Google ForTy Maps Six Forest and Land-Use Classes at 10-Meter Resolution

Forest Typology maps what the green pixels mean

Google DeepMind, Google Research, the World Resources Institute and IIASA announced Forest Typology 2020, or ForTy, on October 22. The global map classifies land at 10-meter resolution into six forest and land-use types. It distinguishes primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land. A separate early-release time series applies a related approach to annual changes from 2017 through 2025. The 2020 map and the time series have different maturity labels; a user should not silently treat the early series as a fully validated historical record.

The distinction matters because a patch can remain green after a natural forest is replaced by a timber plantation or a tree-crop farm. A binary tree-cover map can miss the ecological and regulatory difference. The partners link this to deforestation-free supply chains and the European Union's deforestation regulation. The mapping product is evidence for screening and investigation, not legal certification of any shipment or proof of land ownership.

Training and validation figures

The original research announcement says the team used roughly 1.7 million labeled locations from surveys, regional datasets and targeted review. AlphaEarth Foundations embeddings summarize multiple years of satellite observations for each 10-meter pixel and help the model detect planting patterns, growth cycles and harvest. Those inputs support classification; they do not reveal every on-the-ground land-use decision or replace local records.

The partners say they evaluated ForTy at more than 9,000 randomly selected sites, each interpreted by two trained experts using high-resolution imagery and satellite history. They report the right forest type 90 percent of the time and forest-versus-nonforest discrimination 95 percent of the time. These are the authors' reported validation numbers for the cited sampling and category definitions. A single overall accuracy hides class imbalance and regional variation, so a buyer should inspect per-class errors and local ground truth before using a pixel label for a high-stakes decision.

ForTy provides probability scores for each class, allowing different thresholds in different contexts. This is useful when plantation versus naturally regenerating forest is uncertain. A policy team should preserve those probabilities rather than flattening them into a confident yes-or-no. A map may show where field inspection is most valuable; it cannot alone settle a dispute about a particular farm or the date of conversion.

Availability and limits

The partners say the 2020 map and the early time series are freely available through Google Earth Engine under an open license. They also link data downloads and an explorer. Access to a dataset does not imply every user can reproduce training, and a 10-meter pixel is not a survey marker. A responsible use compares map versions, dates, local geography and the decision's error costs.

This October 26 issue reports the October 22 release. It treats the reported accuracy as a team result, not an independent test by this publication, and keeps the early-release status of the annual series visible.