Spatial Uncertainty Quantification in Wildfire Forecasting for Climate-Resilient Emergency Planning

AI summary

Overview

• Research area: Machine learning for Earth observation, specifically spatial uncertainty quantification in wildfire spread forecasting.
• Technical level: Intermediate. The paper explains its uncertainty quantification methods clearly, but assumes some familiarity with deep learning, probabilistic prediction, and remote sensing concepts.
• Scope: A 12-fold evaluation of three uncertainty quantification techniques applied to a lightweight spatiotemporal transformer for next-day wildfire spread prediction on the WildfireSpreadTS dataset, plus a new distance metric for characterizing where prediction errors occur spatially.

What This Paper Is About

Machine learning models can forecast how wildfires will spread using satellite and weather data, but they typically output a single prediction with no indication of where the model is confident or unreliable. The author argues this is a critical gap for emergency responders, who need to know which parts of a forecast to trust. The paper systematically measures, calibrates, and spatially maps predictive uncertainty in wildfire spread forecasting, and shows that uncertainty is not random noise but clusters in consistent zones around predicted fire perimeters.

Key Contributions

1. First systematic analysis of spatial uncertainty in high-resolution, Earth-observation-based wildfire forecasting, a setting the author states had no prior uncertainty quantification work.
2. A novel centroid-oriented boundary distance metric that measures spatial offset between predicted and ground-truth firelines directionally, and is compared against Average Surface Distance (ASD) and Hausdorff Distance (HD).
3. Empirical buffer zone characterization: the finding that high-uncertainty regions form consistent bands around predicted firelines (reported as 20–60 m in the abstract and discussion, with metric-specific peaks of 28–35 m centroid, 47–64 m ASD, and 148–166 m HD).
4. A benchmark of three UQ methods (MC Dropout, Deep Ensembles, Bayesian Neural Networks) on calibration metrics (ECE, Brier Score, NLL) and a feature-ablation study showing vegetation and active-fire inputs outperform weather, topography, and land cover groups.

Main Findings

• Vegetation features dominate predictive skill: Vegetation plus active fire indicators achieved the highest mean Average Precision (AP) of 0.378 ± 0.083 across 12 folds, versus a persistence baseline of 0.191 ± 0.063. Weather + active fire reached 0.323 ± 0.078, land cover + active fire 0.319 ± 0.092, topography + active fire 0.317 ± 0.082, and all features + active fire 0.319 ± 0.077.
• Deep Ensembles calibrate best: Deep Ensembles produced the lowest error on all three calibration metrics — ECE 0.512 ± 0.018, Brier Score 0.265 ± 0.009, and NLL 0.731 ± 0.023. MC Dropout scored ECE 0.536 ± 0.015, Brier 0.294 ± 0.012, NLL 0.805 ± 0.020, and BNN scored ECE 0.525 ± 0.014, Brier 0.283 ± 0.019, NLL 0.794 ± 0.054.
• Uncertainty has coherent spatial structure: Rather than appearing as scattered noise, uncertainty concentrates near predicted fire perimeters. It is sharply localized in larger fires and more diffuse in smaller or fragmented fires, illustrated across three events of large, medium, and small size (approximately 125.6 acres and 5.2 acres at the extremes).
• Buffer zone sizes depend heavily on the metric: Peak distances from KDE histograms were 32.19 m (vegetation) and 33.48 m (all features) for the centroid boundary distance; 64.15 m and 55.86 m for Average Surface Distance; and 165.78 m and 155.67 m for Hausdorff Distance.
• Feature attribution confirms drivers: Integrated Gradients on a CNN surrogate model (R² = 0.81 fidelity) identified active fire presence as the dominant influence, followed by vegetation indices (NDVI, EVI2), with thermal reflectance bands contributing less.
• UTAE outperforms heavier alternatives: The author reports UTAE beats ConvLSTM and standard U-Net by up to 3.9 AP points on this dataset, and notes Swin Transformer approaches carry roughly 27M parameters versus UTAE's roughly 1M (described as approximately 1M in the methods section and approximately 1.1M in the appendix), making them more expensive and prone to overfitting on current wildfire dataset sizes.

Methodology in Plain English

The study uses the publicly available WildfireSpreadTS dataset, which contains 64×64 pixel patches centered on active fires, with 5 days of multimodal inputs per sample and a binary burn mask for a future day as the target. The dataset spans 607 wildfire events across the western United States from January 2018 to October 2021, totaling 13,607 daily images.

The forecasting model is UTAE, a U-Net variant with a temporal attention encoder that applies simplified multi-head self-attention across the time dimension and uses those attention weights on skip connections. Inputs include VIIRS reflectance bands, NDVI, EVI2, meteorological variables, topography, land cover classes, fire masks, and day-of-year. Data is resampled to 375 m spatial and 24-hour temporal resolution, numerical features are standardized, and training uses random crops to 128×128 pixels, flips, 90-degree rotations, and angle-aware adjustments to wind direction and aspect. The model is trained with AdamW (β1 = 0.9, β2 = 0.999, learning rate 0.01, weight decay 0.01) using weighted binary cross-entropy for 10,000 steps with batch size 32.

For uncertainty, the author tests three approaches: MC Dropout with 20 stochastic forward passes at test time; Deep Ensembles of independently seeded UTAE models, each also using 20 MC Dropout passes; and a Bayesian Neural Network using Bayes-by-Backprop variational inference with multiple sampled weight distributions. Calibration is measured with Expected Calibration Error, Brier Score, and Negative Log-Likelihood.

To locate errors in space, the author introduces a centroid-aligned boundary distance metric: it finds the centroid of the true burn mask and the centroid of the false-positive region (predicted pixels not in the ground truth), draws a line between them, and measures the distance between where the true and predicted firelines cross that line. This is compared against Average Surface Distance and Hausdorff Distance, each capturing a different aspect of boundary disagreement.

Why This Matters

Impact on research: The paper reframes predictive uncertainty as spatially informative rather than as an obstacle or noise term. It provides a reproducible evaluation protocol (three UQ methods, three calibration metrics, three distance metrics) for a task where deterministic prediction has been the norm, and it introduces a distance metric designed for directionally meaningful fireline offsets.

Real-world applications:

• Emergency managers can use spatial uncertainty maps to prioritize which parts of a forecast perimeter to treat as reliable and where to hedge resource allocation.
• Fire suppression teams can interpret buffer zone estimates as a practical cue for where predictions are least dependable, informing containment line placement.
• Climate adaptation planners can fold uncertainty estimates into long-term risk assessments and land-use decisions as fire risk rises.
• Decision support system builders can integrate the 20–60 m buffer zones, which the author notes correspond to tactically relevant scales for incident management teams.

Industry relevance: Insurance, utilities, and disaster-response organizations increasingly rely on geospatial risk models. A model that reports calibrated confidence and a spatial footprint of likely error is more usable in operational workflows than one that outputs only a probability map.

Future Directions

• Extending to aleatoric uncertainty: The paper explicitly limits itself to epistemic uncertainty; separating irreducible data noise from model uncertainty remains open.
• Broadening geographic coverage: All experiments are U.S.-only, leaving transferability to other fire regimes (Mediterranean, Amazon, Australia) untested.
• Reconciling weather and topography inputs: Adding these feature groups reduced performance, which the author attributes to temporal and spatial resolution mismatches — better alignment or higher-resolution inputs could change this.
• Stress-testing the centroid metric: The author notes it may be unreliable with disconnected fire fronts or misaligned centroid axes, suggesting a need for metrics robust to fragmented fire geometry.

Target Audience

Wildfire and emergency management practitioners who need to interpret forecast reliability, remote sensing and geospatial machine learning researchers working on uncertainty quantification, climate adaptation planners incorporating model confidence into risk assessments, and machine learning engineers building decision support tools for hazard forecasting. Readers with a basic grasp of probabilistic prediction and satellite imagery will get the most from it.

Authors’ abstract

Climate change is intensifying wildfire risks globally, making reliable forecasting critical for adaptation strategies. While machine learning shows promise for wildfire prediction from Earth observation data, current approaches lack uncertainty quantification essential for risk-aware decision making. We present the first systematic analysis of spatial uncertainty in wildfire spread forecasting using multimodal Earth observation inputs. We demonstrate that predictive uncertainty exhibits coherent spatial structure concentrated near fire perimeters. Our novel distance metric reveals high-uncertainty regions form consistent 20-60 meter buffer zones around predicted firelines - directly applicable for emergency planning. Feature attribution identifies vegetation health and fire activity as primary uncertainty drivers. This work enables more robust wildfire management systems supporting communities adapting to increasing fire risk under climate change.