Climate change is closely linked to changes in soil organic carbon (SOC) content, which affects the terrestrial carbon cycle. Consequently, it is essential for carbon accounting and sustainable soil management to predict SOC content accurately. Although there has been an extensive utilization of optical remote sensing data and environmental factors to predict SOC content, few studies have explored their applicability in karst areas. Therefore, it remains unclear how SOC content can be accurately simulated in these areas. In this study, 160 soil samples, 8 environmental covariates and 14 optical remote sensing variables were used to build SOC content prediction models. Three machine learning models, i.e., support vector machine (SVM), random forest (RF) and extreme gradient boosting (XGBoost), were applied for each of three land use classes, including the entire study area, as well as farmland and forest areas. The variables with the greatest influence were the optical remote sensing bands, derived indices, as well as precipitation and temperature for forest areas, and optical remote sensing band11 and Pop-density for farmland. The results from this study suggest that RF and XGBoost are superior to SVM in prediction accuracy. Additionally, the simulation accuracy of the RF model for the forest areas (R² = 0.32, RMSE = 6.81, MAE = 5.63) and of the XGBoost model for farmland areas (R² = 0.28, RMSE = 4.03, MAE = 3.27) was the greatest. The prediction model based on different land use types could obtain a higher simulation accuracy than that based on the whole study area. These findings provide new insights for the estimation of SOC content with high precision in karst areas.

Soil organic carbon (SOC) is generally thought to act as a carbon sink; however, in areas with high spatial heterogeneity, using a single model to estimate the SOC of the whole study area will greatly reduce the simulation accuracy. The earth surface unit division is important to consider in building different models. Here, we divided the research area into different habitat patches using partitioning around a medoids clustering (PAM) algorithm; then, we built an SOC simulation model using machine learning algorithms. The results showed that three habitat patches were created. The simulation accuracy for Habitat Patch 1 (R² = 0.55; RMSE = 2.89) and Habitat Patch 3 (R² = 0.47; RMSE = 3.94) using the XGBoost model was higher than that for the whole study area (R² = 0.44; RMSE = 4.35); although the R² increased by 25% and 6.8%, the RMSE decreased by 33.6% and 9.4%, and the field sample points significantly declined by 70% and 74%. The R² of Habitat Patch 2 using the RF model increased by 17.1%, and the RMSE also decreased by 10.5%; however, the sample points significantly declined by 58%. Therefore, using different models for corresponding patches will significantly increase the SOC simulation accuracy over using one model for the whole study area. This will provide scientific guidance for SOC or soil property monitoring with low field survey costs and high simulation accuracy.

Soil organic carbon (SOC) changes affect the land carbon cycle and are also closely related to climate change. Visible-near infrared spectroscopy (Vis-NIRS) has proven to be an effective tool in predicting soil properties. Spectral transformations are necessary to reduce noise and ensemble learning methods can improve the estimation accuracy of SOC. Yet, it is still unclear which is the optimal ensemble learning method exploiting the results of spectral transformations to accurately simulate SOC content changes in the Three-Rivers Source Region of China. In this study, 272 soil samples were collected and used to build the Vis-NIRS simulation models for SOC content. The ensemble learning was conducted by the building of stack models. Sixteen combinations were produced by eight spectral transformations (S-G, LR, MSC, CR, FD, LRFD, MSCFD and CRFD) and two machine learning models of RF and XGBoost. Then, the prediction results of these 16 combinations were used to build the first-step stack models (Stack1, Stack2, Stack3). The next-step stack models (Stack4, Stack5, Stack6) were then made after the input variables were optimized based on the threshold of the feature importance of the first-step stack models (importance > 0.05). The results in this study showed that the stack models method obtained higher accuracy than the single model and transformations method. Among the six stack models, Stack 6 (5 selected combinations + XGBoost) showed the best simulation performance (RMSE = 7.3511, R² = 0.8963, and RPD = 3.0139, RPIQ = 3.339), and obtained higher accuracy than Stack3 (16 combinations + XGBoost). Overall, our results suggested that the ensemble learning of spectral transformations and simulation models can improve the estimation accuracy of the SOC content. This study can provide useful suggestions for the high-precision estimation of SOC in the alpine ecosystem.