DeepMind releases ForTy map: 6 forest types at 10-meter resolution

Google DeepMind and Google Research, together with the World Resources Institute (WRI) and the International Institute for Applied Systems Analysis (IIASA), have published Forest Typology 2020, or ForTy. It is a global map that divides the land surface into 10×10 meter patches and assigns each one to one of six forest and land-use categories. The model identifies the forest type correctly in 90% of cases and distinguishes forest from non-forest in 95%. Validation used more than 9000 random points. The map comes with an early time series for 2017–2025 that shows not only where forest disappeared, but also what replaced it.

Why everything looks green to a satellite

Conventional forest cover maps answer a single question: are there trees on this pixel? For climate and biodiversity, that is not enough. Primary forest that has never been logged stores noticeably more carbon and supports more species than a monoculture plantation on a ten-to-twenty-year harvest cycle. From above, though, it all blends into one green blotch.

For regulators, this is already a practical problem. The EU Regulation on Deforestation-free Products (EUDR) bans imports of coffee, cocoa or palm oil from land that was forest in 2020 and later became farmland. For timber, the requirements are stricter: primary or naturally regenerating forests cannot be converted into plantations. Companies and oversight bodies need to know more than whether trees are present. They need the forest type and what grew back after logging. A year ago, the same partners released Natural Forests of the World 2020, which gave a binary answer: natural forest or not. ForTy takes the next step and describes exactly what kind of forest it is.

Six classes and 1.7 million labels

The classes follow UN FAO definitions: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry (palm, rubber, cocoa, fruit), and "other land," which ranges from grassland and cropland to cities and water.

The difficulty is that some categories are defined by how land is used, not by how it looks. To teach the model that difference, the team collected about 1.7 million labeled points from expert surveys, national and regional datasets, and targeted checks in places where the model made mistakes. The architecture is custom and tuned for remote sensing data. The key input is AlphaEarth Foundations embeddings: compact summaries of multi-year satellite observations for each 10-meter pixel. From these, the model picks up planting patterns, growth cycles and the rhythm of harvesting.

Key parameters of the release:

• resolution: 10 meters, global coverage;
• six classes based on FAO definitions;
• training set: about 1.7 million labeled locations;
• validation: more than 9000 random points, with two experts per point;
• forest type accuracy: 90%, forest/non-forest separation: 95%;
• time series: annual, from 2017 to 2025;
• access: Google Earth Engine under an open license.

Probabilities instead of yes or no

ForTy does not output a hard label. For each class, the model returns a probability, and the user chooses the threshold. A company assessing risk under the EUDR can accept only pixels with very high confidence to avoid drawing complaints from the regulator. A researcher estimating total forest area can take the full range of probabilities and get a more complete picture.

The time series closes another gap. When natural forest is cleared and replanted as a plantation, the land quickly turns green again, and standard cover maps do not see the change. ForTy captures that substitution as well as reverse processes, such as abandoned land being overgrown. The series is still an early release: Google promises to refine it together with partners and users.

Who can use this in practice

The first users are obvious: companies with EU supply chains that need to prove where their raw materials come from, and the regulators themselves. The 2020 baseline map matches the EUDR cutoff date, so it can serve as a reference point, while the series through 2025 will show what happened to a given plot afterward. For conservation organizations and forest investment funds, the map can help show where protection is working and where natural forest is being turned into plantation. Researchers will find the probability layers useful for estimating areas.

The data can be downloaded or connected directly in Earth Engine, and there is also an interactive app for browsing at 10-meter resolution. For Google, this is another piece of its Earth AI lineup and the AI for Nature program, which already includes forest loss driver maps, a deforestation risk forecast and species distribution models. ForTy shows how a foundation model for Earth observation becomes a tool for certification and compliance.