How does a 10-meter map tell a plantation from a primeval forest?

What a green pixel hides

From orbit a primary rainforest and a eucalyptus plantation look like the same green pixel. An oil-palm estate does too. They hold vastly different amounts of carbon and biodiversity. The European Union’s regulation on deforestation-free products requires proof that coffee, cocoa, palm oil or wood did not come from land that was forest in 2020 and has since been converted to agriculture. Wood products must also be free of forest degradation: they cannot originate from primary forest turned into plantation or from naturally regenerating forest turned into plantation. Companies and regulators therefore need to know not just where trees stand but what kind of forest they form and what replaces them after harvest.

Last year the same partners released a 10-meter map that answered only one question: was this patch natural forest in 2020 or not? The new Forest Typology (ForTy) 2020 map goes further. It sorts every patch into six categories and adds an early time series that shows year-by-year change from 2017 through 2025.

Six classes built on FAO definitions

The map uses definitions from the UN Food and Agriculture Organization. Primary forest means native forest with no visible signs of human disturbance. Naturally regenerating forest grew back on its own after logging or fire. Planted forest was seeded or planted by people and managed over long cycles. Plantation forest is fast-growing, intensively managed, often a single species. Tree crops and agroforestry grow things other than wood — oil palm, rubber, cocoa, fruit. Other land covers everything else: grassland, cropland, cities, water.

Each square kilometer displays the most common type, but the underlying database stores the 10-meter classification for every pixel. Users can explore it at full resolution in an Earth Engine app.

Training on 1.7 million labeled sites

Teaching a model to separate these classes is hard because the differences depend on land use, not just appearance. The team built a training set of about 1.7 million labeled locations drawn from expert surveys and national datasets. Regional datasets and targeted reviews of problem areas were also included. The model uses a custom architecture designed for remote-sensing data. A key input is AlphaEarth Foundations embeddings — data-compact summaries of years of satellite observations for each 10-meter pixel. Those embeddings let the model spot subtle clues: planting patterns, growth cycles, harvest rhythms.

Validation across 9,000 random sites

With the World Resources Institute and the International Institute for Applied Systems Analysis the team validated ForTy against more than 9,000 randomly chosen sites worldwide. At each site two trained experts studied high-resolution imagery and years of satellite records to decide the true land type. ForTy picked the correct forest type 90 percent of the time. It separated forest from non-forest land 95 percent of the time, outperforming other global forest maps.

Instead of a single yes-or-no label the model outputs a probability score for each class. A company assessing deforestation risk under the EU regulation might set a strict threshold and accept only pixels the model is highly certain about. A researcher estimating total forest extent might use the full spectrum of probabilities to get a fuller picture of the potential landscape.

From snapshot to yearly story

A single-year map is a baseline. Forests change. The new early-release time series applies a similar approach to every year from 2017 to 2025. It shows not only where forests disappear and where they grow back but also what kind of change happened: a natural forest turned into a timber plantation, or abandoned land slowly regenerating into forest. This matters because some of the most important changes are hard to see. When a natural forest is logged and replanted as a plantation the land soon looks green again. Standard tree-cover maps might not pick up the change. ForTy’s time series makes those hidden changes visible — both losses and gains.

Open access and next steps

The Forest Typology 2020 map, described in Scientific Data, and the time series are now freely available on Google Earth Engine under an open license. Governments, companies, conservation groups and scientists can use them to build deforestation-free supply chains, direct forest finance where it is needed, and protect the natural forests that matter most. Over the coming months the partners will keep improving the time series with users. ForTy is part of a broader AI for Nature portfolio that includes mapping drivers of forest loss, forecasting deforestation risk, modelling global species distributions and monitoring biodiversity in the field. It is also the latest addition to Google Earth AI, Google’s collection of geospatial models and datasets for environmental challenges. How many of the 1.7 million training sites remain in active use is not stated in the announcement.