Beyond tree cover: a 10 m global map of forest types

Most global forest maps answer one question per pixel: are there trees here or not? Google DeepMind and Google Research, with the World Resources Institute (WRI) and IIASA, have released a map that answers a harder one: what kind of forest is this, and what happened to it? Forest Typology (ForTy) 2020 classifies every 10-by-10-meter patch of land on Earth into one of six forest and land-use types, and it ships with an early annual time series covering 2017 to 2025.

What shipped

ForTy sorts land into six classes, using definitions from the UN Food and Agriculture Organization:

• Primary forest: native forest with no visible signs of human disturbance.
• Naturally regenerating forest: forest that grew back on its own after logging or fire.
• Planted forest: planted or seeded, managed over long cycles.
• Plantation forest: fast-growing, intensively managed, often a single species.
• Tree crops and agroforestry: oil palm, rubber, cocoa, fruit.
• Other land: grassland, cropland, cities, water.

The 2020 map and the time series are free on Google Earth Engine under an open license, and the 2020 map is described in a paper in Scientific Data. ForTy extends last year's binary Natural Forests of the World 2020 map to six classes and a time dimension.

Why tree cover is the wrong variable

From orbit, an old-growth rainforest, a eucalyptus plantation and an oil palm estate can all look like the same green, yet a primary forest holds far more carbon and biodiversity than a timber plantation harvested every decade or two.

The EU's deforestation regulation (EUDR) requires that coffee, cocoa, palm oil and similar products do not come from land that was forest in 2020 and was later converted to agriculture. Wood products must also avoid forest degradation: primary forest converted to plantation or planted forest, or naturally regenerating forest converted to plantation. A tree-cover mask cannot check that. You need the forest type at the 2020 baseline and what replaced it.

The time series targets the hardest case. When a natural forest is logged and replanted as a plantation, the pixel turns green again within a few seasons. A tree-cover map may register the clearing, but it might not pick up the change in what the land has become. ForTy is built to label that transition as natural to plantation.

How it works

Many of these classes depend on land use, not only on how the land looks. DeepMind's approach has three parts:

1. Inputs: AlphaEarth Foundations embeddings, compact per-pixel summaries of years of satellite observations. A single image cannot show planting rows or harvest cycles; a multi-year embedding can.
2. Labels: about 1.7 million labeled locations from expert surveys, national and regional datasets, and targeted reviews of places where the model was failing.
3. Model: a custom architecture for remote sensing data, not detailed further in the post.

On more than 9,000 randomly sampled validation sites, each labeled by two trained experts, ForTy chose the correct forest type 90% of the time and separated forest from non-forest 95% of the time. DeepMind reports that both figures beat other global forest maps.

The design choice I like most

ForTy outputs a probability for each class at each pixel, not a single hard label. An EUDR compliance team wants precision and can keep only high-confidence pixels. A researcher estimating forest extent wants recall and can integrate over the distribution. Shipping probabilities moves the precision-recall decision to the user, who knows the cost of each kind of error. I would like to see calibration curves next to the headline accuracy, because a threshold is only as useful as the probabilities behind it.

My take

This is an applied geospatial dataset, not a frontier model, and a good example of what foundation embeddings are for. AlphaEarth does the expensive representation learning once; a downstream classifier with well-chosen labels and careful validation then produces a public artifact that regulators, commodity buyers and conservation groups can use directly.

There are caveats. The time series is an early release. 90% accuracy averaged over six classes will be uneven, and I would expect most of the confusion at the planted-versus-plantation boundary. Smallholder plots and mixed agroforestry remain hard at 10 m. Anyone using this for compliance should validate it on their own regions before relying on a threshold.

The direction is right. Tree cover was always a proxy; the useful variables are forest history and land use. The original announcement links to the Earth Engine apps if you want to explore it.