ForTy maps forest types at 10-meter resolution
What changed
Google DeepMind presents a map of six forest and land-use classes, alongside a preliminary series tracking changes from 2017 to 2025. The added information reveals whether an area of green is natural forest, planted forest or tree crops, and how it shifts between these uses.
The takeawayAI-assisted analysis
ForTy moves beyond recording whether trees are present to comparable records of land types and transitions, allowing natural-forest protection, managed forests and restoration to be observed separately. Practical assessments still require error estimates for each class and relevant evidence from the ground.
The breakdown
Connecting a baseline to annual change
On October 22, Google DeepMind presented Forest Typology and made a preliminary annual series for 2017 to 2025 available alongside it. The 2020 map provides a baseline for comparing change; it is not a real-time picture of forests in 2026.
The associated paper was published on October 21, while the Figshare dataset records May 22 as its initial publication date. This article follows the date of the new overview, without presenting existing data or past observations as newly produced today.
SourcesBeyond tree cover: mapping the world’s forest typesGoogle DeepMind · Maxim Neumann and Charlotte Stanton · 2026-10-22Read the sourceSourcesGlobal forest typology at 10-meter resolution for forest and land-use monitoringScientific Data · Neumann et al. · 2026-10-21Read the sourceSourcesForest Typology 2020 (v1)Figshare · Forest Typology research team · 2026-05-22Read the sourceFrom canopy appearance to land use
ForTy divides the land surface into six classes: primary forests, naturally regenerating forests, planted forests, intensively managed plantation forests, tree crops and agroforestry systems, and other land. The model uses representations of satellite observations from AlphaEarth Foundations, learning the differences between classes through multiple stages of training and manually annotated data.
The paper explains that the classes reflect management history and land use, which cannot be determined from a single image of green tree cover alone. The output retains relative confidence scores for each class, rather than providing only the highest-scoring label. These scores have not undergone post-hoc probability calibration.
SourcesGlobal forest typology at 10-meter resolution for forest and land-use monitoringScientific Data · Neumann et al. · 2026-10-21Read the sourceUnderstanding what 90% accuracy measures
From a globally stratified sample of 9,100 validation sites, the paper used the 8,190 sites on which experts had reached consensus to estimate performance. Overall accuracy across the six classes was 90.2%; this does not mean that every region or forest class performed equally well.
The paper specifically notes that the other-land class has a substantial area weight and strongly influences the overall metric. The F1 score for planted forests was 58.2%, compared with 82.9% for primary forests. Assessing whether a particular class is useful requires examining its precision, recall and uncertainty. F1 combines information about false positives and missed cases; it cannot be treated as the probability that an individual pixel is correct.
SourcesGlobal forest typology at 10-meter resolution for forest and land-use monitoringScientific Data · Neumann et al. · 2026-10-21Read the sourceH/ACC perspective
Making changes in green cover meaningful
H/ACC editorial interpretation: The value of this work is a comparable record of what a piece of land changes from and to. When trees are replanted after natural forest is cleared, canopy cover may recover even though the management regime has changed. Distinguishing these classes allows restoration, planting and natural-forest protection to be assessed against their respective goals.
Retaining classification scores also helps users plan verification. They can first examine areas where the classification is uncertain, then check them against local records. If only a color-coded label is retained, that uncertainty disappears from subsequent calculations. The advance is an additional layer of information that can be checked; carbon stocks and biodiversity still require their own measurements.
Annual series need a further test
H/ACC editorial interpretation: Strong performance on a static classification is an important foundation for ongoing monitoring. Establishing when an area changed from natural forest to a plantation also requires distinguishing real change from fluctuations in classification. A different label in adjacent years can provide a lead for investigation, to be checked against a continuous sequence of images and ground records.
The most valuable next step would be published validation of transitions by year and class, with an explanation of how errors affect area estimates. That would help successive map updates form a traceable body of evidence, beyond a new set of colors each year.
Limits of the evidence
- The paper's validation metrics apply to the 2020 baseline. They cannot be applied directly to every year or type of transition in the preliminary annual series.
- Class scores are relative confidence values without post-hoc calibration. The data cover land from 65 degrees south to 84 degrees north, rather than every polar area.
- Forest management practices vary by region, and the paper notes that some classifications rely on operational approximations. We have not independently rerun the model or validated individual sites.
What to watch nextAI-assisted analysis
- Will the 2017 to 2025 series publish independent validation by year and class, especially precision and recall for transitions from natural forest to plantations?
- Can harvesting and regeneration records from the same region be aligned with consecutive annual classifications, with review results for low-confidence areas made public?
Sources
Reporting and research used in this story.
- Beyond tree cover: mapping the world’s forest typesGoogle DeepMind · Maxim Neumann and Charlotte Stanton · Source published 2026-10-22
- Global forest typology at 10-meter resolution for forest and land-use monitoringScientific Data · Neumann et al. · Source published 2026-10-21
- Forest Typology 2020 (v1)Figshare · Forest Typology research team · Source published 2026-05-22