Separating natural forests from other tree cover with AI for deforestation-free supply chains
November 13, 2025
Maxim Neumann, Research Engineer, Google DeepMind, and Charlotte Stanton, Senior Program Manager, Google Research on behalf of the broader research team
Natural Forests of the World 2020 is an AI-powered map that distinguishes natural forests from other tree cover. This critical baseline helps governments, companies, and communities meet deforestation-free goals and protect ecosystems.
—
Forests play a vital role in regulating rainfall, mitigating floods, storing and sequestering carbon, and sustaining most of the planet’s land-based species. Despite this, deforestation continues rapidly. A major challenge in conservation is differentiating centuries-old natural forests from newly planted or plantation forests using satellite data. Most existing maps only show “tree cover,” combining diverse types of woody vegetation. This conflates temporary plantations with permanent loss of biodiversity-rich natural forests.
This distinction is especially important due to new regulations like the European Union Regulation on Deforestation-free Products (EUDR), effective December 31, 2020. It prohibits products (coffee, cocoa, rubber, timber, palm oil) sold in the EU if they originate from recently deforested or degraded land, aiming to protect natural forests such as primary and naturally regenerating forests. Reliable, high-resolution global maps of natural forests in 2020 are crucial to comply. Protecting these forests is key for climate stability and is emphasized by COP30.
—
To support this need, in collaboration with Google DeepMind, we are releasing *Natural Forests of the World 2020*, a new map and dataset published in *Nature Scientific Data*. This collaboration includes the World Resources Institute and the International Institute for Applied Systems Analysis. It provides the first global, 10-meter resolution map that distinguishes natural forests from other tree cover, achieving 92.2% accuracy validated against an independent global dataset. This publicly available baseline can help companies with due diligence, assist governments in monitoring deforestation, and empower conservation groups to focus their efforts.
—
### How AI can separate the forest from the trees
Distinguishing natural forests from complex agroforestry systems or 50-year-old plantations is challenging using a single satellite image. To address this, we developed an AI model that evaluates a 1280 x 1280 meter patch over a full year. It estimates the likelihood that each 10 x 10 meter pixel is natural forest by considering temporal context.
Our multi-modal temporal-spatial vision transformer (MTSViT) model analyzes seasonal Sentinel-2 satellite imagery, topographical data (elevation, slope), and geographic coordinates to distinguish spectral, temporal, and texture patterns that differentiate natural forests from commercial plantations and other land covers.
—
### Building the Natural Forests 2020 map
We sampled over 1.2 million global patch locations to create a large training dataset and trained the MTSViT model to identify natural forests and other land types. The model was applied globally to generate a seamless 10-meter resolution probability map.
For validation, we created an evaluation dataset by updating an independent global forest management dataset from 2015 to focus on natural forests in 2020. Full details can be found in the associated published paper.
—
### What’s next: A new vision for forest understanding
We hope this baseline aids policymakers, auditors, and companies complying with deforestation-free regulations like EUDR. However, forests are dynamic ecosystems. Future efforts will classify more forest types, including primary forest, naturally regenerating forest, planted forest, plantations, tree crops, and other land covers.
We are developing a multi-year series of global forest type maps using next-generation AI models, expected to be released in 2026.
To foster community engagement, we released two large benchmark datasets:
– The *Planted* dataset offers over 2.3 million time-series examples across 64 species/genera of planted forests and tree crops worldwide.
– The *Forest Typology* (ForTy) benchmark includes 200,000 multi-source image patches with detailed per-pixel labels for mapping natural forests, planted forests, and tree crops.
—
### Helping to protect our planet
Achieving climate and nature goals requires transparent, trusted, high-resolution data. We are committed to making these tools widely accessible to support governments, companies, and communities in meeting deforestation-free goals and safeguarding critical ecosystems.
Learn more about AI and sustainability efforts at:
– [Google Earth AI](https://blog.google/technology/research/new-updates-and-more-access-to-google-earth-ai/)
– [Google Earth Engine](https://cloud.google.com/blog/topics/sustainability/look-back-at-a-year-of-earth-engine-advancements)
– [AlphaEarth Foundations](https://deepmind.google/blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/)
—
### Acknowledgments
This research was co-developed by Google DeepMind and Google Research in collaboration with the World Resources Institute and International Institute for Applied Systems Analysis.
We thank our collaborators at Google, WRI/Global Forest Watch, and IIASA: Anton Raichuk, Charlotte Stanton, Dan Morris, Drew Purves, Elizabeth Goldman, Katelyn Tarrio, Keith Anderson, Maxim Neumann, Mélanie Rey, Michelle J. Sims, Myroslava Lesiv, Nicholas Clinton, Petra Poklukar, Radost Stanimirova, Sarah Carter, Steffen Fritz, Yuchang Jiang.
Special thanks to early map reviewers: Andrew Lister (US Forest Service), Astrid Verheggen, Clement Bourgoin, Erin Glen, Frederic Achard, Jonas Fridman, Jukka Meiteninen, Karen Saunders, Louis Reymondin, Martin Herold, Olga Nepomshina, Peter Potapov, Rene Colditz, Thibaud Vantalon, Viviana Zalles.
—
### Quick links
– [Paper](https://www.nature.com/articles/s41597-025-06097-z)
– [Earth Engine data catalog](https://developers.google.com/earth-engine/datasets/catalog/projects_nature-trace_assets_forest_typology_natural_forest_2020_v1_0_collection)
– [Demo script](https://nature-trace.projects.earthengine.app/view/natural-forests-2020)
– [Dataset](https://plus.figshare.com/articles/dataset/Natural_forests_of_the_world_2020_-_probability_maps/28881731)
– [Data generation library (GitHub)](https://github.com/google-deepmind/geeflow)
– [Modeling library (GitHub)](https://github.com/google-deepmind/jeo)
—
### Other posts of interest
– Differentially private machine learning at scale with JAX-Privacy (Nov 12, 2025)
– Forecasting the future of forests with AI: From counting losses to predicting risk (Nov 5, 2025)
– Exploring a space-based, scalable AI infrastructure system design (Nov 4, 2025)
—
**Featured images available:**
– Natural forests and planted forest boundary illustration
– Global extent of natural forests in 2020
– Workflow visualization for Natural Forests map generation
**Note:** To include images or videos on WordPress, upload them to your media library and insert with appropriate captions as described above.
