Understanding the environmental drivers underlying the spatial heterogeneity of soil organic carbon (SOC) in mountainous regions remains a major challenge in digital soil mapping. This study investigated the spatial distribution and driving mechanisms of SOC contents in a typical subtropical mountainous area using an integrated modeling and interpretation framework based on 132 soil samples. The SOC content in Yangshan County ranged from 3.33 to 50.00 g kg⁻¹, with a coefficient of variation of 48.64%, indicating a moderate level of variability across the study area. Six mainstream modeling approaches were compared, including multiple linear regression (MLR), geographically weighted regression (GWR), Cubist, eXtreme Gradient Boosting (XGBoost), random forest (RF), and a hybrid RF-GWR model. The results showed that RF outperformed traditional linear methods and other machine learning approaches, achieving an R² of 0.45 and RMSE of 7.78 g kg⁻¹, while the hybrid model further improved prediction accuracy (R² = 0.48). Then, spatial mapping revealed a clear elevational gradient, with higher SOC values concentrated in forested mountainous areas in the north and lower values distributed across low-elevation cultivated and disturbed zones. SHAP analysis identified intrinsic soil properties, particularly total nitrogen (TN) and cation-exchange capacity (CEC), as dominant controls on SOC contents. When extended to prediction datasets, relative humidity (RH) and mean annual precipitation (MAP) showed greater importance on SOC, suggesting an amplification of climatic factors at the broader scale. Subsequently, hotspot analysis of GeoShapley components further revealed the spatial differentiations in group indicators, with overall contributions ranked as soil physicochemical properties (36.4%) > geographic conditions (21.1%) > climate (17.4%) > organisms (12.9%) > parent material (12.1%). Soil properties formed clustered hotspots overlaid on carbonate-dominated areas, while geographic conditions and climate primarily acted as spatial modulators, generating localized zones of intensified or weakened influence across the landscape. The integrated framework proposed in this study has potential applicability across broader regions. These findings provided a scientific basis for the localized interpretation of environmental drivers of SOC and offered valuable support for region-specific land management and sustainable decision-making.

Agricultural carbon emissions represent a major source of greenhouse gases and play a critical role in achieving global climate mitigation and sustainable agricultural development targets. In China, the rapid transformation of agricultural production systems has led to substantial spatial heterogeneity in emission patterns and driving mechanisms of agricultural carbon emissions, while the underlying processes at the county scale remain insufficiently understood. This study investigated the spatiotemporal evolution and potential influencing factors of agricultural carbon emissions at the county level from 2000 to 2022 in Guangdong Province, China. First, agricultural carbon emissions were estimated based on a multi-source accounting framework covering land management, crop cultivation, animal production, and straw burning based on internationally recognized emission accounting methods and IPCC global warming potentials. Then, spatial clustering characteristics were analyzed using local spatial autocorrelation (LISA) to identify heterogeneous emission patterns. Finally, an interpretable machine learning framework combining Random Forest (RF) and SHapley Additive exPlanations (SHAP) was employed to quantify the nonlinear effects and relative contributions of multiple socioeconomic and agricultural drivers. The results showed that agricultural carbon emissions in Guangdong Province exhibited a fluctuating but overall decreasing trend, declining from 50.89 Mt CO₂-eq in 2000 to 39.24 Mt CO₂-eq in 2022, with an overall reduction of 22.9%. High-emission clusters were primarily concentrated in western and northern Guangdong, while low-emission areas were mainly located in the Pearl River Delta (PRD). The RF models demonstrated satisfactory predictive performance, with spatial cross-validated R² values ranging from 0.75 to 0.91 across different years. SHAP analysis suggested that ploughing area, fertilizer and pesticide usage, agricultural machinery power, and primary industry GDP were the dominant factors associated with agricultural carbon emissions, whereas urbanization consistently showed a negative association. Furthermore, these drivers exhibited pronounced nonlinear responses and distinct regional heterogeneity, particularly between the PRD and the western and northern parts of Guangdong Province. These findings suggested that agricultural carbon emissions are jointly influenced by agricultural production intensity, mechanization, and socioeconomic transition and can provide a scientific basis for developing region-specific low-carbon agricultural policies and promoting the sustainable transformation of agricultural systems.