Monitoring and enhancing soil carbon absorption is essential for optimizing crop yields and combating climate change through carbon sequestration. While machine learning approaches for Soil Organic Carbon (SOC) prediction have gained research attention, there remains uncertainty regarding optimal algorithm selection. Our study introduces a superlearner-based SOC prediction framework, utilizing soil samples from Arkansas, Idaho, Nebraska, and Utah. The model incorporates remote sensing data, including vegetation indices from Sentinel-2 and Digital Elevation Model from ALOS PALSAR, with implemented feature selection techniques. Model validation was performed using independent soil samples from Salt Lake City, Utah. Our findings revealed superior performance of the linear regression-based superlearner (nRMSE: 7.6%, R²: 0.804) over its random forest counterpart (nRMSE: 8.3%, R²: 0.768). The linear regression approach, while more accurate, demanded careful base learner selection and hyperparameter optimization. In contrast, the random forest variant exhibited consistent accuracy across different base learner configurations. When applied to new locations, the linear regression-based model demonstrated better predictive capabilities (nRMSE: 52.9%, RMSE: 0.48% OC).
