Calculation of surface roughness using machine learning algorithms combined with knowledge distillation

A B S T R A C T Surface roughness is a key parameter in meteorological simulations and wind energy assessments, and its spatial distribution is influenced by various factors. However, the complexity of these factors makes it difficult to retrieve roughness. Although machine learning methods have partially addressed this issue, they still face the challenge of insufficient measurement data. To tackle this, the present study proposes a knowledge distillation framework that integrates physical models and machine learning. It establishes a "teacher-student" model to enable knowledge transfer from regions with sufficient data to target regions with zero samples. In the source domain, where abundant ground truth data is available, four models-Random Forest, Support Vector Regression, Multi-layer Perceptron, and Transformer—were trained. The Multi-layer Perceptron, which achieved the best performance (correlation coefficient: 0.81, RMSE: 0.74, MAE: 0.51), was selected as the teacher model. Then, using the knowledge distillation method, soft labels were generated from remote sensing data in the target region to guide the training of the student model. This facilitated cross-domain knowledge transfer. The results show that the student model's training accuracy improved to 0.89, with the RMSE and MAE reduced to 0.62 and 0.33, respectively, significantly outperforming the teacher model. Compared to ERA5 reanalysis data and land surface model results, the student model's inversion of surface roughness in the target region reduced the mean absolute error by approximately 18%, effectively solving the parameter estimation problem under the condition of no measurement samples. This study significantly enhances the accuracy of surface roughness estimation and provides more reliable parameter input for meteorological simulations and numerical weather forecasting. • Integrating physical models with machine learning algorithms. • Achieving cross-region knowledge transfer of surface parameters. • Optimizing surface roughness estimation through knowledge distillation.