Google DeepMind and Google Research, with the World Resources Institute and the International Institute for Applied Systems Analysis, released Forest Typology (ForTy) 2020, a global map that classifies every 10-by-10-metre patch of land into six FAO-based categories: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land. Alongside it the teams published an early time series tracking how forest type changed each year from 2017 to 2025, including natural forest replaced by plantations or farms. According to the post, the model was trained on about 1.7 million labelled locations drawn from expert surveys, national and regional datasets and targeted reviews, and takes AlphaEarth Foundations embeddings as a key input. For validation, two trained experts examined high-resolution imagery and years of satellite records at more than 9,000 randomly chosen sites; ForTy picked the right forest type 90% of the time and separated forest from non-forest land 95% of the time. The map outputs a probability score per class rather than a single label, and is available on Google Earth Engine under an open licence.
Our takeThe interesting design choice is shipping probabilities rather than one hard label: a company facing the EU deforestation rules can demand near-certainty, a researcher can keep the soft scores, and the same map supports opposite precision-recall trade-offs. The accuracy figures come from expert photo-interpretation, so the reference labels are themselves read off imagery; a systematic misread, say of recently replanted land, would be shared by model and validator. The open question is per-year error in the 2017-2025 series, since class changes compound across steps.