High-Resolution Mapping of Soil Organic Carbon Stocks Using Machine and Deep Learning Approaches Across Mediterranean Land Uses

Soils play crucial role as reservoir of organic carbon, reflecting the quality and fertility of terrestrial ecosystems. Consequently, understanding the spatial distribution of soil organic carbon (SOC) stocks and the factors that influence these distributions is imperative for ensuring environmental sustainability and achieving carbon neutrality. This study compares four algorithms namely, Random Forest (RF), Gradient Boosting Machine (GBM), Deep Neural Network (DNN), and Convolutional Neural Network (CNN), which use 29 environmental covariates and 442 soil samples from various land use types to predict and map SOC stocks at a depth of 0–30 cm in the Aix-Marseille-Provence (AMP) Metropolis, France. The results revealed that forests presented the highest SOC content (57 g·kg⁻1) and stock (7.5 kg·m⁻2), while vineyards displayed the lowest values (SOC content: 8.9 g·kg⁻1; stock: 3.4 kg·m⁻2). Urban areas exhibited significant SOC levels, influenced by human activity, with an average content of 45.6 g·kg⁻1 and a stock of 6.2 kg·m⁻2. The Shapley values method revealed that precipitation, elevation, land cover, vegetation index, and temperature were the major factors contributing to the prediction of the SOC stock. The SOC stock prediction algorithms revealed that the RF outperformed the other algorithms (R2 = 0.83, RMSE = 1.41 kg·m⁻2, MAE = 0.97 kg·m⁻2). The area of applicability (AOA) function demonstrated that the RF model was reliable, as almost all the predicted areas fell within the AOA. These results could lead to the development of guidelines for facilitating the sustainable management of carbon sequestration in various land use types within the AMP Metropolis and other Mediterranean regions. This study represents a pioneering effort in the development of advanced artificial intelligence approaches for the high-resolution (10 m) prediction and mapping of soil organic carbon (SOC) stocks in a Mediterranean area: the Aix-Marseille-Provence (AMP) Metropolis in France. It compares the performance of four predictive algorithms — Random Forest (RF), Gradient Boosting Machine (GBM), Deep Neural Network (DNN), and Convolutional Neural Network (CNN) using 29 environmental covariates derived from climatic, topographic, land use, remote sensing, human footprint, physicochemical soil parameters, and geological data. The analysis is based on 442 soil samples from fifteen different land use types, collected from historical archives and recent field campaigns. Initially, we analysed the variation in SOC content and stock according to land use type to understand how different management practices and ecosystems influence carbon storage a particularly critical issue in the Mediterranean context, such as that of the AMP Metropolis, which is highly sensitive to the effects of climate change and anthropogenic pressures. This step is essential for identifying land use types with high carbon sequestration potential and for assessing the impact of land use changes. The analysis revealed that forests presented the highest SOC contents and stocks, while vineyards presented the lowest values. The study also reveals that the RF model outperforms the other models in terms of prediction accuracy and quality of SOC stock mapping. Based on the results of the RF model, the Shapley Values method was applied to identify the main factors contributing to SOC stock prediction, namely, precipitation, altitude, land cover, vegetation index, and temperature. The Area of Applicability (AOA) method was subsequently used to determine the zones where the model’s predictions are considered reliable. These findings provide valuable decision-support tools, offering essential information for sustainable soil management and the development of carbon sequestration strategies tailored to Mediterranean environments. Forests had the highest soil organic carbon (SOC) stock; vineyards had the lowest. Urban area showed notable SOC stocks due to human influence. High-resolution SOC stock prediction and mapping were conducted using sad-vanced artificial intelligence algorithms in a Mediterranean Metropolis. Among the developed algorithms in this study, Random Forest (RF) algorithms demonstrated the best performance in predicting SOC stock (R2 = 0.83). Precipitation, elevation, land cover, vegetation index, and temperature were identi-fied as key predictors of SOC stock using Shapley value. The “Area of Applicability” approach validated the reliability of the RF model, as nearly all the predicted areas fell within the applicable domain.