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

cs.AIOct 16, 2026

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

Authors: Hager Radi Abdelwahed, Toluwaleke Olutayo, Chirag Goel, Jana Osea, Aldo Zaimi, Tom Swystun, Sandy Erni, Francisco Emiliano Lopez Saavedra,

Organizations: Mila: Quebec AI Institute · Natural Resources Canada - Canadian Forest Services

Abstract

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.

Figures & tables

Figure 1: Overview of the proposed AI surrogate framework . (a) Geographic regions (hexels) across Canada are split into training, validation, and test sets (44, 5, and 5 regions, respectively). A standard region (hexel) covers an area of 130,000 $\mathrm{km}^{2}$ . (b) Hexels are divided into square patches for training. (c) A multi-source U-Net model combines spatial and tabular data to predict wildfire hazard in a multi-task fashion. (d) Three spatial outputs are predicted: burn probability, fire intensity, and rate of spread. (e) During inference, predicted patches are reassembled into complete regions using a sliding window approach.

Target	Model	$\downarrow$ Normalized MAE	$\rightarrow 0$ Normalized Bias	$\uparrow$ Spearman	$\uparrow$ CCC	$\uparrow$ AUC IoU Top-10%	$\uparrow$ AUC IoU Full
BP	Mean Baseline	$5.18$	$4.53$	$-$	$-$	$0.10$	$0.55$
	XGBoost	$1.19\pm 0.02$	$0.47\pm 0.06$	$0.70\pm 0.01$	$0.33\pm 0.01$	$0.05\pm 0.01$	$0.59\pm 0.01$
	U-Net Spatial	$0.50\pm 0.02$	$-0.18\pm 0.06$	$0.88\pm 0.00$	$0.67\pm 0.03$	$0.24\pm 0.01$	$0.69\pm 0.00$
	+ Weather	$\mathbf{0.35\pm 0.00}$	$\mathbf{-0.14\pm 0.03}$	$\mathbf{0.91\pm 0.00}$	$\mathbf{0.85\pm 0.00}$	$0.39\pm 0.04$	$\mathbf{0.75\pm 0.00}$
	+ Weather + Fire Size	$0.35\pm 0.02$	$\mathbf{-0.14\pm 0.03}$	$\mathbf{0.91\pm 0.00}$	$0.85\pm 0.01$	$\mathbf{0.40\pm 0.04}$	$\mathbf{0.75\pm 0.00}$
FI	Mean Baseline	$0.64$	$-0.32$	$-$	$-$	$0.10$	$0.50$

Table 1: Main results of AI surrogate model comparing the predictive performance of the baselines and incremental multi-source U-Net models for Burn Probability (BP) , Fire Intensity (FI) , and Rate of Spread (ROS) . The baseline spatial architecture is further augmented with weather features and fire size features. Each value is reported as the mean and standard deviation of three runs with different random seeds (standard deviations less than

0.01

are reported as

0.00

). Metrics are reported on the test set at region level ; patch-level results are reported in Table 8 of Appendix E . Best performance for each specified metric is highlighted in bold. Note: Correlation metrics are undefined for the mean baseline.

Panel of Figure 2 — Figure 2: Comparison of the AI surrogate model predictions against BurnP3+ reference outputs for a representative test region (region 16) across the three targets: BP, FI, and ROS. Masked-out regions represent water bodies or non-fuel areas.

BP3+ Baseline: 38.9 h (140,040 s) \| Memory: 100.8 GB \| 32 CPUs
Hardware Setup	Inference (s)	Total E2E (s)	RAM + Device
GPU (L40S)	$7.1\pm 0.2$	$40.8\pm 0.2$	8 + 29 GB (37%)
MacBook (MPS)	$78.9\pm 2.4$	$93.9\pm 2.5$	30 GB (30%)
CPU (4 Cores)	$138.6\pm 0.7$	$173.6\pm 2.2$	13 GB (13%)
Hardware Setup	Speed-up Factor vs. BurnP3+
GPU (L40S)	$\sim\mathbf{3,432.35\times}$

Table 2: Performance and memory comparison between the simulator baseline (BurnP3+) and the AI surrogate model across hardware variants. Values are shown for region 16, one of 54 distinct geographical zones used to partition the Canadian map. Percentages in parentheses indicate memory usage relative to the baseline’s 100.8 GB.

Metric	Value
Exact 13-Class Accuracy	$0.476\pm 0.008$
Within-1 Bin Accuracy	$0.873\pm 0.004$
Within-2 Bin Accuracy	$0.970\pm 0.001$
High Hazard ( $\geq 10$ ) IoU	$0.850\pm 0.002$

Table 3: AI surrogate model performance on downstream hazard mapping metrics across test regions.

Figure 3: Fuel interventions in region 16: the left panel shows the first intervention: C2 pixels replaced with M1, while the right panel shows the second intervention: pixels within the selected BurnP3+ fire polygons reassigned to D2.

Figure 4: BurnP3+ and AI surrogate model responses for the two fuel interventions in region 16. Rows show the shared baseline (top), the C2-to-M1 intervention (middle), and the post-fire D2 intervention (bottom). Columns are grouped by output: BP, FI, and ROS. Within each pair, BurnP3+ predictions appear on the left and surrogate predictions on the right.

Figure 5: Counterfactual-minus-baseline changes produced by BurnP3+ and the surrogate for the two fuel interventions in region 16. Columns show changes in BP, FI, and ROS. The first two rows show the BurnP3+ and surrogate responses to the C2-to-M1 intervention, respectively. The last two rows show the corresponding responses to the post-fire D2 intervention.

Figure 6: Spatial response of the surrogate to the weather intervention in region 16, obtained by replacing the baseline weather inputs with the corresponding inputs from region 17 in British Columbia. The panels show the resulting counterfactual-minus-baseline changes for BP, FI, and ROS from left to right.

Figure 7: Spatial response of the surrogate to the climate-change intervention in region 16. From left to right, the panels show the counterfactual-minus-baseline changes in BP, FI, and ROS.

Intervention	Model	$\Delta\overline{\mathrm{BP}}$	$\Delta\overline{\mathrm{FI}}$	$\Delta\overline{\mathrm{ROS}}$
C2 $\rightarrow$ M1	BurnP3+	$-70.64\%$	$-63.78\%$	$-38.37\%$
C2 $\rightarrow$ M1	AI Surrogate	$-74.30\%\pm 3.24\%$	$-62.36\%\pm 0.39\%$	$-38.35\%\pm 0.83\%$
Burned area $\rightarrow$ D2	BurnP3+	$-72.38\%$	$-60.53\%$	$-48.29\%$
Burned area $\rightarrow$ D2	AI Surrogate	$-87.59\%\pm 0.64\%$	$-57.78\%\pm 0.37\%$	$-40.27\%\pm 0.67\%$
QC $\rightarrow$ BC weather	AI Surrogate	$206.18\%\pm 59.22\%$	$97.79\%\pm 0.71\%$	$23.67\%\pm 1.74\%$
Climate-adjusted spread duration	AI Surrogate	$271.39\%\pm 58.22\%$	$1.27\%\pm 0.71\%$	$-0.20\%\pm 2.20\%$

Table 4: Aggregate counterfactual responses for region 16. Each entry gives the counterfactual-minus-baseline percentage changes in the region-wide mean.

Scenario A
Target	Model	$\downarrow$ Normalized MAE	$\rightarrow 0$ Normalized Bias	$\uparrow$ Spearman	$\uparrow$ CCC	$\uparrow$ AUC IoU Top-10%	$\uparrow$ AUC IoU Full
BP	U-Net (Base)	$4.620\pm 0.891$	$4.595\pm 0.895$	$0.616\pm 0.044$	$0.076\pm 0.021$	$0.102\pm 0.010$	$0.550\pm 0.015$
	+ Recalibration	$0.500\pm 0.057$	$-0.435\pm 0.090$	$0.616\pm 0.044$	$0.366\pm 0.064$	$0.102\pm 0.010$	$0.550\pm 0.015$
	+ Zone Alignment	$\mathbf{0.335}\pm 0.071$	$\mathbf{-0.260}\pm 0.113$	$\mathbf{0.890}\pm 0.004$	$\mathbf{0.708}\pm 0.093$	$\mathbf{0.266}\pm 0.018$	$\mathbf{0.718}\pm 0.006$
FI	U-Net (Base)	$\mathbf{0.280}\pm 0.004$	$-0.106\pm 0.036$	$\mathbf{0.812}\pm 0.002$	$0.740\pm 0.012$	$\mathbf{0.142}\pm 0.003$	$\mathbf{0.652}\pm 0.002$
	+ Recalibration	$0.325\pm 0.019$	$0.144\pm 0.043$	$\mathbf{0.812}\pm 0.002$	$\mathbf{0.757}\pm 0.011$	$\mathbf{0.142}\pm 0.003$	$\mathbf{0.652}\pm 0.002$

Table 5: Main results comparing the predictive performance of our multi-source U-Net models for Burn Probability (BP) , Fire Intensity (FI) , and Rate of Spread (ROS) on a separate regional study . U-Net model is based on Spatial + Weather input data. Each value is reported as the mean and standard deviation of 3 different-seed runs. Metrics are reported on the test set at region level . Best performance for each specified metric is highlighted in bold.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

ID	Name	Description	fbp_code	FuelType	SeasonState
1	Spruce-Lichen Woodland	C-1	1	C-1	direct
2	Boreal Spruce	C-2	2	C-2	direct
3	Mature Jack or Lodgepole Pine	C-3	3	C-3	direct
4	Immature Jack or Lodgepole Pine	C-4	4	C-4	direct
5	Red and White Pine	C-5	5	C-5	direct
6	Ponderosa Pine - Douglas-Fir	C-7	7	C-7	direct

Appendix

Table 6: Canadian fuel types present in our dataset, according to the FBP system.

Figure 8: FBP-generated initial Rate of Spread (iROS) curves across fuel classes as a function of the Initial Spread Index (ISI). These curves serve as a continuous functional representation of fuel characteristics, replacing discrete fuel codes.

Feature Name	Source	Dimension	Processing
Input: Spatial
Fuel	Forest Fire Behaviour Prediction (FBP) System ( Forestry Canada Fire Danger Group 1992 )	$H\times W\times 4$	Neural network compression of 18-D iROS curves
Ignition	( Erni et al. 2024 )	$H\times W\times 2$	Cause-aggregated layers with mean imputation
Elevation	Digital Elevation Model ( Natural Resources Canada 2024 )	$H\times W\times 1$	Min-max normalization (train-split statistics)
Terrain Derivatives	Derived from Elevation	$H\times W\times 3$	Slope via $\arctan(G)/(\pi/2)$ ; Aspect via unit vector ( $\sin,\cos$ )
Input: Tabular (Weather & Environmental Context)

Appendix

Table 7: Summary of input features and output targets, detailing spatial and tabular dimensions along with exact preprocessing configurations and sources.

Target	Model	$\downarrow$ Normalized MAE	$\rightarrow 0$ Normalized Bias	$\uparrow$ Spearman	$\uparrow$ CCC	$\uparrow$ AUC IoU Top-10%	$\uparrow$ AUC IoU Full
Patch-Level Evaluation
BP	Mean baseline	$9.33$	$8.70$	$-$	$-$	$0.16$	$0.60$
	XGBoost	$1.24\pm 0.02$	$0.55\pm 0.05$	$0.55\pm 0.01$	$0.28\pm 0.00$	$0.12\pm 0.00$	$0.58\pm 0.01$
	U-Net Spatial	$0.52\pm 0.01$	$-0.16\pm 0.07$	$0.78\pm 0.00$	$0.57\pm 0.02$	$0.31\pm 0.00$	$0.67\pm 0.00$
	+ Weather	$\mathbf{0.36\pm 0.01}$	$\mathbf{-0.12\pm 0.03}$	$\mathbf{0.80\pm 0.00}$	$\mathbf{0.69\pm 0.00}$	$0.32\pm 0.01$	$0.68\pm 0.00$
	+ Weather + Fire Size	$0.36\pm 0.02$	$-0.13\pm 0.03$	$\mathbf{0.80\pm 0.00}$	$0.69\pm 0.01$	$\mathbf{0.33\pm 0.00}$	$\mathbf{0.69\pm 0.00}$

Appendix

Table 8: Main results comparing the predictive performance of incremental multi-source U-Net models for Burn Probability (BP) , Fire intensity (FI) , and Fire Rate of Spread (ROS) . The baseline spatial architecture is progressively augmented with weather. Each value is reported as the mean and standard deviation of 3 different-seed runs (standard deviations less than

0.01

are reported as

0.00

). Metrics are reported on the test set at the patch-level. Best performance for each specified metric is highlighted in bold. Note: Correlation metrics are undefined for the mean baseline.

Panel of Figure 9 — Figure 9: Prediction analysis for test Region 16 for the 3 targets: BP, FI, and ROS. Top row: probability-density comparisons using log-scaled histograms. Bottom row: 2D Hexbin Calibration Plot between AI-surrogate model predictions and BurnP3+ reference maps.

Target	Hexel	$\downarrow$ Normalized MAE	$\rightarrow 0$ Normalized Bias	$\uparrow$ Spearman	$\uparrow$ CCC	$\uparrow$ AUC IoU Top-10%	$\uparrow$ AUC IoU Full
BP	Region 01	$0.277\pm 0.019$	$-0.134\pm 0.043$	$0.951\pm 0.003$	$0.894\pm 0.018$	$0.417\pm 0.046$	$0.806\pm 0.009$
	Region 12	$0.378\pm 0.086$	$-0.157\pm 0.157$	$0.887\pm 0.001$	$0.866\pm 0.069$	$0.429\pm 0.145$	$0.734\pm 0.021$
	Region 16	$0.284\pm 0.057$	$-0.194\pm 0.099$	$0.953\pm 0.001$	$0.908\pm 0.045$	$0.450\pm 0.027$	$0.810\pm 0.001$
	Region 39	$0.517\pm 0.012$	$-0.005\pm 0.073$	$0.819\pm 0.009$	$0.683\pm 0.016$	$0.165\pm 0.008$	$0.639\pm 0.004$
	Region 49	$0.279\pm 0.014$	$-0.190\pm 0.033$	$0.944\pm 0.005$	$0.905\pm 0.011$	$0.487\pm 0.008$	$0.769\pm 0.007$
FI	Region 01	$0.213\pm 0.003$	$-0.095\pm 0.014$	$0.883\pm 0.004$	$0.840\pm 0.006$	$0.199\pm 0.010$	$0.710\pm 0.004$

Appendix

Table 9: Region-level performance across individual test hexels for our best model (Spatial + Weather). Values report the mean

\pm

standard deviation across three random seeds.

Hazard class	$\boldsymbol{H_{\mathrm{scaled}}}$ interval
1	$H_{\mathrm{scaled}}<0.01$
2	$0.01\leq H_{\mathrm{scaled}}<0.025$
3	$0.025\leq H_{\mathrm{scaled}}<0.05$
4	$0.05\leq H_{\mathrm{scaled}}<0.1$
5	$0.1\leq H_{\mathrm{scaled}}<0.25$
6	$0.25\leq H_{\mathrm{scaled}}<0.5$

Appendix

Table 10: NRCan hazard bin definitions used to discretize the scaled hazard index

H_{\mathrm{scaled}}

Figure 10: Row-normalized confusion matrix for BurnP3+ reference and surrogate-predicted hazard classes.

Target	Model	$\downarrow$ Normalized MAE	$\rightarrow 0$ Normalized Bias	$\uparrow$ Spearman	$\uparrow$ CCC	$\uparrow$ AUC IoU Top-10%	$\uparrow$ AUC IoU Full
BP	U-Net Spatial	$0.496$	$-0.199$	$0.878$	$0.674$	$0.239$	$0.692$
	+ Weather (Spatial)	$\mathbf{0.345}$	$-0.176$	$\mathbf{0.915}$	$0.847$	$\mathbf{0.375}$	$\mathbf{0.754}$
	+ Weather (Tabular)	$0.371$	$\mathbf{-0.025}$	$0.905$	$\mathbf{0.851}$	$0.362$	$0.739$
FI	U-Net Spatial	$0.301$	$-0.057$	$0.854$	$0.768$	$0.237$	$0.691$
	+ Weather (Spatial)	$\mathbf{0.272}$	$-0.048$	$\mathbf{0.859}$	$\mathbf{0.788}$	$\mathbf{0.246}$	$\mathbf{0.697}$
	+ Weather (Tabular)	$0.290$	$\mathbf{-0.021}$	$0.853$	$0.779$	$0.223$	$0.688$

Appendix

Table 11: Region-level comparison of multi-task U-Net models using spatial and tabular weather encoders for Burn Probability (BP) , Fire Intensity (FI) , and Rate of Spread (ROS) . The U-Net Spatial model does not include weather variables. Best performance for each target and metric is highlighted in bold, with normalized bias evaluated according to proximity to zero.

Target	Model	$\downarrow$ Normalized MAE	$\rightarrow 0$ Normalized Bias	$\uparrow$ Spearman	$\uparrow$ CCC	$\uparrow$ AUC IoU Top-10%	$\uparrow$ AUC IoU Full
BP	Single-task	$0.404$	$\mathbf{-0.137}$	$0.893$	$0.783$	$\mathbf{0.427}$	$0.744$
	Multi-task	$\mathbf{0.345}$	$-0.176$	$\mathbf{0.915}$	$\mathbf{0.847}$	$0.375$	$\mathbf{0.754}$
FI	Single-task	$\mathbf{0.259}$	$\mathbf{0.009}$	$\mathbf{0.873}$	$\mathbf{0.814}$	$\mathbf{0.277}$	$\mathbf{0.712}$
	Multi-task	$0.272$	$-0.048$	$0.859$	$0.788$	$0.246$	$0.697$
ROS	Single-task	$\mathbf{0.204}$	$\mathbf{-0.026}$	$\mathbf{0.799}$	$\mathbf{0.747}$	$\mathbf{0.248}$	$\mathbf{0.665}$
	Multi-task	$0.216$	$-0.044$	$0.782$	$0.717$	$0.225$	$0.652$

Appendix

Table 12: Region-level comparison between a multi-task U-Net that jointly predicts all three targets and three single-task U-Nets trained separately for Burn Probability (BP) , Fire Intensity (FI) , and Rate of Spread (ROS) . Both configurations use identical inputs, architecture, and training settings. Best performance for each target and metric is highlighted in bold.

Figure 11: Post-fire D2 intervention in region 16, showing fuels replaced within selected BurnP3+ fire polygons.

Figure 12: BurnP3+ and surrogate outputs under the baseline and post-fire D2 intervention in region 16.

Figure 13: Counterfactual-minus-baseline changes in BP, FI, and ROS for BurnP3+ and the surrogate under the post-fire D2 intervention.

Scenario A
Target	Model	$\downarrow$ Normalized MAE	$\rightarrow 0$ Normalized Bias	$\uparrow$ Spearman	$\uparrow$ CCC	$\uparrow$ AUC IoU Top-10%	$\uparrow$ AUC IoU Full
BP	U-Net (Base)	$2.605\pm 1.108$	$2.517\pm 1.161$	$0.543\pm 0.077$	$0.121\pm 0.026$	$0.085\pm 0.018$	$0.524\pm 0.028$
	+ Recalibration	$0.666\pm 0.111$	$-0.645\pm 0.117$	$0.544\pm 0.076$	$0.182\pm 0.122$	$0.085\pm 0.018$	$0.524\pm 0.028$
	+ Zone Alignment	$\mathbf{0.652}\pm 0.122$	$\mathbf{-0.631}\pm 0.128$	$\mathbf{0.588}\pm 0.094$	$\mathbf{0.209}\pm 0.148$	$\mathbf{0.085}\pm 0.017$	$\mathbf{0.538}\pm 0.034$
FI	U-Net (Base)	$\mathbf{0.285}\pm 0.005$	$\mathbf{-0.095}\pm 0.041$	$\mathbf{0.809}\pm 0.005$	$0.730\pm 0.011$	$\mathbf{0.143}\pm 0.011$	$\mathbf{0.649}\pm 0.003$
	+ Recalibration	$0.330\pm 0.023$	$0.158\pm 0.050$	$\mathbf{0.809}\pm 0.005$	$\mathbf{0.747}\pm 0.019$	$\mathbf{0.143}\pm 0.011$	$\mathbf{0.649}\pm 0.003$

Appendix

Table 13: Main results comparing the predictive performance of our multi-source U-Net models for Burn Probability (BP) , Fire Intensity (FI) , and Rate of Spread (ROS) on a separate regional study . U-Net model is based on Spatial + Weather + Fire Size input data. Each value is reported as the mean and standard deviation of 3 different-seed runs. Metrics are reported on the test set at region level . Best performance for each specified metric is highlighted in bold.

Surrogate Modeling

Surrogate modeling builds faster or more interpretable approximations of expensive simulations and black-box models.

See all 229 papers

Sep 1, 2026

Random Forest-Informed Cellular Automaton for Large-Scale Wildfire Spread Modelling

Accurate large-scale wildfire spread modelling requires models that capture both the environmental conditions associated with fire occurrence and the local dynamics of fire propagation. We propose a three-stage framework that combines a Random Forest (RF) model with a cellular automaton (CA). First, an RF model trained on the 2021 Canadian fire season estimates daily pixel-level fire-occurrence probabilities. Second, quantile gradient boosting models provide optional spread-rate priors for sensitivity analysis. Third, an RF-informed CA combines the RF probability layer with neighbourhood-driven spread on a 5 km grid. The RF model achieved AUC values of 0.725--0.795 on the 2022--2024 datasets, while the RF-informed CA achieved substantially higher spatial overlap than the evaluated CA-only baselines in the 2023 simulation. A higher-resolution simulation provides an additional qualitative assessment of local spatial errors. These results suggest that combining RF-derived probabilities with local CA spread can improve large-scale wildfire simulations under the tested conditions.

Siyu Chen, Esha Saha, Hao Wang

Cellular Automata Wildfire Forecasting

Ask PDF

Jun 10, 2026

Neural-Parameterized Cellular Automata for Wildfire Spread

Traditional wildfire models rely on rigid, low-dimensional parameters and static fuel maps, frequently underpredicting fire spread. To address this weakness, we introduce a hybrid deep-learning parameterized Probabilistic Cellular Automata (CA) framework implemented in JAX. Our approach employs a Multi-Scale Convolutional Neural Network to dynamically generate spatially varying parameters that govern fire-spread probability, wind alignment, and slope influence. This hybrid design captures complex, nonlinear environmental interactions while preserving the physical interpretability of the underlying three-state CA. The JAX implementation enables hardware acceleration and gradient-based parameter calibration. Evaluated on six large-scale wildfires in the western United States, the model maintains IoU > 0.6 over 72-hour forecast horizons after a 10-day data assimilation window during which the model is fitted incrementally to observed perimeters; the resulting forecast is a conditional projection of fire growth under the suppression regime already ncoded in those observations.

Maksym Zhenirovskyy, Ion Matei, Rohit Vuppala +2

Fujitsu Research of America, 4655 Great America Pkwy, Santa Clara, 95124, CA, USA

Neural Cellular Automata Cellular Automata Wildfire Forecasting Spatiotemporal Forecasting

Ask PDF

Sep 16, 2026

Interpretable Patch-Based Deep Learning for Wildfire Spread Prediction from Ensemble Simulations

Wildfire spread is traditionally predicted using physics-based simulators, which are physically interpretable but whose cost increases with each additional ensemble member. We ask how well deep learning surrogates can reproduce these simulations at a fraction of this cost, training them on 10,584 fire spread simulations at 2m resolution for the Rectoret region in Catalonia, Spain. Four architectures are compared: a patch-based U-Net, a transfer-learned ResNet-50, a physics-informed network constrained by the wind-driven advection equation and a Swin-Unet transformer. Among the terrain and vegetation variables, only surface fuel load predicts burn probability with any strength (r = 0.27) and including it lowers prediction error by 21%. The remaining variables correlate weakly and are highly duplicative. Next, an experiment with saliency, occlusion and rotation demonstrates the models' learning. Convolutional models rely primarily on distance from the current fire front, while Swin-Unet assigns more weight to fuel and terrain, a finding also noted in an unrelated wildfire dataset. When applied without retraining to the second region, Pedriza, all three convolutional models still predict fire spread, losing accuracy by a small but systematic margin.

Marcin Lawenda, Aleksandra Krasicka, David Caballero +2

Poznan Supercomputing and Networking Center, Poznań, Poland · MeteoGrid, Madrid, Spain · Technical University of Czestochowa, Częstochowa, Poland

Neural Surrogate Modeling Wildfire Forecasting Physics-Informed ML Convolutional Neural Networks

Ask PDF

Abstract

Surrogate Modeling

Explore similar work

Random Forest-Informed Cellular Automaton for Large-Scale Wildfire Spread Modelling

Neural-Parameterized Cellular Automata for Wildfire Spread

Interpretable Patch-Based Deep Learning for Wildfire Spread Prediction from Ensemble Simulations