Comprehensive comparison of gap-filling techniques for eddy covariance net carbon fluxes
Comprehensive comparison of gap-filling techniques for eddy covariance net carbon fluxes
复制标题
DOI：
10.1016/j.agrformet.2007.08.011
复制
发表时间：
2007-12-10
影响因子：
6.2
通讯作者：
Stauch, Vanessa J.
中科院分区：
文献类型：
--
作者：
Moffat, Antje M.;Papale, Dario;Stauch, Vanessa J.
我们回顾了估计涡流协方差时间序列中净生态系统二氧化碳交换 (NEE) 缺失值的 15 种技术，并基于来自欧洲六个森林站点的一组 10 个基准数据集评估了它们在不同人工间隙情景下的性能。间隙填充的目标是再现 NEE 时间序列，因此本工作重点是估计缺失的 NEE 值，而不是编辑或删除这些时间序列中的可疑值 由于测量中的系统误差（例如夜间通量、平流）。通过为每个基准数据集生成 50 个带有人工间隙（长度范围从单个半小时到连续 12 天）的辅助数据集来检查间隙填充，并使用各种统计指标评估性能。间隙填充的性能因站点而异，并取决于聚合水平（本地半小时时间步长与每日时间步长），长间隙比短间隙更难填充，并且白天技术之间的差异比夜间更明显。非线性回归技术（NLR）、查找表（LUT）、边际分布抽样（MDS）和半参数模型 （SPM）总体表现良好。基于人工神经网络的技术（ANN）通常（即使只是稍微）优于其他技术。平均日变化（MDV）的简单插值技术表现出中等但一致的性能。几种复杂的技术，例如双无味卡尔曼滤波器 (UKF)、多重插补方法 (MIM)、陆地生物圈模型 (BETHY)，以及其中一种 ANN 和一种 NLR，都显示出高偏差，导致年度总和的可靠性较低，表明可能需要额外的开发。将 10 个基准数据集中的估计随机误差与人为间隙残差进行比较的不确定性分析表明，这些技术已经达到或非常接近测量的噪声极限。根据此处检查的技术和现场数据，间隙填充对 NEE 年度总量的影响不大，大多数技术落在 +/- 25 g C m(-2) 年(-1.) (c) 2007 Elsevier B.V 保留所有权利。
We review 15 techniques for estimating missing values of net ecosystem CO2 exchange (NEE) in eddy covariance time series and evaluate their performance for different artificial gap scenarios based on a set of 10 benchmark datasets from six forested sites in Europe.The goal of gap filling is the reproduction of the NEE time series and hence this present work focuses on estimating missing NEE values, not on editing or the removal of suspect values in these time series due to systematic errors in the measurements (e.g., nighttime flux, advection). The gap filling was examined by generating 50 secondary datasets with artificial gaps (ranging in length from single half-hours to 12 consecutive days) for each benchmark dataset and evaluating the performance with a variety of statistical metrics. The performance of the gap filling varied among sites and depended on the level of aggregation (native half-hourly time step versus daily), long gaps were more difficult to fill than short gaps, and differences among the techniques were more pronounced during the day than at night.The non-linear regression techniques (NLRs), the look-up table (LUT), marginal distribution sampling (MDS), and the semiparametric model (SPM) generally showed good overall performance. The artificial neural network based techniques (ANNs) were generally, if only slightly, superior to the other techniques. The simple interpolation technique of mean diurnal variation (MDV) showed a moderate but consistent performance. Several sophisticated techniques, the dual unscented Kalman filter (UKF), the multiple imputation method (MIM), the terrestrial biosphere model (BETHY), but also one of the ANNs and one of the NLRs showed high biases which resulted in a low reliability of the annual sums, indicating that additional development might be needed. An uncertainty analysis comparing the estimated random error in the 10 benchmark datasets with the artificial gap residuals suggested that the techniques are already at or very close to the noise limit of the measurements. Based on the techniques and site data examined here, the effect of gap filling on the annual sums of NEE is modest, with most techniques falling within a range of +/- 25 g C m(-2) year(-1.) (c) 2007 Elsevier B.V All rights reserved.