DeepMind maps global forest types at ten meter resolution

What this is about

Google DeepMind, Google Research, the World Resources Institute and the International Institute for Applied Systems Analysis released the global Forest Typology 2020 map, known as ForTy, on October 22, 2026. It divides land into ten by ten meter cells and distinguishes six classes, including primary natural forest, naturally regenerating forest, plantations, tree crops and agroforestry.

The distinction matters in practice. A timber plantation can look as green from space as an established forest, while storing carbon differently and usually supporting less biodiversity. Earlier global maps often showed only whether trees were present. ForTy aims to show what kind of forest or tree based land use is there.

What ForTy actually does

The system combines satellite imagery, vegetation changes over time and reference data from more than 9,000 validation sites. A machine learning model assigns every grid cell to one of six classes. According to the accompanying dataset paper, the map achieved about 90 percent overall accuracy in validation.

The baseline map describes 2020. The team also provides an early annual time series for 2017 through 2025. This makes it possible to inspect not just a snapshot, but whether land may have shifted from natural forest to a plantation or agriculture. The data is available through Google Earth Engine for researchers, authorities and civil society organizations.

Why it matters

For carbon accounting and biodiversity protection, saying “there are trees here” is not enough. An old, diverse forest performs different ecological functions from a uniform plantation. Measuring deforestation, checking supply chains or judging restoration therefore requires a typology rather than a simple green area.

The dataset was designed around definitions from the United Nations Food and Agriculture Organization and requirements in the EU Deforestation Regulation. It could therefore help identify suspicious changes at scale. It does not replace legal evidence for an individual plot.

In plain language

A conventional tree map is like a city map that only distinguishes built and unbuilt land. ForTy adds building types such as house, factory, school or warehouse. That second layer reveals whether two areas that look similar actually serve the same function.

A practical example

A trader checks 12,000 supply plots for cocoa and timber. Instead of inspecting every site immediately, its team compares the 2020 ForTy baseline with the time series through 2025. On 180 plots, the data suggests a possible shift from natural forest to tree crops or open land.

Those 180 signals are then checked with sharper satellite images, supply records and field inspections. The map narrows the search area. It does not automatically decide whether a shipment is unlawful, but points people to places that deserve closer review.

Scope and limits

First, about 90 percent overall accuracy still leaves misclassifications. Clouds, seasons and similar canopy structures can confuse classes.

Second, the baseline represents 2020. The additional time series is described as an early version and may need later corrections.

Third, a grid label proves neither ownership nor the cause of a change. Sanctions, certification and purchasing decisions still require local evidence and a documented review process.

SEO and GEO keywords

Google DeepMind, ForTy 2020, Forest Typology, global forest map, satellite imagery, machine learning, deforestation, EU Deforestation Regulation, Google Earth Engine, primary forest, forest plantations, biodiversity