This method has not been reviewed by an outside body or tested for accuracy against independent reference data, so no accuracy figure exists yet. This page says so wherever it matters. The EU does not approve or accredit satellite tools or methods.
Status and scope
VoxTrace reports what named, public satellite datasets record on a submitted farm plot: forest at the EUDR cut-off of 31 December 2020, and tree-cover loss since. Each report is evidence for the responsible operator’s due diligence file. It is not a compliance determination, a due diligence statement, a certificate, or evidence of legality or traceability; operators and EU competent authorities make compliance determinations.
This page describes method version open_source_v1.2 with report presentation plot-evidence/1.1, as run by ATI’s self-hosted engine, with tree-cover loss observed through 31 December 2025. A Global Forest Watch Pro adapter is built but switched off. The two engines are not interchangeable, so switching between them is a methodology change handled under change control.
The method suits plot-level evidence for crop commodities such as cocoa, coffee, oil palm, rubber and soya. It does not cover wood’s forest-degradation condition or the multi-site history required for cattle. It is product documentation, not legal advice.
Rule by rule
VoxTrace measures two things the Regulation turns on: forest on the plot in 2020, and tree-cover loss since. Everything else in an operator’s due diligence sits outside the report. The table says where each answer stops. Legal references follow the consolidated text of 26 December 2025.
| EUDR requirement | What VoxTrace measures | Where it stops |
|---|---|---|
| Deforestation-free: no conversion of forest to agricultural use after 31 December 2020 (Articles 2(3), 2(13) and 3) | Hectares of 2020 forest in the plot (JRC), and hectares and years of tree-cover loss from 2021 to 2025 (Hansen). JRC Tropical Moist Forest is shown beside them as an independent cross-check through 2025. | Loss is not conversion: harvests, fire, storms and roads register too. Cause and later land use are not established, so the report says “change indicators”, never “deforestation”. |
| Forest: over 0.5 ha, trees over 5 m, canopy over 10%, not mainly agricultural or urban (Article 2(4)); agroforestry with crops under trees is an agricultural plantation, not forest (Article 2(6)) | Hectares JRC maps as forest in 2020 on a ~10 m grid. JRC leaves tree crops out, in line with Article 2(6). The cocoa and natural-forest cross-checks show what else was on the plot in 2020. | A map approximates the legal definition: tree height and in-situ potential are not measured plot by plot, and a map can miss real forest. A zero baseline with loss is therefore never reported as clean. |
| Geolocation of all production plots (Article 9); a polygon for non-cattle plots over 4 ha (Article 2(28)) | A completed report needs the actual plot polygon, whatever its size. One result per plot, up to four plots per request. Requests without coordinates are accepted for field verification. | A point cannot give hectares, and no circle or buffer is substituted for the real boundary. The report cannot show that every contributing plot was submitted, or link a plot to a shipment. |
| Produced under the relevant law of the country of production (Article 3) | Not measured. The live engine does not check protected areas. | Land rights, permits, labour and tax evidence stay with the operator. |
| Wood: the forest-degradation condition (Articles 2(7) and 2(13)(b)) | Not assessed. The TMF cross-check prints any degradation it records, for context only. | Out of scope for wood. |
| Country classification (Implementing Regulation (EU) 2025/1093) | Printed separately on the records page, never blended into the plot result. | A country category says nothing about one plot. Low risk is not proof that a plot is deforestation-free. |
| Risk assessment, mitigation and the due diligence statement (Articles 10 and 11) | One documented input to the operator’s assessment. | The conclusion and the statement are the operator’s. |
| Records kept for five years | Each report carries an ID, analysis time, dataset versions and the observation-through date, and complete reports a signed verification code. | ATI does not store reports durably. The operator keeps the file. |
Definitions the method relies on. Checked against the consolidated text on 7 October 2026. The cut-off date is in Article 2(13)(a).
- Deforestation (Article 2(3))
- The conversion of forest to agricultural use, whether human-induced or not.
- Forest (Article 2(4))
- Land of more than 0.5 hectares with trees higher than 5 metres and canopy cover of more than 10%, or trees able to reach those thresholds in situ, excluding land predominantly under agricultural or urban use.
- Agricultural plantation (Article 2(6))
- Includes agroforestry systems where crops are grown under tree cover, and all plantations of relevant commodities other than wood. Agricultural plantations are excluded from “forest”.
- Forest degradation (Article 2(7))
- Structural change to forest cover: primary or naturally regenerating forest converted into plantation forest or other wooded land, or primary forest into planted forest.
- Deforestation-free (Article 2(13))
- Produced on land not subject to deforestation after 31 December 2020; for wood, also harvested without inducing forest degradation after that date.
Data sources
Three locked datasets decide every outcome. Three more are measured beside them as independent cross-checks. Each is pinned to one version, and every report prints the version, publisher and access date of each dataset it read.
| Dataset | Publisher and version | What it records | Role in VoxTrace |
|---|---|---|---|
| Hansen Global Forest Change | University of Maryland (Hansen et al., 2013), GFC-2025-v1.13 | Year of tree-cover loss, 2001–2025, and a land/water data mask; ~30 m grid | Decides the outcome: loss inside the plot for 2021–2025, by year. The mask confirms full data coverage. |
| JRC Global Forest Cover 2020 | European Commission Joint Research Centre (Bourgoin et al., 2025), GFC2020-V3 | Forest or not forest in 2020, produced to support EUDR; ~10 m grid | Decides the outcome: hectares of 2020 forest inside the plot, the baseline. |
| SBTN Natural Lands Map | Science Based Targets Network (Mazur et al., 2025), v1.1 | Natural or non-natural land, 2020; ~30 m grid | Audit record only. It decides no outcome and no area figure from it is printed: its authors advise against using the map to quantify areas. |
| JRC Tropical Moist Forest | European Commission Joint Research Centre (Vancutsem et al., 2021), v1_2025 | Forest state in 2020; first year of deforestation or degradation, through 2025; ~30 m grid | Cross-check. Its post-2020 deforestation also counts as a change indicator when Hansen records none (see Outcomes). |
| Cocoa probability | Google for the Forest Data Partnership, model 2026a | Probability that a pixel was cocoa in 2020; ~10 m grid | Cross-check only: hectares with a probability of 0.90 or more. That is the threshold the JRC applied to the earlier 2025a version of this model when it removed cocoa from its 2020 forest map (GFC2020 V3); VoxTrace uses the newer 2026a version. The publisher’s balanced threshold is 0.54. |
| Natural Forests of the World | Google (Neumann et al., 2025), v1.0 | Probability of natural forest in 2020; ~10 m grid | Cross-check only: hectares with a probability of 0.52 or more. |
Cross-checks are shown only when read. A cross-check layer appears in a report only once it has been loaded for the plot’s area; a layer not read is left out, never shown as zero. The report’s dataset register lists exactly the layers used. The cocoa and natural-forest maps are published only through Google Earth Engine, whose commercial use is paid; ATI will serve them from its own server, like the others, once it has the data. Until then reports leave them out.
Not used. Hansen’s 2000 canopy layer is held but does not enter the calculation. The World Database on Protected Areas needs a commercial licence, so no report states protected-area overlap. Sentinel-2 imagery appears only as supporting pictures.
Coverage. Tiles are held for latitude 0–10°N and longitude 10°W–10°E. That covers the cocoa-growing areas of Côte d’Ivoire, Ghana and Togo and of southern Nigeria. A plot outside that area cannot be analysed.
How a plot is analysed
Each plot goes through the same steps on ATI’s server, and the same input always gives the same numbers.
- Intake. The boundary is drawn on a map or imported as GeoJSON (Polygon or MultiPolygon, no interior holes, up to 2 MB): up to four plots per request and 100 vertices each. A request without coordinates goes to field verification instead.
- Geometry checks. The shape must be valid, use WGS84 coordinates and cover no more than 50 ha. Area is computed after projecting the plot into its own UTM zone.
- Tiles. The engine finds the dataset tiles under the plot. A plot crossing longitude 0°, as in eastern Ghana, is split per tile and the counts are summed exactly.
- Pixel rule. A pixel counts when its centre lies inside the boundary; partial pixels are not counted. Pixel area comes from each dataset’s native grid at the plot’s latitude.
- Baseline. JRC 2020 forest pixels × pixel area gives hectares of 2020 forest.
- Change. Hansen pixels with a loss year from 2021 to 2025 × pixel area gives hectares of loss, by year and in total. Every such pixel counts; no canopy-density threshold is applied.
- Coverage. Hansen’s data mask must show every pixel in the plot as valid, or the report is not marked complete.
- Natural-land record. SBTN’s share of loss pixels on natural land and, where JRC shows no forest, its natural pixels across the plot are recorded for audit. Neither decides an outcome.
- Cross-checks. TMF gives 2020 forest (its undisturbed and degraded classes) and hectares first deforested or degraded in each year from 2021 to 2025. The cocoa and natural-forest layers give hectares at or above their thresholds, computed from the share of each ~10 m pixel above threshold within each ~30 m cell.
- Record. The outcome rules run. The result stores dataset versions, tiles read, coverage, pixel areas, the analysis time and the method version.
A report is marked complete only when the polygon is valid, all three locked dataset versions are verified, coverage is complete and every dataset’s source details are filled in. Anything less is shown as needing more evidence, not as a result.
Outcomes
The report gives one of four plain outcomes, never a grade, score or colour.
| Measured inside the plot | What the report says |
|---|---|
| No tree-cover loss from 2021 to 2025, and no TMF deforestation after 2020 where TMF was read | No post-2020 forest-change indicators detected |
| Any measured loss, with 2020 forest mapped in the plot, however small the loss | Tree-cover change indicators detected, with the measured value |
| No Hansen loss, but TMF records deforestation after 2020 | Tree-cover change indicators detected, naming TMF and the years |
| Loss measured where JRC maps no 2020 forest | Additional evidence required: a zero baseline proves neither bare ground nor conversion before 2020 |
| No polygon, incomplete coverage, or missing dataset details | Additional evidence required |
| Invalid geometry, a plot outside coverage, or an engine error | Analysis did not complete |
No tolerance. Any deforestation or forest degradation on a plot automatically disqualifies the commodities and products from that plot. (Article 9(1)(d).) VoxTrace therefore hides no measured loss. The engine also assigns an internal review level from fixed thresholds (loss under 0.1 ha, or under 0.5 ha and under 5% of the 2020 forest; over 0.5 ha or 5%; over 20%, or over 1 ha in a single year). These are ATI’s operating rules for review, not legal tolerances, and they do not change the outcome shown. They have not yet been calibrated against reference data.
The TMF rule (new in v1.1). When Hansen records no loss but TMF records deforestation after 2020 inside the plot, the report says change indicators were detected and names TMF and the years. A named dataset recording post-2020 deforestation should not sit under a “no change” headline. The rule only ever adds caution: TMF never clears a plot that Hansen flags.
Not applied to customers. The engine also computes an SBTN rule that would treat loss mostly on non-natural land as less serious. It is kept for audit but changes no outcome, because there is not yet enough evidence to calibrate it.
Supporting imagery
Reports can carry up to two Sentinel-2 Level-2A pictures of the plot from the Copernicus Data Space Ecosystem, for human context. They never change the measured result.
- The baseline picture is searched only within 2020, so no post-cut-off pixels can appear in it. The current picture is searched over the 90 days before the report.
- Candidate dates are screened with Sentinel’s scene classification, and one date is chosen per picture. That date and the share of the plot that was clear are printed under it.
- Cloud, shadow, uncertain and missing pixels are greyed out. If no date is clear enough, no picture is forced: the report shows a labelled “context image unavailable” panel, or leaves the pictures out.
- Each report credits the imagery as “Contains modified Copernicus Sentinel data”, with the acquisition years.
Known limitations
The biggest limit for West African cocoa is that the EU’s own 2020 forest map does not count cocoa agroforestry as forest. Each limit is handled by reporting less, not by guessing more.
| Limitation | Effect | How it is handled |
|---|---|---|
| JRC excludes tree crops from “forest” | On 15 of 15 measured Ghanaian cocoa farms, JRC showed 0 ha of 2020 forest where Hansen implied 0.6–35.8 ha of standing canopy. | A zero baseline with loss gives “additional evidence required”, never a clean result. The cocoa and natural-forest cross-checks show what else was there in 2020. |
| Tree-cover loss is not deforestation | Harvests, fire, storms, roads and cocoa replanting all register as loss. | The report says “change indicators” and asks for dated land-use records and field evidence. |
| Pixel size against smallholder plots | A ~30 m pixel is about 0.077 ha at 7°N, so a 2 ha plot holds about 26 of them; edge pixels can belong to a neighbour. | A pixel counts only when its centre is inside the boundary, and every measured value is shown. |
| Loss data ends in 2025 | Hansen loss after 2025 is not yet visible. | The observation-through date is printed. The recent Sentinel-2 picture is visual context only. |
| Loss counted at any canopy density | Loss in sparse tree cover counts too, which is more cautious than the common 30% canopy convention. | Stated here and in every report’s method note. |
| Geographic coverage | Plots outside 0–10°N and 10°W–10°E cannot be analysed, including northern Ghana above 10°N. | “Analysis did not complete”, never a partial result. |
| No protected-area check | Overlap with protected areas is not reported. | Left out of every report until a licensed dataset is in place. |
| The datasets disagree | Hansen, JRC and TMF use different definitions and sensors, and differ on some real plots. | Disagreement is shown, not resolved: each layer’s figure is printed beside the others. |
| No measured accuracy | ATI cannot yet say how often a result is right. | Said plainly here and under Evidence to date. |
Evidence to date
The method has been tested for internal consistency and measured on a small set of real farms. It has not been tested for accuracy against independent reference data.
- JRC on Ghanaian cocoa (July 2026). On 15 of 15 real farm polygons, JRC recorded 0 ha of 2020 forest while Hansen implied 0.6–35.8 ha of standing canopy. None of those zeros was corroborated as bare land, which is why a zero baseline with loss is never reported as clean.
- The SBTN refinement. On 18 polygons, the SBTN share of loss rests on a median of about 13 pixels per farm: too few to calibrate, so the refinement stays off.
- A public reference set was tried and rejected. A Ghana cocoa polygon set of 21,031 plots had a median plot of 0.244 ha, against about 11.4 ha for the farms ATI analyses. Sampling it would have biased the results, so the study stopped at its quality gate.
- Consistency tests. Eleven reference polygons with known expected outcomes are written into the test suite, and a parity check proves the two places the decision rules run agree. These tests show the rules are applied as written; they do not show the rules are right.
- JRC’s new V4 forest map (October 2026). Around ATI’s reference plots, V4 agrees with V3 on about 97% of forest pixels in forest reserves but only 45–85% in farming areas, where it maps up to a quarter less forest. ATI keeps V3, whose accuracy is published, until V4’s accuracy is.
- Comparison with FAO’s Whisp (October 2026). On 216 public cocoa plots from the CERSGIS Ghana reference dataset, VoxTrace’s own measurements match those Whisp computes on Google Earth Engine: JRC 2020 forest within 0.1 ha on every plot, Hansen loss since 2020 on 98% of plots and TMF deforestation on every plot, with near-identical totals. The outcomes differ by design. Whisp rated every plot low risk, mostly because its 2020 cocoa map shows the land as cocoa already; VoxTrace, which has no cocoa map in its method yet, reported no change on 133 plots, change on 33 and asked for evidence on 50. The run also found and fixed a bug that had marked zero-loss reports incomplete. Agreement between two tools is not accuracy against the ground.
An accuracy study needs a few hundred real farm boundaries, shared with consent. Until then, this page makes no accuracy claim.
Report integrity
Every complete report carries a verification code: a record of its key figures signed by ATI. Anyone can check it at ati-platform.com/verify, in their own browser, and compare the signed figures with the PDF. Request records and incomplete analyses carry no code, because there is no measurement to vouch for.
Change control
The thresholds, datasets and decision rules are locked. Any change that could alter a customer’s outcome is a methodology change with its own approval, and every report records the versions it used, so an older report can always be traced to the method that produced it.
What counts as a methodology change: any change to a threshold, a dataset or its version, the pixel rule, the decision rules, or the engine in use.
What a change requires: evidence that it is right, such as reference polygons with known ground truth for the case it affects; a review by ATI’s founder of every customer outcome it would shift; identical changes in both places the decision rules run, proven by the parity tests; and a new method version, recorded in ATI’s decision log. Tests pin known divergences on purpose, so an accidental methodology change fails the build.
Version history.
- open_source_v1.2 (current). Hansen GFC-2025-v1.13, observation through 31 December 2025. On ATI’s seven reference plots it records exactly the same 2021–2024 loss as GFC-2024-v1.12 and adds 2025.
- open_source_v1.1 (October 2026). Adds the TMF, cocoa and natural-forest cross-checks, and the TMF rule. Corrects the report’s method note, which had described a 30% canopy threshold the engine never applied.
- open_source_v1 (July 2026). The self-hosted engine on Hansen, JRC and SBTN, with the locked outcome rules.
Next steps
The most valuable steps are an accuracy study and an outside review; both depend on farm boundaries shared with consent.
| Step | What it shows a reviewer | Depends on |
|---|---|---|
| Repeat the Whisp comparison on whole farms shared with consent | Agreement on the plots customers actually submit, not only public samples | Owners’ consent |
| Obtain the cocoa and natural-forest layers | The commodity and natural-forest cross-checks in every report | A data extract from the publishers; Earth Engine export is paid for commercial use |
| Adopt JRC GFC2020 V4 once its technical report is published | The EU’s latest forest map, with documented accuracy | The JRC report, due by December 2026 |
| Run an accuracy study on a few hundred independently labelled plots | A stated accuracy figure, the standard test for map-based results | Consented farm boundaries and a partner |
| Commission an outside review of this statement | The closest thing to an audit opinion | The steps above, so there is something to review |
| Add tiles north of 10°N if customers need them | Coverage of northern Ghana and the wider Sahel | Customer demand |
| License a protected-area dataset | Protected-area overlap can be reported | Budget |
Data licences
Every dataset is used under its publisher’s terms, and each report’s dataset register names the publisher and version it read.
| Dataset | Licence or terms | Attribution |
|---|---|---|
| Hansen Global Forest Change | Creative Commons Attribution 4.0 | Source: Hansen/UMD/Google/USGS/NASA |
| JRC Global Forest Cover 2020 | Free use by anyone without permission or royalty; citation requested | Bourgoin et al., European Commission, Joint Research Centre |
| JRC Tropical Moist Forest | Free of charge, without restriction of use; acknowledgement and citation required | Vancutsem et al. (2021), European Commission, Joint Research Centre |
| SBTN Natural Lands Map | Creative Commons Attribution-ShareAlike 4.0 | Science Based Targets Network. ATI does not redistribute the map or maps made from it; reports carry only per-plot figures |
| Cocoa probability | Creative Commons Attribution 4.0 | Produced by Google for the Forest Data Partnership |
| Natural Forests of the World | Creative Commons Attribution 4.0 | This dataset is produced by Google |
| Sentinel-2 imagery | Copernicus free, full and open data policy | Contains modified Copernicus Sentinel data, printed under each picture |
Sources
Law and guidance
- Regulation (EU) 2023/1115, consolidated text of 26 December 2025
- Commission Notice C/2026/3896, EUDR guidance
- EUDR country classification list and Implementing Regulation (EU) 2025/1093
Datasets
- Hansen, M. C., et al. (2013). High-resolution global maps of 21st-century forest cover change. Science 342: 850–853. Data version GFC-2025-v1.13.
- Bourgoin et al. (2025). GFC2020: a global map of forest land use for year 2020. Earth System Science Data. Data version GFC2020-V3.
- Bourgoin, C., et al. (2026). Maps of Global Forest Cover 2020 Version 3 and Global Forest Type 2020 Version 1 supporting the EU Deforestation Regulation. JRC146622, Publications Office of the European Union. Lists the commodity maps and thresholds behind GFC2020 V3.
- Mazur et al. (2025). SBTN Natural Lands Map v1.1. Science Based Targets Network.
- Vancutsem, C., et al. (2021). Long-term (1990–2019) monitoring of forest cover changes in the humid tropics. Science Advances 7: eabe1603. Data version v1_2025.
- Cocoa probability model 2026a, produced by Google for the Forest Data Partnership, CC BY 4.0.
- Neumann, M., et al. (2025). Natural forests of the world: a 2020 baseline for deforestation and degradation monitoring. Scientific Data. Produced by Google, CC BY 4.0.
- Sentinel-2 Level-2A imagery, Copernicus Data Space Ecosystem.
Tools referred to
- Whisp, FAO Open Foris.