The time between a wildfire igniting and the first aerial resource arriving on scene determines whether an incident is controlled at a quarter of a hectare or allowed to grow into a disaster covering thousands of hectares.
Every minute of detection delay costs. Not in abstract terms. In hectares burned, in structures threatened, in lives at risk.
Why existing systems are slow
Traditional satellite fire detection pipelines work like this: a satellite captures thermal data, stores it, transmits it to a ground station during the next available pass, where it is processed by servers, reviewed by analysts, and routed to operations centres. In good conditions this takes 45 minutes. In real conditions it often takes considerably longer.
The fundamental problem is that all the intelligence sits on the ground. The satellite is used only as a sensor. Everything else happens after the fact.
What we do differently
Our approach puts a lightweight inference model directly in the data processing pipeline. Rather than downloading full resolution thermal imagery and processing it on the ground, we detect thermal anomalies consistent with fire ignition at the point of ingest, before full image volumes are transmitted.
The system then applies three contextual checks before producing an alert. It examines the land cover classification at the anomaly location. It retrieves the current wind field from ERA5 atmospheric reanalysis to model the likely spread direction. It checks the live fuel moisture index from Sentinel-2 data for the affected area.
Only when all three checks have been completed does the system produce a ranked alert with severity, confidence intervals, recommended actions, and a 24 hour spread trajectory forecast.
The result
In operational testing across the 2025 fire season, the system consistently produced alerts in under 60 seconds from Meteosat scan acquisition to alert delivery. The false positive rate across over 1400 detected thermal anomalies was below 12 percent. After cognitive filtering, fewer than 1 percent of delivered alerts were false positives.
The system detected fires of less than a quarter hectare on seven separate occasions. All seven were confirmed by subsequent satellite imagery or ground crews.
This is the difference between catching a fire at a manageable size and watching it become a regional emergency.