ForTy v1 scored 95.5% overall accuracy when researchers grouped its four forest classes together and tested them against non-forest. In the study’s comparison on the same validation dataset, ForTy had the highest overall accuracy among the maps tested, including the European Commission Joint Research Centre’s GFC2020. ForTy’s full six-class map scored 90.2%, measuring a different task. The peer-reviewed ForTy study reports both results.
The researchers used 8,190 sites with expert-consensus labels, drawn from an initial stratified sample of 9,100. Each site’s result was based on its central 10-by-10-meter pixel.
| Classification task | ForTy v1 | EU Joint Research Centre map |
| Forest versus non-forest | 95.5% overall accuracy | GFC2020 was among the maps compared on the same validation dataset; ForTy had the highest overall accuracy |
| Four forest-type classes | 90.9% overall accuracy | ForTy’s overall accuracy was 3.1 percentage points higher than GFT2020 |
What ForTy’s accuracy scores measure
Overall accuracy is the share of evaluated sites assigned the correct category. The number changes with the categories being tested: 95.5% combines all four forest classes and compares them with non-forest, while 90.2% covers the complete six-class scheme. A third result, 94.8%, measures natural forest against all other land cover.
For readers coming from our earlier explanation of what ForTy’s 90.2% accuracy means, the key point is that these scores answer different classification questions. None is a measure of every class on its own.
The six-class results vary by forest type. ForTy’s F1 score was 82.9% for primary forest, 72.6% for naturally regenerating forest, 66.7% for plantation forest and 58.2% for planted forest. F1 combines a class’s precision, or the share of mapped examples that are correct, with its recall, or the share of actual examples the map detects. Planted forests were the most difficult of those forest classes to separate using this measure.
The separate four-class comparison with JRC GFT2020 groups planted and plantation forest together, and combines tree crops and agroforestry with other land. ForTy scored 90.9% on that task, with an overall-accuracy advantage of 3.1 percentage points over GFT2020 in the study. It is a different comparison from forest versus non-forest.
What ForTy’s confidence scores tell map users
ForTy’s pixel scores are relative confidence values, not calibrated probabilities. The default map assigns a pixel to the class with the highest score, but users can change the threshold for a class to suit an application.
A higher threshold favors precision: fewer mapped pixels are accepted, reducing commission errors. A lower threshold favors recall, capturing more of the actual class while reducing omissions. The study recommends evaluating threshold choices with local validation data when an application calls for a particular balance.
