Monitoring soil organic carbon from earth observation in the era of cover cropping and no-tillage

The estimation of Soil Organic Carbon (SOC) from optical image spectroscopy typically relies on the availability of bare-soil conditions, which are increasingly rare due to the widespread adoption of conservation agriculture practices. This study evaluates alternative strategies for SOC prediction under limited bare-soil availability by comparing four methodological approaches based on Sentinel-2 imagery and related products: (i) bare-soil multispectral composites, (ii) vegetation indices, (iii) AlphaEarth Satellite Embeddings, and (iv) a hybrid geostatistical-machine learning model (KpR-Cubist). These methods were tested across three cropland regions with contrasting pedoclimatic conditions: Italy, France, and Taiwan. The evaluation relied on >1800 topsoil samples collected between 2020 and 2024. Results show that bare-soil availability varies significantly by region, with cloud cover and vegetation/farm management being the main limiting factors. Models using Satellite Embed-dings consistently achieved the highest predictive accuracy (RPIQ up to 2.24) , outperforming conventional bare-soil composite and vegetation-based models. Incorporating spatial coordinates further improved model performance, revealing strong spatial autocorrelation in SOC distribution. The hybrid kriging-Cubist approach achieved comparable accuracy to the embedding-based models, confirming the value of integrating spatial dependence into data-driven frameworks. Overall, the study demonstrates that deep-learning-derived satellite embeddings models provide effective alternatives for SOC estimation in croplands where bare-soil imagery is increasingly unavailable due to sustainable soil management practices.