The date you picked is probably not the date of your satellite image

Ask a satellite tool for a picture of a farm on 15 August, and you will get a picture. It may not have been taken anywhere near 15 August, and most tools will not tell you.

The thing nobody explains first

A satellite is not a camera you can point. The Sentinel-2 satellites that most free forest monitoring runs on pass over any given place roughly every five days, on a fixed orbit, at a fixed time of morning. You cannot ask for a picture of Tuesday. You can only take whichever passes already happened.

And most of them are useless. The satellite photographs the top of whatever is there, so if there is cloud, you get cloud. In the wet tropics — which is where cocoa, palm, coffee and rubber grow — it is normal for weeks or months of consecutive passes to be too cloudy to read the ground through.

So "show me this plot on 15 August" frequently has no answer. There is no image. Every satellite tool has to decide what to do at that point, and the decision is almost always invisible to you.

What tools do instead

They widen the search. If nothing usable exists near your date, look a week either side. Then a month. Then three months. Eventually something clear enough turns up, and that is what you get, labelled with the date you typed.

Each step is a reasonable response to "no usable image here." Together they will hand you a picture of a different season and present it as an answer to your question.

We know because ours did it. Asked for 15 August over a cocoa-growing area in Côte d'Ivoire, our software returned imagery from 3 May — 104 days earlier, a different season — and reported the result without mentioning the substitution. That was not a competitor's tool and it was not deliberate. Nobody had looked.

What it actually costs, measured

We ran the same three dates over four areas and recorded the gap between the date requested and the scene returned:

Two things stand out.

The drift is worst exactly where the commodities are. Temperate Europe lands within days. Côte d'Ivoire, Indonesia and Myanmar — cocoa, palm, coffee, rubber — drift by two to three months. The regions with the most compliance exposure have the least reliable imagery, and it is the same weather that causes both.

It goes in both directions. The Myanmar request for 15 August returned 3 October, seven weeks later. So "the most recent image" is not a safe assumption either.

Why a seasonal gap is not just imprecision

If the gap were random noise this would be a rounding problem. It is not: a 90-day gap in a monsoon climate means comparing the wet season with the dry one.

Vegetation indices are seasonal by construction. NDVI, NDMI and NBR all move with rainfall, leaf flush and senescence, and in seasonal tropics that movement is large — comparable to, and sometimes larger than, the change caused by clearing land. So a seasonal mismatch does not merely blur an answer. It can manufacture apparent change where nothing happened, or hide real change behind a greener comparison image.

For EUDR screening the stakes are specific. A plot is judged by comparing forest cover near the 31 December 2020 cutoff against forest cover now. If your "2020 baseline" is actually a dry-season scene and your "current" image is a wet-season one, the difference you measure is partly phenology and partly land use, and nothing in the output tells you the ratio.

What we changed

A hard cap on how far the search may wander. Forty-five days either side of the date you asked for. That is roughly a season's edge, and about nine chances for the satellite to catch a clear sky, so genuine cloud gaps still resolve.

Declining instead of substituting. Past that, we return an error rather than an answer:

No clear Sentinel-2 image within 45 days of 2025-08-15 for this area. Every scene in that window was too cloudy, and imagery from further away is not substituted because it would likely show a different season. Try a date in a drier part of the year.

We would rather say we cannot answer than answer a question nobody asked.

Saying so when the gap is real but acceptable. Between 7 and 45 days the result carries a notice with both dates and the gap in days, above the result rather than in a footnote.

One deliberate exception. EUDR screening still searches wider, because its baseline has to reach across a defined window around the 2020 cutoff and is then constrained back into it. Capping that would turn workable plots into errors. It is the right behaviour there and the wrong behaviour everywhere else, which is why it is an exception rather than a default.

What to check in whatever tool you use

This is not a GeoTown problem, and pointing at ours is not the purpose. If you are acting on satellite analysis of tropical land, four questions are worth asking of any tool, ours included:

1. What is the acquisition date of the image behind this result? Not the date you requested — the scene date. If the interface will not tell you, that is your answer.
2. How far can it wander from the date I asked for, and does it stop?
3. When it compares two dates, are they in the same season? Same month in different years is usually a reasonable proxy.
4. What happens when there is no usable image — does it tell me, or does it quietly find something else?

The fourth is the one that matters most, and it is the one almost no tool answers on screen.

The reason any of this surfaced is that we started checking our own outputs against independent references. Dates were the first thing that did not line up, and they were the easiest to check.

If you are evaluating a satellite compliance tool this year, ask what it does when the sky is not clear. The answer tells you more about the product than any accuracy figure will.

Contains modified Copernicus Sentinel data. You can check scene dates yourself on any analysis at geotown.io/explore, and our methodology documents scene selection in full.