ForTy v1’s 90.2% overall accuracy applies to its six-class 2020 forest map, not to every class or location. The study in Scientific Data calculated the score using 8,190 sites with expert-consensus labels. Class-level results vary, and planted forest had the lowest F1 score among the forest classes.
What ForTy’s 90.2% accuracy measures
Overall accuracy is the share of evaluated sites that the map classified correctly across all six categories. ForTy v1 maps land between 65° south and 84° north at 10-meter resolution. Its six classes are primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land.
Researchers sampled 9,100 sites. Two trained experts assessed each site using high-resolution imagery, geographic context and satellite time series; they resolved disagreements by consensus. The reported accuracy uses the 8,190 sites with consensus labels and the central 10-meter pixel at each site.
That 90.2% is an aggregate across the six categories. It does not mean a particular forest type, parcel or pixel has a 90.2% chance of being correct. The class-level scores show why the distinction matters.
How the class-level F1 scores compare
F1 combines precision, or how often a mapped class is correct, with recall, or how much of that class the map detects. In the study’s six-class results, planted forest had the lowest F1 score among the forest classes, while the “Other” category scored highest.
| Forest or land class | Six-class F1 score |
| Primary forest | 82.9% |
| Naturally regenerating forest | 72.6% |
| Planted forest | 58.2% |
| Plantation forest | 66.7% |
| Other land | 96.4% |
The study says the strong result for “Other” substantially influences the 90.2% overall figure. So the aggregate alone can hide weaker performance in a specific class, particularly planted forest.
What ForTy’s confidence scores mean
ForTy’s class scores are ensemble-averaged relative confidence values, not probabilities calibrated to match real-world correctness rates. A score of 0.8, for example, should not be read as an 80% chance that the mapped label is right.
For decisions that depend on a particular region or on balancing false positives against missed areas, the study recommends evaluating thresholds with local validation data. A higher threshold favors user’s accuracy, reducing commission errors; a lower threshold favors producer’s accuracy, reducing omissions.
The 2020 map and the announced annual series
The 90.2% result belongs to ForTy v1’s 2020 baseline map. On October 22, 2026, Google DeepMind separately announced an early annual ForTy series covering 2017–2025. The announcement describes that series as an early release being improved with partners and users: its year-by-year layers have a different time scope from the study’s evaluation of the 2020 map. Google DeepMind
