

Source-linked AI summary

Accelerating Wildfire Hazard Mapping Across Canada with an AI Surrogate Modelling Framework

Hager Radi Abdelwahed, Toluwaleke Olutayo, Chirag Goel, Jana Osea, Aldo Zaimi, Tom Swystun, Sandy Erni, Francisco Emiliano Lopez Saavedra, Yan Boulanger, Gaétan Marceau Caron

arXiv:2610.30170v2cs.AI

TL;DR

Canada’s increasing wildfire activity heightens the need for frequent hazard mapping, while process-based simulations are computationally intensive. This paper presents a national-scale AI surrogate framework for BurnP3+ outputs and reports rapid inference and coherent counterfactual responses. Its dataset is controlled-access, despite publicly accessible code and models.

  • Problem

    Increasing wildfire activity creates a need for frequent national hazard mapping, while process-based simulation workflows can take days to weeks for a regional unit.

    Source: §1 · pp. 1–2

  • Method

    A national-scale multi-task deep learning surrogate emulates BurnP3+ hazard outputs, alongside a framework for training, evaluation, and regional fine-tuning.

    Source: §1 · p. 2, Problem Formulation · p. 4

  • Results

    The surrogate runs approximately 806× faster than BurnP3+ on CPU and responds coherently to tested counterfactual input changes.

    Source: §8 · p. 13

  • Takeaways & Limitations

    The framework makes rapid scenario testing feasible for scientists and emergency planners, supporting timely risk assessment.

    Source: §8 · p. 13

    The underlying dataset is controlled-access, although the code and models are publicly accessible.

    Source: §1 · p. 2

Abstract

from arXiv · show

Canada's increasing wildfire activity requires frequent national hazard mapping to guide proactive mitigation, including fuel management, land-use planning, infrastructure protection, and evacuation routing. However, process-based wildfire simulation frameworks (e.g., BurnP3 and BurnP3+) rely on computationally intensive Monte Carlo simulations to model burn probability, fire intensity, and rate of spread. These workflows can take days to weeks to process a single regional modelling unit. To overcome these computational bottlenecks, we present three contributions developed through collaboration between AI researchers and wildfire modelling experts: (a) a standardized framework for training, fine-tuning, and evaluating AI surrogate models against process-based wildfire simulators; (b) a multi-task deep learning surrogate trained on national-scale data to emulate burn probability, mean fire intensity, and mean rate of spread at a fraction of the computational cost (seconds); and (c) domain-specific metrics and counterfactual evaluation protocols to probe out-of-distribution generalization. Integrating geospatial data (fuel, terrain, ignition likelihood) and regional weather, our model approximates BurnP3+ simulation outputs, achieving strong performance across agreement, rank-correlation, error, and spatial-overlap metrics. It preserves downstream hazard classifications and produces physically plausible responses to expert-informed counterfactual interventions. Furthermore, we demonstrate zero-shot transfer to a separate regional BurnP3+ study and show that lightweight fine-tuning can substantially improve regional calibration and predictive agreement. On a representative region, the surrogate achieves 806-fold end-to-end CPU acceleration and uses 13% of BurnP3+'s host memory, facilitating rapid scenario analysis.

1 Introduction

Increasing wildfire activity makes regularly updated, high-resolution hazard maps important, but process-based BurnP3+ simulations can take days or weeks. The paper introduces a national-scale AI surrogate framework to emulate BurnP3+ outputs and evaluate transfer and counterfactual responses.

Motivation

  • BurnP3+ hazard maps quantify burn probability, mean fire intensity, and mean fire spread rate, but simulations take days or weeks even on high-performance computers.Source: §1 · p. 1
  • Frequent, accurate, high-resolution hazard assessments support planning for fuel treatment, utility safety, zoning, critical infrastructure, and emergency evacuation.Source: §1 · p. 1

Contributions

  • The framework trains, evaluates, fine-tunes, and deploys regional deep-learning surrogates, with a national multi-source, multi-task model approximating BurnP3+ burn probability, fire intensity, and rate of spread from days to seconds.It also uses domain-specific metrics and counterfactual protocols to probe spatial generalization and out-of-dataset interventions.Source: §1 · p. 2

Data access

  • The underlying dataset is controlled-access, despite public availability of the code and models, to prevent misinterpretation and improper use in high-stakes planning.Source: §1 · p. 2

Contributions

  • The national surrogate transfers to an independent regional BurnP3+ study, and lightweight fine-tuning corrects out-of-distribution regional miscalibration.Source: §1 · p. 2

2 Related Work

Prior wildfire-surrogate work demonstrates substantial computational savings, but regional generalization remains a concern. This paper situates its national-scale framework as a way to address that constraint.

2 Related Work

  • Prior wildfire emulators report large speedups, including over 99% less computation time than JULES-INFERNO and 0.6% of original simulation time in South Australia.Other cited approaches include spatio-temporal neural networks for fire-front simulations and forecasts reported as 10^2 to 10^4 times faster than direct simulations.Source: §2 · pp. 2–3
  • Regional generalization remains a concern for wildfire surrogates trained on narrow geographic areas, which can inherit simulation biases.The paper contrasts this with leveraging simulation outputs across Canada’s diverse fuels, terrain, and fire regimes.Source: §2 · p. 3
  • Neural emulators can generate large ensembles and probabilistic fire-arrival maps efficiently, providing information within the narrow decision window before fire spreads.The passage links this use to uncertainty estimation.Source: §2 · p. 3

3 Dataset

The dataset pairs national BurnP3+ hazard simulations with geospatial and weather inputs, then prepares regional patches and targets for surrogate training. Regional splitting is used to prevent spatial data leakage.

BurnP3+ Simulations

  • The simulation environment used baseline 2020 FBP fuel maps, 1980–2024 meteorological records, a DEM, and human- and lightning-caused ignition probability surfaces, partitioning Canada into 54 regional “hexels” of 130,000 km2.The AI surrogate was provided the same spatial drivers as the Monte Carlo simulations, including weather, topography, fuel, and ignition variables.Source: BurnP3+ Simulations · p. 3
  • BurnP3+ simulations produce national 100-meter rasters for burn probability, mean fire intensity, and mean rate of spread.The study used 20,000–150,000 Monte Carlo iterations per region and generated approximately 22 million simulated fires across Canada.Source: BurnP3+ Simulations · p. 3

Machine Learning Dataset Preparation

  • The surrogate dataset combines 100-meter spatial features with regional weather and fire-size features, using FBP-derived fuel features, terrain, and aggregated ignition maps.Fuel classes are represented through 18-dimensional initial rate-of-spread curves compressed into a 4-dimensional embedding; weather and fire-size sources and processing are also specified.Source: Spatial Inputs · p. 3, Tabular Inputs · pp. 3–4
  • Burn probability targets are min-max normalized, while fire intensity and rate of spread targets receive log(1 + x) transformation and z-score standardization.The training-split statistics are used for target preprocessing.Source: Reference Targets · p. 4
  • Regional grids are divided into overlapping 256 × 256 patches and split by region into approximately 80/10/10 training, validation, and test sets to prevent spatial data leakage.After filtering, the splits contain 15,422 training, 2,146 validation, and 1,736 test samples.Source: Data Sample Generation · p. 4

4 Methodology

The methodology formulates wildfire hazard prediction as joint pixel-wise mapping from geospatial and regional tabular inputs to three hazard maps. It uses a U-Net surrogate and evaluates predictions with complementary error, agreement, rank, and spatial-overlap metrics.

Problem Formulation

  • One multi-task model maps geospatial and regional weather inputs to aligned pixel-wise maps of burn probability, fire intensity, and rate of spread.The spatial input includes fuel, terrain, and ignition likelihood features alongside tabular weather data.Source: Problem Formulation · p. 4

Model Design

  • A U-Net combines local landscape patterns with regional weather context, while tabular features are projected across each patch and stacked as input channels.The authors chose channel stacking after tests found minimal gains from more complex bottleneck fusion relative to its architectural overhead.Source: Model Design · p. 4

Evaluation Metrics

  • Evaluation combines normalized error and bias, Spearman correlation and CCC, and top-K IoU measures for extreme-risk areas and overall spatial accuracy.The Top-10% AUC-IoU targets extreme-risk areas; AUC-IoU Full measures overall spatial accuracy.Source: Evaluation Metrics · p. 5

5 Experiments

Across BP, FI, and ROS, the multi-task U-Net outperforms the evaluated baselines, with weather improving BP and more modest gains for FI and ROS. The surrogate also substantially reduces runtime and memory use while preserving downstream hazard patterns, though errors in extreme hazard classes remain an operational priority.

5 Experiments

  • The multi-task U-Net outperforms all evaluated baselines across BP, FI, and ROS; adding weather substantially improves BP and yields more modest gains for FI and ROS.Fire-size percentiles yield slight improvements on some regional metrics and are retained because they are critical for BurnP3+ calibration.Source: §5 · p. 5, Table 1 · p. 6

Time and Memory Performance Comparison

  • ∼806× CPU acceleration and 13% of BurnP3+ host memory: the surrogate’s end-to-end pipeline is substantially faster and uses less host memory on region 16.BurnP3+ required 38.9 hours and 100.8 GB across 32 CPU cores; the comparison covers one representative region.Source: Time and Memory Performance Comparison · pp. 5–6, Table 2 · p. 6

Hazard Mapping Evaluation

  • ∼48% exact 13-class agreement rises to ∼87% within 1 bin and ∼97% within 2 bins; IoU is ∼85% for high-hazard classes 10–13.Hazard is derived from burn probability and capped fire intensity, then rescaled and discretized into 13 ordinal classes; residual errors in extreme classes remain an operational priority.Source: Hazard Mapping Evaluation · p. 6

6 Counterfactual Scenario Analysis and Validation

The study tests whether the surrogate reproduces BurnP3+ responses to fuel interventions and probes its sensitivity to weather and climate counterfactuals in region 16. Fuel-response comparisons use simulator validation, while weather and climate interventions are assessed through surrogate sensitivity analysis.

Fuel Interventions

  • For C2-to-M1, surrogate reductions in BP, FI, and ROS were 74.30%, 62.36%, and 38.35%, versus BurnP3+ reductions of 70.64%, 63.78%, and 38.37%.The surrogate reproduced the broad spatial pattern; agreement scores (CCC, ρs, NMAE) were (0.826, 0.924, 0.362) for BP, (0.834, 0.861, 0.323) for FI, and (0.725, 0.781, 0.424) for ROS.Source: Fuel Interventions · p. 10
  • For the post-fire D2 intervention, surrogate reductions in BP, FI, and ROS were 87.59%, 57.78%, and 40.27%, versus BurnP3+ reductions of 72.38%, 60.53%, and 48.29%.The surrogate predicted a larger BP reduction but smaller FI and ROS reductions, and its response was smoother than BurnP3+'s.Source: Fuel Interventions · p. 10
  • Across both fuel interventions, reductions in BP, FI, and ROS match expected responses to lower-flammability fuels, but ROS agreement is weaker than BP and FI.The surrogate captures the principal spatial responses tested, while reproducing spread-rate changes remains a limitation.Source: Fuel Interventions · p. 10

Weather Intervention

  • Imposing the BC weather regime increased region-wide mean BP by ∼206%, FI by ∼98%, and ROS by ∼24%, though BP decreased in some areas.FI and ROS increases were widespread, whereas the BP response was more heterogeneous.Source: Weather Intervention · p. 10

Climate-Change Intervention

  • Under the climate-change intervention, mean BP increased by ∼271%, mean FI by ∼1%, and mean ROS decreased by ∼0.2%.Spread-event durations were shifted using projected 2050s RCP 4.5 changes; baseline weather lists were retained.Source: Climate-Change Intervention · p. 10

Summary of Counterfactual Responses

  • Across the four counterfactuals, fuel responses broadly match BurnP3+ patterns, while weather and climate changes produce aggregate responses consistent with expected effects.Together, the findings support rapid counterfactual assessment under alternative fuel and environmental conditions.Source: Summary of Counterfactual Responses · p. 11

7 Cross-Study Model Transfer

The paper evaluates a nationally trained surrogate on a separate BurnP3+ regional study with differing input configurations and simulation characteristics. Regional recalibration, zone alignment, and lightweight fine-tuning improve BP performance, while FI and ROS are comparatively more stable.

Cross-Study Model Transfer

  • The regional benchmark uses four scenarios with distinct inputs and BurnP3+ outputs, including fuel classes absent from national training and differences in simulation parameterization and calibration.The national and regional study areas overlap geographically, but none of the regional scenarios reproduces the national training input configuration.Source: §7 · pp. 11–12
  • Regional target recalibration reduced Scenario A BP MAE from 4.620 to 0.500 and corrected bias, while Spearman correlation and IoU remained unchanged.The passage reports that BP was more affected by regional domain shift than FI or ROS.Source: Results · p. 12
  • Zone alignment increased Spearman from 0.616 to 0.890, CCC from 0.366 to 0.708, and full AUC IoU from 0.550 to 0.718.The adjustment mapped the misaligned regional fire-weather zone's input features to neighboring zone features aligned with the training distribution.Source: Results · p. 12

1. Under Scenario B, BP again exhibits the largest domain

Across out-of-distribution scenarios, burn probability is most sensitive to regional shifts and benefits strongly from fine-tuning, while fire intensity and rate of spread generally transfer more robustly. This target-dependent pattern persists across model configurations.

1. Under Scenario B, BP again exhibits the largest domain

  • Under Scenarios C and D, fine-tuning substantially improves burn probability, while fire intensity worsens or remains more robust without adaptation and rate of spread changes little.Under Scenario C, BP MAE falls from 3.812 to 0.302 and Spearman rises from 0.416 to 0.835; under Scenario D, BP MAE falls from 3.862 to 0.284 and Spearman rises from 0.419 to 0.835.Source: §1 · p. 13
  • Regional fine-tuning improves burn probability across scenarios, but its effects on fire intensity and rate of spread vary by target and scenario.For Scenario B, burn-probability MAE drops from 5.115 to 0.278, Spearman rises from 0.584 to 0.901, and full AUC IoU rises from 0.530 to 0.730.Source: §1 · p. 13
  • With the spatial + weather + fire size model, the base model generalizes less reliably when regional fire-size distributions differ, but fine-tuning largely recovers performance.Under Scenario B, burn probability reaches 0.901 Spearman and 0.732 full AUC IoU after adaptation.Source: §1 · p. 13
  • Across scenarios, fine-tuning reduces burn-probability MAE by over 90% and raises Spearman correlation above 0.83, while fire intensity and rate of spread transfer more robustly zero-shot.The reported pattern spans distinct out-of-distribution scenarios and model configurations.Source: §1 · p. 13

8 Conclusion

The paper presents a multi-task surrogate for BurnP3+ that rapidly approximates three wildfire hazard outputs and responds coherently to controlled input changes. It transfers to a separate regional study, with lightweight fine-tuning improving predictions where domain shift is strongest.

8 Conclusion

  • On a representative region, the surrogate runs about 806× faster than BurnP3+ on CPU, shifting a regional scenario from days to seconds.Counterfactual prediction shifts were directionally consistent with BurnP3+ outputs in verified scenarios.Source: §8 · p. 13
  • The surrogate approximates burn probability, fire intensity, and rate of spread from spatial, weather, and fire-size inputs, and lightweight regional fine-tuning improves performance under domain shift.The model is a simulator surrogate, inherits BurnP3+ limitations, and is intended to complement expert judgment in high-stakes decisions.Source: §8 · p. 13

A Codebase and Trained Models

  • The dataset is restricted to reduce risks of misuse and misinterpretation; access may be considered upon reasonable research request.The paper notes that raw hazard maps could be misinterpreted by commercial entities, with possible adverse socioeconomic impacts on communities and public safety.Source: A Codebase and Trained Models · p. 17

B Dataset

  • Regional rasters are reprojected and aligned to a shared Canada Albers grid before clipping into model-ready patches for national-scale mosaicking and training.Raw regional spatial datasets are provided at 100-m resolution in locally centered coordinate systems.Source: B Dataset · p. 17
  • Fuel classes are represented by FBP-derived initial-spread curves, transformed into an 18-dimensional vector and compressed to a learned 4-dimensional embedding.The curve-based representation maps discrete fuel categories into a continuous, physically grounded parameter space that may facilitate transfer to related fuel classes absent from training.Source: B Dataset · p. 17

C Evaluation Metrics

D Experimental Details

  • The U-Net was trained for 25 epochs on A100 GPUs, with three runs reported; longer 50- and 100-epoch runs did not improve performance.The selected multi-task loss weights burn probability, fire intensity, and rate of spread at 0.5, 0.25, and 0.25, respectively.Source: D Experimental Details · p. 19

E Additional Results

Across five Canadian test regions, burn probability generally shows strong agreement with BurnP3+, while agreement varies more for fire intensity and rate of spread. In Region 16, fire intensity and rate of spread preserve low-to-moderate spatial ordering but underrepresent their upper tails.

Region 16 analysis

  • In Region 16, fire intensity and rate of spread retain low-to-moderate ordering but regress toward the mean and underrepresent long upper tails.Burn probability has the closest agreement, though its highest values are systematically underestimated.Source: E Additional Results · pp. 20–21
  • For fire intensity and rate of spread, reasonable correlation alongside lower CCC reflects better capture of relative spatial variation than of extreme-value magnitude and variability.Source: E Additional Results · pp. 20–21

Regional performance

  • Burn probability achieves CCC above 0.8 in four of five regions, exceeding 0.89 in Regions 01, 16, and 49.Fire intensity reaches CCC 0.84 in Region 01; rate of spread exceeds 0.8 only in Region 01.Source: E Additional Results · p. 21

F Hazard Evaluation Details

The hazard evaluation derives a pixel-wise index from burn probability and capped fire intensity, rescales it using the reference maximum, and assigns 13 ordinal classes. The surrogate agrees strongly across elevated-hazard areas, but class 13 has zero IoU because extreme reference values raise its effective threshold.

Hazard calculation

  • The pixel-wise hazard procedure caps average intensity at 10,000 kW · m−1, multiplies it by burn probability, rescales to [0, 100] using the reference maximum, then discretizes into 13 classes.Source: F Hazard Evaluation Details · p. 21

Hazard classification results

  • Errors concentrate near the diagonal at moderate-to-high hazard levels, while narrow low-hazard bins make those classes more sensitive to small continuous-value differences.Source: F Hazard Evaluation Details · p. 21
  • Approximately 85% IoU for classes 10–13 indicates strong spatial agreement with BurnP3+ in delineating elevated-hazard areas.Source: F Hazard Evaluation Details · p. 21
  • Class 13 has zero IoU because extreme BurnP3+ values raise the scaled threshold and surrogate-underestimated peaks fall into lower classes.This reflects failure to reproduce the most extreme magnitudes, not a broader failure to identify high-hazard areas.Source: F Hazard Evaluation Details · p. 21

G Additional Experiments

Additional experiments find that spatial weather encoding generally performs better than tabular encoding, while joint multi-task learning benefits burn probability but underperforms separate models for fire intensity and rate of spread. The unified model is more computationally efficient, but reducing negative transfer remains future work.

Architecture Ablation

  • Spatial weather encoding achieves stronger performance across most metrics and all three targets than tabular encoding.Tabular encoding has biases closer to zero for burn probability and fire intensity and slightly higher burn-probability CCC, but trails on remaining correlation, error, and spatial-overlap metrics.Source: G Additional Experiments · p. 21

Single-Task vs. Multi-Task Models

  • Multi-task learning outperforms single-task models on most burn-probability metrics but underperforms them across all fire-intensity and rate-of-spread metrics.The comparison uses the same backbone, inputs, and training settings.Source: G Additional Experiments · p. 21
  • The unified multi-task model offers computational and deployment advantages, while future work will investigate strategies to reduce negative transfer and match or exceed task-specific performance.Source: G Additional Experiments · pp. 21–22

H Operationally Representative Post-Fire Fuel Intervention

A localized post-fire fuel intervention produces closely matching region-wide mean responses from the surrogate and BurnP3+, with spatial agreement strongest near modified fuels. Additional results report aligned cross-study trends, and the framework can evaluate existing simulations or train models on users’ own data.

H Operationally Representative Post-Fire Fuel Intervention

  • The intervention modifies fuels in polygons from about 84 fires affecting 243,100 ha, or 1.45% of Region 16, representing changes plausible after one to two fire seasons.Source: H Operationally Representative… · p. 22
  • For the post-fire D2 intervention, surrogate mean changes closely match BurnP3+: BP −7.745% ± 0.078 versus −7.803%, FI −2.842% ± 0.037 versus −3.058%, and ROS −1.979%±0.048 versus −2.076%.Changes concentrate within and around modified fuel areas, with comparatively limited response elsewhere.Source: H Operationally Representative… · p. 22
  • Within edited areas, Spearman correlations are 0.87, 0.82, and 0.67 for BP, FI, and ROS, respectively, falling with distance as the intervention signal weakens.Only 1.45% of the region is directly modified, making region-wide rankings sensitive to small differences where changes are very small.Source: H Operationally Representative… · p. 22
  • The surrogate closely reproduces BurnP3+’s region-wide mean response and localized response around modified fuels under this operationally plausible, spatially limited intervention.Source: H Operationally Representative… · p. 22

I Cross-Study Model Transfer Additional Results

  • Cross-study evaluation reports that performance trends align closely between the two model configurations.Table 13 presents predictive performance for the Spatial + Weather + Fire Size model on a separate regional study, with means and standard deviations across three seed runs.Source: I Cross-Study Model Transfer… · p. 23, Table 13 · p. 27

J How to Use Our AI Framework

  • Users with an existing BurnP3 or BurnP3+ simulation can evaluate provided models or train new ones using spatial rasters, weather and fire-size data, simulation-output maps, and auxiliary inputs.The repository README provides instructions for preparing inputs, training and evaluation, and inference.Source: J How to Use Our AI Framework · pp. 23, 25

Related papers

Loading 2610.30170v2…