Google DeepMind announced ForTy on October 22, 2026, a global map developed with Google Research, the World Resources Institute (WRI) and the International Institute for Applied Systems Analysis (IIASA). The 10-meter map classifies land into six forest and land-use categories for 2020. Google also announced a separate early annual series spanning 2017 through 2025; the detailed accuracy results concern the 2020 map.
What ForTy classifies
A view of tree cover alone cannot distinguish natural forest from a managed plantation or an agricultural tree crop. ForTy separates those landscapes into six categories:
- Primary forest: native-species forest with no visible signs of human-caused disturbance.
- Naturally regenerating forest: forest established predominantly through natural regrowth after human-caused disturbance.
- Planted forest: forest established predominantly by planting or seeding. ForTy uses a rotation period longer than 40 years as an operational proxy for this class.
- Plantation forest: intensively managed forest, typically made up of one or two species of similar age, with a rotation of 40 years or less.
- Tree crops and agroforestry: agricultural tree crops such as oil palm, rubber and fruit orchards, including agroforestry systems.
- Other land: non-tree uses and vegetation that does not meet the map’s forest-class criteria. Other wooded land is grouped here rather than mapped separately.
The 40-year cutoff is a practical simplification, not a universal description of forest management; the paper cautions that it may not fit every management system.
What the 10-meter resolution and coverage mean
Each raster cell represents a 10-by-10-meter area. The map covers land between 65°S and 84°N. In the paper’s overview display, the majority class is shown for each 1 km² cell, a broader summary than the underlying 10-meter raster.
Resolution describes the size of the mapped cells, not certainty at every location. Mixed pixels and boundaries between classes can make classification uncertain, and satellite imagery does not directly reveal a forest’s management history.
What ForTy’s accuracy figures measure
The 2020 validation assessed 9,100 stratified sites. Two trained experts resolved 8,190 by consensus; 738 were pending consensus and 172 were marked uncertain. The paper reports 90.2% overall accuracy across the six classes on the resolved sites. The two binary accuracy figures measure different tasks, while class-level F1 scores show how results vary by category. An F1 score combines precision and recall for a class.
| Validation measure | Result | What it measures |
| Overall accuracy | 90.2% (SE 0.3%) | Correct classifications across all six classes on 8,190 consensus-resolved sites |
| Natural forest vs. non-natural land | 94.8% (SE 0.3%) | A binary distinction between natural forest and non-natural land |
| Forest vs. non-forest | 95.5% | A binary distinction between forest and non-forest |
| Primary forest F1 score | 82.9% | Class-specific F1 score |
| Naturally regenerating forest F1 score | 72.6% | Class-specific F1 score |
| Plantation forest F1 score | 66.7% | Class-specific F1 score |
| Planted forest F1 score | 58.2% | Class-specific F1 score |
| Other F1 score | 96.4% | Class-specific F1 score |
The spread matters when a question concerns a particular class: the reported F1 score is 58.2% for planted forest and 96.4% for Other. The 90.2% overall figure does not describe the performance of every category equally.
Confidence scores and thresholds
ForTy provides relative class-confidence scores, not calibrated probabilities. Its categorical layer assigns each pixel the class with the highest score. The paper recommends a higher threshold when users want to favor user’s accuracy and reduce commission errors, or false positives. A lower threshold favors producer’s accuracy and reduces omission errors, or false negatives.
These choices let analysts tune how the map is used, but they do not turn a class score into a precise probability. The 40-year rotation cutoff is also a proxy; the map does not directly observe management history from satellite imagery.
Where to explore ForTy
Open the ForTy 2020 map in Google Earth Engine or Google’s separate early time-series view. The 2020 dataset is licensed under CC BY 4.0. ForTy aligns with definitions used in the European Union Deforestation Regulation (EUDR) and is a monitoring resource; a map classification is not a regulatory compliance ruling for an individual product or parcel.
