Scientists Released 2,000 Wildfire Photos to Teach AI to Spot Fires
Earlier fire detection could mean faster evacuations — and more homes saved.
A team of researchers has released a free, public collection of 2,148 satellite images of 25 California wildfires, each one paired with a digital map showing exactly which pixels were on fire. The images come from Sentinel-2, a pair of European Space Agency satellites that photograph the whole planet a few times a week. They were chosen using infrared light, which cuts through smoke better than ordinary cameras. Each pixel covers about 20 meters — roughly the length of a tennis court.
The point is to give AI a practice test. Teaching software to spot fires from orbit is harder than it sounds, because fire is astonishingly rare in a satellite image. In this dataset, actual burning pixels make up just 0.0766% of all the cells — fewer than one in a thousand. A lazy AI can simply say "no fire anywhere" and look 99.9% accurate while being useless. By packaging fire and non-fire images together, with matching maps, the researchers make it possible to fairly test which approaches actually work. They also split the data by fire incident, so the AI is judged on blazes it has never seen.
Here is the honest catch: those fire maps were generated by a computer rule using infrared brightness, not drawn by human eyes. Only 233 of the test images got a manual review. So AI trained on this data is partly learning from a machine's best guess, mistakes included. And this is a research dataset, not a product — no fire department will get alerts from it tomorrow.
Still, the release matters. Wildfire response is a race against time, and every minute shaved off detection means firefighters reach a blaze while it is still manageable and residents get a warning sooner. Insurance companies, emergency managers, and utilities that shut off power lines during high winds all depend on knowing where fire is right now. Because the data, the maps, and the training code are all openly available, anyone — a grad student, a startup, a fire agency — can build on it without starting from zero.
- 2,148 satellite images of 25 California wildfires (2020–2026) are now free for anyone to use, each with a map of exactly where flames were burning
- Burning pixels are only about 0.0766% of the data — fewer than 1 in 1,000 — so an AI that always says 'no fire' would look 99.9% right and be completely useless
- The fire maps were made by computer rules rather than human eyes, with only 233 test images manually checked, so the labels aren't perfect
Why It Matters
Faster space-based fire detection could mean quicker evacuations, earlier firefighter response, and homes saved.