Google DeepMind described an early annual ForTy series covering 2017–2025. The study’s 90.2% overall accuracy measures a separate product: the 2020 map, evaluated at 8,190 sites resolved by expert consensus. The distinction matters when using the series to track changes over time: that score describes the 2020 evaluation, not each year in the series.
What the ForTy forest map classifies
ForTy v1 maps six forest and land-use classes for the 2020 reference year at 10-meter resolution: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land. Its mapped coverage extends from 65°S to 84°N. The study describes a model trained on about 1.7 million labeled locations using satellite embeddings from AlphaEarth Foundations.
Some class distinctions depend on management history and intent, which satellites cannot directly observe. To distinguish planted forest from plantation forest, the study uses a 40-year rotation threshold as an operational proxy. The paper cautions that this threshold may not fit every biome or management system.
What the 90.2% accuracy measures
The study selected 9,100 stratified sites and used the 8,190 resolved through expert consensus to calculate accuracy. It reports 90.2% overall accuracy for the 2020 six-class map, with a standard error of 0.3%. The assessment used the central 10-meter pixel at each site.
That overall figure combines results across classes. The study’s reported class-level F1 scores vary considerably. F1 is a separate metric from overall accuracy, combining precision and recall.
| Class | F1 score in the 2020 six-class evaluation |
| Primary forest | 82.9% |
| Naturally regenerating forest | 72.6% |
| Plantation forest | 66.7% |
| Planted forest | 58.2% |
| Other land | 96.4% |
ForTy also supplies relative confidence scores for its classes. The study says these are not calibrated probabilities of a class being correct. Users can set thresholds that trade precision against recall; for regional applications, the paper recommends checking local score distributions and validating thresholds against local reference data.
Why forest-area estimates differ by method
The study reports both map-based pixel counts and sample-based estimates, in Mha, or million hectares. They are different ways of estimating area, so the figures should be read in their respective columns.
| Region and measure | Map-based estimate (Mha) | Sample-based estimate (Mha) |
| Global total forest | 4,037 | 4,142 ± 105 |
| Asia primary forest | 859 | 661 |
| Asia total natural forest | 1,243 | 1,238 |
| Africa total forest | 601 | 695 |
In Asia, the primary-forest estimates differ, while the total natural-forest estimates are close. The study says the gap largely reflects a different allocation between primary and naturally regenerating forest, rather than a similar-sized difference in total natural forest. For Africa, it associates the lower map-based estimate mainly with sparse woodland and savanna forests near the forest-canopy threshold.
The 2020 map is licensed under CC BY 4.0 and available through Google Earth Engine. The classification layer, validation data, and code are also public.
