内容来源：https://research.google/blog/separating-natural-forests-from-other-tree-cover-with-ai-for-deforestation-free-supply-chains/

内容总结：

【科技前沿】谷歌发布全球首张AI绘制的“天然森林地图”，助力实现“零毁林”供应链

本报讯 在全球森林保护形势日益严峻、欧盟《零毁林产品法规》等新政陆续实施的背景下，谷歌深度思维与谷歌研究团队于2025年11月13日联合发布全球首张10米分辨率“2020年世界天然森林”地图。这一突破性成果通过人工智能技术精准区分天然森林与人工林，为构建“零毁林”供应链提供关键科学依据。

该地图采用新型多模态时空视觉变换器模型，通过分析哨兵2号卫星全年观测数据、地形坐标等多维信息，使AI能像专业林业工作者一样，依据植被季节性光谱特征与空间纹理，准确识别具有百年生态价值的天然森林与短期轮作的经济林。经全球独立数据集验证，其识别准确率高达92.2%，显著解决了传统卫星地图将天然森林与人工植被混为一谈的“苹果与橘子对比”难题。

研究团队表示，这项成果将直接助力企业履行欧盟新规要求——自2021年起在欧销售的咖啡、可可、橡胶等商品必须证明其原料未导致天然森林破坏。同时，该地图也为即将召开的联合国气候变化大会COP30提供了重要的森林保护基准数据。

据悉，科研团队正研发新一代全球森林类型动态监测系统，计划于2026年发布涵盖原始森林、自然再生林等六类生态系统的系列图谱。目前已同步开放230万组植树造林时序数据与20万组多源影像样本，推动全球科研机构共同完善森林精准监测网络。

专家指出，这项突破标志着AI技术正成为全球生态治理的重要工具，通过提供透明可靠的高清地图，将有效推动各国落实《巴黎协定》森林保护承诺，守护维系地球气候稳定与生物多样性的绿色命脉。

中文翻译：

利用人工智能区分天然林与其他林木覆盖，构建零毁林供应链
2025年11月13日
谷歌DeepMind研究工程师马克西姆·诺伊曼、谷歌研究院高级项目经理夏洛特·斯坦顿（代表全体研究团队）

《2020年世界天然林分布图》是一份运用人工智能技术绘制的地图，能够准确区分天然林与其他林木覆盖。这一关键基准数据将助力各国政府、企业和社区实现零毁林目标，守护生态系统。

森林通过调节降雨、缓解洪涝、储存固存碳汇以及维系地球绝大多数陆生物种，对地球生态至关重要。尽管森林作用重大，全球毁林速度仍触目惊心。保护行动面临的核心挑战在于：如何通过卫星数据区分具有数百年历史的天然生态系统与新生人工林或经济林。现有地图大多仅显示“林木覆盖”这一基础指标，导致不同森林类型被混为一谈——这无异于将短期经济林采伐与不可再生的生物多样性宝库的永久消失划上等号。

随着欧盟《零毁林产品法规》等全球新规出台，厘清森林类型的需求愈发迫切。该法规要求在欧盟销售的咖啡、可可、橡胶、木材和棕榈油等产品，其原料不得来自2020年12月31日后遭受毁林或退化的土地，旨在保护原始林和自然再生林等天然林。这一政策亟需一份可靠、高精度、全球统一的2020年天然林基准地图。这些森林的保护也是COP30的核心议题，国际社会公认其对气候稳定与人类福祉的关键作用。

为响应这一需求，我们与谷歌DeepMind合作，在《自然·科学数据》期刊发布了《2020年世界天然林分布图》及配套数据集。该项目与世界资源研究所、国际应用系统分析研究所协同开发，为毁林与退化监测提供了关键基准。我们首次实现了全球统一的10米分辨率天然林识别地图，经独立全球数据集验证准确率高达92.2%。这份公开基准数据将助力企业开展尽职调查，辅助政府监测毁林动态，赋能环保组织精准守护核心生态区。

人工智能如何实现“见林又见树”
仅凭单张卫星图像，难以区分天然林与复合农林业系统或树龄达50年的人工林。为此我们开发了模拟林业专家工作方式的AI模型：将1280×1280米地块分割为10×10米像素单元，通过持续观测地块全年变化，逐像素研判天然林概率。这种时空关联分析突破了单一时相观测的局限。新型多模态时空视觉变换器（MTSViT）模型通过整合哨兵2号卫星时序影像、地形数据（如高程与坡度）及地理坐标，识别出复杂天然林与整齐划一的速生商业林等地类在光谱特征、时序规律和纹理结构上的本质差异。

为构建2020年全球天然林地图，我们在全球范围内采样超过120万个1280×1280米地块，以10米分辨率创建海量多源训练数据集。利用该数据训练MTSViT模型掌握天然林复杂特征后，将其应用于全球陆地范围，生成无缝拼接的10米分辨率概率分布图。我们重构了2015年全球森林管理独立数据集并更新2020年天然林标签，据此建立专项评估数据集完成严格验证。更多技术细节详见论文。

未来展望：森林认知新维度
我们期待这份2020年天然林基准数据成为政策制定者、审计机构和企业应对欧盟新规的重要工具。但森林始终处于动态变化中。要真正支撑全球生态保护，需进一步细化森林分类并追踪其变迁轨迹，重点区分以下类型：天然林（碳储量高、生物多样性丰富）、人工林、种植园及经济林木（如生态友好型咖啡/可可农林系统）。

为此我们正基于新一代AI模型，研发全球森林类型多年序列地图。该系列地图将陆地划分为六类：原始林、自然再生林、人工林、种植园、经济林木和其他土地覆盖，预计于2026年发布。为推动学界共同参与，我们同步发布两大基准数据集：“人工林数据集”涵盖230万条多传感器长时间序列样本，助力AI识别全球64种人工林与经济林木；“森林类型学（ForTy）基准”提供20万组多源多时相图像斑块及像素级标签，专为天然林/人工林/经济林木的核心分类任务设计。

守护地球家园
将气候与自然承诺转化为行动，需要透明、可靠、高精度的数据支撑。我们致力于推动这些工具的广泛普及，期待新数据集与工具能促进政府、企业及社区协同行动，实现零毁林目标，守护人类赖以生存的关键生态系统。

了解更多谷歌人工智能与可持续发展项目，请访问谷歌地球AI、谷歌地球引擎及AlphaEarth基础知识平台。

致谢
本研究由谷歌DeepMind与谷歌研究院联合世界资源研究所、国际应用系统分析研究所共同完成。
特别感谢谷歌、世界资源研究所/全球森林观察、国际应用系统分析研究所的协作同仁：安东·赖丘克、夏洛特·斯坦顿、丹·莫里斯、德鲁·珀维斯、伊丽莎白·高德曼、凯特琳·塔里奥、基思·安德森、马克西姆·诺伊曼、梅拉妮·雷、米歇尔·J·西姆斯、米罗斯拉娃·莱西夫、尼古拉斯·克林顿、佩特拉·波克卢卡尔、拉多斯特·斯塔尼米罗娃、莎拉·卡特、史蒂芬·弗里茨、蒋玉常。

诚谢早期地图审阅专家（按机构字母排序）：
美国林务局安德鲁·利斯特、联合研究中心阿斯特里德·费尔赫芬/克莱门特·布罗尼/弗雷德里克·阿沙尔/雷内·科尔迪特、世界资源研究所艾琳·格伦/维维安娜·扎勒斯、瑞典农业科学大学乔纳斯·弗里德曼、芬兰国家技术研究中心尤卡·梅蒂宁、世界自然基金会加拿大分会卡伦·桑德斯、国际生物多样性联盟-国际热带农业中心路易斯·雷蒙丁/蒂博·万塔兰、德国地学研究中心马丁·赫罗德/奥尔加·涅波姆希纳、马里兰大学/世界资源研究所彼得·波塔波夫。

英文来源：

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 are vital for our planet as they regulate rainfall, mitigate floods, store and sequester carbon, and help sustain the majority of the planet’s land-based species. Despite their importance, deforestation continues at an alarming rate. A key challenge in conservation efforts is differentiating centuries-old natural ecosystems from newly planted forests or tree crop plantations with satellite data. Most existing maps simply show "tree cover," a basic measure of any woody vegetation, leading to an "apples-to-oranges" comparison. This conflates the harvesting of a short-term plantation with the permanent loss of an irreplaceable, biodiversity-rich natural forest.
The need for this distinction is more important than ever due to new global regulations, like the European Union Regulation on Deforestation-free Products (EUDR). This regulation mandates that products like coffee, cocoa, rubber, timber, and palm oil sold in the EU cannot come from land that was deforested or degraded after December 31, 2020, with the goal of protecting natural forests, like primary and naturally regenerating forests. This policy creates a need for a reliable, high-resolution, and globally-consistent map of natural forests as they existed in 2020. The protection of these forests is also a central pillar for COP30, which recognizes their crucial role in climate stability and human well-being.
In an effort to help meet this need, together with Google DeepMind, we’re releasing Natural Forests of the World 2020, a new map and dataset, published in Nature Scientific Data. This project stems from a collaboration with the World Resources Institute and the International Institute for Applied Systems Analysis, and provides a critical baseline for deforestation and degradation monitoring. We provide the first globally consistent, 10-meter resolution map that differentiates natural forests from other tree cover and achieves a best-in-class accuracy of 92.2% when validated against a global independent dataset. We hope that this publicly available baseline can help companies conduct due diligence, support governments in monitoring deforestation, and empower conservation groups to target their efforts to protect what matters most.
How AI can separate the forest from the trees
Distinguishing a natural forest from a complex agroforestry system or a 50-year-old planted forest is difficult using a single satellite image. To overcome this, we developed an AI model that acts like a forester, observing a patch of land over the course of a year, segmenting a 1280 x 1280 meter patch and estimating the likelihood that each 10 x 10 meter pixel within it is a natural forest. This allows the model to make assessments based on the surrounding context, rather than a single snapshot. This novel multi-modal temporal-spatial vision transformer (MTSViT) model analyzes seasonal Sentinel-2 satellite imagery and topographical data (e.g., elevation and slope), along with the sample’s geographical coordinate. By observing satellite imagery over time, the model identifies distinct spectral, temporal, and texture signatures (i.e., data patterns used to recognize different forest types) that differentiate complex, natural forests from uniform, fast-growing commercial plantations and other land use and land cover.
To build the Natural Forests of the World 2020 map, we sampled over 1.2 million global 1280 x 1280 meter patch locations at 10-meter resolution to create a massive, multi-source training dataset. We used this data to train the MTSViT model to recognize complex patterns of natural forests and other land types. We then applied the trained MTSViT model across all land on Earth, generating a seamless, globally consistent 10-meter probability map. To rigorously validate the map, we created an evaluation dataset by repurposing an independent dataset focused on global forest management for 2015 and updating its labels to focus on natural forests for 2020. See more details in the paper.
What's next: A new vision for forest understanding
We hope that the Natural Forests of the World 2020 baseline proves to be a valuable resource for policymakers, auditors, and companies seeking to comply with new deforestation-free regulations such as the EUDR. But forests are not static. To truly support global conservation and sustainability, we need to distinguish between more classes of forest and, crucially, understand how they change over time. This involves differentiating between and locating key forest types: natural forests (carbon-dense and biodiversity-rich forests), planted forests, plantations, and commercial tree crops (such as ecosystem-friendly coffee and cocoa agroforestry systems).
To advance this effort, we’re developing a new multi-year series of global forest type maps, powered by next-generation AI models. These maps will categorize the world's land into six distinct types: Primary Forest, Naturally Regenerating Forest, Planted Forest, Plantation Forest, Tree Crops, and Other Land Cover. We expect to release these comprehensive maps in 2026.
To encourage the broader research community to contribute to this effort, we have also released two large-scale benchmark datasets. These datasets are important for developing and rigorously testing the next generation of AI models designed to analyze the world’s forests. The Planted dataset is a global, multi-sensor long-temporal collection featuring over 2.3 million time-series classification examples. It is specifically designed to help AI models recognize 64 different (species or genera) types of planted forests and tree crops worldwide. The Forest Typology (ForTy) benchmark provides a truly global-scale dataset with 200,000 multi-source and multi-temporal image patches with per-pixel labels for semantic segmentation models. This resource is tailored for the core task of mapping the key classes: natural forest, planted forest, and tree crops.
Helping to protect our planet
Turning climate and nature ambitions into action requires transparent, trusted, and high-resolution data. We are committed to making these tools as accessible as possible. We hope these new datasets and tools will help governments, companies, and communities work together to meet their deforestation-free goals and protect the critical ecosystems on which we all depend.
Learn more about our AI and sustainability efforts by checking out Google Earth AI, Google Earth Engine, and AlphaEarth Foundations.
Acknowledgments
This research was co-developed by Google Deepmind and Google Research in collaboration with WRI and IIASA.
We thank our collaborators at Google, World Resources Institute (WRI) / Global Forest Watch (GFW), and 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 thanks to early map reviewers: 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).