New benchmark WILDFIREIA predicts wildfire escape risk with public data

Researchers from the University of Virginia and collaborating institutions have released WILDFIREIA, the first nationwide benchmark for predicting wildfire initial attack (IA) failure using only publicly available environmental and contextual data. The benchmark compiles 38,128 naturally caused wildfire events from the FPA-FOD database, aligning them with FIRMS/VIIRS thermal satellite detections, gridMET weather and fire-danger variables, LANDFIRE vegetation and fuel layers, OpenStreetMap access features, and WorldPop population density. To ensure reproducibility, the authors fix event units, size-based label rules, chronological splits, and explicitly forbid using final fire size, containment timestamps, or post-discovery satellite data. This prevents data leakage and simulates real-time prediction at fire discovery.

Under this strict protocol, 16 models from tabular, temporal, spatial, and spatiotemporal families were evaluated. XGBoost achieved the best performance with an AUPRC of 53.3%, revealing that public discovery-time data provides useful but incomplete signal for IA failure prediction. The analysis also found that FIRMS/VIIRS thermal detections are the least redundant data source, and fuel characteristics are the strongest static predictors when dynamic observations are unavailable. The authors have released preprocessing outputs and model-ready caches to support reproducible research. This work is a significant step toward scalable, data-driven wildfire early warning systems that do not rely on proprietary agency response records.