How Sylva screens a plot for deforestation

Sylva does not ask you to trust a black box. Every plot verdict traces to named, public datasets and to the 31 December 2020 cutoff. Here is what the screening looks at and how it reaches a result.

The data it draws on

• WHISP (Open Foris / FAO): the authoritative screen, a convergence-of-evidence approach that combines many forest and disturbance datasets into a per-plot verdict against the cutoff.
• Sentinel-2 (Copernicus): true-colour imagery per year and an NDVI vegetation layer, for inspecting land use before and after the cutoff.
• JRC EU Forest Observatory: the Commission's reference forest map for 2020.
• Global Forest Watch / Hansen: tree-cover-loss and near-real-time alert layers, as context.

How it reaches a verdict

A plot lands in one of three states, keyed off WHISP's perennial-crop verdict and the post-cutoff signals:

• Deforestation-free: no post-cutoff loss on land that was forest at the cutoff, and a low commodity-risk verdict.
• Inconclusive: an ambiguous signal, such as a more-information-needed verdict or a post-cutoff loss intersecting forest on the parcel. The plot is not failed; it is sent for review.
• Deforestation detected: a high commodity-risk verdict or a new post-cutoff plantation. The batch cannot proceed without resolution.

What decides whether a post-cutoff loss reading counts is the 2020 forest baseline, not the size of the reading. A loss on land the EU references show as non-forest in 2020 (already agricultural, or predominantly agricultural under Art. 2(4)) is not the conversion of forest that Art. 2(3) defines as deforestation: the plot reads low, with the reading kept as a visible “noted” flag rather than hidden. Where a plot was forest in 2020 and carries an alert, an overlap test splits the parcel into hectare-sized cells through WHISP to localise whether the alert actually falls on the 2020 forest.

Two principles guide the logic: it never sits milder than WHISP, and the discriminator is the 2020 forest baseline, not raw tree-cover loss, because tree-cover loss is not by itself deforestation (Art. 2(3)). The exact operating lines are part of the product; the inputs and the philosophy are open.

Where the human takes over

An inconclusive verdict is a prompt, not a conclusion. The operator reviews it on the map, comparing a pre-2020 Sentinel-2 year with a recent one, checking NDVI, and reading the per-source figures. They can capture the inspected view and add pre-2020 land-use evidence, then record an Article 10 determination, all kept in the case file with a timestamped trail an authority can request.

GPS points and the analysis buffer

Smaller farms often deliver a single GPS point plus a declared area instead of a boundary, which the regulation permits up to 4 ha (5th FAQ, section 1.8). Screening a bare point evaluates almost no ground: in our verification run, WHISP measured 0 ha on the point alone and 2.98 ha on the buffer for the same 3-hectare plot. Sylva therefore derives a circle of the declared area around the point and screens that instead. The circle is an analysis aid, drawn dashed and labelled as derived everywhere it appears; the due diligence statement and every geolocation export carry the supplier's point, never the circle, and a plot declaring more than 4 ha on a point is flagged because a polygon is required at that size.

What Sylva will not do

It will not treat a single dataset as the last word, present tree-cover loss as deforestation, or sign off a plot the evidence does not support. Better an honest inconclusive than a confident wrong answer.