Google DeepMind, Google Research, the World Resources Institute (WRI), and the International Institute for Applied Systems Analysis (IIASA) have released Forest Typology (ForTy) 2020: dividing global land into a 10×10 meter grid classified into six categories—primary forests, naturally regenerating forests, plantations, fast-growing timber forests, economic forests and agroforestry (oil palm, rubber, cocoa, etc.), and other land—while also releasing an early time series of annual changes from 2017 to 2025, which reveals whether natural forests have been converted to plantations or farms. The model uses satellite embeddings from AlphaEarth Foundations as input, and the paper was published in Scientific Data on 10 21.
Confidence: High (95) · Event time 2026-10-22
Reasoning
- The EU Deforestation Regulation (EUDR) requires that coffee, cocoa, palm oil, and other commodities do not come from agricultural land converted from forest after 2020, and that wood products do not come from plantations established on primary forests. Companies and regulators need to know the forest type of each plot in 2020, but most global maps only show "whether trees are present."
- Officials state that experts verified multiple random points across 9000, with forest type accuracy at approximately 90% and forest/non-forest accuracy at approximately 95%, outperforming other global forest maps; each pixel provides confidence scores for each of the six categories, allowing users to adjust thresholds based on their use case, such as selecting only high-confidence pixels for compliance reviews.
Usage
- The map and validation data are freely available on Google Earth Engine and Figshare under a CC BY 4.0 license; training, inference, and evaluation code are available on GitHub at google-deepmind/jeo and google-deepmind/geeflow. The DeepMind blog post "Beyond tree cover" includes an online map and time series browser.
Risks
- According to the paper and The Decoder, economic forests and plantations are only recognized at about 50%, and fast-growing timber forests at about 60%; one-quarter to one-third of areas labeled as fast-growing timber forests or economic forests are misclassified—precisely the boundaries most critical for EUDR compliance. The map overestimates primary forests in Asia by approximately 2 billion hectares and misses approximately 9400 thousand hectares of forest in Africa.
- The overall accuracy of approximately 90% is largely inflated by easily classified samples such as "other land" (as noted by the authors); the time series is an early version.
- Officially released on 10 22, this is not news from the past 24 hours, but is included in this issue due to new media coverage.