Google DeepMind's ForTy Maps Every Forest Type on Earth at 10-Meter Detail

Google DeepMind, Google Research, the World Resources Institute and IIASA have released ForTy, a global dataset that classifies land at a nominal 10-meter resolution into six forest and land-use types. It includes a 2020 reference map and annual layers from 2017 through 2025, with access through Google Earth Engine and downloadable grids.

The distinction between forest types affects carbon accounting, biodiversity research and supply-chain audits. A primary rainforest and an intensively managed plantation may both register as tree cover, even though their ecological value, management history and regulatory treatment differ substantially.

Why canopy is an incomplete signal

Conventional tree-cover products often assign the same green signal to forests with different origins and uses. A naturally regenerating forest can store more carbon and support more species than a single-species plantation harvested on a short rotation. Tree cover can also return quickly after clearing, obscuring a conversion when maps track canopy presence alone.

The European Union Deforestation Regulation relies on a December 31, 2020 cutoff for cattle, cocoa, coffee, oil palm, rubber, soy, wood and many derived products. Agricultural goods must be free from deforestation after that date, while wood products are also subject to forest-degradation requirements. ForTy can help analysts identify the type of forest present at the cutoff and screen for later conversion to agriculture or plantation forestry.

Six labels behind the canopy

The taxonomy follows definitions from the UN Food and Agriculture Organization, giving each pixel one of the following classes:

How pixels pick up land use

Forest categories encode appearance, management and history, which makes them harder to classify than basic tree cover. Planting rows, synchronized growth, harvest cycles and recovery patterns can reveal management, while a single satellite image may show only a continuous canopy.

The model uses AlphaEarth embeddings as its main input. These compact numerical vectors summarize satellite observations for each location over time. They preserve signals associated with vegetation structure, seasonality, disturbance and regrowth, giving the classifier a multiyear fingerprint for each 10-meter pixel.

The training set contains about 1.7 million labeled locations assembled from expert surveys, national and regional datasets, and targeted reviews of areas where earlier model versions struggled. That mix helps the model handle forest types whose appearance varies across climates and regions.

Global accuracy needs local checks

The published validation used more than 9,000 randomly selected sites worldwide. Two trained experts reviewed each site using high-resolution imagery and historical satellite records. The study reports about 90% accuracy for forest-type classification and 95% accuracy for distinguishing forest from non-forest, exceeding the comparison products evaluated in the paper.

ForTy also supplies a probability score for each class. A strict decision threshold produces fewer, higher-confidence flags, while a lower threshold captures more candidates for review. Researchers can retain the complete probability distribution when uncertainty matters to an analysis.

Global accuracy figures can conceal variation among regions, classes and landscapes, so production workflows should account for several limitations:

• Local performance: Results should be checked against regional reference data before operational use.
• Mixed pixels: A nominal 10-meter cell can contain several land uses near boundaries or on small farms.
• Temporal noise: An isolated annual class change may reflect uncertainty, making neighboring years and source imagery useful for confirmation.
• Regulatory use: The map provides screening evidence. Compliance decisions also require plot geolocation, traceability, legal information and documented due diligence.

Load the map and inspect the schema

The 2020 reference layer and 2017 to 2025 time series are freely available under an open license. The project provides several access routes for exploration, analysis and bulk processing:

The 2020 collection can be opened in the Earth Engine JavaScript editor with its catalog asset ID:

const forty2020 = ee.ImageCollection(
  'projects/nature-trace/assets/forest_typology/forest_typology_2020_v1_0_collection'
);

print(forty2020);
print(forty2020.first());

When green returns after clearing

A clear-cut natural forest may regain dense tree cover after being replanted as a monoculture. A conventional canopy map can record loss followed by recovery, while ForTy’s annual classifications can surface the underlying shift from naturally regenerating forest to plantation or tree-crop agriculture.

That transition is relevant to commodity sourcing because the EUDR distinguishes agricultural conversion from forest degradation. Analysts can intersect ForTy with farm boundaries, concessions or sourcing areas, use class probabilities to rank uncertain cases, and send high-risk locations for imagery review or field verification.

Carbon and biodiversity researchers can use the same time series to separate natural recovery from managed planting, compare habitat types and study changes that aggregate tree-cover totals would merge into a single category.