Summary
Google Research says its work with Muon Space and the Earth Fire Alliance’s FireSat programme is developing an AI-assisted satellite system to detect wildfires as small as 5 by 5 metres. The planned constellation aims to image the entire Earth every 20 minutes once fully deployed, with full operational capacity expected around 2030.
Google Research is supporting a satellite programme designed to detect wildfires while they are still small, using specialised cameras and artificial intelligence to distinguish fires from other heat sources and visual artefacts. The Earth Fire Alliance’s FireSat constellation is intended to identify fires as small as 5 by 5 metres—roughly the size of a car—from orbit.
Google describes the work in an interview with research scientist Chris Van Arsdale, who leads the company’s climate and energy research group. The project is being developed with Muon Space, which is working with Google to support the FireSat programme.
Why early wildfire detection is difficult
Fire authorities need information before a fire has spread beyond a manageable area. Google says agencies it consulted estimated that a fire covering roughly 50 to 100 square metres may still offer a chance of containment, leading the project to set a more ambitious detection target of 25 square metres.
Existing satellite observations can be difficult to use for this purpose. Images may be around 11 hours old or too low-resolution to reveal a rapidly developing fire when it is still small. A satellite image can also contain signals that resemble fire, including sunlight reflected by clouds, industrial heat sources such as smokestacks and backyard grills.
One way to improve precision would be to build a large, expensive high-resolution satellite that spends a long time observing the ground. FireSat instead uses smaller, lower-cost satellites in low-Earth orbit and relies on repeated observations and machine-learning models to improve the interpretation of those images.
How the AI is trained to separate fires from false alarms
Google says it built a camera optimised for wildfire detection and initially trained its models using images collected during controlled burns. Researchers have also flown the camera over California, recording scenes that appear to be fires from above and observing how they change over time.
That time-series information is important because not every visible fire should trigger an emergency response. Agricultural burns, for example, are controlled fires. By following how different events develop, the models can be trained to distinguish fires that become wildfires from those that do not.
Once the full FireSat constellation is deployed, its satellites are intended to image the Earth every 20 minutes. The repeated measurements would allow the system to detect new fires earlier and track their development in near real time. Google also describes the resulting observations as a potential “ground-truth” record showing which small fires grow into large wildfires.
In this context, ground truth means a record of observed events that can be compared with the predictions made by an AI system. More such data could help refine the detection models and give researchers a better picture of how fire behaviour varies with local conditions.
A phased path to global coverage
Google says the first batch of operational satellites is already in orbit and that the next batch is expected the following year. The complete system will require approximately 50 satellites, with a goal of imaging the world every 20 minutes around 2030.
Before reaching that stage, the programme aims to capture a complete image of the Earth’s surface every hour within the next two years. The project therefore combines a current operational phase with longer-term coverage targets rather than offering the final global observation rate immediately.
For first responders, the intended benefit is additional time: earlier alerts could help crews act while a fire is small, while repeated observations could help them forecast how it is spreading. The dataset could also support longer-term planning. Google gives firebreak placement as one example, because planners could use historical information about fire activity and local conditions to decide where barriers may be most useful.
The full 5-by-5-metre detection capability and 20-minute global revisit rate depend on completing the planned constellation. Google says reaching full operational capacity will take several years, with the current target set at around 2030.