From a satellite, an old rainforest, a eucalyptus plantation and an oil palm estate can all look like the same patch of green, yet they hold very different amounts of carbon and wildlife. Google DeepMind and Google Research, working with the World Resources Institute and the International Institute for Applied Systems Analysis, have released a map that tries to tell them apart everywhere on Earth. Forest Typology 2020, or ForTy, sorts every 10-by-10-metre patch of land into six classes: five forest and tree-crop types, from primary forest to timber plantations and tree crops, plus land that is not forest.
The timing is practical. The EU's regulation on deforestation-free products requires that goods such as coffee, cocoa and palm oil do not come from land that was forest in 2020 and has since been converted, and wood products must not come from primary forest later turned into plantations. Most global maps only show whether trees are present, which cannot answer that question. ForTy extends last year's Natural Forests of the World map by saying what kind of forest is there, and an early time series tracks changes each year from 2017 to 2025, including natural forest replaced by plantations, which looks green again soon after replanting and is easy to miss.
The model was trained on about 1.7 million labelled locations from expert surveys, national datasets and targeted reviews, using AlphaEarth Foundations embeddings, compact summaries of years of satellite data for each pixel, so it can pick up planting patterns and harvest cycles. Checked against more than 9,000 randomly chosen sites assessed by two trained experts each, it picked the right forest type 90% of the time and separated forest from non-forest 95% of the time, which the team says beats other global maps. Rather than a single label, it gives a probability for each class, so a company managing regulatory risk can set a strict threshold while a researcher uses the full range.
The 2020 map is described in Scientific Data, and both it and the time series are free on Google Earth Engine under an open licence.
The caveats matter for anyone relying on it for compliance. A 90% global accuracy still leaves one site in ten misclassified, and the announcement does not break accuracy down by region or class. Enforcement happens plot by plot, so users should check local accuracy before treating a pixel as evidence. The time series is labelled an early release, and Google says it will keep revising it with partners.