Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Aditya Chakravarty

TL;DR

This work addresses the need for uncertainty-aware, high-resolution wildfire forecasting by leveraging multimodal Earth observation inputs and three epistemic-UQ approaches (MC Dropout, Deep Ensembles, Bayesian NNs) within a lightweight UTAE transformer. It demonstrates that uncertainty concentrates along fire perimeters, yielding interpretable 20–60 m buffer zones that can inform emergency planning and resource allocation. Vegetation-related features and recent fire activity drive predictive performance and uncertainty patterns, while a centroid-aligned boundary distance metric provides a practical tool for quantifying spatial offsets between predictions and reality. The findings enable more risk-aware wildfire management under climate change by integrating calibrated uncertainty maps into operational workflows.

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.

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

TL;DR

This work addresses the need for uncertainty-aware, high-resolution wildfire forecasting by leveraging multimodal Earth observation inputs and three epistemic-UQ approaches (MC Dropout, Deep Ensembles, Bayesian NNs) within a lightweight UTAE transformer. It demonstrates that uncertainty concentrates along fire perimeters, yielding interpretable 20–60 m buffer zones that can inform emergency planning and resource allocation. Vegetation-related features and recent fire activity drive predictive performance and uncertainty patterns, while a centroid-aligned boundary distance metric provides a practical tool for quantifying spatial offsets between predictions and reality. The findings enable more risk-aware wildfire management under climate change by integrating calibrated uncertainty maps into operational workflows.

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.

Paper Structure

This paper contains 22 sections, 7 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Climate Impact and Pathway
  3. Methods
  4. Dataset
  5. Model Architecture
  6. Uncertainty Quantification
  7. Results
  8. Feature Importance and Model Performance
  9. Uncertainty Calibration
  10. Spatial Uncertainty Structure
  11. Quantitative Buffer Zone Analysis
  12. Discussion and Conclusion
  13. Appendix
  14. Feature Attribution
  15. UTAE Model Selection Rationale and Procedural Details
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Qualitative comparison of model predictions for three fire events of varying size: large (top), medium (middle), and small (bottom). Each row shows the NDVI input, mean prediction from a Deep Ensemble, and the ground-truth burn mask. The influence of vegetation features on the model's mean predictions is clearly evident. The events span approximately 125.6 acres (large), and 5.2 acres (small).
  • Figure 2: Example input channels from a single sample at prediction time, including Sentinel-2 bands, NDVI, EVI2, and active fire features. These inputs are provided as a 5-day sequence to the model.
  • Figure 3: Feature importance scores computed using Integrated Gradients. Active fire presence dominates attribution, followed by vegetation indices (NDVI, EVI2). Thermal bands are less influential.
  • Figure 4: Schematic of boundary distance computation between predicted and ground truth fire masks.