Concise summaries of RAI wildfire studies. Each entry highlights the problem setting, data
sources, evaluation choices, and operational implications, with links to the full paper
and supporting materials.
Wildfire-FM is a wildfire-specific foundation model trained with weather, active-fire observations, topography, vegetation, and static environmental data. The study focuses on model evaluation as much as model design: transfer results can look different when matching rules, scoring thresholds, or output contracts change.
Research question
When a wildfire model scores well, how much of that performance comes from learned fire behavior, and how much comes from the evaluation contract?
Wildfire transfer results change when matching rules change, even when the output is held fixed.
Head-selection metrics can reward ranking quality while missing the decision threshold that operators care about.
Fixed contracts make comparisons easier to audit across occupancy, spread, retrieval, and regression tasks.
A Nationwide Benchmark for Wildfire Initial Attack Failure Prediction with Public Environmental Data
Runyang Xu, Xueqi Cheng, Yushun Dong
This study defines initial attack failure prediction as a reproducible national benchmark. It aligns 38,128 naturally caused FPA-FOD events with public discovery-time signals from FIRMS/VIIRS, gridMET, LANDFIRE, OpenStreetMap, and WorldPop, while excluding outcome-derived clues that would leak the answer.
Research question
How far can public information available at discovery time go in flagging fires that may escape early control?
Public discovery-time data contains useful early-risk signal, but it does not fully determine suppression outcome.
FIRMS/VIIRS is the least redundant source under full input; fuel is the strongest static fallback signal.
The benchmark fixes event units, labels, time splits, forbidden features, and metrics before comparing models.
