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, an advanced AI-driven map, precisely delineates natural forests from other tree cover. This foundational dataset empowers governments, corporations, and communities to achieve ambitious deforestation-free targets and safeguard vital ecosystems.
Forests are indispensable global assets, regulating precipitation patterns, mitigating flood risks, acting as vital carbon sinks, and supporting the vast majority of terrestrial species. Despite their paramount importance, relentless deforestation continues to threaten these critical ecosystems. A significant hurdle in effective conservation is the precise differentiation of ancient natural forests from recently established tree plantations or agricultural land using satellite imagery. Many existing maps broadly categorize all woody vegetation as 'tree cover,' leading to misleading comparisons that conflate the loss of unique, biodiversity-rich natural forests with the managed harvesting of short-rotation plantations.
This crucial distinction is amplified by emerging global legislation, such as the European Union Regulation on Deforestation-free Products (EUDR). This regulation mandates that specified commodities, including coffee, cocoa, rubber, timber, and palm oil, traded within the EU must originate from land not subjected to deforestation or degradation post-December 31, 2020, aiming to preserve primary and naturally regenerating forests. Consequently, a precise, high-resolution, and globally standardized map of natural forest extent as of 2020 is now an urgent necessity. The protection of these invaluable forests also stands as a cornerstone objective for COP30, acknowledging their indispensable role in climate stability and human well-being.
Gemini generated image showing natural forest (left) bordering a planted forest (right). Global satellite-based models struggle to distinguish between them, complicating efforts to protect the more biodiversity-rich natural forest.
Addressing this critical requirement, Google DeepMind and Google Research are proud to introduce Natural Forests of the World 2020. This comprehensive new map and dataset, meticulously published in Nature Scientific Data, is the result of an extensive collaboration with the World Resources Institute and the International Institute for Applied Systems Analysis. It establishes an indispensable benchmark for monitoring deforestation and degradation. This groundbreaking achievement represents the first globally consistent, 10-meter resolution map capable of differentiating natural forests from other tree cover, boasting a state-of-the-art accuracy of 92.2% validated against a robust global independent dataset. We are confident that this freely accessible resource will significantly aid corporations in their due diligence processes, assist governments in robust deforestation monitoring, and empower conservation organizations to strategically direct their efforts toward protecting irreplaceable natural assets.
The global extent of natural forests in 2020 (originally at 10-meter resolution).
How AI can separate the forest from the trees
Precisely distinguishing between a pristine natural forest, a complex agroforestry system, or even a mature 50-year-old plantation presents a significant challenge using isolated satellite images. To surmount this, we have engineered an advanced AI model that functions akin to an experienced forester. It meticulously analyzes a specific land parcel over an entire year, dissecting 1280 x 1280 meter segments and calculating the probability of each 10 x 10 meter pixel belonging to a natural forest. This contextual approach allows the model to render assessments based on encompassing environmental factors, transcending the limitations of single-frame analysis. This innovative multi-modal temporal-spatial vision transformer (MTSViT) model synergistically integrates seasonal Sentinel-2 satellite imagery and detailed topographical data, including elevation and slope, with geographic coordinates. By observing satellite data across distinct temporal periods, the model discerns unique spectral, temporal, and textural signatures—critical data patterns that effectively differentiate intricate natural forests from uniform, fast-growing commercial plantations and diverse land use types.
play silent looping videopause silent looping videounmute videomute videoTo construct the Natural Forests of the World 2020 map, we meticulously curated a vast, multi-source training dataset by sampling over 1.2 million global 1280 x 1280 meter patch locations at a precise 10-meter resolution. This extensive data fueled the training of our MTSViT model, equipping it to accurately identify intricate patterns characteristic of natural forests and other land classifications. Subsequently, we deployed the fully trained MTSViT model across the Earth's landmass, yielding a seamless and globally uniform 10-meter probability map. Rigorous validation was achieved by developing a dedicated evaluation dataset, repurposing an independent global forest management dataset from 2015 and adapting its labels to specifically reflect natural forest conditions in 2020. Further details are comprehensively presented in the accompanying scientific paper.
End-to-end workflow of the Natural Forests map generation (annotating data generation, processing, model training, map generation, and validation steps).
What's next: A new vision for forest understanding
We envision the Natural Forests of the World 2020 baseline serving as an indispensable tool for policymakers, auditors, and businesses striving for compliance with stringent deforestation-free regulations, such as the EUDR. However, forests are dynamic entities. True global conservation and sustainability demand a more granular understanding, involving the precise differentiation of numerous forest classes and, critically, the tracking of their temporal changes. This entails distinguishing between and accurately mapping key forest types: dense, biodiversity-rich natural forests; established planted forests; managed plantations; and commercial tree crops, including ecologically beneficial coffee and cocoa agroforestry systems.
To propel this critical endeavor forward, we are developing a pioneering multi-year series of global forest type maps, powered by cutting-edge AI models. These forthcoming maps will meticulously classify the Earth's landmass into six distinct categories: Primary Forest, Naturally Regenerating Forest, Planted Forest, Plantation Forest, Tree Crops, and Other Land Cover. We anticipate the release of these comprehensive maps in 2026.
To foster widespread innovation within the research community, we are also releasing two large-scale benchmark datasets. These datasets are pivotal for developing and rigorously evaluating next-generation AI models dedicated to analyzing global forest ecosystems. The Planted dataset offers a global, multi-sensor, long-temporal collection encompassing over 2.3 million time-series classification examples, specifically engineered to enable AI models to identify 64 distinct types of planted forests and tree crops worldwide. Furthermore, the Forest Typology (ForTy) benchmark provides a truly global-scale dataset comprising 200,000 multi-source and multi-temporal image patches with per-pixel labels, optimized for semantic segmentation models. This invaluable resource is tailored for the fundamental task of mapping core classes: natural forest, planted forest, and tree crops.
Helping to protect our planet
Translating ambitious climate and nature objectives into tangible actions necessitates transparent, reliable, and high-resolution data. We are steadfastly committed to ensuring the broadest possible accessibility of these transformative tools. Our sincere hope is that these novel datasets and advanced tools will galvanize collaborative efforts among governments, corporations, and communities, enabling them to achieve their deforestation-free mandates and diligently protect the vital ecosystems upon which all life depends.
Explore our advancements in AI and sustainability by visiting Google Earth AI, Google Earth Engine, and AlphaEarth Foundations.
Acknowledgments
This groundbreaking research was meticulously co-developed by Google DeepMind and Google Research, in close collaboration with WRI and IIASA.
We extend our sincere gratitude to our esteemed collaborators at Google, the World Resources Institute (WRI) / Global Forest Watch (GFW), and the International Institute for Applied Systems Analysis (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 recognition is given to our early map reviewers whose insightful feedback was invaluable: Andrew Lister (United States Forest Service), Astrid Verheggen (Joint Research Centre), Clement Bourgoin (Joint Research Centre), Erin Glen (WRI), Frederic Achard (Joint Research Centre), Jonas Fridman (Swedish University of Agricultural Sciences), Jukka Meiteninen (VTT), Karen Saunders (World Wildlife Fund Canada), Louis Reymondin (Alliance Bioversity International - CIAT), Martin Herold (GFZ Helmholtz Centre for Geosciences), Olga Nepomshina (GFZ Helmholtz Centre for Geosciences), Peter Potapov (University of Maryland/WRI), Rene Colditz (Joint Research Centre), Thibaud Vantalon (Alliance Bioversity International - CIAT), and Viviana Zalles (WRI).