Google DeepMind and WRI Unveil ForTy, an AI Map Distinguishing Global Forest Types at 10-Meter Resolution

Google DeepMind and the World Resources Institute (WRI) have developed ForTy (short for Forest Typology), an AI-driven global land cover map designed to differentiate between primary forests, timber plantations, and agricultural tree crops worldwide.

While conventional satellite maps typically indicate only whether an area has tree canopy cover, ForTy classifies global land area for the year 2020 into six distinct categories at a 10-meter spatial resolution. The categories comprise primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops (such as oil palm, rubber, and cocoa), and other land. The underlying data provides 10-meter detail, while the visualized map presents the dominant land use class within each square kilometer. The researchers describe ForTy as the first global map of its kind, improving on existing datasets that only span specific geographic regions.

Driven by Stricter Deforestation Regulations

The primary catalyst for the project is the European Union Regulation on Deforestation-free Products (EUDR). The regulation bars commodities linked to deforestation or forest degradation occurring after December 31, 2020, including cases where primary forests were converted into commercial plantations. Because crops like cocoa and oil palm are categorized as agriculture, enterprises evaluating supply chain compliance require an accurate historical baseline of land use as of the end of 2020.

Beyond EUDR tracking, ForTy is slated to supply foundational data for conservation initiatives such as the Tropical Forest Forever Facility (TFFF) and the United Nations objective to protect 30 percent of the planet's land area by 2030.

Establishing a reliable 2020 baseline remains technically challenging because land classification depends heavily on historical context rather than surface appearance alone. From orbit, a mature planted mixed forest can look nearly identical to an untouched natural canopy, while a freshly cleared plantation still technically qualifies as forestry land despite appearing as bare soil. To address these ambiguities, ForTy tracks planting layouts, biological growth cycles, and harvesting timelines rather than relying solely on single-snapshot imagery.

A Three-Stage Architecture Built on AlphaEarth

ForTy relies on Google DeepMind's AlphaEarth Foundations model, a virtual satellite architecture introduced in July 2025. AlphaEarth compresses multi-satellite streams—encompassing optical satellite photography, radar, laser scans, and climate simulations—into dense pixel-level descriptions.

ForTy refines these descriptions through three sequential modeling phases:

1. An initial image-recognition model trains on verified baseline examples.
2. Predictions from this model merge with external datasets covering tree height, surrounding infrastructure, and recorded forest loss to form an expanded training corpus for a second model.
3. A third model evaluates several consecutive years of data to smooth out temporary anomalies and misreadings.

Google's Gemini language model assisted the team in reinterpreting older training data to align with formal EU regulatory definitions. In addition, the development cycle incorporated multiple active learning rounds, during which human specialists contributed approximately 43,000 corrections. These targeted edge cases, including European vineyards erroneously flagged as tree crops, Central American natural pine stands mislabeled as plantations, and unflagged oil palm tracts across Africa and Indonesia.

Accuracy Benchmarks and Known Weaknesses

In an evaluation of 8,190 randomly chosen sites reviewed independently by two experts, ForTy demonstrated an overall accuracy of 90.2 percent. However, researchers noted that the broad "other land" category heavily boosts this aggregate score.

Performance diverges substantially across individual forest categories:

• Primary forests: The model correctly identified 89 percent of ground-confirmed primary forest areas.
• Tree crops and planted forests: The model captured only around 50 percent of existing sites.
• Plantations: Detection hovered around 60 percent.
• False positives: Approximately one-quarter to one-third of regions labeled as tree crops or plantations were misclassified, and one-third of areas designated as planted forest were incorrect.

Despite these blind spots, ForTy outperformed the European Commission’s dedicated EUDR datasets on the identical evaluation set. It achieved 95.5 percent accuracy versus 93.6 percent for general forest-versus-nonforest separation, and 90.9 percent versus 87.8 percent when distinguishing specific forest subcategories.

The model also exhibits geographic discrepancies. It overestimates primary forest in Asia by roughly 200 million hectares—primarily within Russia's boreal conifer belt—while undercounting approximately 94 million hectares across Africa due to challenges in identifying open dry forests and savannas. It also faces difficulties identifying agroforestry systems such as cocoa or coffee cultivated beneath natural tree canopies.

Global Forest Distribution and Open Availability

Based on the mapping data, the researchers estimate Earth's forest coverage at roughly 4.1 billion hectares, containing nearly 3.9 billion hectares of natural forest (with about 2 billion hectares of primary forest). Planted forests represent approximately 250 million hectares, while tree crops span nearly 200 million hectares, predominantly in Asia.

The findings also highlighted regional compositions: Brazil, Indonesia, and the Democratic Republic of the Congo possess the largest relative shares of natural forest. By contrast, planted forests comprise roughly two-thirds of forest land in Sweden and Finland, leaving natural forest across only about one-fifth of Sweden's territory.

The ForTy map layer, classification models, and benchmark evaluation data have been released publicly under a Creative Commons CC BY 4.0 license, accessible via Google Earth Engine. Code implementations have also been made available across the GeeFlow and JEO repositories as part of Google's broader Earth AI initiative.