Google ForTy Explained: A 10-Meter AI Map of Forest Types

Google announced Forest Typology, or ForTy, on October 22: a 10-meter AI-derived map that distinguishes different kinds of forest from agricultural tree cover. Alongside the 2020 baseline, an early annual series covers 2017 to 2025, helping users examine what replaced a forest rather than only where trees disappeared.

The work comes from Google DeepMind, Google Research, the World Resources Institute and IIASA. The announcement follows the scientific paper's publication on October 21. The distinction matters: this is a newly announced mapping resource, not a claim that the underlying 2020 landscape is current.

What ForTy classifies

ForTy has six classes: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land. That is more useful for some environmental questions than a simple tree-cover label.

A plantation can become green again after harvesting. A tree crop can also look like forest from above. If a system only tracks the presence of trees, it can miss a change in how land is being used. ForTy aims to make those distinctions visible.

The earlier Natural Forests of the World map separated natural forest from other cover. This release adds a more detailed classification, while the early time series extends the question from the type of land to changes over time.

How the AI model works

The Scientific Data paper describes a cascaded deep-learning pipeline trained on roughly 1.7 million samples. It combines weak supervision with active learning, using geospatial satellite embeddings rather than expecting one image to reveal a forest's history.

Weak supervision makes use of imperfect existing labels. Active learning helps focus additional labeling effort on cases where the model needs better examples. AlphaEarth Foundations embeddings provide compact representations of satellite observations that the classification system can use.

The paper reports 90.2% overall accuracy across the six classes, 94.8% for natural-forest classification and 95.5% for distinguishing forest from non-forest. Those results use 8,190 resolved sites from a 9,100-site expert validation sample. They describe the published baseline evaluation, not a separate validation of every year in the early time series.

ForTy access, format and license

The Earth Engine catalog lists the 2020 dataset under CC BY 4.0. Its coverage runs from 65 degrees south to 84 degrees north. It is available as an Earth Engine ImageCollection.

Five bands hold relative confidence scores. The sixth class, other land, is their complement.

A team can choose its own classification threshold. The dataset is open, but that does not mean unlimited free commercial Earth Engine processing.

Where the map needs caution

A global accuracy score hides uneven performance. The catalog lists planted-forest F1 at 58.2%. Shaded coffee or cocoa systems can also be mistaken for natural forest.

The confidence scores are uncalibrated softmax outputs. They are not automatically real-world probabilities of correctness; quantitative risk assessment may need regional recalibration.

That is especially relevant because land-use categories depend on management history and intent, which satellites cannot directly observe. Ten-meter output is useful detail, but it does not turn an inferred category into verified ground truth.

What changes for geospatial AI products

ForTy gives developers a richer starting point for conservation tools, supply-chain screening and land-use analysis. The practical opportunity is to use uncertain classifications to decide where better evidence is needed, rather than hiding uncertainty behind a clean map.

The most useful product built on this data may be the one that helps a reviewer find the ambiguous locations quickly, not the one that gives every location a confident-looking answer.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain complex AI research and products clearly, so developers and buyers understand both the capability and its limits.

Book a 30-minute call.