A Comparative Study of EO Data Fusion and Satellite Embeddings for Soil Organic Carbon Estimation in Mainland Portugal

ABSTRACT Reliable information on soil organic carbon (SOC) is essential for sustainable land management, climate change mitigation, and the conservation of soil biodiversity. However, conventional SOC assessment, which generally involves field based soil sampling followed by laboratory analyses such as dry combustion, wet oxidation, is labor intensive, time consuming, and costly, particularly when applied over large areas. The integration of Earth Observation (EO) data with Machine Learning (ML) algorithms has the potential to predict SOC on a large scale. This study estimates the SOC in mainland Portugal using two datasets; Dataset‐I incorporates Sentinel‐1/2, with ancillary climate and topographic covariates while Dataset‐II contains the Google Satellite Embeddings (GSE) dataset. The Recursive Feature Elimination with Cross Validation (RFECV) was used to identify and remove redundant variables. Three state‐of‐the‐art ML algorithms, i.e., Random Forest (RF), Gradient Tree Boost (GBT) and Classification and Regression Tree (CART) were used for SOC prediction. The results indicate that RF achieved the highest predictive performance using Dataset‐II, with an R 2 of 0.34, RMSE of 19.2 , MAE of 13.2 , and RPIQ of 1.27 under spatial cross validation, while independent validation on the 43 withheld samples yielded an R 2 of 0.44, RMSE of 19.89 , MAE of 15.50 , and RPIQ of 1.18. The calibrated RF uncertainty was generally lower across central and southern Portugal, with greater spatial variability in the northern region. The findings contribute to the advancement of SOC measurements for sustainable development.