Detecting a wildfire when it is still only the size of a car could completely change how firefighters and authorities respond. Google Research is working with Muon Space and the Earth Fire Alliance on FireSat, a satellite constellation that will combine specialized cameras and artificial intelligence to monitor fires in near real time.
The goal is ambitious: once the system is complete, it will be able to identify fires measuring approximately 5 by 5 meters anywhere on the planet. Why does detecting such a small flame matter so much? Because the first few minutes can determine whether a fire is contained or becomes a large-scale emergency.
The challenge of detecting fires from space
Satellites already observe Earth’s surface, but their images are not always useful for finding fires in their early stages. Some may be up to 11 hours old or have too low a resolution to distinguish a small fire.
In addition, a satellite image can contain many misleading signals: clouds reflecting sunlight, industrial chimneys, hot surfaces, or even a backyard grill. For emergency teams, an incorrect alert is not a minor issue. Too many false positives can overwhelm their response capacity.
Google evaluated two paths. One involved building enormous, highly precise and expensive satellites capable of observing every area for a long time. The other focused on launching many smaller satellites and using machine learning models to compensate for some of their limitations.
FireSat follows the second strategy: more coverage, frequent observations and algorithms trained to interpret the data.
How AI learns to recognize a real fire
The training began with a camera designed specifically to detect fires. Researchers used it during controlled burns to create a reliable reference for what a fire looks like from the air.
They then flew the camera over California and recorded scenes that could be mistaken for fires. By comparing how those scenes changed over time, the models learned to distinguish a real fire from a harmless thermal signal.
Time is a key signal
A single photograph does not always tell the whole story. A controlled agricultural burn may look like a wildfire in one image, but its progression will be different. That is why AI needs to analyze sequences of observations rather than individual pixels alone.
Once the constellation is deployed, FireSat plans to observe Earth every 20 minutes. This frequency will make it possible to build a reference record, known in research as ground truth, showing which types of fires go on to grow and which ones die out or remain controlled.
With this data, researchers will be able to adjust the models to prioritize fires that are most likely to spread. In other words, the system will not just look for fire; it will also try to understand what could happen next.
A camera optimized for small fires
Emergency agencies have indicated that a reasonable opportunity to contain a fire usually exists when it covers between 50 and 100 square meters. FireSat has set an even more demanding goal: detecting fires covering 25 square meters, equivalent to an area measuring 5 by 5 meters.
That figure does not mean every fire of that size will be detected under all circumstances. Visibility, smoke, clouds, terrain and the satellite’s position still play a role. The advantage comes from combining a specialized camera with frequent observations and models capable of interpreting the context.
This approach also highlights something important about AI: it often does not replace sensors. It makes them more useful by finding patterns in large volumes of information and turning them into alerts that people can review and act on.
What changes for emergency teams
For those fighting fires on the ground, the main benefit is time. An early alert can help mobilize resources before the flames spread, while tracking the fire’s development can improve predictions about its direction and speed.
The information could also support urban planning decisions. For example, city officials could analyze where to build a firebreak—a barrier designed to slow the spread of flames—using historical data rather than estimates alone.
FireSat’s promise is not that AI will put out fires on its own. Its value lies in providing faster, more accurate information so people can act sooner.
FireSat aims to observe the planet every 20 minutes
Google says it has already placed the first group of operational satellites into orbit and that the next batch will arrive next year. Full deployment will require approximately 50 satellites and is expected to reach its full capacity around 2030.
In the meantime, the goal for the next two years is to obtain a complete image of Earth’s surface every hour. That improvement alone could change how authorities detect and track fires during high-risk seasons.
The project still faces an extensive phase of testing, calibration and validation. AI will have to prove that its alerts are fast, reliable and useful enough in regions with very different conditions. But the central idea is clear: observe more often, understand changes more effectively and deliver that information while it can still make a difference.
Original source
https://blog.google/innovation-and-ai/models-and-research/google-research/wildfire-tracking-ai