Five machine learning (ML) algorithms were employed for gap-filling surface fluxes of CO2, water vapor, and sensible heat above three different ecosystems: grassland, rice paddy field, and forest. The performance and limitations of these ML models, which are support vector machine, random forest, multi-layer perception, deep neural network, and long short-term memory, were investigated. Firstly, the accuracy of gap-filling to time and hysteresis input factors of ML algorithms for different ecosystems is discussed. Secondly, the optimal ML model selected in the first stage is compared with the classic method—the Penman–Monteith (P–M) equation for water vapor flux gap-filling. Thirdly, with different gap lengths (from one hour to one week), we explored the data length required for an ML model to perform the optimal gap-filling. Our results demonstrate the following: (1) for ecosystems with a strong hysteresis between surface fluxes and net radiation, adding proceeding meteorological data into the model inputs could improve the model performance; (2) the five ML models gave similar gap-filling performance; (3) for gap- filling water vapor flux, the ML model is better than the P–M equation; and (4) for a gap with length of half day, one day, or one week, an ML model with training data length greater than 1300 h would provide a better gap-filling accuracy.

flux gap-filling,Penman-Monteith equation,Machine learning techniques