Keep forest types visible in parcel summaries
New forest maps distinguish land types that a single parcel label can hide. Preserve the class breakdown, smaller patches and unresolved coverage when turning map outputs into a sourcing review.
Keep development and test groups separate before comparing performance. No measured results are shown.
A forest map becomes useful to a sourcing team when it helps decide which part of a supplier’s land needs closer examination. Turning every parcel into its most common class can discard precisely the smaller area that prompted that question.
Google DeepMind’s 22 October announcement of Forest Typology, or ForTy, introduces a global 2020 baseline at 10-metre resolution, separating forest types, tree crops and other land. The opportunity is a more discriminating first inspection than a simple tree-cover overlay. For a data owner, the corresponding design decision is how those distinctions survive the summary presented to a reviewer.
Preserve the parts that change the investigation
Consider a hypothetical sourcing review. A supplier submits a boundary containing a large cultivated area and a smaller strip beside a watercourse. The map assigns different classes within that boundary. A dashboard selecting the most common class displays only the cultivated area’s label.
That summary answers which class dominates the mapped parcel. It does not answer whether a potentially relevant forest area occurs inside it. An investigator could need the location of that strip even when it contributes little to the total area. Conversely, a small cluster of questionable pixels should not automatically determine the status of the entire supplier.
The proposed output should therefore retain a spatial breakdown, the assignment rule and a route back to the underlying map. Let the reviewer distinguish a coherent patch, scattered uncertain assignments and pixels intersecting the supplied boundary. Those patterns suggest different next steps: inspect land-use history, obtain a more reliable boundary or ask for specialist interpretation.
This is a proposal for investigation support. It is not a finding about a real supplier, an automated exclusion rule or an executed ForTy experiment.
Make the aggregation rule visible
The research paper, published on 21 October, describes a cascaded learning system using satellite embeddings and refined training labels. Its distributed class scores support different classification rules. The authors also warn that the scores are not calibrated probabilities, that mixed boundary pixels remain difficult, and that management history is only indirectly visible to the model.
For the proposed parcel service, separate the map’s output from the organisation’s summary rule. Store the source release and reference year, then record which rule assigns each pixel to a class. Define how the calculation handles pixels cut by the parcel edge and how it measures area. A reproducible map overlay requires these choices even when the user sees only a concise table.
The following is a proposed report structure. It contains no measured areas or suggested universal threshold.
| Parcel summary field | What the reviewer can inspect |
|---|---|
| Mapped area by retained class | Smaller classes that a dominant label would hide |
| Locations of relevant patches | Whether the signal is concentrated or scattered |
| Unassigned or excluded area | Which parts were omitted by the chosen rule |
| Boundary and assignment method | Whether a change in geometry or threshold changed the result |
| Reference year and later evidence | Whether the question concerns the mapped period or a subsequent change |
Keep excluded area visible in the total. If the service reports shares only among pixels it accepted, a parcel with extensive unresolved coverage could look reassuringly uniform. Label that denominator explicitly and provide the whole-parcel coverage beside it. Do not silently treat an unassigned pixel as evidence that the relevant class is absent.
Carry the smaller patch into the review
Swipe or scroll for the full diagram →
A parcel summary should preserve the areas that can change the next investigation.
- Observed
- Hypothetical input: the supplied overlay contains a dominant class and a smaller differently classified strip.
- Inference
- Proposed action: show the class breakdown and strip location so a reviewer can request the relevant history.
- Evidence boundary
- The strip’s actual land use, later changes and supplier acceptability remain unresolved.
A constructed reporting example, not a ForTy result, a measured parcel or a regulatory decision.
A fictional parcel contains a dominant mapped class and a smaller differently classified strip. Preserve both locations for investigation. The summary does not decide supplier acceptance.
The figure separates the hypothetical input, the proposed investigation and the unresolved decision. Its purpose is to show what a parcel summary should preserve, not to suggest that a particular mapping rule has been validated.
Match the summary to the question
A reviewer asking whether a parcel contains any area worth investigating needs a different summary from someone estimating the predominant land type across a portfolio. Agree which question the service answers before choosing the aggregation rule. A convenient database field called “forest status” would conceal that distinction.
For an initial sourcing screen, show the retained classes and their locations. For a portfolio report, document how parcel results combine and avoid implying that a map-derived total is a corrected estimate of actual land area. The paper distinguishes pixel-counted areas from estimates corrected using validation samples, and identifies class-dependent errors. (Area estimates and limitations)
The 2020 reference period also needs a visible place in the report. A review of later activity requires additional dated evidence. DeepMind’s announcement separately describes an early annual series for 2017–2025. That does not make the baseline itself a current observation, and a service should identify which product supplied each result. (Release announcement)
This article makes no conclusion about an organisation’s legal obligations. The proposed summary supports deciding where further evidence is needed. A map class alone does not settle supplier acceptance.
Test the information lost at the summary step
The earlier geospatial-agent article asks which tools and orchestration improve an analysis. Here the decision concerns the transformation after a map is available: which information reaches the person reviewing a parcel.
Before commissioning a production integration, ask the delivery team for a small set of constructed parcel reports. Include a dominant class with a smaller relevant patch, a boundary crossing a mixed area, extensive unassigned coverage and a parcel whose later records conflict with the historical map. These fixtures test the reporting contract. They do not measure classification accuracy.
Have a domain reviewer identify the next investigation from both the full overlay and the proposed summary. Record when the summary hides a location or implies more certainty than the overlay supports. Separately assess classification on suitable local reference evidence before using the service to change sourcing decisions.
The data owner should approve the summary only when it preserves the distinctions needed for that investigation. Retain a dominant label where it is useful for navigation, alongside the class breakdown and unresolved coverage. If the smaller patch changes the next action, it belongs in the report.