有限样本下土壤有机碳密度空间分布预测模型对比分析.

[Objective] The aim of this study is to explore the accuracy and applicability of different machine learning models for predicting the spatial distribution of surface soil organic carbon density (SOCD) with limited samples, which can provide a references for the study of watershed scale carbon pool in the Chinese Loess Plateau. [Methods] In this study, we compared the accuracy and stability of the predicted SOCD in topsoil (0-20 cm) by four machine learning models, namely Multiple Linear Stepwise Regression (SR), Random Forest (RF), Extreme Gradient Boosting (XGB) and Support Vector Machine (SVM), based on the limited measured samples in a sub-watershed of Yanhe River in the Chinese Loess Plateau. [Results] (1) Under the condition of limited samples, all models successfully and appropriately predicte the spatial distribution of SOCD, among which the SVM model has the best model performance, and the average RMSE, R² and MAE of 50 predictions is 0.74. 0.43 and 0.64, respectively. (2) The average SOCD of different land use types are consistent between measured and predicted values but shows significant difference among land use types. SOCD decreases in the order: shrubland forestland > grassland > cropland. The total organic carbon of cultivated land in the study area is 2.39×106t (0-20 cm). (3) The evaluation of feature importance shows that terrain factors, NDVI max, near-infrared surface reflectance (B5) and Brightness index have significant contributions to the accuracy of predictions. [Conclusion] Under the condition of limited samples, the machine learning model combined with controlling features can be effectively applied to the prediction of the spatial distribution of topsoil SOCD at the watershed scale in the Chinese Loess Plateau. [ABSTRACT FROM AUTHOR]