Table 3 Qualitative analysis of the reference sample sites with full agreement.

Tropical forests

• Locations with no images and “no forest” (<5% of tree canopy) – no issues detected.

• “Forest with no or very low human impact” – we found 2% of locations with signs of human activities nearby and 1% of degraded forests.

• “Forest with signs of human activities nearby” – no issues detected.

• “Naturally regenerating forest” – no issues detected.

• “Plantation forests” – no issues detected.

• “Fruit trees (olives, apples, nuts, cocoa, etc.)” were sometimes confused with young oil palm plantations, which is a separate class in our legend.

• “Oil palm” plantations – no issues detected.

• “Tree shelter belts, small forest patches” were sometimes confused with naturally regenerating forests and with agroforestry.

• “Agroforestry or sparse trees on agricultural fields” were sometimes confused with fruit plantations.

• “Trees in urban/built-up areas” were confused with fruit plantations and “Agroforestry or sparse trees on agricultural fields”, and “naturally regenerating forests”.

• There were many mixed pixels with fruit trees plantations, agroforestry, tree shelterbelts and small forest patches.

• All locations with the following classes were revised by experts: “fruit trees (olives, apples, nuts, cocoa, etc.)”, “tree shelter belts and small forest patches”, “agroforestry or sparse trees on agricultural fields”, and “Trees in urban/built-up areas”;

• Merged “forest with no or very low human impact”, “forest with human impact nearby”, and “degraded and disturbed” forests into one class called “Naturally regenerating forest without any signs of management”.

• Merged “fruit trees (olives, apples, nuts, cocoa, etc.)”, “tree shelter belts, small forest patches”, “Agroforestry or sparse trees on agricultural fields”, and “Trees in urban/built-up areas” into one class “Agroforestry”.

Temperate forests

• Locations with no images – no issues detected.

• Locations with “no forest” – we found only a few misclassifications, which included a poplar plantation that was not visible on the Microsoft Bing Maps image, degraded forest dominated by snags, and abandoned fields that are reverting to forests.

• “Forest with no or very low human impact” and “forests with human impact nearby” were correctly classified, with the exception of shrubland in Australia, which was partly misclassified as forest. “Naturally regenerating forest” had only a few misclassifications such as “planted forests” mapped as “naturally regenerating forest”.

• “Planted forest” were correctly classified with the exception of planted forests in the USA and China that were confused with “naturally regenerating forests”.

• “Plantation forests” – no issues detected.

• “Fruit trees (olives, apples, nuts, cocoa, etc.)” were sometimes confused with pine nut plantations.

• “Tree shelter belts, small forest patches” had mistakes related to belts between young plantations or naturally regenerating forests.

• “Agroforestry or sparse trees on agricultural fields” – this category was not understood very well. Many misclassified points were either sparse natural forests or naturally regenerating forests take place.

• Revisited and replaced if necessary locations classified as being “no forest”, with “forest with no or very low human impact” in Australia, revisited and replaced if necessary “naturally regenerating forest” with “planted forest” in the USA and China, revisited and replaced if necessary “fruit trees (olives, apples, nuts, cocoa, etc.)” with “Agroforestry or sparse trees on agricultural fields”.

• Classes were merged similarly to the tropic’s category, to ensure global map consistency.

Boreal forests

• Locations with no image available are classified correctly.

• “No forest” – no issues detected.

• “Forest with no or very low human impact” confused with degraded forest.

• “Forest with human impact nearby” confused with “naturally regenerating forests”.

• “Naturally regenerating forest” confused with “planted forest” in Sweden and Finland.

• “Planted forest” were correct, except for Moscow region, Russia.

• No issues detected with agroforestry, tree shelterbelts and trees in urban areas.

• Revisited and replaced, if necessary, “forest with human impact nearby” with “naturally regenerating forests” in Finland and Sweden, and planted forests around Moscow, Russia.

• Classes were merged similarly to the tropics, to ensure global map consistency.