Satellite maps usually show whether trees are present in an area, but not what kind of forest or crop it is. Google DeepMind and the World Resources Institute built ForTy, short for Forest Typology, a global map that sorts the world's land in 2020 at 10-metre resolution into six classes: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops such as oil palm, rubber or cocoa, and other land. The researchers say it is the first global map of its kind, since earlier datasets, including the European Commission's, cover only parts of the world.
The practical driver is the EU deforestation regulation (EUDR), which bars goods linked to forest cleared or degraded after 31 December 2020, including goods made after primary forest was turned into a plantation. Oil palm and cocoa farms count as farmland, so companies assessing supply chains need to know how a site was used in 2020. ForTy provides that reference point and could also feed other efforts, such as the Tropical Forest Forever Facility and the UN goal of protecting 30% of land by 2030.
The core difficulty, DeepMind explains, is that a forest's type depends on its history rather than its appearance. Google DeepMind's global mapping ambitions have run into map-accuracy questions before, as UN calls for Equal Earth projections in September kept Google Maps on Mercator. An old mixed plantation can look like natural forest from orbit, while a plantation cleared last month still counts as forest by definition even though satellite images show bare ground.
To tell the classes apart, the system relies on AlphaEarth Foundations, which condenses data from several satellites into a compact description of each pixel. Several models work in sequence on top of it: the first learns from verified examples; its predictions, combined with data on tree height, buildings and forest loss, become training material for a second model; a final step analyses several years together to smooth out isolated errors. Gemini helped adapt older training data to EU definitions.
During correction rounds experts added about 43,000 examples to fix errors such as European vineyards labelled as tree crops, natural pine forests in Central America tagged as planted forest, and oil palms missed in Africa and Indonesia. Some distinctions remain out of reach for satellites, which rarely show how land is managed. The researchers separate plantations from other planted forests using a 40-year rotation threshold, which does not suit every region.
In testing, two experts reviewed each of 8,190 randomly chosen sites and found the classification correct in 90.2% of cases. That overall figure is inflated by the easy 'other land' class. By type the picture changes: primary forests confirmed on the ground were recognised 89% of the time, while tree crops and planted forests were picked up only about half the time and plantations 60% of the time. Where the map flags plantations or tree crops, a quarter to a third of those areas are wrong, and for planted forest the error rate is one in three.
Despite these limits ForTy beats the European Commission maps built for the EUDR on the same test data: 95.5% against 93.6% for forest versus non-forest, and 90.9% against 87.8% for forest types. It also inherits a flaw from its training data, which rarely labelled forests near roads or settlements as primary, even when they looked untouched.
The researchers estimate the world holds about 4.1 billion hectares of forest, nearly 3.9 billion of it natural and about 2 billion primary. Planted forests cover roughly 250 million hectares and tree crops nearly 200 million, mostly in Asia. Regional errors are significant: the map overstates Asian primary forest by about 200 million hectares, largely in Russian conifers, and misses about 94 million hectares in Africa, probably dry open woodland and savanna.
At country level Sweden stands out: about two-thirds of its forest area is planted, and natural forest covers only about a fifth of the country. The authors say overall tree cover may be adequate, but the share of natural forest, which matters for biodiversity, could be critically low. The highest shares of natural forest are in the Democratic Republic of the Congo, Indonesia and Brazil.
Hard cases remain. In some areas cocoa or coffee grown under natural canopy is still labelled forest, and selectively logged or burned forests pose problems too. There is also no separate class for open woodland, which the EUDR treats apart, though the researchers plan to add one. Users can set their own thresholds using the confidence scores attached to each class. For Italian forests, the startup Ruma launched EcoSense in August, digital twins certifying CO2 absorption through the new national registry.
No comments yet. Be the first!