BCWildfire: A Long-term Multi-factor Dataset and Deep Learning Benchmark for Boreal Wildfire Risk Prediction

AI summary

Overview

Research area: Geospatial deep learning / time-series forecasting applied to wildfire risk prediction, with remote-sensing data curation and benchmarking.

Technical level: Intermediate. Readers need basic familiarity with time-series forecasting architectures (CNN, Transformer, Mamba), remote-sensing covariates, and classification metrics, but no specialist wildfire-science background.

Scope: The paper introduces BCWildfire, a 25-year, daily-resolution, 38-variable wildfire dataset covering 240 million hectares of boreal forest in British Columbia and adjacent regions (2000–2024), and benchmarks six time-series forecasting models on it for next-day wildfire risk prediction.

What This Paper Is About

Wildfire risk depends on long chains of interacting factors — fuel, weather, terrain, and human activity — that accumulate over seasons and years, but the public datasets available for training models are mostly small, short-window, and built around tracking a fire's spread after it has already started. This paper builds a long-term, large-scale, multimodal dataset for the boreal forest of British Columbia and uses it to test whether modern time-series deep learning architectures can predict where fire will occur the next day. The goal is to give machine-learning researchers a realistic, standardized disaster-risk benchmark and to expose where current models still fail.

Key Contributions

1. A large-scale long-term dataset (BCWildfire): 2.4 million samples across 240 million hectares of boreal landscape over 25 years, integrating 38 multimodal wildfire-driving factors — addressing gaps in geographic coverage, feature completeness, and long-term temporal representation in existing resources.
2. A dataset designed for time-series forecasting rather than spread modeling: Unlike existing resources focused on spatial fire-spread behavior, BCWildfire is structured to support temporal reasoning over cumulative driver interactions, at a uniform 1 km spatial and daily temporal resolution.
3. A systematic benchmark across four architecture paradigms: CNN-based (SCINet), linear-based (TSMixer, CrossLinear), Transformer-based (Crossformer, FEDformer), and Mamba-based (S_Mamba) models are compared for next-day wildfire risk prediction.
4. Analysis of drivers and design choices: SHAP attribution identifies the most influential fire-driving factors, and the effect of spatial positional embedding is measured across all architectures.

Main Findings

• Dataset scale and composition: The final training/validation/test pool contains 1,015,275 samples — 338,425 positive (fire) and 676,850 negative (non-fire) samples. The full dataset covers 38 daily variables over a spatial domain of 2,782 × 1,302 km² at 1 km resolution, spanning 2000–2024.
• All models plateau below 92% recall: Under the 10-day input configuration without positional embedding, the best-performing models — CrossLinear (recall 87.59%), Crossformer (87.49%), and FEDformer (91.18%) — plateau below 92% recall, while precision stays in a narrow 83–88% range. The paper attributes this ceiling to extreme class imbalance, strong spatiotemporal heterogeneity, and complex nonlinear driver interactions.
• Architecture differences are modest and pattern-dependent: Transformer-based models (Crossformer, FEDformer) achieve higher recall and stability, reflecting an advantage in capturing long-range temporal dependencies; linear models TSMixer and CrossLinear remain competitive and efficient when temporal correlations are strong; S_Mamba achieves balanced results (precision 84.21%, recall 86.44%, F1 85.31%, PR-AUC 94.83%).
• Positional embedding usually helps, with exceptions: Adding spatial positional embedding improves most models, with the largest recall and F1 gains for Transformer-based models. S_Mamba shows a clear recall improvement of 3.659% and an F1 improvement of 2.15%. Table 2 also shows SCINet's F1 declining slightly (86.38 to 85.90) with positional embedding, so the improvement is not universal across every metric.
• SHAP: recent fire activity dominates: The most influential feature is fire detection, whose strong negative SHAP contribution indicates that the absence of recent fire activity is a key signal for identifying potential new ignitions, while persistent fire signals correspond to already-burned areas.
• Soil moisture and surface energy flux promote ignition: Soil moisture L3 (28–100 cm depth) and surface latent heat flux show strong positive effects, indicating that intermediate soil moisture and high surface energy exchange promote rapid drying of near-surface fuels. Snow cover exerts a pronounced negative influence, suppressing ignition risk.
• Spectral and topographic variables contribute secondarily: Band 3 (459–479 nm reflectance), Band 32 (11.77–12.27 µm emissivity), EVI, and FPAR contribute positively; hillshade, slope, and aspect rank within the top 15 predictors, but their effects remain secondary to biophysical and thermal factors.
• Drivers of ignition in the distributional analysis: Temperature increases, fuel accumulation, precipitation decreases, and surface soil moisture reductions are strongly associated with wildfire ignition. However, substantial distributional overlap remains between burned and unburned areas, underscoring intrinsic stochasticity — fires may ignite or remain dormant under similar conditions. Regions that did burn tend to exhibit slightly higher deep-layer (100–289 cm) soil water content.
• Quantitative metrics hide spatial differences: Figure 4 shows models producing distinct spatial patterns of false positives and high-risk regions even when their metrics in Table 2 differ only marginally, suggesting that ensembling could help and that metrics on sampled test data may be biased.
• Qualitative limitations remain: Most wildfire occurrences fall within areas classified as high risk, but model limitations remain evident in predicting new ignitions and small-scale fires.

Methodology in Plain English

The authors assembled a daily data cube for British Columbia and surrounding regions from 2000 to 2024. Inputs come from four source families: MODIS satellite products (fuel/vegetation, land surface temperature, active fire detections), ERA5-Land reanalysis (weather and soil moisture), OpenStreetMap (water bodies, settlements, power grid), and the ASTER digital elevation model (terrain). Fuel conditions use the 500 m, 4-day MCD15A3H LAI/FPAR product plus daily 500 m MOD/MYD09GA reflectance in Bands 1, 2, 3, and 7 and derived NDVI and EVI. Meteorological variables from ERA5-Land (approximately 11 km) include 2 m air temperature, 10 m eastward and northward wind components, atmospheric pressure, total precipitation, surface latent heat flux, 2 m dewpoint temperature, snow cover, and soil moisture across the 0–289 cm depth range. Because ERA5-Land is spatially smoothed, MODIS thermal products MOD/MYD11A1 (approximately 1 km) and MOD/MYD09CMG (approximately 5.6 km) were added to capture fine-scale thermal anomalies. Terrain variables (slope, aspect, hillshade) derive from 30 m ASTER DEM, plus distance to the nearest water body. Human-activity variables use the yearly 500 m MCD12Q1 land-use product and distances to the nearest settlement or power-grid infrastructure.

Wildfire activity labels come from merging daily MOD/MYD14A1 active fire detections (1 km, mid-infrared Bands 20 and 21, up to four observations per day), retaining only high-confidence pixels, with pixel values reflecting the most recent observation time that day.

Preprocessing addressed three problems: cloud contamination in MODIS optical imagery (mitigated with MODIS quality-control bands and temporal filling from historical records, deliberately without introducing future data leakage), inconsistent coordinate systems and resolutions (all data reprojected to WGS84, resampled to a uniform 1 km grid using nearest-neighbor interpolation for categorical data and bilinear for continuous variables), and temporal mismatch (non-daily sources forward-filled into a unified daily sequence).

Wildfire risk prediction is framed as binary time-series classification at the pixel level: predict the probability of fire at time t from driver and ignition conditions at t−1 through t−n. To handle class imbalance, negatives within a 60 km and 3-day spatiotemporal buffer around positives were excluded so the model does not sample from high-risk regions. For each land-cover type and year, negatives were held at a fixed multiple of positives — two times in training and validation, and equal in the test set. The previous 10 days of drivers were used as input to predict the following day.

Six models were trained with binary cross-entropy loss and the Adam optimizer on two NVIDIA A6000 GPUs, batch size 128, 50 epochs, and a learning rate scheduled to 1 × 10⁻⁵. Years 2000–2020 form the training set, 2021–2022 the validation set, and 2023–2024 the test set. Performance is reported with precision, recall, F1-score, and precision-recall area under curve (PR-AUC). SHAP analysis on Crossformer was used to attribute importance to individual drivers.

Why This Matters

Impact on research: The paper argues that the scarcity of publicly available, standardized disaster-prediction datasets limits how far time-series forecasting research can be tested in real-world risk settings, while wildfire experts continue to rely on traditional models such as Random Forest and XGBoost. BCWildfire provides a unified benchmark that spans 25 years and 38 multimodal drivers, letting machine-learning researchers evaluate CNN, linear, Transformer, and Mamba architectures on a stochastic, imbalanced natural-hazard task rather than on conventional traffic, energy, or weather forecasting datasets. It also documents concrete failure modes — sub-92% recall ceilings, distributional overlap between burned and unburned areas, and metric-insensitive spatial differences between models — that define open problems.

Real-world applications:

• Proactive wildfire management and suppression planning, where ignition risk must be anticipated rather than reacted to after detection.
• Resource allocation and evacuation readiness, using next-day risk maps to position crews and warn communities in boreal regions.
• Utility and infrastructure protection, since the dataset includes distance to power-grid infrastructure and land-cover types that shape ignition likelihood.
• Forest and carbon management in boreal landscapes, where the paper notes substantial carbon storage and high wildfire susceptibility.

Industry relevance: The dataset and benchmark are directly relevant to insurers and catastrophe modelers, electric utilities operating in fire-prone boreal terrain, forestry and carbon-accounting firms, and government fire-management agencies. Because spatial positional embedding consistently improved most models and different architectures produce different spatial error patterns, the paper suggests that ensemble approaches — rather than a single best model — are a practical near-term direction for operational deployment.

Future Directions

• Extending the temporal horizon: The authors state that future work will extend modeling beyond the 10-day input window to better exploit the dataset's 25-year long-term information and to examine how extended temporal drivers influence wildfire occurrence and evolution.
• Breaking the performance ceiling: The consistent plateau below 92% recall across architectures raises the question of what model design or training objective could handle the extreme class imbalance and spatiotemporal heterogeneity better than current approaches.
• Ensembling across paradigms: Since models with nearly identical metrics produce visibly different spatial error patterns, combining outputs from Transformer, linear, CNN, and Mamba models is an open avenue the paper explicitly raises.
• Better quantitative evaluation: The paper warns that comparisons based solely on sampled test data may be biased and fail to capture true predictive capability, implying a need for evaluation protocols that account for the spatial structure of predictions.

Target Audience

This paper primarily serves machine-learning and time-series forecasting researchers looking for a challenging, real-world natural-hazard benchmark, and remote-sensing or geospatial deep-learning practitioners interested in multimodal data fusion. Wildfire scientists and fire-management analysts will find the SHAP driver analysis and the dataset itself useful, though the modeling sections assume some deep-learning familiarity. Practitioners in insurance, utilities, and forestry who need data-driven risk assessment may benefit from the results, but will need to work through the technical benchmarking sections or rely on the accompanying code and data release (https://github.com/SynUW/BCWildfire).

Authors’ abstract

Wildfire risk prediction remains a critical yet challenging task due to the complex interactions among fuel conditions, meteorology, topography, and human activity. Despite growing interest in data-driven approaches, publicly available benchmark datasets that support long-term temporal modeling, large-scale spatial coverage, and multimodal drivers remain scarce. To address this gap, we present a 25-year, daily-resolution wildfire dataset covering 240 million hectares across British Columbia and surrounding regions. The dataset includes 38 covariates, encompassing active fire detections, weather variables, fuel conditions, terrain features, and anthropogenic factors. Using this benchmark, we evaluate a diverse set of time-series forecasting models, including CNN-based, linear-based, Transformer-based, and Mamba-based architectures. We also investigate effectiveness of position embedding and the relative importance of different fire-driving factors. The dataset and the corresponding code can be found at https://github.com/SynUW/mmFire