A comparison of gap-filling algorithms for eddy covariance fluxes and their drivers

Abstract. The errors and uncertainties associated with gap-filling algorithms of water, carbon, energy fluxes data have always been one the main challenges global network microclimatological tower sites that use eddy covariance (EC) technique. To address these concerns find more efficient algorithms, we reviewed eight to estimate missing values environmental drivers nine for three major typically found in EC time series. We then examined algorithms' performance different scenarios utilising from five towers during 2013. This research's objectives were (a) evaluate impact gap lengths on each algorithm (b) compare traditional new techniques data, fluxes, separately their corresponding meteorological drivers. was evaluated by generating windows lengths, ranging a day 365 d. In scenario, period chosen randomly, removed dataset accordingly. After running variety statistical metrics used performance. showed levels sensitivity lengths; Prophet Forecast Model (FBP) revealed most sensitivity, whilst artificial neural networks (ANNs), instance, did not vary as much changing length. generally decreased increasing length, yet differences significant smaller than 30 No between recognised However, linear slight superiority over those machine learning (ML), except random forest (RF) estimating ground heat flux (root mean square – RMSEs 28.91 33.92 RF classic regression CLR, respectively). ML MDS other algorithms. Even though ANNs, (RF), eXtreme Gradient Boost (XGB) comparable provided consistent results slightly less bias against indicated no single outperforms all situations, but is potential alternative ANNs regards gap-filling.