Sentinel-2: ten metres

Sentinel-2 does not measure fire but reflected sunlight. What it delivers is a true-colour image at ten metres of ground resolution. For the first time an individual fire edge becomes visible rather than just a heat blob, the difference between “it is burning there” and “this is how far it has got”.

Alongside the true-colour image, samoEKO carries the near-infrared (10 metres) and short-wave infrared (20 metres) bands for every selected scene. They are not needed for the image layer; they are the stock for the burn severity calculation.

There is no ready-made, keyless tile service for Sentinel-2. Three routes were examined and ruled out: NASA GIBS carries only orbit tracks for Sentinel-2 and no image product. The Copernicus Data Space Ecosystem provides the imagery but requires an account and a configured instance identifier for image output. And the freely reachable “Sentinel-2 cloudless” service from EOX states in its own capabilities that it is under a non-commercial licence, samoEKO is a product of samoLabs, so the layer is unusable.

What does exist without registration are the raw scenes as Cloud Optimized GeoTIFF in the AWS Open Data programme, catalogued through a STAC catalogue. For each query day samoEKO selects the best scene per acquisition tile and cuts the map tiles through its own tiling service. The browser sees none of this: neither the catalogue nor the storage nor the tiling service ever appears in a browser request.

The sub-area is the satellite's acquisition tile in the military grid (MGRS), roughly 110 by 110 kilometres; the working area contains about a hundred of them. A self-drawn grid would inevitably leave holes.

Only scenes with less than 60 per cent cloud and at the same time less than 60 per cent missing data enter the selection. The second filter is not incidental: a scene at the edge of the acquisition strip can be 57 per cent empty and still report 0.05 per cent cloud, because the cloud fraction refers to the valid pixels. Filtering on cloud alone selects precisely those empty scenes.

Among the remainder, the best mix of usability and freshness wins. An age penalty of 1.5 percentage points per day answers exactly one question: how much more cloud may an image carry and still win on freshness? Measured across the working area it lowers the mean image age from 4.3 to 2.5 days and raises the highest cloud fraction from 20 to 30 per cent a trade that points the right way for a situation picture.

Reflected light means daytime only, and cloud-free only. And the image does not have one date but one per acquisition tile. On 3 August 2026 the working area was covered by 101 tiles with acquisitions from 25 July to 3 August, at a highest cloud fraction of 29.7 and a mean of 0.2 per cent.

That is why the legend states the scene date and cloud fraction for each view. A detail image that is several days old depending on the area, without showing its date, would be a trap in operational use.

Below zoom level 10 the layer deliberately stays empty: a map tile there is 78 kilometres wide and would almost always straddle two scenes, of which it could show only one. At that scale the daily images carry the overview.