Abstract. The errors and uncertainties associated with gap-filling algorithms of water, carbon and energy fluxes data, have always been one of the prominent challenges of the global network of microclimatological tower sites that use eddy covariance (EC) technique. To address this concern, and find more efficient gap-filling algorithms, we reviewed eight algorithms to estimate missing values of environmental drivers, and separately three major fluxes in EC time series. We then examined the performance of mentioned algorithms for different gap-filling scenarios utilising data from five OzFlux Network towers during 2013. The objectives of this research were (a) to evaluate the impact of training and testing window lengths on the performance of each algorithm; (b) to compare the performance of traditional and new gap-filling techniques for the EC data, for fluxes and their corresponding meteorological drivers. The performance of algorithms was evaluated by generating nine different training-testing window lengths, ranging from a day to 365 days. In each scenario, the gaps covered the data for the entirety of 2013 by consecutively repeating them, where, in each step, values were modelled by using earlier window data. After running each scenario, a variety of statistical metrics was used to evaluate the performance of the algorithms. The algorithms showed different levels of sensitivity to training-testing windows; The Prophet Forecast Model (FBP) revealed the most sensitivity, whilst the performance of artificial neural networks (ANNs), for instance, did not vary considerably by changing the window length. The performance of the algorithms generally decreased with increasing training-testing window length, yet the differences were not considerable for the windows smaller than 60 days. Gap-filling of the environmental drivers showed there was not a significant difference amongst the algorithms, the linear algorithms showed slight superiority over those of machine learning (ML), except the random forest algorithm estimating the ground heat flux (RMSEs of 30.17 and 34.93 for RF and CLR respectively). For the major fluxes, though, ML algorithms showed superiority (9 % less RMSE on average), except the Support Vector Regression (SVR), which provided significant bias in its estimations. Even though ANNs, random forest (RF) and extreme gradient boost (XGB) showed close performance in gap-filling of the major fluxes, RF provided more consistent results with less bias, relatively. The results indicated that there is no single algorithm which outperforms in all situations and therefore, but RF is a potential alternative for the ANNs as regards flux gap-filling.

中文翻译：

---

涡动协方差通量及其驱动器的间隙填充算法比较

摘要。与水，碳和能量通量数据的填空算法相关的误差和不确定性一直是使用涡度协方差（EC）技术的全球微气候塔台网络的主要挑战之一。为了解决此问题，并找到更有效的缺口填充算法，我们回顾了八种算法来估计环境驱动因素的缺失值，并分别评估了EC时间序列中的三个主要通量。然后，我们在2013年期间利用来自5个OzFlux网络塔的数据，研究了上述算法在不同缺口填充情况下的性能。本研究的目标是（a）评估训练和测试窗口长度对每种算法性能的影响；（b）比较EC数据的传统和新的填补空白技术的效果，用于通量及其相应的气象驱动因素。通过生成九种不同的训练测试窗口长度（从一天到365天不等）来评估算法的性能。在每种情况下，通过连续重复这些差距来覆盖整个2013年的数据，在每个步骤中，这些值都是使用较早的窗口数据建模的。在运行每种情况之后，将使用各种统计指标来评估算法的性能。该算法对训练测试窗口显示出不同程度的敏感性。先知预测模型（FBP）显示出最高的敏感性，而例如，通过更改窗口长度，人工神经网络（ANN）的性能并没有显着变化。算法的性能通常随着训练测试窗口长度的增加而降低，但是对于小于60天的窗口，差异并不明显。填补环境驱动因素的差距表明，算法之间没有显着差异，线性算法显示出比机器学习（ML）略有优势，除了随机森林算法估计地面热通量（RMSE分别为30.17和34.93） RF和CLR）。但是，对于主要通量，ML算法显示出优势（平均RMSE减少9％），但支持向量回归（SVR）除外，后者在估计中提供了明显的偏差。尽管人工神经网络，随机森林（RF）和极端梯度增强（XGB）在主要通量的缺口填充方面表现出接近的表现，相对而言，RF提供了更一致的结果，偏差更小。结果表明，没有任何一种算法能在所有情况下都胜于其他算法，因此，就通量间隙填充而言，RF是ANN的潜在替代方案。