Comparison of Soil Organic Carbon Prediction Accuracy Under Different Habitat Patches Division Methods on the Tibetan Plateau

ABSTRACT Soil organic carbon (SOC) plays an important role in soil fertility and the global carbon cycle. Therefore, accurate estimation of SOC is of great significance in carbon sink accounting and carbon sequestration increase. The accuracy and stability of models estimating SOC density (SOCD) tend to decrease because of the high spatial heterogeneity of environmental factors and SOC. However, research on how to improve model stability is limited. Therefore, this study investigated a strategy to divide a study area into different habitat patches using partitioning around medoids (PAM) clustering, land use type, and climate trend. In this approach, we selected optimal environmental covariates using recursive feature elimination (RFE). We then used three machine‐learning models to predict SOCD on the Tibetan Plateau. The results showed that (1) average SOCD in the 0–20 cm soil surface layer on the Tibetan Plateau was 4.85 kg C m −2 and SOCD increased from northwest to southeast, which was consistent with previous reports. Areas with high SOCD tended to have high uncertainty. (2) The RFE feature selection method reduced the number of input variables used in the SOCD estimation model and improved the accuracy of predictions by combining machine‐learning models. Compared with the SVM model, the RF and XGBoost models performed better for SOCD estimation. (3) Habitat patches division based on land use type and PAM clustering did not perform as well as expected. The simulation accuracy based on climate trend division was slightly higher than that of global modeling for the whole study area. (4) Biological and climatic factors had a higher impact on the prediction of SOCD than other variables. This study characterized the spatial heterogeneity of SOCD well and can provide a valuable reference for regional carbon stock estimation and carbon management on the Tibetan Plateau.