Integrating Natural and Agricultural Management Predictors to Improve the Digital Mapping Accuracy of Soil Organic Carbon Density in Arid Zone Cropland

ABSTRACT Mapping the spatial distribution of soil organic carbon density (SOCD) is essential for estimating soil organic carbon stocks and sequestration potential. The choice of environmental variables affects the accuracy of digital soil mapping (DSM). Previous studies on mapping SOCD in cropland have primarily focused on natural factors while giving insufficient consideration to agricultural management factors. This has led to limited interpretive power and low predictive accuracy in DSM models. In recent decades, large‐scale cropland reclamation and sustained agricultural management in the arid zone of Northwest China have significantly affected the spatial heterogeneity of regional SOCD. In this study, we selected Manasi County in the Xinjiang Uygur Autonomous Region as the study area. We chose irrigation capacity ( IC ), multiyear accumulation of the normalized difference vegetation index ( MA‐NDVI ), and multiyear accumulation of the enhanced vegetation index ( MA ‐ EVI ) as agricultural management factors. Remote sensing and meteorological data were using to spatialize these agricultural management factors. Subsequently, we developed a random forest (RF) mapping model jointly driven by natural and agricultural management factors to enhance the prediction accuracy of cropland SOCD. Among all environmental factors, IC was confirmed as the most important factor for SOCD prediction. Compared to using natural factors alone, the inclusion of agricultural management factors improved the mapping accuracy of cropland SOCD, increased the R 2 from 0.33 to 0.57, and decreased the root mean square error (RMSE) from 0.63 to 0.50 kg/m 2 . In addition, the highest standard deviation of predicted SOCD values decreased from 0.52 to 0.40 kg/m 2 . This study confirmed the effectiveness of incorporating agricultural management factors in enhancing prediction accuracy and reducing model uncertainty in cropland SOCD mapping.