Google ForTy Maps Forest Types Beyond Tree Cover

Short answer

Google's ForTy announcement gives conservation teams a more specific question than whether land is covered by trees: what kind of forest is there? The practical opportunity is better screening of places that need investigation. A map classification alone should not settle a consequential land-use decision.

What Google announced on October 22

In its October 22 announcement, Google DeepMind describes Forest Typology 2020, developed with Google Research, WRI and IIASA. It distinguishes six land-use classes at 10-meter resolution and accompanies that baseline with an early annual series for 2017–2025.

The release distinguishes naturally occurring forests, planted forests, plantations, tree crops and other land. Google reports evaluation against more than 9,000 sites, with 90% forest-type accuracy and 95% forest/non-forest accuracy. Those are reported aggregate results; DiscoverAI has not reproduced them. The announcement links to open data and Earth Engine access.

Why the distinction changes the research question

Our editorial interpretation is that a nonprofit could use this type of map to prioritize where to request local evidence. For example, a conservation officer might compare a proposed project boundary with a historical baseline and flag locations where a broad “green area” description hides a change in land use. This is a proposed application, not a verified deployment outcome.

Keep the research question narrow. “Which locations deserve a closer look?” is a reasonable pilot question. “Which supplier is responsible for an unlawful conversion?” requires additional evidence and a separate decision process. Avoid letting the apparent precision of a map substitute for proof of the underlying claim.

What a local pilot should measure

Choose a small area for which your team already has reliable records. Before opening model predictions, document the known land history, boundary uncertainties and the decisions the pilot is intended to support. Ask someone familiar with the location to identify where the available records are weakest.

Compare the mapped classifications with those records. Log disagreements, mixed areas and boundary cases rather than averaging them into one reassuring number. Record the time needed to resolve a disputed location. A useful screening layer should improve the order of investigation without creating more unreviewable alerts than the team can handle.

Use separate columns for what the map suggests, what a local source confirms and what remains unresolved. That prevents a tentative model label from turning into an unqualified statement in a donor report or supplier email.

What not to infer from this release

An aggregate accuracy figure is not a confidence guarantee for an individual property. A historical dataset is also not a live observation of today's conditions. These are general evaluation cautions, not a finding that ForTy failed a particular site.

The operational decision belongs with people who understand the local evidence and the consequences of a mistake. Escalate disputed cases rather than assigning an automatic pass or failure. Readers seeking regulatory advice should consult the relevant current rules separately; this article evaluates a research release.

A useful next step for nonprofits

Start by deciding what fieldwork or documentary review a map would help prioritize. Then assign an owner to check a small sample and write an explicit stop rule for unresolved disagreements. Expand only if the pilot reduces the time to a defensible research decision.

For a broader starting stack, see our Top 5 AI tools for nonprofits. Use the Decision Workspace to record an evaluation plan and evidence gaps. This is an October 25 analysis of an October 22 announcement, not a new launch today.