This week: Google DeepMind released a free map that can tell an old rainforest from a palm oil farm, one 10-meter square at a time.

Elsewhere: How well the map actually scores, where it still gets confused, and how to poke at it yourself.

The New Prime Signal

Satellite maps are great at answering one question: are there trees here? They are much worse at the follow-up: what kind of trees, and who put them there? As Google DeepMind's announcement puts it, "From space, an old-growth rainforest, a eucalyptus plantation, and an oil palm estate can look like the same patch of green." Those three patches hold very different amounts of carbon and wildlife, and for anyone trying to prove their coffee or furniture didn't come from a cleared forest, the difference matters a lot.

What happened: On October 22, Google DeepMind and Google Research, with the World Resources Institute (WRI) and the International Institute for Applied Systems Analysis (IIASA), released Forest Typology (ForTy) 2020. It's "a global map that categorizes every 10-by-10-meter patch of land on Earth into six distinct forest and land-use types." The six are primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land, based on definitions from the UN Food and Agriculture Organization. To teach the model the difference, the team "built a large-scale collection of forest-type examples with about 1.7 million labeled locations." Alongside the 2020 map, they're sharing "an early time series that tracks how these forests changed each year from 2017 to 2025, including where natural forests were replaced by plantations or farms." The 2020 map is the finished product, described in a peer-reviewed Scientific Data paper. The yearly series is an early release that Google says it will keep improving over the coming months.

Why it matters: The European Union's deforestation-free products rule (EUDR) says things like coffee, cocoa and palm oil can't come from land that was forest in 2020 and has since been turned into farmland. Wood products such as furniture also can't come from primary forests that were later converted into plantations. A plain tree-cover map can't check that, because, as the blog notes, "When a natural forest is logged and replanted as a plantation, the land soon looks green again, so standard tree-cover maps might not pick up the change." ForTy is meant to fill that gap: Google calls the 2020 map "a key baseline for the EU's deforestation regulation." It's a tool for regulators, companies and researchers. It doesn't enforce the rule or stop a single chainsaw by itself.

The fine print: Google says that, checked against more than 9,000 randomly chosen sites, "ForTy picked the right forest type 90% of the time, and told forest apart from non-forest land 95% of the time, outperforming other global forest maps." Those are the team's own validation numbers, run with its partners. The paper is more specific: 8,190 of 9,100 sites were fully resolved by expert consensus, giving 90.2% overall accuracy on the six classes and 95.5% for forest versus non-forest. It also says the headline number is "substantially influenced" by the easy "other land" class. The hardest class, planted forest, scores an F1 of just 58.2%, and shaded cocoa and coffee farms "can still be misclassified as natural forest in some regions." About half of the 21 authors work at Google, so this is largely a company grading its own map. And again, the 2017 to 2025 series is an early release, so treat any year-by-year change you see as a draft.

This is the unglamorous kind of AI that might matter most: no chatbot, no demo video, just a very large, very careful spreadsheet of the planet's trees. It won't settle every argument about a specific hillside, and the authors are refreshingly upfront about where it stumbles. But "is that a forest or a farm?" is a question regulators, buyers and conservation groups now have to answer constantly, and a free, open map that shows its own confidence beats squinting at green pixels.

The tricky part is that forest types depend on how land is used, not just how it looks. A planted forest and a wild one can look nearly identical from above, so the model has to pick up quieter clues.

How it works: ForTy takes as a key input AlphaEarth Foundations embeddings, which the blog describes as "data-compact summaries of years of satellite observations for each 10-meter pixel, so it can spot subtle clues like planting patterns, growth cycles and harvest rhythms." Instead of a yes-or-no label, it gives each patch a probability score for every class. That lets people pick their own strictness. Google's example: a company checking EUDR risk "might set a strict threshold," while a researcher estimating total forest extent could use the full range. One fun result from the paper: in Sweden, a very green country, about two-thirds of the forest area is classified as planted or plantation. Lots of trees, not much wilderness.

Try it: The 2020 map is free under a CC-BY 4.0 license. Developers can load it from the Earth Engine data catalog with a few lines of code, the raw files are on Figshare, and the training code is in DeepMind's JEO repository on GitHub. If you're not a map person, the takeaway is simpler: next time a company says its product is "deforestation-free," maps like this are how someone can start checking.