← Back to overview

How this actually works, and where it falls short

The datasets we query, how a flag gets decided, how we tested it before showing it to anyone, and the honest limits of what a satellite can see.

What we check

Every point is checked against the same 31 December 2020 cutoff the regulation itself uses, against a small set of named, publicly documented satellite datasets. Nothing proprietary or hidden: you can look up any of these directly.

DatasetWhat it measuresRole in your flag
JRC Global Forest Cover 2020 A 10m global forest map fixed at 31 Dec 2020, the EU's own reference year. The "before" baseline every point is compared against. Named directly in the Commission's own Guidance Document.
Hansen Global Forest Change Tree-cover loss detected each year since 2000. The main signal behind a RED flag: loss after the cutoff, above the noise floor described below.
JRC Tropical Moist Forest Annual forest-cover transitions, split into clear deforestation and partial degradation. Confirms deforestation independently of Hansen, and separately drives the AMBER flag for wood/timber, where forest degradation is its own legal concept.
JRC Global Map of Forest Types A 10m map of primary, naturally regenerating, and planted/plantation forest at the 2020 baseline. Non-binding context only. It shows the forest-type mix in 2020, not a later conversion, and never changes the screening flag.
RADD & GLAD alerts Near-real-time radar and optical disturbance alerts, updated weekly to quarterly depending on the sensor. Not part of the compliance flag itself. Feeds a separate "nearby pressure" tier showing recent activity in the wider area, since the flag datasets above only update annually.
WWF HydroSHEDS free-flowing rivers Mapped watercourse locations. Context only: flags proximity to a known watercourse. Doesn't affect the RED/AMBER/GREEN decision.

Full source attribution, including license terms, is on our Terms page.

How a flag gets decided

Four possible outcomes per supplier. None of them is a compliance decision on its own.

REDForest loss detected after the cutoff, above the noise floor. Investigate this plot first.
AMBERA degradation signal, not a confirmed clear-cut. Applies to wood/timber only, matching the regulation's own definition.
GREENNo loss or degradation signal found. Lower priority to check, not a guarantee of compliance.
CANNOT_SCREENNo usable location data. A data gap, not a satellite finding, and ranked above RED, never dropped silently.

GREEN does not mean compliant. It means we found no forest-loss signal above the threshold we check for. It's a signal to move on to other suppliers first, not proof that a plot is clean.

Why very small clearings don't trigger a flag

RED and AMBER only fire once a loss signal covers roughly half a hectare of the checked area. That floor is anchored to the regulation's own minimum mapping unit for what counts as forest at all (Regulation (EU) 2023/1115, Art. 2(4)), not tuned to make the tool look more accurate. A real clearing smaller than that can exist without triggering either flag, which is one reason a RED or GREEN result is a starting point, not a verdict.

How we tested it before showing it to anyone

A single-country regression check that only confirms the tool agrees with itself isn't proof it's actually right. So before relying on any of this, we built a separate set of locations where the correct answer was established independently first, then checked whether the tool matched it.

20 hand-picked edge cases
All 20 matched. Each one was checked against a fresh, independent re-implementation of the same satellite data, direct visual inspection of before/after imagery, and, where the location is in Brazil, a cross-check against the national deforestation monitoring system. The set deliberately includes hard cases: a pre-2020 legacy clearing, a clearing that straddles the 2020/2021 boundary, and a real protected forest-coffee reserve.
228 broader-volume cases
All 228 matched, at a scale closer to a real customer portfolio than the 20 hand-picked cases above, using the same exact-match standard.
1,500-row stress test
This one checks speed and memory at volume, not accuracy on its own. A 300-row correctness sample seeded inside it also matched.

That's 248 locations across the two accuracy tiers where the right answer was pinned down independently before the tool ever saw them, not 248 arbitrary rows it happened to get right. This is a permanent, versioned test suite, not a one-time check: it gets re-run every time a dataset, threshold, or flag rule changes.

What this suite has not yet found

None of the verified cases is a confirmed example of a real agroforestry false positive: an actively managed shade-grown farm that the tool incorrectly flags. We specifically looked for one inside a real forest-coffee biosphere reserve, and at that location the tool read correctly. That's a useful data point, but it isn't proof the risk is low. Finding an actual failing example would take either a disputed real customer flag or farm-boundary ground-truth data we don't have access to yet. Treat the agroforestry limitation below as a live, unresolved gap, not a solved problem.

What satellites cannot see

This is the part a sales page usually leaves out. Read it before you trust a flag either way.

Agroforestry and shade-grown canopy

A coarse forest-cover baseline can misread the layered canopy of a shade-grown coffee or cocoa farm as loss, or the opposite: mask real clearing that happens underneath an intact-looking canopy. See "what this suite has not yet found" above.

Annual update lag

The two datasets that drive RED/AMBER update roughly once a year. Very recent clearing may not be visible yet, which reads as a false GREEN. The weekly-to-quarterly nearby-pressure alerts partly cover this gap, but they're a coarser, less precise signal, not a substitute.

Cloud cover

Optical imagery needs a clear sky. Persistently cloudy regions can mean lower-confidence before/after thumbnails, which we mark on the report as a data-quality signal rather than hide.

Point precision

A single coordinate stands in for an area. A geocoded address is far less precise than a GPS point, and a plot of 4 hectares or more legally needs a full boundary polygon, not a point at all.

Occasional data-coverage gaps

Every satellite dataset has small gaps in coverage. Where that happens, the report shows it as a lower data-quality tier rather than presenting a gap as a confirmed reading.

Questions a skeptical buyer should ask

How do I know a GREEN result is actually right, not just unchecked?

Because the underlying method has been checked against locations where the correct answer was established independently first, not just against itself. See "how we tested it" above for exactly what that means and doesn't mean. GREEN still isn't a compliance guarantee: it means no loss signal was found above the checked threshold, on the specific point or polygon submitted.

What can satellites not see?

Shade-grown canopy structure, very recent clearing before the annual datasets catch up, anything under a persistent cloud, and anything smaller than roughly half a hectare. The full list is above, and it's the same list we'd want to see before buying a tool like this ourselves.

Why should I trust a tool built by one person?

You shouldn't have to take that on faith, which is the point of this page. Every dataset above is named and public, the same ones the European Commission's own guidance points to, not a proprietary model you can't inspect. The flag logic is tested against a versioned answer key, not just "it ran without an error." And when something looks wrong, you're emailing the person who wrote the code, not filing a ticket into a queue.

Does a RED flag mean my supplier is non-compliant?

No. RED means a loss signal was found above the noise floor at that location after the cutoff. It's a reason to look at that specific plot before others, not a finding you can submit on its own. Real causes range from genuine illegal clearing to a misplaced coordinate to normal land use that predates the cutoff but shows up at the edge of the checked area.

Is this the same data EU authorities themselves rely on?

The forest baseline (JRC Global Forest Cover 2020) is the dataset the Commission's own Guidance Document names directly. The other layers are the standard public forest-monitoring datasets used across the sector, not something we built in isolation. That doesn't make our output an official government determination; it means the underlying inputs aren't a black box.

How often is the accuracy suite re-checked?

Every time the underlying satellite data, a threshold, or the flag logic changes. It's a standing regression suite, kept and re-run going forward, not a one-time check done once and left behind.

Found a flag you think is wrong, or have a different question? Email contact@eudrscreening.com, we reply personally and want to hear about it.

Screening, not certification

Nothing on this page changes that boundary. This tool gives you a fast, transparent first read across your supplier list so you know where to look first. It feeds your own Article 10 risk assessment; it doesn't replace it, and it isn't a Due Diligence Statement.

See pricing, walk through a sample report, or read the plain explainer on what EUDR requires.