Land-use change strongly regulates soil physical, chemical, and biogeochemical properties, with major implications for soil organic carbon (SOC) storage in the Brazilian Amazon to a depth of 30 cm. This study aimed to model and map the spatial variability of SOC stocks across the Brazilian Amazon, Brazil. A total of 649 georeferenced surface soil samples were analysed for soil texture, pH, total nitrogen, SOC, cation exchange capacity (CEC), base saturation (BS), and bulk density. To enhance predictive accuracy, machine learning (ML) algorithms—Random Forest (RF), Support Vector Machine (SVM), Multiple Linear Regression (MLR), and Artificial Neural Network (ANN)—were applied. In parallel, structural equation modeling (SEM) was employed as a complementary analytical method to quantify both direct and indirect effects among predictors and SOC, thereby enabling causal interpretation beyond predictive performance. Model performance was validated through repeated 10-fold cross-validation using the coefficient of determination (R²) and root mean square error (RMSE). Among predictive models, RF achieved the highest accuracy (R² = 0.998; RMSE = 0.068), followed by SVM (R² = 0.992) and ANN (R² = 0.987). Variable importance analysis consistently identified CEC and BS as the dominant SOC predictors, while temperature contributed minimally. SEM explained 92% of SOC variability, with positive effects of CEC (β = 0.42) and negative effects of BS (β = −0.48). These findings highlight the reliability of ML for SOC prediction in developing climate-resilient land management strategies and targeted conservation planning across the Amazon basin. Future research should integrate high-resolution spatial data, deeper soil profiles, and time-series observations to improve SOC stock estimation under changing land use.
Tiruneh et al. (2026) studied this question.