A natural tropical forest, a managed eucalyptus stand and an oil-palm plantation can all look green from space. Their habitats, histories and uses are very different. Google's new Forest Typology 2020 map, known as ForTy, gives those differences a common set of labels at a resolution of 10 metres, helping researchers look beyond the amount of tree cover to the kind of landscape beneath it.[1]
Google DeepMind and Google Research announced the work with the World Resources Institute and the International Institute for Applied Systems Analysis on 22 October. The release centres on a validated map for 2020, alongside an early-release annual series covering 2017–2025. That distinction matters when comparing a well-assessed baseline with the project's newer historical sequence.[1]
Six labels for different landscapes
ForTy uses categories based on the Food and Agriculture Organization's forest definitions. Primary natural forest is separated from forest regenerating after disturbance, while planted forest and intensively managed plantation forest receive distinct labels. Tree crops and agroforestry form another class. This makes it possible to identify a landscape that retains a canopy while changing from natural forest to agricultural or commercial tree cover.[1]
The six ForTy classes
| Class | What the label describes |
|---|---|
| Primary natural forest | Native forest without clearly visible signs of human disturbance |
| Naturally regenerating forest | Forest regrowing naturally, including after logging or fire |
| Planted forest | Trees established through planting or seeding, typically on longer cycles |
| Plantation forest | Intensively managed, often fast-growing stands with limited species diversity |
| Tree crops and agroforestry | Agricultural tree cover, including oil palm, rubber, cocoa and fruit trees |
| Other land | The remaining landscape, including grassland, cropland, settlements and water |
How the map was built and checked
The classifier draws on approximately 1.7 million labelled locations and satellite-derived representations of the Earth's surface, including AlphaEarth Foundations embeddings. These representations condense information from observations into features that a model can use to distinguish landscape types. Human-labelled examples then connect those features to the six categories.[1]
For independent validation, more than 9,000 randomly sampled sites were examined by two expert annotators each. The collaborators report 90% accuracy across all six classes and 95% for the simpler forest-versus-non-forest distinction. Those are results for the sampled evaluation; accuracy can vary by region and class. The dataset also provides class probabilities, allowing an analyst to examine uncertainty instead of treating every label as equally reliable.[1]
Following a forest through time
The annual maps open a useful line of enquiry: did an area lose trees, or did its trees change category? A transition from natural forest to commercial tree cover can preserve a green appearance while altering the ecosystem. Regrowth can also become visible as a different process from the establishment of a plantation. Reading those transitions requires consistent classifications and checks against local reference information, especially where probabilities are close.[1]
At 10 metres, the map can expose smaller landscape features that coarser global products blur. A pixel can still mix vegetation, paths and buildings, and a forest-type label does not supply a parcel's ownership history or a complete ecological assessment. Those are concrete limits on using a global classification for a local decision.[1]
For conservation researchers and organisations reviewing sourcing areas, the practical gain is a better starting point for investigation. They can distinguish types of tree cover, locate areas of change and decide where closer imagery or field evidence would be valuable. ForTy turns a familiar question about how much forest remains into a more informative one about which forests remain and how they are changing.[1]
Sources & further reading
- Beyond tree cover: mapping the world's forest types Google DeepMind · 22 Oct 2026 · Accessed 27 Oct 2026