On October 22, 2026, Google DeepMind and Google Research, working with the World Resources Institute (WRI) and the International Institute for Applied Systems Analysis (IIASA), released Forest Typology (ForTy) 2020. It is a global map that assigns every 10-by-10-meter patch of land to one of six forest and land-use types. The announcement puts the problem simply: "From space, an old-growth rainforest, a eucalyptus plantation, and an oil palm estate can look like the same patch of green."
Why the categories matter
The map is built around the EU Regulation on deforestation-free products (EUDR). Products such as coffee, cocoa and palm oil must not come from land that was forest in 2020 and has since been converted to agriculture. Wood products also cannot come from primary forest later converted to plantation. A tree-cover map cannot answer that, because a cleared and replanted plot looks green again within a few years. ForTy uses six classes based on UN FAO definitions: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land. It builds on Natural Forests of the World 2020, released last year, which only answered whether land was natural forest or not.
How it was built
The model was trained on about 1.7 million labelled locations from expert surveys, national and regional datasets and targeted reviews of places where it struggled. Its key input is AlphaEarth Foundations embeddings, which the post calls "data-compact summaries of years of satellite observations for each 10-meter pixel". This lets the model pick up planting patterns, growth cycles and harvest rhythms instead of single-date colour. Google says that against more than 9,000 randomly chosen sites, each assessed by two trained experts, ForTy picked the right forest type 90% of the time and separated forest from non-forest 95% of the time, "outperforming other global forest maps."
Probabilities, not verdicts
ForTy publishes a probability for each class rather than a single label. The public mirror's README describes five probability bands scaled 0 to 250, with the "other land" probability implied by the remainder, and a categorical map defined as the highest-probability class. The data is CC-BY 4.0. The Earth Engine catalog entry names the weak spot plainly: "Planted forests exhibit the lowest classification accuracy among all forest classes (F1 = 58.2%)", because they resemble naturally regenerating forest on one side and plantations on the other. The post suggests that a company assessing EUDR risk might accept only high-certainty pixels, while a researcher estimating total forest extent might use the full probability range.
An early-release time series applies the same approach to each year from 2017 to 2025 to show what kind of change happened, for example natural forest turning into timber plantation. Google says it will keep improving that series with partners.
What remains uncertain. The time series is labelled early release, and its accuracy has not been reported in the same way as the 2020 map. The headline accuracy comes from Google and its partners' own validation. The planted-forest class matters most for EUDR wood products and is also the least reliable.
Analysis: ForTy shows a pattern that is becoming standard in applied AI. A large pretrained model turns raw observations into embeddings once, and many small supervised models are trained on top for specific decisions. For teams building compliance tools, the per-class probabilities are the useful design choice. They make the threshold an explicit, documented business decision instead of a hidden model default. The 58.2% F1 for planted forest is the number to show auditors: any decision near the planted/natural boundary needs field evidence, not a map pixel.
Google DeepMind and Google Research released ForTy 2020 on October 22 with WRI and IIASA, a CC-BY 10-meter global map that gives each pixel a probability for primary, naturally regenerating, planted and plantation forest, tree crops, or other land, plus an early 2017-2025 change series; Google reports 90% forest-type accuracy on more than 9,000 expert-checked sites, while the planted-forest class has the lowest F1 at 58.2%.