Boundary-Aware Uncertainty Quantification for Wildfire Spread Prediction

Reliable wildfire spread prediction is critical for emergency planning, but most deep learning models lack principled uncertainty quantification (UQ). Existing methods often evaluate UQ using global metrics, which miss boundary-sensitive cases like wildfire fronts. Jonas V. Funk's new paper introduces the Fire-Centered Evaluation Region (FCER) framework, a spatially conditioned protocol that focuses UQ evaluation on critical fire zones. This shift makes model assessment more operationally relevant for real-world firefighting and evacuation decisions.

Using the WildfireSpreadTS dataset, the study compares an ensemble model against a distilled single-pass student model under FCER. Surprisingly, the lighter student model achieves comparable calibration and complementary uncertainty ranking in boundary-relevant areas, suggesting that efficient models can still provide reliable uncertainty estimates. The work includes open-source code on GitHub for reproducibility. This could improve trust in AI-driven wildfire predictions by providing more meaningful uncertainty metrics near fire edges.