Natural Forests of the World 2020

Natural Forests of the World 2020, provides a global map of natural forest probability for the year 2020 at a 10-meter resolution. It was developed to support initiatives like the European Union's Deforestation Regulation (EUDR) and other efforts for forest conservation and monitoring. The map represents the likelihood of an area being a natural forest, which includes primary, naturally regenerating secondary, and managed natural forests. It distinguishes these from planted forests, tree crops, and other land cover types. The dataset was created using a multi-modal temporal-spatial vision transformer model that analyzed seasonal multi-temporal Sentinel-2 satellite imagery and topographical data (elevation, slope, aspect based on TanDEM-X's Copernicus GLO-30 DEM). The data is provided as a probabilistic map, allowing users to apply a probability threshold to create a binary natural forest map tailored to their specific needs.

Limitations: While this map provides a valuable global baseline, users should be aware of several limitations in our map (assessed at the OA optimal probability threshold of 0.52):

• The model can have difficulty distinguishing complex agroforestry systems (e.g., with shaded tree crops), and smallholder agricultural mosaics can be difficult to distinguish from natural forests.
• Differentiating planted forests from naturally regenerating ones can be challenging, especially in boreal and some temperate zones, where natural forests have lower species diversity and are harvested with longer rotation times compared to the tropics.
• Sparse natural forests, such as savanna, are often at the threshold of natural forest definition for the tree canopy height and coverage ratios. Forest type assignment immediately after a disturbance event (e.g., fire, logging) is inherently ambiguous. It may not be clear from satellite imagery whether the forest will regenerate naturally or if the land will be converted to another use (e.g., plantation, agriculture).
• Other areas of potential confusion could include large parks within urban areas, or planted tree belts that meet forest definition criteria but are not natural.
• The accuracy of our natural forest map is intrinsically linked to the quality and consistency of the various input datasets used for training label generation. These datasets were created using different methodologies, spatial resolutions, temporal ranges, and definitions. Some label layers were the outputs of other models, and are therefore limited by the quality of those models. While our approach aimed to harmonize sources and mitigate the impact of individual dataset errors, inconsistencies and inaccuracies in the underlying data could still influence the final map.