Machine learning-based prediction of soil organic carbon using environmental covariates from the harmonized world soil database

Soil organic carbon (SOC) is an essential indicator of soil health and plays a key role in climate regulation, soil fertility, and sustainable land management. Traditional laboratory methods for SOC determination are often expensive and time-consuming, limiting their application in data-scarce regions. This study developed a machine learning algorithm for SOC prediction based on environmental and soil variables in the Harmonized World Soil Database (HWSD) for soils in Ghana. A dataset of 231 soil samples from three depth intervals and seven other predictors was utilized. Three machine learning models, Multiple Linear Regression (MLR), K-Nearest Neighbors (KNN), and Support Vector Regression (SVR), were evaluated using R², root mean squared error (RMSE), and mean absolute error (MAE). SVR achieved the best predictive performance (Test R² = 0.9683; Mean CV R² = 0.8119), outperforming KNN (Test R² = 0.9642; Mean CV R² = 0.7920) and MLR (Test R² = 0.7041; Mean CV R² = 0.5912). Soil depth was identified as the most influential predictor. Despite the limitations associated with secondary data and the HWSD resolution, the results demonstrate that machine learning provides a reliable and cost-effective approach for SOC estimation using global HWSD data. Unlike studies relying on intensive field sampling, this study demonstrated the feasibility of using global soil datasets for SOC prediction in Ghana, supporting digital soil mapping and climate change mitigation in data scarce regions.