Variability analysis of soil organic carbon content across land use types and its digital mapping using machine learning and deep learning algorithms

Soil organic carbon (SOC) plays a crucial role in carbon cycle management and soil fertility. Understanding the spatial variations in SOC content is vital for supporting sustainable soil resource management. In this study, we analyzed the variability in SOC content across eleven different types of land use in the mining basin of Provence in southeastern France. We modelled this variability spatially using machine and deep learning regression. Four algorithms were tested: random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), and deep neural networks (DNNs). These integrated 162 soil samples and 21 environmental covariates, including climatic parameters, lithology, topographical features, land cover, remote sensing data, and soil physicochemical parameters. The results clearly show a large variability in SOC content across land use types, with forests revealing the highest values (mean of 69.3 g/kg) and arable land the lowest (mean of 8.9 g/kg). The Pearson correlation coefficients (R) indicate that land cover, topography, lithology, environmental indices, and clay content are the main factors influencing the SOC content. The XGBoost model generated the best result (R² = 0.73), closely followed by RF (R² = 0.68) and DNN (R² = 0.60), while SVM showed the weakest performance (R² = 0.36). XGBoost and RF remain the best options for obtaining reliable results with a limited number of soil samples and reduced calculation time. The results of this study provide vital insights for managing soil organic carbon in southeastern France and for climate change mitigation in sustainable land management.