Machine Learning-Based Spatial Mapping of Soil Organic Carbon and Its Climatic and Topographic Drivers in the Arid Regions

Digital soil mapping (DSM) is an effective approach for assessing soil organic carbon (SOC) at regional scales. With the increasing availability of geospatial datasets, a growing number of environmental covariates have been incorporated into SOC mapping. However, the influence of covariate selection on model predictive performance remains poorly understood and has not been systematically quantified. In this study, 2865 observations of SOC stocks from 0–100 cm soil profiles and 147 environmental covariates were compiled across northwestern China. The extreme gradient boosting (XGBoost) algorithm was employed to evaluate the effects of covariate quantity on model performance and to generate a 30 m resolution SOC stock map. The results show that model performance increased rapidly with the inclusion of additional covariates before gradually reaching a plateau. The model achieved 90% of the maximum Lin’s Concordance Correlation Coefficient (LCCC) using 13 covariates and 90% of the maximum coefficient of determination (R2) with 26 covariates. Further increases in covariate numbers continued to improve error metrics, with root mean square error (RMSE) and mean absolute error (MAE) reaching stable levels. Considering multiple performance metrics, model accuracy reached a high and stable level when 26–47 covariates were included. In this study, 47 environmental covariates were selected to construct the SOC stock prediction model. Our SOC map outperformed existing SOC products in capturing the spatial heterogeneity of SOC stocks across northwestern China’s mountain–oasis–desert ecosystems, particularly over complex mountainous terrain. Furthermore, our model further enabled the identification of the regional-scale drivers of temperature and precipitation on the spatial distribution of SOC stocks, as well as the quantification of the local-scale effects of topographic factors, including elevation and slope aspect, on SOC stocks. This study establishes a methodological framework for high-resolution regional SOC stocks mapping, highlighting the importance of optimizing environmental covariate selection to balance predictive accuracy and model complexity. The findings provide practical guidance for efficient and reliable DSM and generate valuable high-resolution SOC stocks data for further carbon cycle studies in arid regions.