Google DeepMind and WRI release ForTy 2020, first global 10-meter forest typology map

Google DeepMind and Google Research, working with the World Resources Institute (WRI) and IIASA, have released Forest Typology 2020, known as ForTy — the first global map that classifies land into six forest and land-use categories at 10-meter resolution. According to a dev.to summary of the Google DeepMind announcement dated 31 October 2026, the underlying dataset paper is published in Scientific Data.

The six classes follow FAO definitions: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land. Because the taxonomy matches the definitions governments and companies already use to report forest data, the map's output can slot into existing reporting and accounting workflows without a translation layer.

How the model was built

A key input to the classifier is AlphaEarth Foundations embeddings — data-compact summaries of years of satellite observations for each 10-meter pixel. These embeddings allow the model to pick up subtle clues such as planting patterns, growth cycles and harvest rhythms, which are what separate, for instance, a timber plantation from a naturally regenerating forest.

Training relied on roughly 1.7 million labeled locations drawn from expert surveys, national and regional datasets, and targeted reviews. The labels fed a cascaded deep-learning pipeline that also produces per-class confidence maps, so each pixel carries an indication of how certain the model is about its assigned class rather than a bare label.

Measured accuracy

The team, together with WRI and IIASA, validated ForTy against 8,190 expert-labeled random sites. The dev.to summary reports 90.2 percent overall accuracy across the six classes, 94.8 percent on natural forest, and 95.5 percent on the broader forest-versus-non-forest question. On the same test data, the model reportedly outperformed the European Commission's comparable map.

A time series and a regulatory anchor

Beyond the static 2020 snapshot, the release includes an early time series tracking how forests changed year by year from 2017 to 2025, including what land cover replaced natural forest that was lost. The project was built explicitly as a reference layer for the EU Deforestation Regulation, whose baseline cutoff is 31 December 2020 — the same date the main map represents. That alignment means compliance work that depends on establishing forest status at the regulation's cutoff has a free, standardized layer to work from.

The dataset is available on Google Earth Engine under an open license, with no application process or waiting list, according to the source.

Why it matters

Most open forest data to date has been limited to tree cover — whether canopy exists — rather than what kind of forest it is. A free, global, 10-meter typology layer changes that. Geospatial engineers, conservation researchers and supply-chain teams can now build class-level forest monitoring without commissioning custom classification work, and the per-class confidence scores let them tune decision thresholds by region instead of accepting a single binary output.

For teams already working in Earth Engine, the practical shift is immediate: pipelines that consumed tree-cover-only inputs can move to typology-aware logic, distinguishing, say, an established plantation from primary forest. And because the classes follow FAO definitions, results should be comparable with official statistics. For deforestation-regulation compliance, biodiversity accounting and land-use research, having a consistent global baseline with quantified uncertainty is a meaningful step up from the status quo.

One caveat is worth noting: the accuracy figures come from the project's own validation, and the time series is described as an early release, so anyone relying on it for legal or regulatory decisions should expect to cross-check against national datasets where those exist.