Coupling Interpretable Feature Selection with Machine Learning for Evapotranspiration Gap Filling

Evapotranspiration (ET) plays a pivotal role in linking the water and carbon cycles between the land and atmosphere, with latent heat flux (LE) representing the energy manifestation of ET. Due to adverse meteorological conditions, data quality filtering, and instrument malfunctions, LE measured by the eddy covariance (EC) is temporally discontinuous at the hourly and daily scales. Machine-learning (ML) models effectively capture the complex relationships between LE and its influencing factors, demonstrating superior performance in filling LE data gaps. However, the selection of features in ML models often relies on empirical knowledge, with identical features frequently used across stations, leading to reduced modeling accuracy. Therefore, this study proposes an LE gap-filling model (SHAP-AWF-BO-LightGBM) that combines the Shapley additive explanations adaptive weighted fusion method with the Bayesian optimization light gradient-boosting machine algorithm. This is tested using data from three stations in the Heihe River Basin, China, representing different plant functional types. For 30 min interval missing LE data, the RMSE ranges from 17.90 W/m2 to 20.17 W/m2, while the MAE ranges from 10.74 W/m2 to 14.04 W/m2. The SHAP-AWF method is used for feature selection. First, the importance of SHAP features from multiple ensemble-learning models is adaptively weighted as the basis for feature input into the BO-LightGBM algorithm, which enhances the interpretability and transparency of the model. Second, data redundancy and the cost of collecting other feature data during model training are reduced, improving model calculation efficiency (reducing the initial number of features of different stations from 42, 46, and 48 to 10, 15, and 8, respectively). Third, under the premise of ensuring accuracy as much as possible, the gap-filling ratio for missing LE data at different stations is improved, and the adaptability of using only automatic weather station observation is enhanced (the improvement range is between 7.46% and 11.67%). Simultaneously, the hyperparameters of the LightGBM algorithm are optimized using a Bayesian algorithm, further enhancing the accuracy of the model. This study provides a new approach and perspective to fill the missing LE in EC measurement.