PurposeSoil organic carbon (SOC) in arid regions necessitates high-resolution spatial modeling to support effective carbon management. However, prediction reliability is often challenged by complex environmental interactions and limited sampling.Materials and methodsThis study focuses on the Tarim River Basin, an extremely arid region spanning approximately 1.02 x 106 km2, where 861 surface soils were systematically collected from 2023 to 2024. The Boruta algorithm was employed for feature selection, and Extreme Gradient Boosting (XGBoost) and Random Forest (RF) models were utilized to predict SOC through 100 bootstrap iterations. Prediction uncertainty was quantified with a 90% confidence interval, and SHAP analysis was applied to interpret environmental variable interactions.ResultsXGBoost outperformed RF in prediction accuracy (R2 = 0.54 vs. 0.50) and exhibited lower uncertainty (mean prediction interval: 12.17 gkg(-)1 vs. 13.58 gkg(-)1). SOC exhibited extreme spatial variability (mean +/- SD: 5.71 +/- 8.34 gkg(-)1, range: 0.10-63.40 gkg(-)1), with higher values observed in the northern Tianshan foothills and peripheral oases, and lower values found in the central Taklimakan Desert. SHAP analysis identified mean annual precipitation (MAP) as the primary driver of SOC, followed by enhanced vegetation index, net primary productivity, and potential evapotranspiration.ConclusionThis study establishes a scalable framework for SOC mapping in arid ecosystems, highlighting the critical role of MAP and other environmental factors in SOC dynamics, thereby improving carbon management strategies and deepening our understanding of key environmental feedback mechanisms.