农田土壤有机碳反演的变量筛选及算法组合优化 ——以泾河流域为例

：To achieve rapid and high-precision inversion of soil organic carbon（SOC）in farmland and to explore suitable combinations of different variable screening methods and machine learning algorithms, this study focused on the farmland of the Jinghe River basin, a typical hilly-gully region of the Loess Plateau. Based on the SOC sample data, this study integrated 126 variables, including MODIS multispectral data and related environmental variables. The study explored the impacts of three variable sets：the full variable set, the competitive adaptive reweighted sampling（CARS）screened variable set, and the correlation coefficient screened variable set when combined with four machine learning algorithms：support vector machine（SVM）, random forest（RF）, back propagation neural network （BPNN）, and extreme gradient boosting（XGBoost）on model accuracy. The results showed that models based on the 11 variables selected via correlation coefficient significantly outperformed those using the full or CARS-selected variable sets. Among these combinations, the SVM algorithm performed best（validation set：R²=0.86, RMSE=0.95, MAE=0.76, RPD=2.66）, followed by the XGBoost, BPNN, and RF models. The SHAP analysis revealed that topographic variables（DEM, terrain relief, profile curvature）contributed 28.41% to the inversion of farmland SOC in the Jinghe River basin, where the topography was complex and variable and the cultivated land was relatively fragmented, followed by the remote sensing index（23.18%） and latitude / longitude（20.21%）. The inversion results based on the combination of correlation coefficient selection and SVM algorithm revealed that the SOC of farmland in the basin ranged from 0.00 g·kg^-1 to 39.94 g·kg^-1, with high values concentrated in the western part of the basin and medium-to-low values interspersed in the central and southern areas. The average SOC content was 8.38 g·kg^-1, which is lower than the national average for farmland SOC.