【Objective】Accurate prediction of the spatiotemporal distribution of cropland soil organic carbon density (SOCD) is essential for assessing regional carbon sequestration potential. Although existing studies have extensively explored the spatiotemporal dynamics of SOCD using geostatistical methods, machine learning, and process-based models, further improvements are still needed in reconstructing continuous spatiotemporal changes in cropland SOCD, identifying stage-specific patterns, and quantifying uncertainties under complex terrain conditions. 【Method】In this study, Guizhou Province was selected as the study area. Based on 2 295 soil samples collected from four periods between 2006 and 2024, together with 67 covariates, a machine learning model was developed using a space-for-time substitution approach to reconstruct the spatiotemporal dynamics of SOCD in the topsoil layer (0-20 cm). 【Result】SOCD in the cropland topsoil of Guizhou increased from 33.48 to 36.89 t·hm-2 during 2006-2024, with an average sequestration rate of 0.18 t·hm-2·a-1 (95% CI: 0.15-0.23 t·hm-2·a-1). The trend showed clear stage differences, with a faster increase during 2006-2012 (0.22 t·hm-2·a-1) and a slower rate during 2013-2024 (0.14 t·hm-2·a-1). Over the past two decades, about 95% of cropland acted as a carbon sink. Higher sequestration rates were mainly found in Zunyi (0.21 t·hm-2·a-1) and Qianxinan (0.22 t·hm-2·a-1). Among all variables, soil temperature, soil moisture, and organic fertilizer input played the most important roles, with relative contributions of 14.76%, 9.81%, and 9.08%, respectively. 【Conclusion】These results provide a clear picture of how SOCD has changed over time in Guizhou’s farmland and offer a practical approach that can be applied to similar studies in other regions.