Spatial Uncertainty in Wildfire Forecasting Using Multi-Modal Earth Observation

TMLR Paper4994 Authors

Abstract: Accurate wildfire forecasting from remote sensing data is essential for climate resilience and emergency planning. Beyond predictive performance, understanding where and why uncertainty arises is critical for operational trust. We analyze the spatial structure of predictive uncertainty in wildfire spread forecasts using multimodal Earth observation (EO) inputs, including Sentinel-2 vegetation indices and VIIRS thermal reflectance. Using Monte Carlo dropout and deep ensembles, we show that predictive entropy maps exhibit coherent spatial patterns aligned with fire boundaries, unlike randomized baselines. We introduce a novel and interpretable centroid-oriented distance metric that reveals high-uncertainty regions consistently form 20–60 meter buffer zones around predicted firelines. Feature attribution using integrated gradients highlights vegetation condition and recent fire activity as primary drivers of model confidence. Deep ensembles further confirm that these uncertainty estimates are probabilistically well-calibrated across multiple folds. Together, these results suggest that spatial uncertainty in EO-based wildfire forecasting is structured, interpretable, and operationally actionable. The code for all experiments is available on GitHub.\footnote{\url{https://github.com/roloccark/wildf-UQ}}

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Evan_G_Shelhamer1

Submission Number: 4994