RAI Hazard Intelligence

From the lab

Research Notes

A closer look at the science behind wildfire prediction.

01 Wildfire-FM

Models & evaluation

Does Your Wildfire Prediction Model Actually Work, or Just Score Well?

Yangshuang Xu, Yuyang Dai, Liling Chang, Qi Wang, Yushun Dong

A wildfire model can earn a strong score and still fall short when the rules of evaluation change. Wildfire-FM brings weather, active-fire observations, terrain, and vegetation into a shared model, then asks what its predictions really tell us.

Read the paper Model release

arXiv: 2605.18911

Wildfire-FM connects environmental observations with a shared model and multiple evaluation tasks.

The same forecast, different scores

A small shift in where or when a fire is predicted can change whether it counts as a correct forecast. Matching rules make that distinction explicit.

Key findings

  • The same forecast can receive different scores depending on how predicted and observed fires are matched.
  • Ranking fires well does not guarantee reliable decisions at a chosen alert threshold.
  • Keeping the data, tasks, and scoring rules fixed makes model comparisons more meaningful.

02 Wildfire Initial Attack

Data & benchmarks

A Nationwide Benchmark for Wildfire Initial Attack Failure Prediction with Public Environmental Data

Runyang Xu, Xueqi Cheng, Yushun Dong

Which fires are likely to escape early control? This study brings together 38,128 naturally caused fires and public environmental data to examine what can be learned at the moment a fire is discovered, before the outcome is known.

Read the paper Code & data

arXiv: 2606.15529

Public data describes conditions at discovery time, before a fire escapes initial control.

What was known when the fire began

FPA-FOD events are aligned with FIRMS/VIIRS, gridMET, LANDFIRE, OpenStreetMap, and WorldPop to build a consistent picture of each fire.

Key findings

  • Public observations offer useful early warning signals, but cannot fully explain the outcome of suppression.
  • Satellite fire detections add distinct information; fuel provides a strong signal when other observations are missing.
  • The benchmark fixes event definitions and time splits, and excludes information learned only after the fire.

Beyond the papers

Explore the data in context.

Open FireEye PyHazards documentation