Unraveling non-linear environmental controls on soil organic carbon in arid China: An interpretable machine learning approach

Accurate prediction of soil organic carbon (SOC) content is essential for soil management, ecosystem sustainability, and climate change mitigation, particularly in ecologically fragile arid and semi-arid regions. Traditional statistical approaches often struggle to capture the complex and nonlinear relationships between SOC and environmental drivers. In this study, multiple machine learning models—including Random Forest (RF), Support Vector Regression (SVR), Partial Least Squares Regression (PLSR), Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Extreme Gradient Boosting (XGBoost)—were evaluated for SOC prediction using a large dataset of 8,621 soil samples collected from north-central and north-western China. A comprehensive set of environmental covariates was incorporated, including terrain attributes, climate variables, soil properties, vegetation indices, soil moisture, and erosion indicators. Model performance was assessed using five-fold spatial cross-validation to mitigate potential spatial dependence between training and validation samples. Prediction accuracy was evaluated using multiple metrics, including adjusted R², MAE, MAPE, MSE, RMSE, and RPIQ. Among the evaluated models, XGBoost achieved the highest predictive performance across land-use types, with an adjusted R² of 0.775 and an RMSE of 0.610 in the spatial validation set. Rather than emphasizing algorithmic superiority alone, this study further explored model interpretability by integrating Shapley Additive Explanations (SHAP) with generalized additive models (GAMs). This combined interpretive framework was used to characterize land-use–specific nonlinear response patterns and model-derived breakpoints of key environmental predictors. The results indicate that temperature, soil pH, vegetation activity (NDVI), and elevation consistently emerged as important predictors associated with SOC variation, although their relative importance and response patterns differed among forestland, grassland, farmland, and unutilized land. These findings highlight the context dependency of SOC–environment relationships in arid and semi-arid landscapes. Overall, this study provides a spatially robust modeling and interpretation framework for SOC prediction at sampling locations, offering transferable insights into how environmental gradients shape SOC variability across land-use types. The approach establishes a methodological basis for future large-scale SOC mapping and uncertainty assessment.