An AI forest map is challenging a familiar shortcut: treating every green patch on a satellite image as though it were the same kind of forest. A new global research dataset distinguishes natural forests, managed woodlands, plantations and tree crops, because the type of trees and how the land is used can matter as much as whether the ground looks covered. The distinction could help regulators, researchers and businesses ask better questions about land change.

The project, called Forest Typology or ForTy, was described by Google DeepMind and its research partners in October 2026. It maps land in squares measuring ten metres by ten metres, using a 2020 baseline and a classification scheme built around six categories. The output is a modelled map, not an on-the-ground census. Each classification represents an inference from observations and reference data, with uncertainty that varies by place and forest type.

What the AI forest map can actually distinguish

The six classes are primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land. The final category is important: it includes places that are not forest at all. It would therefore be misleading to say that the system identifies six different kinds of forest. It actually distinguishes five forest or tree-related categories plus everything else.

Primary forest in this classification is native forest without visible signs of significant human disturbance. Naturally regenerating forest has grown back without being planted, perhaps following logging or fire. Planted forests may be managed over relatively long periods, whereas plantations are often managed intensively for production. Tree crops and agroforestry include activities such as fruit growing or other farming with trees. These distinctions reflect land use as well as appearance.

Someone looking down at a plantation and a naturally regenerating forest might see two green areas. Yet they can differ in ecological structure, management and the pressures that affect them. A simple tree-cover measurement cannot tell the whole story. That is why mapping land management, even imperfectly, could prove useful for decisions about conservation and supply chains.

A satellite picture does not reveal everything

The central challenge is that some categories look very similar from above. A plantation can become visually complex, while a recovering forest can show regular patterns left by human activity. A canopy image also conceals what is happening beneath the leaves. Researchers therefore need evidence beyond a single photograph, such as seasonal observations, growth patterns and known reference examples.

The project combines satellite-derived information with a large set of labelled training locations. DeepMind says the training effort used approximately 1.7 million labelled sites drawn from several kinds of reference material and targeted checks. One input is AlphaEarth Foundations, which creates compact descriptions from repeated Earth observations. Such information may reveal patterns that would be difficult to see in one snapshot, but it does not make the computer infallible.

LiveAIWire has previously explored how AI satellite representations can help map the planet. The new forest project applies that wider idea to a particular policy and environmental problem. It needs to distinguish not merely whether land has trees, but whether those trees belong to a type that matters under a particular conservation or sourcing rule.

What the published accuracy figures do and do not mean

The peer-reviewed Scientific Data paper reports overall accuracy of approximately 90.2% for the six-class scheme in its validation assessment. That sounds reassuring, but a single headline percentage can hide very different results across categories. The study itself warns readers to examine performance for the individual classes relevant to their intended use.

Planted forest was particularly difficult to distinguish, with a reported F1 score of 58.2%. F1 is a measure combining the system’s ability to find a category and avoid incorrectly assigning it. By contrast, the dominant ‘other’ land class achieved a much higher score. That contrast demonstrates why a strong overall result does not mean every plantation, recovering woodland or agricultural tree area has been correctly labelled.

The problem also changes from one continent to another. Some land types are uncommon in particular regional reference samples, making local accuracy difficult to estimate confidently. Differences between a modelled map and another dataset do not automatically reveal which source is wrong. They may point to a classification mismatch, a different year, an unusual landscape or insufficient observations.

For a general reader, the practical lesson is simple. A coloured map can reveal broad patterns and identify places worth investigating. It should not be used as conclusive proof about an individual property without checking the original data, relevant definitions and local evidence. The smaller and more consequential the decision, the less comfortable people should be with an unexplained pixel label.

Why businesses and regulators may care

Environmental rules increasingly ask companies to understand the origins of products linked to land use. It is not enough to know that a farm, plantation or forest exists today. A buyer may need to establish what was there before, whether land use changed and which legal definitions apply. A global classification can help narrow the search and make data more consistent between regions.

But a map does not replace documentary proof or legal judgement. It cannot independently establish who owns land, whether a particular producer followed local requirements or when every change occurred. The evidence needed for an individual supply chain may include farm locations, historical records, government information and independent field assessments. A model can support that work while leaving the final responsibility with people and organisations.

The same distinction applies to public environmental policy. An area recorded as planted forest may still provide important habitat or economic value. A primary forest may require particular protection. Neither classification alone provides a complete moral, ecological or legal verdict. Land cover, biodiversity, community rights and human livelihoods are connected but separate questions.

A baseline is not a live forest monitor

The core ForTy map describes forest types for 2020. The project also discusses work with annual information covering subsequent and surrounding years. That does not mean every square in a browser reflects the current day, nor does it prove that all historical changes can be reconstructed without error. The date of the underlying observation is as important as the colour assigned to a location.

That timing matters after a fire, storm or harvest. A person looking at a landscape in 2026 should not assume a 2020 map describes its present condition. New imagery or annual datasets may help, but only if the particular layer and its date are identified correctly. Environmental monitoring can become misleading when maps with different reference periods are compared as though they were simultaneous.

LiveAIWire has covered another AI effort to identify changes across the Earth. That research addresses the detection of change, whereas ForTy asks a different question about the type and management of vegetation. Both can contribute to understanding landscapes, but neither should be mistaken for a complete, real-time account of the ground.

What a better map might change

The most valuable consequence may be improved questions. Instead of asking how much green area remains, a researcher can ask whether it is natural, recovering or intensively managed. A regulator may identify locations needing closer inspection. A company may notice that a supply chain depends on areas whose classification differs from its assumptions. Those uses are promising because they focus attention where additional evidence matters.

For anyone comparing two places on the map, the appropriate reference may be a local forestry inventory, official land-use record or carefully collected field observations. These sources can use different definitions from the six classes in the research. A disagreement may reflect a real change on the ground, but it may also reflect the way each dataset defines woodland, crops or managed plantations. Careful comparison requires matching the dates and definitions before accusing either source of being wrong.

There are also limits to what the underlying categories describe. Two naturally regenerating forests may differ hugely in age, species composition and ecological condition. A planted forest is not automatically harmful, and green cover is not a substitute for biodiversity. A map designed for forest typology provides a specific lens, not a universal measure of environmental health.

LiveAIWire has reported on satellite analysis used to uncover otherwise hard-to-see fishing activity. The connection is methodological rather than a claim that the projects share data: remote sensing can reveal patterns that traditional observation misses, while still requiring validation and sensible interpretation of uncertainties.

ForTy is therefore a useful step beyond a binary picture of trees and no trees. Its significance lies in the possibility of distinguishing land uses that matter to real environmental decisions. The published limitations are part of the result, not an embarrassment to hide. A forest map becomes more trustworthy when people understand what each colour means, when it was measured and where the model may still be wrong.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.