Worked cases: how the method reads real coffee and cocoa plots.

Last updated: 27 August 2026

What these cases show, and what they don’t

Any deforestation check can be called accurate. Few are shown working on the plots that are genuinely hard to get right. The cases below are real producer-scale plots, drawn from public, independently labelled ground-truth datasets and run through the same read glebora uses in a filing, with the results shown as they came out. None was invented or tuned for this page.

This is a worked case study on real ground truth, not a benchmark. The sample is a small, hand-adjudicated set, so no accuracy claim is made and no accuracy figure is published. The purpose is to show how the read behaves on the cases that matter, including where it holds back.

A note on the outcomes. Pass means the read found no deforestation signal that required escalation for that plot; it is not a statement that a shipment is compliant. Needs review is not a failure. It is the read doing its work: placing an ambiguous plot in front of a reviewer with the evidence assembled, rather than guessing. A determination and its filing remain the responsibility of the operator.

The read, in brief

A defensible read for these two crops never rests on one map. It combines a 2020 baseline (the Copernicus/JRC GFC2020 forest map, which fixes what counted as forest at the cut-off), an annual tree-cover-loss record (Hansen/UMD, which dates disturbance but cannot tell a pruned shade tree from a cleared forest), and a forest-type classification (JRC Tropical Moist Forest, which speaks to whether a disturbance was a conversion at all), corroborated where available by near-real-time radar and optical alerts (RADD and DIST). The full reasoning is set out in the guide on why shade-grown coffee and cocoa get false-flagged, and end to end on the method page. The cases below are that reasoning, applied.

The worked cases

Every figure below is a real reading pulled from a frozen record. Tree-cover loss is shown beside classified deforestation on purpose: loss is not deforestation, and seeing the two in the same row is how the difference becomes visible. Origins are given at the level the public ground-truth source allows; plot coordinates are not published.

1 · Management, not conversion

The headline case, and the one a naive tree-loss-only check gets wrong. Both plots are legally agricultural land already (the 2020 forest baseline reads 0%: a worked shade farm is not forest), so post-2020 canopy loss is husbandry, not a forest turned into a field. The type-aware read registers real loss, sees the deforestation classifier stay silent, and holds the plot for a reviewer rather than flagging it.

Both were adjudicated by a reviewer against dated high-resolution imagery: continuous shade-crop canopy across every year, no clear-cut and no new field. The Ghana plot sits in an independently surveyed cocoa-landscape reference set; the Vietnam plot is a robusta plot located from the public tree-loss layer. A tool that flagged on loss alone would have failed both, wrongly.

2 · Real deforestation, flagged

The read has teeth. On these frontier plots, forest present at the cut-off is converted after it, the classifier confirms the conversion, and radar corroborates the timing. The final row is the case the annual maps missed: they were quiet, but weekly radar had already surfaced a fresh clearing, so the read raised it for review, and the reviewer confirmed it.

The two flagged frontier plots carry independent expert deforestation labels from a public drivers-of-loss dataset; the verdicts agree with those labels. A flag is an evidence-based finding of candidate non-compliance, not a legal ruling.

3 · Clean plots, correctly passed

The counterweight to the flags: the read does not just fire on everything. A settled shade-cocoa parcel and an open coffee belt, both with no post-2020 disturbance, pass cleanly.

What the imagery shows

For the two management cases, the imagery is the deciding evidence. At the scale of a sub-hectare plot, coarse public imagery is too blunt, so the read is carried by dated high-resolution frames examined side by side. Those frames sit behind commercial imagery licences and are not reproduced here; what follows is the read of them.

Ghana, cocoa (shade-grown). Continuous shade-cocoa canopy in every dated frame. The canopy thins over the years, but there is no bare-earth clear-cut, no new crop field and no buildings. The post-2020 tree-cover loss is canopy management inside an existing farm, which is why the read held the plot for a reviewer rather than flagging it.

Vietnam, coffee (robusta). A bare, fallow field inside a plantation matrix at the cut-off; then scrubby regrowth; then a freshly replanted young row plantation. Every dated frame is agricultural land, never forest: a clear-and-replant cycle on existing farmland. The tree-cover loss is real, but it is not a forest turned into a field.

The edge cases, named

The honest part. A read like this has boundaries; concealing them would be the tell of a tool not worth trusting.

• The near-threshold plot. One Cambodian frontier plot sat right on the line the classifier uses to separate a flag from a hold, close enough that a few metres of geometry could tip it either way. The read did not resolve it on that knife-edge. What settled it was the corroborating radar, which confirmed substantial recurring post-2020 clearing in two waves, removing the doubt the borderline classifier score raised. The reviewer signed it as a flag on that combined evidence.
• A point is not a boundary. Several plots were submitted as a single coordinate with a declared area rather than a drawn polygon. That is enough to screen, but on a real filing an ambiguous point prompts a request for the operator’s actual plot boundary before anything is signed.
• Loss the classifier cannot yet type. The management cases exist because tree-cover loss is real but the deforestation classifier stays silent. That silence is informative, not conclusive: it is exactly why those plots go to a reviewer, not to an automatic pass or an automatic flag.

Honest limitations

To keep the last section honest, the same discipline applies to the page as a whole:

• This is a small, hand-adjudicated set, chosen to show behaviour on hard cases, not a representative sample. No accuracy percentage is published, because a handful of worked cases cannot support one.
• Plots were selected by input-side criteria (what the public source data shows about the land), never by the read’s own verdict. Cases the read happened to get right were not cherry-picked.
• Satellite data is a strong screen, not the whole task. It dates and characterises disturbance; it does not replace supply-chain evidence, an operator’s own knowledge, or a reviewer’s judgment.
• The legal responsibility for a Due Diligence Statement stays with the operator. glebora performs defensible due diligence and files as the operator’s Authorised Representative; it does not assume the operator’s liability, and no read is a guarantee of compliance.

Datasets and provenance

Every source below is openly documented and licensed for this use, with attribution. The dataset versions are named deliberately: current data, stated plainly.

Baseline forest

JRC GFC2020, version V3. Copernicus/JRC, free unrestricted use; attribution requested.

Tree-cover loss

Hansen Global Forest Change, 2025 v1.13 (UMD / Google / USGS / NASA). CC-BY 4.0.

Deforestation type

JRC Tropical Moist Forest, v1_2025 (DeforestationYear and DegradationYear). Free unrestricted use; attribution requested.

Radar alerts

RADD Sentinel-1 forest disturbance alerts (WUR / GFW). CC-BY 4.0; Reiche et al. 2021, doi:10.1088/1748-9326/abd0a8.

Optical alerts

OPERA DIST-ALERT (NASA / USGS, HLS Landsat and Sentinel-2). U.S. Government public domain, including modified Copernicus Sentinel data.

Ground truth

CERSGIS Reference Dataset for Land Use Change Mapping in Ghana’s Cocoa Landscape (2024 to 2025), Zenodo, CC-BY 4.0. Sims et al., Global drivers of forest loss (WRI / GFW), Zenodo, CC-BY 4.0.