On the Correction of Near-Surface Turbulent Fluxes over a Farmland Based on Machine Learning Methods

Surface-layer turbulent fluxes are crucial variables for characterizing the land–atmosphere interaction. These fluxes are typically calculated using gradient observations or atmospheric numerical models based on MOST (Monin–Obukhov Similarity Theory). In this study, characteristic quantities for turbulent fluxes derived from MOST are corrected using random forest and XGBoost models, with eddy covariance–derived fluxes serving as reference values. The corrections significantly improve the accuracy of characteristic quantities for fluxes estimate especially the temperature scale and moisture scale. The forward and backward selections were performed for the input variables to reduce their count. The optimal combinations of input variables obtained after these selections maintained a high level of computational accuracy. Furthermore, when these selected combinations were provided as inputs to an artificial neural network, the computational accuracy of the fluxes improved significantly compared with that of the uncorrected MOST-derived fluxes. These findings confirm that selecting input variables is crucial for optimizing machine learning applications in surface-layer-flux correction.