Not one satellite. Many public datasets, read together.
The intuitive picture of deforestation screening is a single satellite photo before and after. Real EUDR screening does not work that way, and it should not. Any single layer fails in predictable, well-documented ways: shade-grown coffee and mature oil palm read as forest, replanting cycles read as loss, clouds hide clearing for months, and a low-resolution map smears a plot boundary across land it never covered.
WHISP, the FAO's open screening tool, answers the question differently. It converges the evidence: for a given plot geometry it reads a stack of independent public datasets, each measuring something different, and asks whether they agree. When tree-cover baselines, annual-loss layers and near-real-time radar alerts all point the same way, the verdict is confident. When they disagree, the plot is surfaced as "needs review" rather than forced into a yes or no. That is the same convergence-of-evidence logic the regulation's own risk-assessment thinking is built on.
The datasets we converge
These are real, public layers WHISP reads per plot. No single one is decisive; the verdict comes from how they line up. The exact set evolves as the science does, which is itself part of the point: it is maintained in the open, not frozen inside a product.
| Signal | Example public layer | What it tells the assessment |
|---|---|---|
| Forest baseline at the cutoff | JRC Global Forest Cover 2020, ESA WorldCover, JRC Tropical Moist Forest | Was the plot forest on 31 December 2020? The question every EUDR assessment starts from. |
| Annual canopy loss | Global Forest Change (Hansen / University of Maryland) | Year-by-year tree-cover loss, so clearing can be dated relative to the cutoff. |
| Radar deforestation alerts | RADD alerts (Sentinel-1) | Near-real-time clearing that sees through cloud, which optical layers miss for months. |
| Primary and intact forest | GLAD primary humid tropical forest, Intact Forest Landscapes | Whether the plot overlaps the forest types the regulation protects most strictly. |
| Commodity and plantation maps | Oil-palm, cocoa, coffee, rubber and soy layers | Tells mature plantation apart from natural forest, so a legal crop is not read as loss. |
| National datasets (opt-in) | PRODES / DETER (Brazil), IDEAM (Colombia) | Producer-country monitoring layered in where the plot country provides one. |
Because every layer is public and named, a screening run is not a score you have to take on trust. It is a set of citable observations, each traceable to a dataset an auditor can look up independently.
Everything tests against 31 December 2020
EUDR defines "deforestation-free" against a fixed date: land not subject to deforestation after 31 December 2020 (and, for wood, no forest degradation after it). Screening is built around that date. The loss and alert layers are read for the post-cutoff window specifically, so the operative question per plot is precise: is there evidence of forest conversion inside this boundary after the cutoff? Land cleared before the date is not in scope; the date, not the regulation's start, is what the satellites are asked about.
Why the institutional reference tool is more defensible
A due diligence file has to survive scrutiny for five years. When a competent authority, a buyer's auditor, or an NGO filing a substantiated concern asks how you concluded a plot was deforestation-free, there is a real difference between two answers:
- "A vendor's proprietary model scored it low risk." The method is a trade secret. It cannot be independently reproduced, its datasets and thresholds are not disclosed, and the reviewer has to trust the black box.
- "The FAO's open WHISP method converged these named public datasets and found no post-2020 loss, on this date." The method is documented, the code is public, the datasets are the same ones institutions use, and anyone can re-run it.
Both can reach the same verdict. Only one is transparent enough to defend line by line. We deliberately build on the public reference implementation rather than a private model, because in a compliance file, reproducibility is worth more than a proprietary edge.
What you get back
For each plot you upload, screening returns a verdict, the indicators behind it, and the dataset names and run date, so the run itself becomes evidence in your file. There are three outcomes, and "needs review" is a feature, not a failure: