ForTy separates natural forests from planted land at 10 metre resolution
The ForTy research dataset classifies the world’s land in 2020 into six forest and land-use categories at 10 metre resolution. Its peer-reviewed description distinguishes natural forests, planted forests, plantations and tree crops, going beyond a simple canopy map. Expert validation found strong overall performance, with larger difficulties between similar managed classes. This historical baseline does not measure today’s deforestation rate.
Science··Evening
Tree cover becomes a land-use map
The ForTy research dataset classifies the world’s land at 10 metre resolution for a 2020 baseline. Its peer-reviewed description separates forest type from simple tree presence by incorporating land use and management history. A green satellite pixel can represent a natural forest, a timber plantation or an agricultural tree crop, with different histories behind similar vegetation.[1]
The map has six categories. It separates primary from naturally regenerating forest, and planted forest from plantations. Tree crops and agroforestry share another category; the remaining class is other land. This historical distribution supplies a common baseline; it does not measure current deforestation rates or the outcomes of conservation work.[1]
Expert comparisons reveal uneven accuracy
Satellite observations supplied numerical representations for staged learning, supported by reference samples. Experts agreed on the classes of 8,190 validation locations. Overall accuracy across six classes reached 90.2 per cent, with the easier other-land class raising the aggregate. The distinction between plantations and planted forests was less reliable. Mature secondary forest can resemble primary forest, while shaded farming can resemble natural vegetation.[1]
Area estimates depend on the counting method
A sample-based estimate put global forest area at 4,142 million hectares. At the 95 per cent confidence level, its uncertainty was 105 million hectares. Counting pixels directly instead yielded an area of 4,037 million hectares. Sparse African woodlands contributed substantially to that difference. Excluding uncertain validation sites may make performance look optimistic, and the model’s confidence scores are not calibrated probabilities. Regional validation and local land-use history remain relevant to interpreting individual classes.[1]
