How the verdict is computed

Article 3(a) of Regulation (EU) 2023/1115 requires that a relevant product be “deforestation-free”: produced on land that has not been subject to deforestation after 2020-12-31. Article 2(3) defines deforestation as the conversion of forest to agricultural use, whether human-induced or not. Article 2(4) defines forest, following FAO, as land spanning more than 0.5 hectares with trees higher than 5 metres and canopy cover above 10%, excluding land that is predominantly under agricultural or urban land use.

Three consequences follow, and every one of them is a place where naive implementations go wrong:

• Land that was not forest at the cut-off cannot be deforested. The overwhelming majority of smallholder cocoa and coffee sits on land converted decades before 2020. Those farms are compliant, and the job of the engine is to prove it cheaply rather than to hunt for a reason to fail them.
• Not all tree-cover loss is deforestation. A rubber plantation being harvested loses tree cover on schedule. Treating every loss pixel as a breach produces a flood of false positives that a compliance team cannot defend and will eventually stop reading.
• Standing forest is not a breach. Shade-grown cocoa under retained canopy is precisely the outcome the Regulation is designed to reward.

Each source answers a question the others cannot, and using one for another’s job is the most common methodological error in this field.

All three are open data and need no API key. That is not a convenience - it is the business model. A verification that depends on a commercial imagery licence can never be given away to the people who most need it.

The JRC baseline covers the globe at 4,320,000 × 1,680,000 pixels - 37.7 GB as a single Cloud-Optimised GeoTIFF, which Plotline never downloads and never opens. The Commission publishes it as 10°×10° granules as well, and those are what the engine reads: for a two-hectare cocoa farm it fetches a few kilobytes out of the middle of the 276 MB granule covering it, over an HTTP range request, in milliseconds.

The granules rather than the mosaic, for a reason we measured rather than assumed: the single global file’s header alone takes about 96 seconds to parse, because a 37.7 GB BigTIFF carries roughly 6.9 million tile offsets. A granule parses in about 1.5 seconds, and the handle is cached, so every subsequent plot in the same region costs nothing.

This is the whole reason a verification costs a fraction of a cent, which is in turn the whole reason a smallholder cooperative can be given one for nothing. The running system counts every byte it transfers and prices it at commercial cloud egress rates - deliberately conservative, since the archives are in fact free to read - so the cost figure shown in the interface is measured rather than claimed.

A typical West African cocoa plot is 1.5 to 3 hectares. On the 30-metre Hansen grid that is between two and three pixels. Decide membership by asking “is the pixel centre inside the polygon?” and the answer for a real farm swings by ±50% depending on where the grid happens to fall. Every reported hectare figure would be noise.

So Plotline does not use a binary mask. It scan-converts the polygon at a much finer sub-grid - up to 16 samples per pixel edge, chosen adaptively so that small plots get the most care - and computes the fraction of each pixel inside the plot. Pixel areas are then integrated exactly on the sphere rather than approximated with a cosine at the pixel centre, because getting that wrong is the most common way to report a wrong hectare figure.

The engine resamples the Hansen loss year onto the JRC 10-metre grid and counts a pixel as deforestation only when both conditions hold: the JRC mapped it as forest on the cut-off date, and Hansen records it as lost after 2020-12-31.

Loss on land that was already agricultural is reported separately as context and never contributes to a breach. This single intersection is what separates a usable tool from one that flags every rubber rotation and every cocoa replant.

This was the most instructive thing the build taught us. The obvious high-resolution check is NDVI - the standard greenness index. Run it on a plot that was clear-felled in 2022 and replanted with cocoa, and by 2026 NDVI comes back higher than the rainforest it replaced. A greenness check does not merely miss that deforestation; it actively certifies it.

So the structural signal is NDMI, the normalised difference moisture index, which contrasts near-infrared with short-wave infrared and therefore responds to canopy water content and vertical structure. Closed tropical forest holds far more water in far more layers than any crop. When forest becomes a tree crop, NDMI falls hard and stays down even after the replacement greens up. A drop past -0.08 is treated as evidence of structural change.

Catalogue metadata reports cloud cover over a 110 km × 110 km tile, which in the tropics is almost always 40–80%. Rejecting those scenes would leave a Ghanaian cocoa farmer with no usable imagery for months. But a scene that is 60% cloudy is very often perfectly clear over one two-hectare plot.

Plotline therefore screens cloud at the plot, using the Sentinel-2 scene classification band over the plot footprint alone, and stacks scenes from the same MGRS tile - identical pixel grid, so no resampling error creeps in - taking the per-pixel median of clear observations. The median is deliberate: it is robust to a bright cloud edge or a dark shadow that a mean would smear across the whole plot. Stacking stops as soon as roughly 92% of the window has a clear look, so an easy plot costs one scene and only difficult plots pay for more.

Independent datasets disagree about the same plot routinely. A tool that silently prefers one has no answer when an auditor cites the other. Plotline reports each evidence line on its own terms, and where they conflict it returns needs review and says why. That third outcome is the one most tools omit, and omitting it is how they end up either blocking compliant farmers or waving through cleared land.

Confidence is a transparent product of named penalties - sample size, dataset availability, cloud, whether the boundary was walked or inferred from a point - rather than a learned score, because an operator has to be able to explain to an auditor why a particular number appeared.