Source-linked AI summary
Interpretable Patch-Based Deep Learning for Wildfire Spread Prediction from Ensemble Simulations
Marcin Lawenda, Aleksandra Krasicka, David Caballero, Luis Torres, Łukasz Szustak
TL;DR
Repeated physics-based simulations make burn-probability mapping costly, motivating deep-learning surrogates. The paper compares four architectures on Rectoret simulations and analyzes their predictions; surface fuel load stands out among covariates, while the architectures differ in the cues they emphasize.
Problem
Physics-based simulators are interpretable and well validated, but their cost rises with each simulation needed for a burn-probability map.
Source: §1 · p. 1Method
The study compares four deep-learning architectures for predicting fire spread on patches, alongside covariate, interpretability, and second-region analyses.
Source: §1 · p. 2Results
Surface fuel load accounts for most explainable variance in burn probability, and convolutional models and Swin-Unet differ in the cues they emphasize.
Source: §9 · p. 13, §10 · p. 13Takeaways & Limitations
The authors suggest fuel management focus specifically on surface fuel loading rather than the full suite of continuity metrics.
Source: §9 · p. 13The Pedriza comparison is descriptive; establishing generalization would require a second region mapped at the same 2 m resolution and additional regions.
Source: §8 · p. 12
Abstract
from arXiv · showhide
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.
1 Introduction
Physics-based fire-spread simulators are interpretable and well validated, but generating burn-probability maps multiplies their simulation cost. This work examines deep-learning surrogates trained on Rectoret simulations and compares four architectures alongside analyses of covariates, model behavior, and transfer to Pedriza.
1 Introduction
- Reliable burn-probability maps require many complete simulations, multiplying the cost of a single run.Each run uses a new combination of wind and ignition conditions.Source: §1 · pp. 1–2
- A trained machine-learning surrogate can reproduce simulator output in one forward pass, making inference virtually free.The dataset records fire-arrival times across Rectoret simulations with varying wind and ignition conditions, plus terrain and fuel descriptors.Source: §1 · p. 2
- The study analyzes covariate relationships, compares four models for 30-minute-ahead prediction on 128 × 128 patches, and examines learned cues and Pedriza performance.The models are a custom U-Net, transfer-learned ResNet-50, physics-informed variant, and Swin-Unet.Source: §1 · p. 2
2 Related Work
Prior work spans physics-based fire simulators and trained surrogates, including CNN/U-Net, physics-informed, and transformer approaches. This paper brings these model families into one dataset and patching scheme, and compares interpretability and performance across two regions.
Physics-based simulators
- Physics-based simulators use Rothermel equations [23], including FARSITE and FlamMap, while cellular automata offer a grid-based alternative.Differentiable GPU-accelerated CA simulators have shown better interregion transfer than purely supervised surrogates.Source: §2 · p. 2
CNN/U-Net surrogates
Transfer learning and physics-informed networks
- Transfer learning motivates the ResNet-50 baseline, while physics-informed networks penalize violations of differential equations, including fire-spread constraints.The cited fire-spread model uses mass and energy conservation constraints with simulated and real thermal data.Source: §2 · p. 3
Transformer-based architectures
Contribution
- This study compares data-driven CNN, transfer-based, physics-informed, and transformer models under one dataset and patching scheme, with interpretability analyses and a two-region comparison.The cited literature more often studies a single model family or region.Source: §2 · p. 3
3 Study Area and Data
The study uses an ensemble of 10,584 stochastic fire-spread simulations for Rectoret and terrain and fuel descriptors as model inputs. Pedriza provides a second, coarser-resolution region for descriptive comparison and model transfer testing.
Simulation ensemble
- The Rectoret dataset contains 10,584 stochastic fire-spread simulations at 2 m resolution, each recording cell-level fire arrival time capped at 180 minutes.Simulations vary wind speed, wind direction, and ignition offset.Source: §3 · p. 3
- The four simulation parameters form a balanced full-factorial design: 3 wind speeds, 8 directions, and a 21 × 21 grid of ignition points.Each of 24 wind configurations is paired with all 441 ignition points, yielding 10,584 runs.Source: §3 · p. 3, Table 1 · p. 4
Terrain and fuel covariates
- Inputs include terrain and vegetation-continuity descriptors, surface fuel load, and other fuel measures; crown fuel is excluded from deep-learning inputs.Surface fuel load is derived from the same fuel-model classification used by the simulator and is normalized as one of 11 input channels.Source: §3 · p. 4
Second region: Pedriza
- Pedriza covers about 128 km² at 10 m resolution, versus Rectoret’s 2 m grid, so equal-sized patches span 25 times more ground area there.The region is used for descriptive aggregation and as a native-resolution model-transfer test bed.Source: §3 · p. 4
4 Exploratory Analysis
Surface fuel load is the clearest covariate-level predictor of burn probability and improves XGBoost prediction when included. Other vegetation-continuity descriptors largely duplicate one another, while wind sensitivity depends on simulation-level direction and speed.
Correlation and spatial autocorrelation
- Vegetation-continuity and fuel descriptors have moderate spatial clustering, with Moran’s I around 0.25–0.35, while slope is 0.10 and aspect is −0.02.The reported values indicate little spatial autocorrelation for slope and aspect at this scale.Source: §4 · p. 4, Figure 1 · p. 5
- At the pixel level, surface fuel load correlates with BP at r = 0.27, while aspect, slope, and continuity descriptors remain at |r| ≤ 0.07.It is the one clear outlier among the listed covariates.Source: §4 · p. 4, Figure 2 · p. 5
- At the coarser aggregated-cell level, vegetation-continuity descriptors correlate mostly above r = 0.95, indicating that they carry largely the same information.Aspect maxima correlate more moderately with continuity and biomass maxima, at r ≈ 0.73–0.83.Source: §4 · p. 5
Wind sensitivity
- Burn probability rises for wind from 315°, and higher wind speeds enlarge and merge high-BP regions, despite the models’ small wind-importance scores.Wind speed and direction are constant within each simulation and therefore cannot enter the pixel-level analysis directly.Source: §4 · p. 5
Feature importance via gradient boosting
- Including surface fuel load reduces BP prediction MSE from 386.19 to 305.32, a 21% reduction.The comparison uses XGBoost with and without the covariate.Source: §4 · p. 5, Figure 3 · p. 6
- With surface fuel load included, its XGBoost importance is 0.700; without it, importance shifts toward CONT_NORM_2_20 (0.260) and aspect (0.210).Crown fuel has effectively zero importance when surface fuel load is included.Source: §4 · p. 5, Figure 4 · p. 6
5 Methods
The methods frame spread prediction as patch-based image-to-image translation and compare convolutional, transfer-learned, physics-informed, and transformer models. The Swin-Unet comparison isolates encoder type, while the PINN adds a wind-advection penalty to the Custom U-Net.
5 Methods
- The Custom U-Net predicts a 128 × 128 fire-state patch from eleven channels describing fire state, front distance, terrain and fuel, and wind.The selected configuration uses Leaky ReLU, base filter width 32, depth 2, and a 0.5 BCE + 0.5 Dice loss.Source: §5 · pp. 6–7
- The ResNet-50 baseline projects the eleven input channels to three, then decodes the pretrained backbone output without encoder–decoder skip connections.The backbone is frozen for 40 epochs before fine-tuning.Source: §5 · pp. 6–7
- The PINN adds a wind-driven advection penalty and spatial-smoothness term to the Custom U-Net, but its λ2 sweep found no significant test Dice-loss effect (p = 0.46).Across three seeds, test Dice loss ranged from 0.086 to 0.090, with seed-to-seed variation larger than the spread between settings.Source: §5 · p. 7
- The Swin-Unet uses a SwinTransformerV2-Tiny encoder with a convolutional decoder, and differs from the ResNet-50 baseline only in its encoder.Both use the same channel projection, five-stage decoder, no skip connections, and 40-epoch backbone freeze.Source: §5 · pp. 7–8
6 Experimental Setup
The setup splits simulations by ignition location, keeps patches from each simulation within one split, and evaluates reconstructed full-domain predictions with five metrics and fire-size stratification.
6 Experimental Setup
- The test set contains 2,400 simulations (22.7%) and the training set 8,184 (77.3%), split by ignition offset on a 200 m grid.Patches are extracted separately within each set, preventing any simulation from contributing patches to both.Source: §6 · p. 8
- Five metrics are reported: MAE, MSE, Dice loss, MCC, and a custom weighted error whose fixed weights make model comparisons possible but not absolute-scale interpretation.MCC remains informative under class imbalance; the weighted error allows false positives and false negatives to receive different penalties.Source: §6 · p. 8
- Evaluation is stratified into four bins by final burned-area fraction, from 0–0.54% through 1.70–2.24%.Overlapping patch predictions are averaged into a full-domain map and thresholded at 0.5.Source: §6 · p. 8
7 Results
Custom U-Net and PINN lead the model comparison, while interpretability tests reveal different channel reliance across architectures. Input ablations show that fire-state information dominates accuracy, though environmental inputs retain a measurable signal.
7.1 Quantitative model comparison
- Custom U-Net and PINN lead every metric, with no significant difference between them; their weighted errors are 0.277 and 0.284, versus 0.861 for ResNet-50 and 2.353 for Swin-Unet.The two leading models use 0.47 M parameters each, against 34.6 M for ResNet-50 and 32.7 M for Swin-Unet.Source: §7.1 · p. 8, Table 2 · p. 9
7.2 Interpretability analysis
- Occlusion ranks current fire-front distance highest for Custom U-Net and PINN, while Swin-Unet prioritizes terrain and fuel and ResNet-50 relies chiefly on the fire mask.For Custom U-Net and PINN, current-front distance importance is 0.326 and 0.324; Swin-Unet's highest channel, cont10m, is 0.403, and ResNet-50's fire-mask importance is 0.289.Source: §7.2 · p. 9, Figure 5 · p. 9
- Fuel-load perturbations produce the strongest response for Custom U-Net and PINN, Swin-Unet responds broadly across static channels, and ResNet-50 is comparatively insensitive.Wind rotation affects every model, with response increasing monotonically through 180°; Swin-Unet is most sensitive and ResNet-50 least sensitive.Source: §7.2 · p. 10, Figure 6 · p. 10
7.3 How much comes from the fire front?
- Dropping either distance channel costs almost nothing, but removing the fire mask as well halves accuracy; terrain, fuel, and wind alone still reach MCC 0.42.The fire mask and distance layers partly encode the same fire-state information, while occlusion importance shifts to the mask when a distance channel is removed.Source: §7.3 · p. 11, Table 3 · p. 11
8 Comparison with a Second Region
Rectoret-trained models retain a similar cross-region performance pattern on Pedriza, but with a systematic zero-shot gap. The comparison also changes spatial resolution, so it cannot isolate a regional transfer effect.
8 Comparison with a Second Region
- The comparison confounds region and scale: Pedriza is mapped at 10 m versus Rectoret’s 2 m, and the data cannot separate their effects.The regions’ covariate distributions have mean histogram overlap 0.63, ranging from 0.47 to 0.77.Source: §8 · pp. 11–12
- Across fire stages, Pedriza curves follow Rectoret’s shape with a roughly constant offset in both metrics rather than collapsing.The figure compares MAE and boundary Dice loss; Swin-Unet is absent from the region logs.Source: §8 · p. 11, Figure 7 · p. 12
- After 50 fine-tuning epochs, Custom and PINN both reach 0.481 Dice loss, while ResNet-50 peaks at epoch 2 before degrading to 0.519.ResNet-50 transfers best zero-shot despite being the weakest model on Rectoret.Source: §8 · p. 12
- Zero-shot Dice loss rises from 0.469–0.487 on Rectoret to 0.552–0.637 on Pedriza.Table 4 compares performance averaged over fire stages 30–120 min.Source: §8 · pp. 11–12, Table 4 · p. 12
- Other transfer studies link performance to ecological similarity [16] and report better transfer with window-local than pairwise self-attention.The Rectoret–Pedriza similarity has not yet been quantified.Source: §8 · p. 12
9 Discussion
Surface fuel load accounts for most explainable pixel-level variation in burn probability, while wind has a real but secondary effect in the rotation experiment. The discussion also identifies independent fuel data and broader, observationally validated evaluation as important next steps.
9 Discussion
- Surface fuel load alone accounts for most explainable pixel-level variance in burn probability; continuity descriptors add little because they largely repeat one another.Its apparent dominance may partly reflect that the fuel raster and simulator share the BEHAVE-Anderson classification; an independently measured fuel product is a next test.Source: §9 · p. 13
- Wind reversal shifts the forecast by about a fifth compared with removing distance from the fire front.Wind’s earlier low channel-specific importance was partly an encoding artifact: it is constant across the area, weakening methods that seek spatial variability.Source: §9 · p. 13
- The training data are entirely synthetic, cover one fuel type, and lack validation against observed fires.The discussion proposes probabilistic surrogates and a multi-scene train/test protocol as future directions; the cited protocol trains on 11 wildfire scenes and tests on 4 held-out scenes [9].Source: §9 · p. 13
10 Conclusions and Future Work
The paper compares four deep-learning architectures and examines their interpretability and cross-region performance. Surface fuel load is the strongest burn-probability signal, and the models show a consistent Pedriza shift that does not establish transferability.
10 Conclusions and Future Work
- Surface fuel load is the strongest burn-probability driver, while convolutional models emphasize distance from the fire front and Swin-Unet gives more weight to terrain and fuel.The Swin-Unet interpretation is consistent with a finding from an unrelated dataset.Source: §10 · p. 13
- The Pedriza comparison shows a consistent shift rather than a decrease in accuracy, but does not prove model transferability.The current comparison cannot separate region from spatial-scale change.Source: §10 · p. 13
- The study compares a custom U-Net, transfer-learned ResNet-50, physics-informed model, and Swin-Unet on high-resolution wildfire simulations.Exploratory analysis and interpretability accompany the model comparison.Source: §10 · p. 13
- The authors identify two next steps: test fuel load using an independently measured product and analyze another 2 m region to separate area from scale effects.Source: §10 · p. 13