Machine learning-based estimation of soil total carbon stocks and their depth-dependent drivers in the northwestern China dryland region

Drylands play an increasingly important role in the global carbon cycle, yet the vertical distribution and controls of soil total carbon stocks (STCS), including soil organic carbon (SOC) and soil inorganic carbon (SIC), remain poorly constrained in dryland regions. In particular, how environmental gradients and SOC–SIC relationships vary with depth is still unclear. Here, we mapped STCS across six soil layers (0-200 cm) in the northwestern China dryland region using five machine-learning models and a suite of environmental covariates. Model skill was assessed using 10-fold cross-validation and an independent hold-out test set, and model interpretation was supported by SHapley Additive exPlanations (SHAP) and structural equation modelling (SEM). Among the tested algorithms, XGBoost consistently achieved the best performance across depths, with the highest accuracy in the 60-100 cm layer. STCS showed strong vertical heterogeneity, with high stocks in the surface layer (0-5 cm), a decline in the mid-profile (15-30 cm), and a modest increase toward deeper layers (30-200 cm); spatially, hotspots were mainly located in the northern and central-eastern parts of the region. Environmental controls of STCS shifted with depth: climatic gradients, topography, and soil properties were most influential in surface layers (0-15 cm), whereas soil moisture and soil physical attributes became increasingly important below 15 cm, especially in deep layers (60-200 cm). Regionally, SIC was the dominant component of STCS, while SOC contributed a smaller fraction but showed depth-dependent associations with SIC. SEM suggested that potential SOC-related priming pathways were associated with SIC differently between surface and deeper layers, indicating possible depth-specific organic–inorganic carbon relationships. Overall, our depth-resolved mapping and interpretation highlight pronounced vertical differentiation in environmental controls and SOC-SIC associations, providing a basis for improving regional carbon accounting and informing dryland carbon management under climate change.