AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Google Deepmind creates a global map of forests with satellite data and AI

Collected Oct 31, 2026

Google DeepMind and the World Resources Institute have released ForTy, short for Forest Typology, a global satellite-derived land cover map pegged to the year 2020. The map assigns every land area on Earth to one of six classes at 10-meter resolution: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops such as oil palm, rubber or cocoa, and other land. The authors describe it as the first global map of its kind; earlier datasets, including one produced by the European Commission, cover only parts of the world. The map, classification layer and test data are freely available under a CC BY 4.0 license, including through Google Earth Engine, and DeepMind has published code in the GeeFlow and JEO repositories.

The immediate driver is the EU Regulation on Deforestation-free Products (EUDR), which bars goods tied to forests cleared or degraded after December 31, 2020. That includes goods produced after primary forest was converted to a plantation, while oil palm and cocoa farms are treated as agricultural land. Companies assessing supply chain risk therefore need a defensible picture of how a given site was used in 2020, and ForTy is built to supply that reference point. The researchers also point to uses in the Tropical Forest Forever Facility (TFFF) and the UN goal of protecting 30 percent of the world's land by 2030.

Separating the classes is hard because a forest's type depends on its history rather than its appearance from orbit. An old planted mixed forest can resemble a natural one, and a plantation cleared only recently still counts as forest by definition even when the imagery shows bare ground. ForTy works from AlphaEarth Foundations, the model DeepMind introduced in July 2025 as a kind of virtual satellite that fuses optical imagery, radar, laser scans and climate simulations into a compact annual description of each 10-by-10-meter area, with embeddings from 2017 onward available in Earth Engine. Several image-recognition stages sit on top. A first model learns from verified examples; its predictions are combined with outside data on tree height, buildings and forest loss to expand the training set for a second model; a final step analyzes several years together to smooth outliers. Gemini helped adapt older training data to EU definitions, and experts added roughly 43,000 examples during active-learning rounds targeting errors such as European vineyards classed as tree crops, Central American pine forests labeled planted, and missed oil palms in Africa and Indonesia.

Accuracy is the crux. Two experts reviewed each of 8,190 randomly selected sites and judged the map correct in 90.2 percent of cases. The authors caution that this headline number is lifted substantially by the easy "other land" class. By type, the map caught 89 percent of primary forest confirmed on the ground, but only about half of tree crops and planted forests and 60 percent of plantations. In the other direction, roughly a quarter to a third of areas mapped as plantations or tree crops were misclassified, and a third of areas labeled planted forest were wrong. The evaluation may flatter the map slightly: experts were most often unsure about managed areas, and the team excluded those uncertain cases from its assessment. Against European Commission maps built for the EUDR on the same test data, ForTy scored 95.5 percent versus 93.6 percent on forest versus non-forest and 90.9 percent versus 87.8 percent on forest types.

Some limits are structural. Satellites rarely reveal management, so the researchers use a 40-year rotation period to split plantations from other planted forests, a cutoff that will not fit every region. The map inherited a training-data weakness: annotators seldom marked forests near roads or settlements as primary even when they looked untouched. Shaded cocoa or coffee grown under a natural canopy can still read as forest, and selectively logged or burned forest is difficult. There is no separate class for open woodlands, which the EUDR treats distinctly, though future versions are expected to add one. Per-class confidence scores let users set their own thresholds and make their own classification calls.

The global totals are about 4.1 billion hectares of forest, including nearly 3.9 billion hectares of natural forest, of which about 2 billion are primary. Planted forests cover about 250 million hectares and tree crops nearly 200 million, mostly in Asia. Regional error is real: the map overestimates primary forest in Asia by roughly 200 million hectares, largely Russia's conifer forests, and misses about 94 million hectares in Africa, likely because of open dry forests and savannas. Tree crops and plantations dominate Southeast Asia, while planted forest makes up most forest area in Sweden and Finland, where about two-thirds of Sweden's forest is planted and natural forest is only around a fifth; the authors say tree cover there may be adequate but the share of natural forest, which matters for biodiversity, could be critically low. The Democratic Republic of the Congo, Indonesia and Brazil have the highest natural-forest shares.

This suggests the map is best read as a strong baseline plus a confidence layer, not a verdict. For commodity traceability, the weak recall on plantations and tree crops is precisely where EUDR due diligence bites: a low-confidence designation is not proof either way, and the asymmetric errors mean a site mapped as "other land" or "forest" may not be what it seems. The combination of open licensing, Earth Engine availability and per-class confidence scores is what makes the release practically useful, since teams can tune thresholds to their risk tolerance.

Why it matters: Sustainability, sourcing and geospatial teams now have a free, global, 2020-dated reference they can query at parcel scale, with published accuracy figures so they can judge where it is trustworthy. Producers and traders exposed to EUDR will likely treat ForTy as one input among several — inference — pairing it with field audits and supplier records rather than relying on any single class label. The numbers to watch are recall on plantations and tree crops and the absent open-woodland class; improvements there, rather than overall accuracy, will decide how much weight the map carries in compliance work.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Google Deepmind and the World Resources Institute have released ForTy, a global map that classifies forests and other land into six categories, including primary forest, plantations, and tree crops. The map could help track deforestation under EU rules, but it identifies only about half of planted forests and tree-crop areas, and 60% of plantations. The article Google Deepmind creates a global map of forests with satellite data and AI appeared first on The Decoder .