Drylands host critical reservoirs of soil organic carbon (SOC) that are highly sensitive to climate change and anthropogenic disturbances. However, the spatiotemporal dynamics and underlying regulatory mechanisms of SOC in drylands remain poorly understood. Understanding the drivers of SOC in drylands is therefore critical for predicting global carbon-climate feedbacks and developing effective land management strategies. Here, we addressed this knowledge gap by analyzing a comprehensive dataset of 14,222 surface soil samples (0-20 cm) from southern Xinjiang, China, a representative dryland region, alongside key environmental variables (climate, topography, and vegetation). Using an advanced machine learning approach based on extreme gradient boosting (XGBoost), we identified the dominant drivers of SOC variation and quantified changes in SOC storage over the past four decades. Our results revealed that the SOC content, SOC density (SOCD), and SOC storage (SOCS) in the regions were 5.95 g center dot kg- 1, 1.70 kg center dot C center dot m- 2, and 96.85 Tg, respectively. Notably, SOC content was highest in paddy fields (10.32 g center dot kg- 1) and wetlands (16.73 g center dot kg- 1), underscoring the impact of land-use practices on SOC dynamics. Climate, soil mineral, and vegetation together explained 47.2% of the variation in SOC. Soil mineral was more important than other variables in controlling the changes in SOC, with Fe2O3 (proxy for Fe oxides) being identified as the most important driving factor. Alarmingly, the SOC storage in the region has declined by 20% since the 1980s, highlighting the fragility of these carbon pools under global change pressures. Our findings underscore the urgent need to integrate drylands into global carbon management frameworks and inform policies aimed at mitigating climate change and land degradation.