TerrAlert Team · 2026-09-09
This is a straightforward walkthrough of what TerrAlert actually does, step by step, describing only functionality that exists in the product today.
You draw a boundary, an Area of Interest, on a map in the dashboard, or submit one programmatically as GeoJSON through the REST API if you are on a plan with API access. You also choose which categories of change to watch for at that location: deforestation, construction and development, flooding, fire, vegetation loss, or urban expansion.
TerrAlert uses Sentinel-2 optical satellite imagery, sourced through the Copernicus Data Space Ecosystem, the European Space Agency's public, open-access programme. As new imagery of your area is captured and becomes available (which, per Sentinel-2's revisit schedule and typical cloud cover, is usually every several days for a genuinely clear, usable image; see How Fast Are Satellite Alerts, Really?), TerrAlert retrieves and processes it.
Once two clear images of the same area exist, a "before" and an "after," an AI model (a Siamese U-Net architecture, for anyone curious about the technical detail) compares them and identifies areas that differ. The result is a set of candidate change polygons, each with an estimated area and a model score from 0 to 1. The score ranks detections; it is not a calibrated probability that the change is real. The model finds that an area changed, not what changed, so each detection is labelled with the category you asked us to watch for when the area watches one category, and as "change" when it watches several.
TerrAlert checks each detection against an independent data source that can see that kind of change, rather than relying on the optical model alone: NASA's FIRMS active-fire product for fire, the GLAD and RADD forest-loss alert systems for deforestation and vegetation loss, and NASA flood maps for flooding. A signal of a different kind nearby does not count. No such source is in use for construction or urban expansion. Each detection says whether it is verified, partly verified, not verified yet or cannot be verified. See Understanding False Alarms for why that matters more than a single headline accuracy number.
Every Area of Interest has a configurable minimum model score and an alert policy. Under the default policy, "confidence", a change above that score alerts once the next clear image still shows it, compared with an earlier one. Under the "corroborated" policy, it alerts only once an independent source has verified it, which means construction and urban expansion alert only if you mark them real yourself. A verified change alerts under either policy. Fires and forest loss that NASA FIRMS, GLAD or RADD report themselves also alert once their score reaches your minimum, labelled as that source's report. An alert is delivered by email or by HMAC-signed webhook, including automatic detection of Slack and Microsoft Teams incoming-webhook URLs, which get a plain chat message instead of the raw signed payload.
On every plan, every model detection includes a before and after pair of still images, the estimated area affected, the model score and its label (the category the area watches, or "change" when it watches several). On Pro and Enterprise you can also open an interactive split view of the two scenes. From there, the decision about what to do next, investigate, escalate, dismiss, is yours. TerrAlert surfaces the finding; it does not make the call for you.
Do I need to be technical to use TerrAlert? No. Drawing an area and choosing categories to watch requires no technical background. The REST API is there for teams that want to integrate it into their own systems, on Pro and Enterprise.
What happens if no clear imagery is available for a while? Nothing gets reported for that area until a new clear image exists, which is correct behavior, not a system failure.
Ready to see it on a real area? Start free or read Getting Started With TerrAlert.
All posts