Using Machine Learning Algorithms for Spatial Prediction of Soil Organic Carbon Based on Environmental Variables and Soil Physicochemical Parameters in the Mediterranean Region

Abstract

Soil plays a key role in storing organic carbon, which is a critical indicator of soil fertility and overall quality. Understanding the spatial distribution of soil organic carbon stock (SOCS) and its influencing factors is essential for promoting sustainable land management. This study applied four machine learning models such as Random Forest (RF), k-nearest neighbors (kNN), Support Vector Machine (SVM), and Cubist to enhance SOCS prediction in the Srou catchment, part of the Upper Oum Er-Rbia watershed in Morocco. A dataset of 120 samples was collected, with 80% used for model training and 20% for validation. Boruta’s feature selection and multicollinearity tests identified nine key factors influencing SOCS. Spatial maps generated from the models were validated using statistical indicators. The RF model showed the highest predictive accuracy (R² = 0.76, RMSE = 0.52 Mg C/ha), followed by SVM and Cubist, while kNN had the lowest performance (R² = 0.31, RMSE = 0.94 Mg C/ha). Key predictors for SOCS included bulk density, pH, electrical conductivity, and calcium carbonate. The proposed machine learning approach demonstrates significant potential for mapping SOCS in similar semi-arid environments.