Soil Organic Carbon Mapping Through Remote Sensing and In Situ Data with Random Forest by Using Google Earth Engine: A Case Study in Southern Africa

This study, conducted within the SteamBioAfrica project, assessed the potential of Digital Soil Mapping (DSM) to estimate Soil Organic Carbon (SOC) across key regions of southern Africa: Otjozondjupa and Omusati (Namibia), Chobe (Botswana), and KwaZulu-Natal (South Africa). Random Forest (RF) models were implemented in the Google Earth Engine (GEE) environment, integrating multi-source datasets including real-time Sentinel-2 imagery, topographic variables, climatic data, and regional soil samples. Three model configurations were evaluated: (A) climatic, topographic, and spectral data; (B) topographic and spectral data; and (C) spectral data only. Model A achieved the highest overall accuracy (R2 up to 0.78), particularly in Otjozondjupa, whereas Model B resulted in the lowest RMSE and MAE. Model C exhibited poorer performance, underscoring the importance of multi-source data integration. SOC variability was primarily influenced by elevation, precipitation, temperature, and Sentinel-2 bands B11 and B8. However, data scarcity and inconsistent sampling, especially in Chobe, reduced model reliability (R2: 0.62). The originality of this study lay in the scalable integration of real-time Sentinel-2 data with regional datasets in an open-access framework. The resulting SOC maps provided actionable insights for land-use planning and climate adaptation in savanna ecosystems.