Google DeepMind and the World Resources Institute have released a free map that sorts all of the world's land into six types: forest that has never been cleared, forest that grew back, planted forest, timber plantation, tree farms such as oil palm, rubber and cocoa, and everything else. Most earlier maps only said whether trees were there. The same teams released one last year that separated natural forest from other tree cover, which gave a yes or no answer. This one says what the other trees are.
The reason it exists is an EU rule that bans cocoa, coffee, palm oil and wood linked to forest cleared after the end of 2020. The rule applies to large and medium companies from 30 December 2026. Whether a plot counts depends on what it was in 2020. A palm farm that was already there is fine, and primary forest that later became a palm farm is not.
The map is strong on forests and weak on farms
The map picks the right type 90.2% of the time, but that score is lifted by the easiest class, ordinary land that is not forest. Class by class, it finds 89% of primary forests but only about half of tree crops. A quarter to a third of the land it labels as plantation or tree crop turns out to be something else. It still beats the European Commission's own maps built for this rule on the same test, so the weakness is the difficulty of the problem, not a worse tool than buyers already had.
The difficulty comes from history. An old planted forest of mixed trees looks natural from space, and a plantation that was just harvested counts as forest even though the image shows bare ground. Shade-grown cocoa and coffee, which grow under a natural canopy, are still labelled forest in some regions. A satellite sees the trees, and the rule asks what the land was used for.
A confidence score turns into a rule
For every class the map gives a probability, and users choose their own cutoff. A company that wants no regulatory risk will accept only land the map is very sure about.
Take a chocolate maker that buys from 1,000 farms. If the map recognises only about half of tree-crop farms with confidence, around 500 farms come back unsure or wrongly labelled. Each one needs a call to the cooperative, dated photos and perhaps a visit from a local agent. At half a day each, that is about 250 working days, roughly a year of one person's work.
The map still helps, because it turns a list of 1,000 into a list of 500 and says which farms to look at first. But the buyer now has a choice between paying for that year of checking and dropping the uncertain suppliers. Dropping them is cheaper. The farms most likely to be dropped are the ones the map finds hardest to read, such as shade-grown cocoa on small plots, and those belong to the growers with the least money to argue. The yearly version of the map, which shows what replaced cleared forest, will make the same pressure sharper, because it gives buyers more flags to act on.
What to do with a map that is honest about its doubts
The map is free in Google Earth Engine, so a cooperative can look up its own plots before its buyer does. For any plot the map shows as forest, the cooperative can collect proof of what grew there in 2020: dated photos, planting records, sales receipts. A buyer should ask its suppliers for the confidence score and the cutoff behind each flag, and treat a flag as a reason to check, not as a verdict.