Turbulent fluxes in the surface layer are significant variables that characterize the interaction between the land and the atmosphere, which are typically calculated based on the Monin-Obukhov Similarity Theory (MOST) in gradient observations or atmospheric numerical models. To further enhance the calculational accuracy of turbulent fluxes in the surface layer, in this paper, the turbulent fluxes derived from the MOST are corrected by Random Forest and XGBoost, with fluxes obtained by the eddy covariance method as reference values, which significantly improves the calculational accuracy of the temperature scale and moisture scale. In order to reduce the number of input variables, this study further performs forward and backward variable selection on the input variables The optimal combinations of input variables after selecting still maintain a high level of computational accuracy. Furthermore, when the selected combinations of input variables are applied to another machine learning model, the artificial neural network, their computational accuracy also show a significant improvement in comparison with the results of MOST. This demonstrates that the variable selection method in this paper of input variables required for machine learning method.