Beyond tree cover: mapping the world’s forest types

What it claims

ForTy 2020 is a per-pixel forest-type map built as a baseline for the EU Deforestation Regulation (EUDR). The EUDR bars coffee, cocoa and palm oil from land that was forest in 2020 and later turned to agriculture. It also bars wood from primary forest later converted to planted or plantation forest, and from naturally regenerating forest later converted to plantation. Last year's Natural Forests of the World 2020 map only said natural or not. ForTy separates the classes those rules depend on. The 2017 to 2025 time series also records what replaced lost forest, including logged forest replanted as plantation. Tree-cover maps can miss that change because the land turns green again.

How it works

The 1.7 million labels come from expert surveys, national and regional datasets, and targeted reviews of places where the model performed poorly. The model uses a custom architecture for remote sensing data. Its key input is AlphaEarth Foundations embeddings, which are compact summaries of several years of satellite observations for each 10-meter pixel. The authors say these embeddings let the model pick up planting patterns, growth cycles and harvest rhythms. They need this because the classes depend on land use, not only on how the land looks. The time series applies a similar approach to each year.

Evidence

The 9,000+ validation sites were chosen at random worldwide. At each site, two trained experts labelled the land using high-resolution imagery and years of satellite records. Against these labels, ForTy scored 95% on forest versus non-forest. The page says ForTy outperforms other global forest maps but gives no figures for them. It reports no per-class or per-region accuracy.

How to use it

The 2020 map and the time series are free on Google Earth Engine under an open license. The page does not name the license. The method is described in a Scientific Data paper. Each pixel has a probability for every class, not only a hard label, so users choose their own thresholds. The page gives two examples. An EUDR compliance check might keep only pixels with high confidence. A study of total forest extent might use the full probability range.

Limits

The 2017 to 2025 time series is an early release. The authors plan to keep revising it over the coming months, so year-by-year change results may shift. The reported accuracy applies to the 2020 map, and the page gives no accuracy figure for the time series. Users should check the actual license terms on Earth Engine before relying on the data.