Improving gap-filling methods for CH4 fluxes of eddy covariance data by combining marginal distribution sampling and machine learning algorithm over paddy fields

Accurate gap-filling CH₄ fluxes from eddy covariance measurements over paddy fields is important in assessing agriculture carbon balance and greenhouse gas emission and therefore contribute to agriculture cleaner production. However, as a type of highly managed wetland, the mechanisms of CH₄ emission from paddy field are different from those from natural wetlands. This study improved the methods to process CH₄ fluxes data from paddy fields. Adding gross primary production (GPP) into the Moving Point test (MP) method enables it to calculate the threshold of friction velocity (U_c*) in the paddy fields. The order of calculated U_c* for all sites is: TWT (0.27 m s^-1) > MSE (0.20 m s^-1) > CRK (0.12 m s^-1) > HRC (0.11 m s^-1) > RIP (0.10 m s^-1) > HRA (0.09 m s^-1) > NC (0.08 m s^-1) > CAS (0.05 m s^-1). GPP is 2.5 hours ahead of CH₄ fluxes, incorporating the asynchrony phenomenon improves the gap-filling performance at all 8 sites, with the R² increases by 6.3%-16.1%, RMSE decreases by 8.0%-26.8%, MAE decrease by 10.5%-25.2%. Replacing soil moisture condition with the ‘on-off switch’ effect of water table only improves the gap-filling performance at CRK, HRA, HRC, MSE and NC sites, with the R² increase by 7.2%, 8.1%, 10.4%, 3.8%, 7.1%, and the improvement of gap-filling accuracy was positively correlated with times of alternate wetting and drying cycles during the growing season. The gap-filling effect of six machine learning algorithms were in order: Adaboost> XGBoost > RF > BPNN > SVM and KNN, with the MDS method worse than the RF. By solving the uneven distribution of CH₄ fluxes, integrating the MDS method into machine learning algorithm could improve the performance of gap-filling, with R² increased by 0.02-0.06, MAE decreased by 0.13%-7.88%. Overall, the study is benefit for data processing of CH₄ fluxes over paddy fields.