Leveraging satellite embeddings and national expert annotations for country-level forest land use mapping

Timothée Stassin, Diego Marcos, Adolfo Kindgard, René Colditz, Anssi Pekkarinen, Martin Herold

Spatially explicit information on forest as a land use can support climate change mitigation and biodiversity conservation actions, as well as supply-chain regulations, but many countries lack recent forest maps consistent with international definitions. Here, we assess a simple and accessible workflow for country-level Forest/Non-Forest mapping using expert reference labels from the FAO Forest Resources Assessment Remote Sensing Survey (FRA RSS), AlphaEarth Foundations satellite embeddings, and Random Forest classifiers. At global scale, Random Forest models trained on satellite embeddings and expert labels achieved state-of-the-art performance for broad Forest/Non-Forest classification (overall accuracy of 91.3 +- 0.4%), when compared to a benchmark of human experts, a deep learning model, and a reference global map (GFC 2020). Performance remained high for coarse land use classification (4 classes) but was lower for finer-grained classification (16 classes). Country-level models trained and evaluated with national subsets of the reference data produced plausible 10 m forest base maps showing broad agreement with GFC 2020 and similar forest areas as reported to FRA 2025. However, country-level accuracy estimates were often affected by limited validation sample sizes and wide confidence intervals. These results suggest that lightweight models using satellite embeddings can provide a practical starting point for country-led uncertainty-aware forest land use mapping in accessible (cloud-)computing environments. Their operational use, however, depends on sufficient high-quality national reference data, expert review, and independent and statistically-robust validation.