Global forest, non-forest and tree crop mapping at 10 m resolution using satellite embeddings

Forests, which cover around one-third of the Earth’s land surface, play a crucial role in climate regulation, the global carbon cycle and biodiversity conservation. Forest-related datasets derived from Earth observation (EO) data often serve as baseline layers in various applications, ranging from biosphere monitoring to policy- and decision-making. However, their accuracy and temporal availability vary across regions and climatic zones, and trees used for agricultural purposes are frequently misclassified as forests. Therefore, there is a need for an accurate, up-to-date, and globally consistent forest cover layer that clearly distinguishes forests from tree crops and is available across multiple years. The current advent of big EO data with a combination of advances in Artificial Intelligence lead to the development of geospatial / EO embeddings as ready-to-use products for local to global applications, including forest monitoring. In this study, we develop a highly accurate global forest/non-forest (F/nF) classification at 10 m spatial resolution, while explicitly classifying tree crops as a sub-class of the non-forest category. Our approach implements Google Alpha Earth Foundation’s Satellite embedding dataset in an automatic training process through simple machine learning approaches, including linear Support Vector Machine (SVM), k-nearest neighbors (kNN) and random forest (RF). Automation is achieved through the generation of training data by intersecting multiple forest-related, land cover, plantation and agroforestry datasets across more than 200 training areas, proportionally representing all global biomes. Classification accuracy is assessed through ~21,000 global F/nF reference samples for the year 2020, complemented by several open-access tree crop and plantation validation datasets for 2019-2021. Our F/nF map for the year 2020 achieves an overall accuracy (OA) of 92% and macro F1-score of 0.91, with balanced omission and commission errors for the forest class of 14% and 13%, respectively. Validation of the tree crop sub-class showed high accuracies with OAs exceeding 90% for oil palm, while additional tree crop classes are still being assessed. Among the evaluated classifiers, both the linear SVM and kNN outperform more complex models, including non-linear SVM variants and fine-tuned RFs. In comparison to other global F/nF layers and widely-used land cover datasets, our F/nF dataset’s performance is better or comparable to these alternatives, while it additionally provides information on tree crops. Moreover, the initial transferability tests demonstrate that the trained models produce accurate and spatially consistent results for the period 2017-2024, showing their strong potential for global multi-year change detection analysis at 10 m spatial resolution. These results can support decision-making for policies and regulations, including the European Union Deforestation Regulation (EUDR). The open-access availability of both the resulting dataset and trained models enables global applicability and encourages further testing, adaptation and development by the EO and forest monitoring communities.