Google DeepMind, Google Research and the World Resources Institute (WRI) have developed a new artificial intelligence-powered map that classifies forests and other land areas across the world using satellite data. Called Forest Typology (ForTy), the project identifies six different land-cover categories at a resolution of 10 metres, helping researchers distinguish natural forests from planted forests, commercial plantations and agricultural tree crops.
The map is designed to address a major limitation in conventional satellite mapping: seeing trees does not necessarily mean identifying a natural forest. A dense natural rainforest, a commercial timber plantation and an oil palm farm can look similar from space, even though they have very different environmental and ecological value. By combining AI models with years of satellite observations, ForTy aims to provide governments, researchers and businesses with more detailed information about how forests are used and how they change over time.
Key Takeaways
- Google DeepMind and the World Resources Institute have developed a global forest classification map called ForTy.
- The map classifies land into six categories at 10-metre resolution.
- It uses Google’s AlphaEarth Foundations model and multiple satellite data sources.
- The system achieved 90.2% overall accuracy in an evaluation involving 8,190 randomly selected sites.
- It helps distinguish primary forests, naturally regenerating forests, planted forests, plantations and tree crops.
- The project supports deforestation monitoring and compliance with the European Union’s deforestation regulation.
- The map and supporting data are freely available under a CC BY 4.0 licence.
What Is Google’s ForTy Global Forest Map?
ForTy stands for Forest Typology. It is an AI-powered geospatial mapping system developed by Google DeepMind and Google Research in collaboration with the World Resources Institute and the International Institute for Applied Systems Analysis (IIASA).
The system classifies the world’s land into six categories:
| Forest or land category | Description |
|---|---|
| Primary forest | Native forest with no visible signs of human disturbance |
| Naturally regenerating forest | Forest that has grown back naturally after events such as logging or fire |
| Planted forest | Forest established through human planting or seeding and managed over longer periods |
| Plantation forest | Intensively managed tree farms, often consisting of fast-growing species |
| Tree crops and agroforestry | Agricultural land growing crops such as oil palm, rubber, cocoa and fruit |
| Other land | Grasslands, croplands, urban areas, water and other non-forest land |
The distinction matters because different types of tree-covered land provide different levels of biodiversity, carbon storage and ecological protection.
For example, a natural rainforest supports complex ecosystems and wildlife habitats, while a commercial plantation may primarily be designed to produce timber. Conventional satellite maps can identify tree cover but may struggle to distinguish between these different land uses.
ForTy attempts to solve this problem by examining patterns across multiple years rather than relying on a single satellite image.
How Google Uses AI and Satellite Data to Map Forests
The system is built using Google’s AlphaEarth Foundations, an AI model introduced in July 2025 to analyse large volumes of Earth observation data.
AlphaEarth combines information from several sources, including:
- Optical satellite imagery.
- Radar observations.
- Three-dimensional laser mapping.
- Climate simulations.
- Other geospatial datasets.
It converts these observations into compact digital representations of individual 10-by-10-metre areas of land. These representations help AI systems identify patterns that may not be obvious in individual satellite images.
ForTy builds additional image-recognition models on top of these representations to identify different forest categories.
The system uses several stages of AI processing. An initial model learns from verified forest examples. Its predictions are combined with external information, including tree height, buildings and forest-loss data, to improve training for subsequent models.
A final stage examines information across multiple years to reduce inconsistent classifications.
Google also used Gemini to help adapt older training data to the definitions required for the European Union’s deforestation regulations.
Researchers added approximately 43,000 additional examples during targeted correction rounds to address errors involving vineyards, pine forests and oil palm plantations.
ForTy Achieves 90.2% Overall Accuracy
Google and its research partners evaluated the map using randomly selected locations across the world.
Two experts reviewed each of 8,190 sites using high-resolution imagery and historical satellite observations to determine the correct land classification.
The evaluation produced an overall accuracy of 90.2%.
However, the researchers caution that this headline figure does not mean every forest category is identified with equal reliability.
| Forest category | Reported identification performance |
|---|---|
| Primary forests | 89% |
| Tree crops | Approximately 50% |
| Planted forests | Approximately 50% |
| Plantation forests | Approximately 60% |
| Overall classification accuracy | 90.2% |
The system performed particularly well at identifying primary forests, but struggled more with plantations, planted forests and agricultural tree crops.
These categories are difficult to distinguish because their appearance can change with harvesting, planting schedules and agricultural management.
The researchers also noted that some errors were related to the limitations of training data and differences between regional forest-management practices.
ForTy nevertheless outperformed European Commission maps on the same evaluation dataset, according to the research team’s comparison.
Why the Map Matters for EU Deforestation Regulations
One of the major motivations behind ForTy is the European Union Regulation on Deforestation-free Products (EUDR).
The regulation requires companies dealing in certain commodities, including coffee, cocoa, palm oil, rubber, timber and soy, to meet deforestation-related requirements.
A key reference date is December 31, 2020. Businesses need to establish whether relevant land was forested at that time and whether subsequent deforestation or degradation has occurred.
This creates a significant data challenge.
A satellite image showing trees today cannot necessarily reveal whether the same land was a natural forest, a timber plantation or an agricultural tree farm in the past.
ForTy provides a detailed 2020 reference map that can help businesses, regulators and researchers examine land-use history.
It could support supply-chain checks, deforestation-risk assessments and conservation planning. However, its classification results should be treated as one source of evidence rather than a substitute for all required due diligence, particularly in areas where the model has lower accuracy.
What the Map Reveals About Global Forests
The research team used ForTy to estimate the distribution of different forest categories worldwide.
Its estimates suggest that the world contains approximately 4.1 billion hectares of forest, including nearly 3.9 billion hectares of natural forest.
The map also estimates:
- Around 2 billion hectares of primary forest.
- Approximately 250 million hectares of planted forests.
- Nearly 200 million hectares of tree crops.
The research highlights significant differences between countries and regions.
Southeast Asia has substantial areas of plantations and tree crops, while planted forests account for a large share of forest areas in Sweden and Finland.
The researchers estimate that roughly two-thirds of Sweden’s forest area is planted, while natural forest accounts for around one-fifth.
This illustrates why total tree cover alone is not a sufficient measure of forest conservation. A country can have extensive tree cover while having a comparatively small area of natural forest.
The map also identifies regional classification weaknesses. It overestimates primary forest in parts of Asia, particularly Russia’s conifer forests, while missing some forest areas in Africa, including open dry forests and savannas.
ForTy Also Tracks Forest Changes From 2017 to 2025
Beyond the 2020 global map, Google DeepMind and its partners have released an early time series covering the years 2017 to 2025.
This allows researchers to examine not only where forest cover has changed but also what type of land replaced it.
For example, a natural forest may be cleared and replaced by a commercial timber plantation. Conventional tree-cover monitoring could show that trees are still present after the change, potentially hiding the loss of a natural ecosystem.
ForTy’s time series aims to make these transitions more visible.
It could help conservation organisations distinguish between natural forest recovery, plantation expansion and the conversion of forests into agricultural land.
The dataset is particularly relevant for long-term environmental monitoring because forest classification can change over time as land is cleared, replanted or allowed to regenerate.
The Bigger Picture
Google’s ForTy project demonstrates how AI is expanding satellite-based environmental monitoring beyond basic land-cover detection. By combining historical imagery, machine learning and geospatial information, researchers can examine not just where trees exist but also what kind of forests they form and how land use changes over time.
This distinction is important for biodiversity protection, climate policy and deforestation-free supply chains. However, the map’s weaker performance in identifying plantations and tree crops shows that AI-generated environmental data still requires expert validation and regional improvements.
Looking Ahead
Google DeepMind and its research partners plan to continue improving the ForTy dataset and its time-series information. Adding more accurate classifications for open woodlands, agricultural tree crops and managed forests could make the system more useful for conservation and regulatory applications.
The availability of the map through Google Earth Engine also creates opportunities for governments, scientists and businesses to build their own analysis tools. Its long-term value will depend on how accurately it can track changes, how frequently the data is improved and whether organisations can combine its findings with reliable field-level information.
Frequently Asked Questions
1. What is Google’s ForTy forest map?
ForTy, or Forest Typology, is an AI-powered global map developed by Google DeepMind, Google Research and partners. It classifies forests and other land into six categories using satellite data.
2. What is the resolution of the ForTy map?
The map uses a 10-metre resolution, meaning each pixel represents a 10-by-10-metre area of land.
3. How accurate is Google’s forest mapping system?
The research team reported 90.2% overall accuracy across 8,190 evaluated sites. However, performance varies considerably between forest categories.
4. What is AlphaEarth Foundations?
AlphaEarth Foundations is Google’s AI model that combines satellite imagery, radar, laser mapping and climate data into compact representations of Earth’s land and coastal areas.
5. Why does the map distinguish natural forests from plantations?
Natural forests generally support different levels of biodiversity and ecological functions compared with managed plantations. Identifying the difference helps conservation organisations and businesses monitor deforestation more accurately.
6. How does ForTy help with EU deforestation regulations?
It provides a detailed 2020 land-use reference that can help companies assess whether commodity production is linked to land that was forested and subsequently converted or degraded.
7. Is Google’s ForTy forest map publicly available?
Yes. Google and its partners have made the map, classification layer and supporting evaluation data available under a CC BY 4.0 licence, including through Google Earth Engine.
8. Does the map track forest changes over time?
Yes. An early time series covers 2017 to 2025, showing changes in forest categories and land use across different years.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
