A 10-meter AI map separates primary forest from plantations and tree crops

DeepMind and WRI's ForTy sorts the world's land into six classes, aiming at deforestation rules, though plantations and tree crops are still often missed.

Google DeepMind and the World Resources Institute have built ForTy (Forest Typology), an AI-generated global land-cover map for 2020. It assigns each 10-meter patch to one of six classes: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops such as oil palm, rubber and cocoa, and other land. The authors describe it as the first global map of this kind. Earlier datasets, including one from the European Commission, cover only parts of the world.

The model behind it is DeepMind's AlphaEarth Foundations, which compresses data from several satellites into a description of each pixel. Three image-recognition stages then refine the classification, and the last smooths out multi-year outliers. Experts added about 43,000 training examples in correction rounds, fixing errors like European vineyards labeled as tree crops and missed oil palms in Africa and Indonesia. Checked against 8,190 random sites, each reviewed by two experts, overall accuracy was 90.2%. On the same test data, ForTy beat the European Commission's EUDR maps: 90.9% vs 87.8% for forest types.

Why it matters: rules like the EU deforestation regulation bar goods linked to primary forest converted to plantation after December 31, 2020. Enforcing that requires telling a native forest from a tree farm, which ordinary forest-cover maps can't do. A consistent global map could make that check cheaper. The authors also point to the Tropical Forest Forever Facility and the UN goal of protecting 30 percent of land by 2030.

The caveats are large. The authors warn the 90.2% score is lifted by the easy 'other land' class. The map found 89% of expert-confirmed primary forests but only about half of tree crops and planted forests and 60% of plantations. It overestimates Asian primary forest by about 200 million hectares, largely Russian conifer forest, and misses about 94 million hectares in Africa. The 40-year rotation cutoff for plantations doesn't fit every region.

Source: The Decoder