We tend to think of bias in data as something in the data we hold: a skewed sample, a mislabelled set, a model trained on the wrong population. The harder bias to see is the data we never collected at all. You can interrogate a dataset you have. You cannot interrogate the one that does not exist, and in geography that missing data is not spread evenly. It thins out exactly where people, money, and infrastructure thin out, and in wildfire that turns out to matter a great deal.
Start with what good data looks like in a well-resourced landscape. Populated and high-value areas are surveyed with airborne LiDAR, which resolves terrain and vegetation in three dimensions down to the sub-metre: canopy height, canopy base height, the ladder fuels that carry a surface fire up into the crown. Move into remote country and that detail disappears. The fallback is coarse and global: elevation from SRTM at roughly 30 m, fuel maps at a similar scale, and, for active fire, satellite feeds measured in hundreds of metres. The workhorse instruments detect fire at 375 m (VIIRS) and 1 km (MODIS) per pixel; the geostationary satellites that watch continuously do so at around 2 km. It is the same ground, seen at resolutions an order of magnitude apart.
This is not a wildfire quirk. It is the general shape of spatial data. A 2023 study in Nature Communications found that OpenStreetMap building data is more than 80% complete for only 16% of the world's urban population, and less than 20% complete for nearly half of it, with completeness tracking closely to a country's wealth and level of development. We map what we value, and we have valued the places where most of us live. The rest is, quite literally, a blur.
The blur has consequences, and two of them are operational. The first is detection. A finer satellite pixel catches a smaller, newer fire; VIIRS can pick up a flaming front of around 40 square metres in daylight, where MODIS needs closer to 100. But resolution is only half the problem. Polar-orbiting satellites do not watch, they pass, a handful of times a day, and a new ignition that starts between overpasses is simply not seen until the next one. In a monitored landscape a camera, an aircraft, or a person on a ridge closes that gap. In the remote dark, nothing does, and the fire is found later and larger.
That delay is expensive in a very particular way. Initial attack, hitting a fire while it is still small, succeeds on the order of 95 to 98% of the time. The problem is the remainder. Analyses of initial attack repeatedly find that the small share of fires which escape early containment go on to account for roughly 85% of the total area burned and a similar share of what is spent fighting them. It’s true that not every large fire is a failure. A growing share of fire is deliberately managed rather than fought, left to do ecological work where it is safe to let it run. The concern here is narrower, as the fires we never got to choose. Detection and response delay is one of the main reasons a fire escapes before anyone can decide what to do with it, removing not just the chance to stop it but the chance to make a considered call at all. So the places we watch least are the places most likely to produce the fires that cost the most.
That is the first of the two failures: we see these fires late. The second is that we forecast them worst exactly where terrain and fuel drive them hardest. Where a fire will run next depends on slope, aspect, and the structure of the fuel, the things airborne LiDAR resolves and a 30 m global model cannot. So the same rugged country that hid the fire from detection also defeats the models meant to predict it, at the moment those predictions matter most. We bring our coarsest tools to the ground that most rewards precision.
Here an honest argument has to slow down, because there is a reason we invest where we do. The UN's framework for disaster risk is worth borrowing: risk is not the hazard alone, but the interaction of hazard, exposure, and vulnerability. A fire in empty country is a hazard with little exposure: few people, few structures, little to lose, and so, by any conventional measure, lower risk. Under that logic, pointing our best sensors and surveys at the places with the most to protect is not negligence. It is triage. The data follows exposure, not hazard, and for defensible reasons.
But that logic leans on a word doing more work than it looks: what counts as something to lose. Exposure, as we measure it, is largely built assets and people. It is quiet on watersheds, on the forests that hold a city's water supply, on carbon, on habitat, on land that is culturally and spiritually central to the people who have lived on it longest. These are not nothing. They are simply harder to put on a balance sheet, and so they sit in the same blind spot as the terrain data. "Low consequence" often means low in the things we bothered to count.
And fire does not respect the triage. This is the part that should give pause even to a reader who values none of what I have just listed, no watershed, no habitat, no carbon. The ignition we found late and modelled poorly, out in the low-exposure dark, is frequently the very one that grows into the fire that reaches the places with everything to lose. You do not have to care about the empty country for its own sake to have a stake in watching it. You only have to live downwind. The seam where the well-mapped meets the unmapped is not fixed, either. In the United States, the wildland-urban interface, the edge where houses meet vegetation, grew by 41% in homes and 33% in area between 1990 and 2010, the fastest-growing land-use type in the country.
"We are building further into the blur, faster than we are mapping it."
None of this argues for the impossible. We are not going to fly LiDAR over every wilderness; the cost is prohibitive, and the wild is, by definition, where we are not. But there is a shift worth making in how we think about closing the gap. Data does not only come from surveying a place in advance. It also comes from operating in it. An aircraft working a fire in remote country is, without trying, taking a measurement of that country: where the fire actually ran, how fast, what the terrain and the wind really did, ground truth from the one moment someone was finally there to see it. A single flight is an anecdote. Thousands of them, gathered over years across the same blank spaces, become a map of exactly the places no survey will ever reach. The deserts on our maps close from the work done in them, not only from the instruments pointed at them from above.
The wider point is not really about fire. When a decision rests on a map, the most important question is often what the map leaves out, and whether it is missing because there is nothing there or because no one looked. It is worth remembering that the phrase "there is no such thing as a natural disaster" is now half a century old; researchers argued in 1976 that disasters are built by human choices about where we live and what we protect, not handed down by nature. The gaps in our data are one of those choices. In wildfire, the cost of that blind spot does not stay in the remote country where it begins.
Sources
Herfort, B. et al. "A spatio-temporal analysis investigating completeness and inequalities of global urban building data in OpenStreetMap." *Nature Communications*, 2023.
Schroeder, W. et al. "The New VIIRS 375 m active fire detection data product." *Remote Sensing of Environment*, 2014 (sensor resolutions and minimum detectable fire size); NASA VIIRS/MODIS active-fire documentation.
Radeloff, V.C. et al. "Rapid growth of the US wildland-urban interface raises wildfire risk." *PNAS*, 2018.
Initial-attack effectiveness and the concentration of area and cost in escaped fires: initial-attack performance literature (FRAMES); Ontario initial-attack analysis, *International Journal of Wildland Fire*.
UNDRR / Sendai Framework for Disaster Risk Reduction: disaster risk as the interaction of hazard, exposure, and vulnerability.