Google DeepMind and WRI launch global forest map using AI

Mapping the world's forests

Google DeepMind and the World Resources Institute have released ForTy, a global satellite map that identifies different types of forest cover. While standard satellite maps often show where trees grow, ForTy goes further by assigning land to six specific categories. These include primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops like rubber or cocoa, and other land.

The project aims to provide a reliable reference point for the year 2020. This specific date is important because of the EU Regulation on Deforestation-free Products, which bans goods linked to forests cleared or degraded after December 31, 2020. Companies need this data to assess risks in their supply chains and ensure they are not sourcing products from converted primary forests.

Beyond trade regulations, the map supports broader environmental initiatives. It can feed into the Tropical Forest Forever Facility and help track progress toward the international target of preserving 30 percent of the planet's land by 2030. By providing a high-resolution view of the world at 10-meter resolution, ForTy offers a tool for both policy makers and conservationists.

How the AI distinguishes types

Identifying forest types from space is difficult because a forest's category often depends on its history rather than its current appearance. For example, an old planted forest might look identical to a natural one from a satellite image. To solve this, ForTy uses a multi-stage process powered by Google DeepMind's AlphaEarth Foundations model. This model compresses data from various satellites into a compact description of each pixel.

The system uses several image recognition models to build on those descriptions. The first model learns from verified examples, and its predictions are combined with external data on tree heights, buildings, and forest loss to train a subsequent model. A final processing stage looks at data from multiple years simultaneously to remove any inconsistencies. Google's Gemini language model also assisted in adjusting older training sets to align with specific EU definitions.

During the development process, human experts added about 43,000 examples to correct specific errors. These corrections addressed issues such as vineyards in Europe being misclassified as tree crops or natural pine forests in Central America being labeled as planted forests. This combination of automated AI processing and human oversight helps the model navigate complex land-use patterns that are hard to see with the naked eye.

Accuracy and regional gaps

In a test of 8,190 randomly selected sites, two experts found the map to be correct in 90.2 percent of cases. However, the researchers noted that this high score is partly due to the easy-to-identify "other land" class. The map's performance varies significantly depending on the specific type of forest being identified. It correctly identified 89 percent of primary forests but found only about half of tree crops and planted forests.

Regional errors also complicate the global picture. The map shows an excess of primary forest in Asia by roughly 200 million hectares, primarily within Russia's conifer forests. Conversely, it misses about 94 million hectares of forest in Africa, likely because it struggles with open dry forests and savannas. These discrepancies highlight the limitations of using satellite imagery alone to understand complex land management across different climates.

Some specific land uses remain very difficult for the AI to categorize accurately. For instance, the map still labels cocoa or coffee grown in the shade of a natural canopy as forest. It also struggles with selectively logged or burned forests. Because satellite images rarely show how land is managed on the ground, the researchers had to use a 40-year rotation period to separate plantations from other types of planted forests.

Availability and future goals

The ForTy dataset, its classification layer, and the test data are all freely available under a CC BY 4.0 license. Users can access these resources through Google Earth Engine. The project is part of a broader effort called Earth AI, which connects Google's mapping capabilities to its Gemini language model to help detect environmental issues like floods and droughts faster.

While the current version of ForTy is a major step forward, the researchers have plans for future updates. They intend to add a separate class for open woodlands, which the EU treats differently from forests. They also want to improve the detection of shaded crops and other managed areas where the AI currently faces uncertainty. These updates will help refine the map's utility for specific legal and conservation requirements.

The underlying AlphaEarth Foundations model was introduced in July 2025 as a kind of virtual satellite. It combines optical images, radar, laser scans, and climate simulations to describe 10-by-10-meter areas. By making these annual embeddings freely available, Google is providing the building blocks for many other large-scale mapping projects that aim to monitor the planet's health in real time.