Abstract
Soil organic carbon (SOC) exhibits distinct spatial heterogeneity across different pedoclimatic regions, yet the underlying regulatory mechanisms and their threshold responses remain poorly understood. In this study, the spatial patterns and underlying region-specific regulatory factors controlling SOC dynamics were investigated across a pedoclimatic gradient represented by the Jiaodong Peninsula (maritime monsoon climate) and Southwest Shandong (continental climate) in Shandong Province, China. Geostatistical analysis coupled with sequential Gaussian simulations provided probabilistic assessment of SOC spatial patterns, while machine learning algorithms (Linear Regression, Random Forest, XGBoost and Support Vector Machine) integrated with SHAP analysis enabled quantification of nonlinear threshold responses and identification of dominant factors governing SOC dynamics. The results showed that SOC in Jiaodong exhibited a west-high-east-low gradient characterized by local-scale structure, whereas Southwest Shandong showed higher SOC contents dominated by macro-scale gradients. The Random Forest model identified distinct regulatory mechanisms in Jiaodong, where NO(3) (-)-N and extractable Fe exhibited a dual-threshold domain (NO(3) (-)-N = 10.0 mg·kg(-1), Fe = 12.0 mg·kg(-1)), with the marginal effect of Fe on SOC shifting from negative to positive when NO(3) (-)-N exceeded its threshold concentration. In Southwest Shandong, total nitrogen (TN) was revealed as the dominant predictor, with a critical threshold at 3.25 g·kg(-1) above which SOC increased by 2.0 g·kg(-1), while NO(3) (-)-N showed negative effects above 27 mg·kg(-1). This study demonstrates that the combination of interpretable machine learning and geostatistical approaches can effectively elucidate region-specific threshold mechanisms and nonlinear controls governing SOC dynamics. This approach is critical for developing spatially-explicit soil carbon management strategies under varying pedoclimatic conditions.