兹Google Deepmind用卫星数据和AI绘制全球森林地图

兹Google Deepmind用卫星数据和AI绘制全球森林地图

关键点

- Google Deepmind和世界资源研究所绘制了2020年土地覆盖情况全球卫星地图，将分辨率10米的地区分为6类，包括原始森林、种植园和树木作物。

- 该系统利用AlphaEarth基金会和多个图像识别阶段，根据生长和收获周期区分森林类型，而不仅仅是外观。

- 地图的总体准确度达到90.2%，并且在同一测试数据上超过了欧洲联盟委员会的地图，但只确定了约一半的树木作物和人工林以及60%的种植园。

Google Deepmind和世界资源研究所开发了ForTy，这是一张AI地图，它区别了世界各地的原始森林，木材种植园和树木作物农场。

卫星地图通常显示树木是否在特定地区生长，而不指明森林或作物类型。 ForTy，简称森林类型学，更进一步，在2020年以10米分辨率将世界陆地面积划为6个等级。 这些森林包括原始森林、自然再生森林、人工林、种植园林、油棕榈、橡胶或可可等树木作物和其他土地。 作者说ForTy是同类的第一张全球地图。 早期的数据集，包括来自欧盟委员会的数据集，只覆盖世界部分地区。

欧盟的毁林规则需要一个可靠的2020年参考点

该项目的主要驱动力是欧盟的无毁林产品条例(EUDR)，该条例禁止2020年12月31日之后与清理或退化的森林有关的商品，包括原生森林转换为种植园后生产的商品。 油棕榈和可可农场算作农业用地，因此评估供应链风险的公司需要知道一个场地在2020年是如何使用的。

ForTy给了他们2020年的参考点，还可以为其他依赖相同数据的举措提供参考，包括热带永远森林基金(TFFF)和联合国到2030年保护世界30%土地的目标。

建立这个参照点很困难，因为森林的类型取决于它的历史，而不只是它看起来的样子，DeepMind 称。 古老的人工混合森林可以看起来像一个天然的天然森林，尽管卫星图像显示裸露的地面，但最近清理的种植园从定义上来说仍然是森林。

ForTy使用三个模型阶段来完善其分类

为了区分这些类别，ForTy运行在Google DeepMind的AlphaEarth基础模型上，将多颗卫星的数据压缩成每个像素的紧凑描述。 一些图像识别模型以这些描述为基础，从一个从经核实的例子中吸取教训开始。

第一个模型的预测，加上关于树木高度、建筑物和森林损失的外部数据，成为第二个模型的扩大培训数据。 最后一步将数年的数据放在一起分析，以平缓异常。 Google的Gemini语言模型也帮助了旧的培训数据适应欧盟的定义。

在这些修正回合中，专家们增加了约43 000个例子，以解决错误，例如欧洲的葡萄园被归类为树木作物、中美洲的天然松林被标记为人工林、非洲和印度尼西亚漏掉的油棕榈。

即使作了这些修改，对卫星来说，仍然无法作出某些区分。 卫星图像很少显示如何管理土地，因此研究人员使用40年的轮转期将种植园与其他人工林分开，这种截断并不适合每个地区。

ForTy对种植园和树木作物的可靠性较低

为了检查地图，两名专家对8 190个随机选址进行了审查，发现在90.2%的案例中正确无误。 作者警告说，这一总体得分从容易识别的"其他土地"班级得到了很大的提升。

根据森林类型，结果大不相同。 地图正确确定了89%的原始森林，专家在实地证实，但只发现了约一半的树木作物和人造林以及60%的种植园。 翻转镜头，大约四分之一到三分之一的植树场或树木作物地图被错划，三分之一的被标记为人工林的地图是错误的。

这些差距影响到对欧洲发展成果评估最重要的界限，评价中的不确定性也可能使结果看起来略好一些。 专家们往往对所管理地区不放心，小组在评估中排除了这些不确定的情况。

根据同样的测试数据，ForTy仍然优于欧洲联盟委员会专门为EUDR绘制的地图，得分95.5%，而森林与非森林分类为93.6%，森林类型为90.9%，而森林类型为87.8%。

该模型还从其培训数据中继承了一个弱点，其注释者很少将道路或定居点附近的森林归类为原始森林，即使这些森林似乎没有受到影响。

瑞典有许多树木，但天然森林很少

研究者估计，世界约有41亿公顷的森林，包括近39亿公顷的天然森林，其中约20亿公顷为原始森林。 植树造林面积约2.5亿公顷，树木作物面积近2亿公顷，主要分布在亚洲。

区域错误使情况复杂化。 该地图高估亚洲原始森林约2亿公顷，主要分布在俄罗斯的针叶林中。 在非洲，它错失了大约9 400万公顷的森林，这可能是由于开放的干燥森林和草原。

瑞典约三分之二的森林面积是种植的，天然林仅覆盖全国的五分之一左右。 作者说，总的树木覆盖面积可能足够，但自然森林所占的份额可能非常低，而自然森林对生物多样性很重要。 刚果民主共和国、印度尼西亚和巴西的天然森林份额最高。

ForTy仍然与荫蔽的作物和开阔的林地斗争

除了这些区域错误之外，一些土地使用仍然难以分类。 在一些区域，地图上仍将天然树冠下种植的荫蔽可可或咖啡标为森林，有选择地砍伐或烧毁的森林也造成了问题。

ForTy还没有为开阔的林地单独设置一个类，EUDR对此分别处理，尽管研究者表示未来的版本应该增加一个。 用户可以使用数据集对每个类的置信分数来进行自己的分类选择，这样可以让他们设定自己的阈值。

地图，分类层和测试数据在CC BY 4.0许可下可自由获取，包括在Google地球引擎中。 Google DeepMind拥有GeeFlow和JEO寄存器中的代码发布。

Deepmind 推出 2025年7月基础的AlphaEarth基金会模型是一类虚拟卫星。 它将光学卫星图像，雷达，激光扫描，以及气候模拟结合成每个10乘-10米区域的紧凑描述，每年的嵌入量覆盖自2017年以来的数年，可免费使用Google地球引擎。

ForTy是最早在这些嵌入物上建造的大型测绘项目之一，Google 称是地球AI的一部分。 由于Google在2025年秋天将地球AI与Gemini连接起来，以帮助探测洪水、干旱和藻类的生长速度更快，因此这一连接将该项目置于更广泛的努力之中。

AI News Without the Hype – 由人类治愈

订阅THE DECODER免费阅读，每周AI通讯，我们独家的"AI雷达"前沿报告每年6次，全面存档访问，访问我们的评论部分。

现在订阅

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Google Deepmind creates a global map of forests with satellite data and AI

Key Points

- Google Deepmind and the World Resources Institute created a global satellite map of 2020 land cover, classifying areas at 10-meter resolution into six categories, including primary forest, plantations, and tree crops.

- The system uses AlphaEarth Foundations and multiple image-recognition stages to distinguish forest types by growth and harvest cycles, not appearance alone.

- The map achieved 90.2% overall accuracy and outperformed European Commission maps on the same test data, but identified only about half of tree crops and planted forests and 60% of plantations.

Google Deepmind and the World Resources Institute have developed ForTy, an AI map that distinguishes primary forests, timber plantations, and tree crop farms worldwide.

Satellite maps usually show whether trees grow in a given area without identifying the forest or crop type. ForTy, short for Forest Typology, goes further by assigning the world's land area in 2020 to 6 classes at 10-meter resolution. These cover primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops such as oil palm, rubber, or cocoa, and other land. The authors say ForTy is the first global map of its kind. Earlier datasets, including one from the European Commission, cover only parts of the world.

EU deforestation rules need a reliable 2020 reference point

The main driver behind the project is the EU Regulation on Deforestation-free Products (EUDR), which prohibits goods linked to forests cleared or degraded after December 31, 2020, including goods produced after primary forest was converted to a plantation. Oil palm and cocoa farms count as agricultural land, so companies assessing supply chain risks need to know how a site was used in 2020.

ForTy gives them that 2020 reference point and could also feed into other initiatives that depend on the same kind of data, including the Tropical Forest Forever Facility (TFFF) and the UN goal of protecting 30 percent of the world's land by 2030.

Building that reference point is hard because a forest's type depends on its history, not just how it looks, DeepMind says. An old planted mixed forest can look like a natural one from space, and a recently cleared plantation still counts as forest by definition even though satellite images show bare ground.

ForTy uses three model stages to refine its classifications

To distinguish these categories, ForTy runs on Google DeepMind's AlphaEarth Foundations model, which compresses data from multiple satellites into a compact description of each pixel. Several image recognition models build on those descriptions, starting with one that learns from verified examples.

That first model's predictions, combined with outside data on tree heights, buildings, and forest loss, become expanded training data for a second model. A final step analyzes several years of data together to smooth outliers. Google's Gemini language model also helped adapt older training data to the EU's definitions.

During those correction rounds, experts added about 43,000 examples to address errors such as vineyards in Europe classified as tree crops, natural pine forests in Central America labeled as planted forests, and oil palms missed in Africa and Indonesia.

Even with those corrections, some distinctions remain out of reach for satellites alone. Satellite images rarely show how land is managed, so the researchers use a 40-year rotation period to separate plantations from other planted forests, and that cutoff does not fit every region.

ForTy is less reliable for plantations and tree crops

To check the map, 2 experts reviewed each of 8,190 randomly selected sites, finding it correct in 90.2 percent of cases. The authors warn that this overall score gets a large boost from the easy-to-identify "other land" class.

Results vary sharply by forest type. The map correctly identified 89 percent of primary forests that experts confirmed on the ground but found only about half of tree crops and planted forests and 60 percent of plantations. Flipping the lens, roughly a quarter to a third of areas mapped as plantations or tree crops turned out to be misclassified, and the map was wrong for a third of areas labeled as planted forest.

These gaps hit the boundaries that matter most for the EUDR, and uncertainty in the evaluation may also make the results look slightly better than they are. Experts were most often unsure about managed areas, and the team excluded those uncertain cases from its assessment.

ForTy still outperforms European Commission maps made specifically for the EUDR on the same test data, scoring 95.5 percent compared with 93.6 percent for forest versus nonforest classification, and 90.9 percent compared with 87.8 percent for forest types.

The model also inherited a weakness from its training data, whose annotators rarely classified forests near roads or settlements as primary forest, even when the forests appeared untouched.

Sweden has plenty of trees but little natural forest

The researchers estimate that the world has about 4.1 billion hectares of forest, including nearly 3.9 billion hectares of natural forest, of which about 2 billion hectares are primary forest. Planted forests cover about 250 million hectares, and tree crops cover nearly 200 million hectares, mostly in Asia.

Regional errors complicate that picture. The map overestimates primary forest in Asia by about 200 million hectares, largely in Russia's conifer forests. In Africa, it instead misses about 94 million hectares of forest, likely because of open dry forests and savannas.

About two-thirds of Sweden's forest area is planted, and natural forest covers only around one-fifth of the country. The authors say overall tree cover there may be adequate, but the share of natural forest, which matters for biodiversity, could be critically low. The Democratic Republic of the Congo, Indonesia, and Brazil have the highest shares of natural forest.

ForTy still struggles with shaded crops and open woodlands

Beyond these regional errors, some land uses remain difficult to classify. In some regions, the map still labels shaded cocoa or coffee grown beneath a natural canopy as forest, and selectively logged or burned forests also pose problems.

ForTy does not yet have a separate class for open woodlands, which the EUDR treats separately, though the researchers say future versions should add one. Users can make their own classification choices using the dataset's confidence scores for each class, which let them set their own thresholds.

The map, classification layer, and test data are freely available under a CC BY 4.0 license, including in Google Earth Engine. Google DeepMind has released the code in the GeeFlow and JEO repositories.

Deepmind introduced the underlying AlphaEarth Foundations model in July 2025 as a kind of virtual satellite. It combines optical satellite images, radar, laser scans, and climate simulations into a compact description of each 10-by-10-meter area, with annual embeddings covering years since 2017 freely available in Google Earth Engine.

ForTy is one of the first large mapping projects built on those embeddings and, Google says, is part of Earth AI. That connection places the project within a broader effort, as Google connected Earth AI to Gemini in fall 2025 to help detect floods, droughts, and algal blooms faster.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

Subscribe now

Auto-fetched and translated by AI Atlas. All rights belong to the original author.