Source-linked AI summary
Random Forest-Informed Cellular Automaton for Large-Scale Wildfire Spread Modelling
Siyu Chen, Esha Saha, Hao Wang
TL;DR
Large-scale wildfire modelling needs to represent both environmental fire occurrence and local propagation under heterogeneous, nonstationary data. The paper combines a Random Forest probability layer with cellular-automaton spread, finding stronger large-scale recovery than CA-only baselines in the tested 2023 simulation, while identifying scale- and season-specific limitations.
Problem
Wildfire models must bridge cell-level fire-occurrence prediction and spatially continuous spread modelling amid heterogeneous, incomplete, and nonstationary data.
Method
A three-stage framework uses Random Forest daily occurrence probabilities, optional quantile spread-rate scenarios, and a probabilistically coupled CA for local spread.
Results
The RF-informed CA recovered the large-scale burned footprint better than CA-only baselines in the 2023 simulation, with α = 50 most stable among tested settings.
Takeaways & Limitations
Combining RF-derived probabilities with local CA spread can improve large-scale wildfire simulations under the tested conditions.
Takeaways & Limitations
Full RF-informed CA performance has only been evaluated for 2023, and same-day meteorological inputs make simulations retrospective rather than operational forecasts.
Abstract
from arXiv · showhide
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.
1 Introduction
Wildfire modelling must connect fire occurrence prediction with spatial spread simulation, while remaining stable under noisy data and changing conditions. Existing cellular automata offer explicit spatial structure but can depend on empirically defined parameters that reduce stability across years.
- 2023’s record burned area and severe smoke impacts highlighted the need for reliable tools anticipating ignition and large-scale spread.
- Wildfire research includes fire danger assessment, occurrence prediction, spread simulation, and fire effects analysis.
- This study focuses on fire occurrence prediction and fire spread simulation because they are often studied separately but closely connected in practice.Occurrence prediction supports fire danger assessment and early resource planning, whereas spread simulation evaluates development across space and time.
- Cellular automata provide clear structure and explicit spatial form, but empirically defined spread parameters can reduce stability with noisy data or across years.
2 Related Work
Wildfire models face heterogeneous data, incomplete observations, and changing spatio-temporal relationships. Prior work motivates combining machine-learning occurrence probabilities with cellular automata that explicitly represent local spread, while retaining interpretability and adjustable components.
- Canadian wildfire modelling is challenged by multisource heterogeneity, incomplete observations, and strong spatio-temporal non-stationarity.
- Statistical ignition models are interpretable but limited in capturing complex nonlinear predictor interactions, motivating increased use of machine learning.
- Fire-occurrence models estimate cell-level ignition likelihood but do not directly simulate spatial continuity or neighbour-driven growth.Spatially explicit spread therefore requires coupling with cellular automata or process-based simulators.
- Traditional cellular automata are interpretable, computationally efficient, and spatially explicit, but can generalize poorly across years and struggle with large or fast-moving fires.Their parameters may fail under nonstationary weather-fuel interactions, restricting continental-scale applicability.
- The proposed RF-informed CA separates daily RF probability fields from local CA propagation and combines them probabilistically.Stage 2 additionally supplies optional quantile-based spread-rate scenarios for sensitivity analysis.
3 Methodology
The study develops a three-stage pipeline for estimating daily fire occurrence and simulating wildfire spread across Canada. It combines RF-derived probability fields, optional quantile spread-rate scenarios, and neighbourhood-driven propagation in an RF-informed CA.
- Three-stage framework: The framework uses a Random Forest for daily pixel-level fire-occurrence probabilities, quantile gradient boosting for conditional spread-rate quantiles, and an RF-informed CA for spatial simulation.Stages 1 and 2 are trained on 2021 data and applied to later years without retraining; Stage 3 is mainly evaluated on 2023.
- Stage 1: Fire occurrence: Stage 1 predicts whether each pixel is recorded as burning on a given day using 16 static landscape and same-day weather or FWI predictors.The response represents daily fire occurrence rather than first ignition, and variables containing realized fire activity or growth are excluded.
- Stage 1: Fire occurrence: Stage 1 RF probabilities are aggregated to a 5 km grid as the probability that at least one constituent pixel is burning.The aggregation assumes pixel-level independence, although spatial correlation can make coarse-grid probabilities large.
- Stage 2: Spread-rate scenarios: Stage 2 predicts the conditional distribution of daily fire spread distance in metres per day using quantile gradient boosting rather than a single spread estimate.The models estimate q50, q75, q90, and q95; q50 and q75 are used as optional spread-strength scenarios in Stage 3.
- Stage 2: Spread-rate scenarios: Stage 2 predictions use information available by the beginning of each day, excluding future fire growth, and are aggregated as optional dynamic spread-rate priors for the CA.Separate models are fitted for each quantile, so quantile crossing can occur and monotonicity violations are reported.
- Stage 3: RF-informed CA: Stage 3 combines the RF daily probability layer with local CA neighbour-driven spread on a 5 km grid using daily time steps.The RF layer is probabilistic rather than a hard constraint, while α controls neighbour-spread strength; varying α is a sensitivity analysis rather than selection of one universal value.
4 Results
The RF occurrence model showed moderate, relatively stable cross-year discrimination, while the RF-informed CA substantially outperformed CA-only baselines in 2023 spatial overlap. Sensitivity results indicate that weak neighbour spread best preserved the RF-derived pattern, whereas stronger settings increasingly overexpanded simulations.
- 0.725–0.795 AUC was achieved across the 2022–2024 test years, indicating moderate and relatively stable discrimination.
- Stage 2 diagnostics: The q50 and q75 outputs from Global-PrevGrow were retained as optional spread scenarios for sensitivity analysis rather than optimal priors.Global showed overcoverage for [q75, q95], while ECO had low [q90, q95] coverage and many ordering violations.
- The RF-informed CA produced much higher spatial overlap than deterministic and probabilistic CA-only baselines, whose final IoU values were below 0.05.All models used the same 5 km grid and initial fireline in the 2023 comparison.
- Small regime: α = 50, 100: α = 50 and α = 100 added limited neighbour spread, with α = 50 remaining closer to the observed daily IoU pattern.Weak neighbour spread was more stable at the 5 km scale, rather than indicating a universal spread parameter.
- Small regime: α = 50, 100: α = 50 produced about 1,500 km2 of final area error, compared with about 4,025 km2 for α = 100.The smaller bias for α = 50 matched its lower spread probability.
- Small regime: α = 50, 100: α = 50 generally yielded lower Hausdorff distances than α = 100, whose more even outward spread could smooth fire edges away from observations.Hausdorff distance measures disagreement between simulated and observed fire edges.
- Medium regime: Medium settings diverged earlier and overestimated burned area, with final area errors of 15,225 km2 for α = 300, 7,200 km2 for q50, and 13,750 km2 for q75.The q50 and q75 quantile priors behaved similarly to medium fixed-spread settings at the 5 km grid scale.
- Stress tests: α = 1000 and α = 2000 caused very large area errors of about 104,000 km2 and 210,000 km2, respectively, as neighbour spread dominated the update.These stress-test settings were used only as upper-bound sensitivity tests.
5 Discussion
The discussion examines how the RF probability layer interacts with local CA propagation, including baseline comparisons, spread-strength sensitivity, and spatial-evaluation trade-offs.
- Baseline comparison: Final IoU values below 0.05 and burned-area underestimation of nearly 6000 km2 characterize both CA-only baselines, whereas RF-informed CA reaches a final IoU of 0.9771 and Dice coefficient of 0.9884.
- Sensitivity to spread strength: α = 50 and α = 100 keep CA spread closer to the RF-derived pattern, while α = 300, q50, and q75 increase area overestimation and α = 1000 and α = 2000 produce unrealistically fast spread.
- Sensitivity to spread strength: Stage 2 quantile scenarios function as alternative spread-strength inputs for sensitivity analysis rather than standalone simulations or consistent improvements over fixed settings.
- Component interaction: RF-informed CA combines a daily RF probability field with neighbourhood-driven CA propagation, allowing either component to activate an unburned cell.
- Spatial evaluation: IoU and Hausdorff distance capture different spatial properties: RF-informed CA has higher overlap but a larger maximum boundary error than CA baselines.
- Baseline comparison: RF-informed CA achieves substantially higher spatial overlap than the evaluated CA-only baselines under matched 2023 conditions.All models use the observed first-day fireline and the same 5 km grid.
6 Conclusion
The study concludes that the RF-informed CA separates fire-occurrence likelihood from local spread and performs best at a tested small nonzero α on the 5 km grid. Its broader validity remains constrained by limited Stage 3 evaluation, unresolved component attribution, boundary errors, retrospective inputs, and resolution dependence.
- Framework: The RF-informed CA combines a daily fire-occurrence probability layer with local neighbour-driven spread through probabilistic updates.
- Framework: At the 5 km grid, α = 50 performed best among the tested settings, while larger values produced stronger area overestimation.
- Framework: Stage 2 scenarios did not consistently outperform fixed settings and are therefore treated as sensitivity cases rather than evidence of improvement.
- Limitations: Stage 3 evaluation is limited mainly to the 2023 fire season, and the comparison does not fully separate CA propagation from the RF layer.
- Limitations: Same-day meteorological inputs make the simulations retrospective rather than operational forecasts.
- Limitations: The 500 m diagnostic reveals more detailed local boundary structure while making fire-front disagreement more visible.
- Limitations: Future work should evaluate an RF-only baseline, examine coarse-grid probability aggregation, and validate Stage 3 across additional years and fire events.
A.1 Predictors used in the models
Table 6 lists the predictors used in the Stage 1 and Stage 2 models.
- Stage 1: Table 6 identifies the predictors used in Stage 1 models.
- Stage 2: Table 6 identifies the predictors used in Stage 2 models.
- Overview: The table presents predictor usage across both modelling stages.
A.2 Variables excluded to avoid information leakage
Table 7 lists key fire-growth fields excluded from the predictor set to avoid information leakage.
- Excluded variables: Table 7 identifies key fire-growth fields excluded from the models.
- Excluded variables: The listed fields were not used as predictors.
- Purpose: The table documents variables excluded to avoid information leakage.
A.3 Construction of the CFSDS burn progression data
The CFSDS pixel-level daily fire progression data were constructed from MODIS and VIIRS active-fire detections accessed through NASA FIRMS.
- CFSDS burn progression data use active-fire detections from MODIS and VIIRS obtained through NASA FIRMS.FIRMS provides near-real-time active-fire data, while the MODIS and VIIRS products differ in spatial resolution.
A.4 Definitions of Evaluation Metrics
This section defines the evaluation metrics used to assess spatial overlap, perimeter discrepancies, extreme deviations, and quantile predictions.
- Intersection over Union (IoU): IoU measures spatial overlap between simulated and observed burned areas.It is defined for predicted burned area A and observed burned area B.
- Dice score: Dice score is an overlap metric that places more weight on shared burned area.
- Hausdorff distance: Hausdorff distance measures the largest geometric discrepancy between simulated and observed fire perimeters.It is defined using the boundaries of the predicted and observed burned areas.
- Maximum distance error: Maximum distance error measures the worst-case pointwise spatial deviation between predicted and observed fire perimeters.
- Pinball loss: Pinball loss evaluates quantile predictions by penalizing under- and over-estimation asymmetrically.The metric is defined for quantile level τ ∈(0, 1).
A.5 Additional Metrics and Supporting Results
Additional metrics and figures provide supporting analyses beyond the main text, including cumulative area error and distance-based measures of extreme deviations.
- The section reports supporting metrics and figures that are omitted from the main text for clarity.
- Cumulative area error measures the difference between simulated and observed burned area over time.It is informative but not used as a primary metric, because the main text focuses on IoU and Hausdorff distance.
- Additional distance-based metrics assess extreme spatial deviations, with results consistent with the Hausdorff distance trends.Maximum distance error is given as an example of these additional metrics.