How accurate is the Fire kit?

This article explains the models, data, validation results and limitations of the Fire kit. Use this information to interpret a simulation result.

Note: The Fire kit is a planning tool for comparing locations and scenarios, such as spread on different slopes, ignition along a power line, or changes in wind and fuel dryness. It is not a forecast of a real fire. Never use it to decide whether to evacuate. For an active fire, follow your local emergency services.

Fire behaviour models

The kit uses the standard fire behaviour science that US land managers use:

• Surface fire: Rothermel's spread model (1972, with Albini's 1976 update), with the 40 Scott and Burgan fuel models and the original 13.
• Crown fire: Van Wagner's model for when a fire climbs into the trees, Rothermel's 1991 model for how fast a crown fire runs, and Finney's method for fires that are partly in the canopy.
• In Canada: the Canadian Forest Fire Behaviour Prediction (FBP) System (Forestry Canada 1992, updated by Wotton and others in 2009), Canada's own model, built from experimental fires in Canadian forests. It works from the fire weather codes Canadian agencies use every day.
• Spread across the map: a minimum travel time method, the same idea used by the US Forest Service's FlamMap. The fire takes the fastest path through every cell, uphill and downwind first.

In the US these are the models behind FARSITE, FlamMap and BehavePlus, the US tools for fire planning; in Canada, the FBP System is the model behind Canada's national fire behaviour maps and Prometheus.

Data sources

The quality of a result depends on the input data:

• Fuels (US): LANDFIRE, the national map of fuel types and tree canopy from the US Department of the Interior and the US Department of Agriculture, at 30 m. The lower 48 states use the newest LANDFIRE update, with the previous release filling any gaps. Alaska and Hawaii use their own LANDFIRE releases.
• Fuels (Canada): Natural Resources Canada's national fuel type map (2024), at 30 m: spruce, pine, aspen, mixedwood, grass and the land that does not burn.
• Terrain: Scribble Maps elevation, from USGS 1 m lidar across most of the US and 30 m elsewhere.
• Wind: in the US the kit starts from the National Weather Service hourly forecast for your area; elsewhere from a steady 20 km/h west wind. You can change it to test any wind.

The report card displays fuel data coverage and confidence for your area. Both maps show fuels as they were when they were made (LANDFIRE 2024 and 2025, the Canadian map 2024). A fire, a harvest, a fuel treatment or new homes since then are not in them.

Model validation

Current validation checks:

• The fire behaviour calculations is checked against an independent open source implementation of the same models (Pyretechnics). On flat ground the spread rate, the fire's intensity and its shape match within 0.5%. With wind and slope together they match within 10%, because the two tools combine wind and slope in slightly different ways that are both used in practice.
• The Canadian model is checked against the published Fire Weather Index tables (the Initial Spread Index), and each fuel type's spread, crowning and fuel consumption against the published FBP equations.
• Known shapes: with no wind on flat ground the fire grows as a circle, and with wind it grows into the ellipse the models predict.
• Model constraints: wet fuel does not burn, roads, water and rock stop a fire (even crossing them diagonally), and a fire never burns beyond the time you set.
• Real fuel data: when we load a new LANDFIRE release, we check its values at known places (forest, shrub, grass, towns) before the kit uses it.

The model checks are part of our automated tests.

Validation against recorded wildfires

We replayed real wildfires 12 hours at a time. Each test starts from the real fire's edge, as mapped from NASA satellite data. It runs the kit for 12 hours with the wind and humidity measured at the nearest airport, and compares where the kit's fire spread with where the real fire spread. The kit used the same terrain and fuel maps you get, mapped before each fire.

• 2026 fires (current maps): 96 twelve-hour periods of 80 wildfires, 60 in the US and 36 in Canada.
• 2018 to 2021 fires (LANDFIRE 2016, the map made before them): 117 periods of 104 US wildfires.

Spread rate. In the US the kit's pace was close to real fires. On average it took about the same time to burn the same area: 1.1 times the real speed in 2026, 1.0 times in 2018 to 2021. About 6 in 10 periods were within a factor of 2. It ran a little slow in timber and a little fast in grass and shrub, and too fast at night, when real fires calm down. In Canada the kit ran about 3 times faster than real fires. It keeps the afternoon fire weather for every hour, but real Canadian fires burn hard for only a few hours a day. Read Canadian times as hours of active afternoon burning.

Spread direction. We compared the shape of the kit's fire with a fire that grows the same amount evenly in every direction. In strong wind, over about 18 km/h, the kit was closer to the real shape in most periods. In light wind it did about as well as even growth, because direction then depends on small wind shifts, terrain and fire behaviour that a steady wind cannot capture. Overall the kit was closer than even growth in about half of US periods and 7 in 10 Canadian periods.

Interpreting these results. Use the kit to compare places and scenarios, and to see how far a fire could run in a few hours under steady conditions. Do not use it to predict exactly where a real fire will be.

Planned validation checks:

• FlamMap comparison: the same landscapes run through FlamMap, the US Forest Service tool, to show how close our results are.
• Regression checks: model changes are released only if they do not reduce these scores.

We will update this page as each benchmark is published.

Simulation limitations

Real fires can be faster and bigger than the simulation. It does not model:

• Embers. Wind can carry burning embers far ahead of a fire and start new spot fires. This is how many homes burn. The kit stops a fire at roads, water and town blocks that a real fire can jump.
• Homes as fuel. LANDFIRE maps most built-up areas as land that does not burn, so the fire stops at the edge of town. Real fires can spread from house to house.
• Changing weather. The wind and fuel dryness stay the same for the whole run. Real fires change with the time of day, wind shifts and fronts.
• Firefighting. Crews, aircraft and fire lines are not modelled.
• Fire making its own weather. Very large fires can create their own winds.

How to read a result

• Compare scenarios. Use it to compare scenarios: which ignition is worse, how much a drier day matters, which slopes and canyons carry fire fastest.
• Check the confidence line. Trust a result more when the card says the fuel data covers the area with high confidence.
• Treat times as a guide. "Burns by hour 6" means "under these steady conditions". Strong wind and dry fuel make real fires faster.
• Consider ember spread near homes. Where the fire reaches the edge of a neighbourhood, assume embers could carry it further.

See also

• Fire - Wildfire Spread Simulation: how to use the kit.