We have carried out an inter-comparison between EddyUH and EddyPro®, two public software packages for post-field processing of eddy covariance data. Datasets including carbon dioxide, methane and water vapour fluxes measured over 2 months at a wetland in southern Finland and carbon dioxide and water vapour fluxes measured over 3 months at an urban site in Helsinki were processed and analysed. The purpose was to estimate the flux uncertainty due to the use of different software packages and to evaluate the most critical processing steps, determining the largest deviations in the calculated fluxes. Turbulent fluxes calculated with a reference combination of processing steps were in good agreement, the systematic difference between the two software packages being up to 2.0 and 6.7 % for half-hour and cumulative sum values, respectively. The raw data preparation and processing steps were consistent between the software packages, and most of the deviations in the estimated fluxes were due to the flux corrections. Among the different calculation procedures analysed, the spectral correction had the biggest impact for closed-path latent heat fluxes, reaching a nocturnal median value of 15 % at the wetland site. We found up to a 43 % median value of deviation (with respect to the run with all corrections included) if the closed-path carbon dioxide flux is calculated without the dilution correction, while the methane fluxes were up to 10 % lower without both dilution and spectroscopic corrections. The Webb–Pearman–Leuning (WPL) and spectroscopic corrections were the most critical steps for open-path systems. However, we found also large spectral correction factors for the open-path methane fluxes, due to the sensor separation effect.
The eddy covariance (EC) technique is the most direct and defensible way to
measure and calculate vertical turbulent fluxes of momentum, energy and gases
between the atmosphere and biosphere. During the last 3 decades, the
number of long-term EC stations all over the world has increased
exponentially, covering a wide range of different ecosystem types (FLUXNET,
In this study, we have performed an inter-comparison between EddyUH and
EddyPro, two public and commonly used software packages for EC data
processing and calculation. The aims are to estimate the flux uncertainty
due to the use of different software packages for half-hour as well as for
cumulative sums, and to assess the most critical processing steps,
determining the largest deviations in the calculated fluxes. We focus not
only on LE and CO
The software inter-comparison was performed using datasets from two field
sites in southern Finland. The first dataset was collected at the Siikaneva fen
site (61
The second dataset was collected between 1 July and 30 September 2010 at the Erottaja
site located in Helsinki city centre (60
The turbulent fluxes of CO
The results presented in this study are based on EddyUH version 1.7 and
EddyPro version 5.2.1. EddyUH is a software package for EC raw data
processing, developed by the Micrometeorology Research Group at the
Department of Physics, University of Helsinki (Finland). It is freely
downloadable from
The EddyUH software was compared against EddyPro, perhaps the most used
software in the EC flux community, developed by LI-COR Biosciences Inc.
(Lincoln, NE, USA). It is freely available and well documented (
The eddy covariance datasets were processed using the reference combination
of processing steps (Fig. 1) and available methods implemented in EddyUH and
EddyPro (Table 1). The applied methods were the same for most of the steps.
However, some differences between software packages were present. In EddyUH
the raw data despiking was done by the difference limit method (Appendix B),
while in EddyPro the Vickers and Mahrt (1997) method was used.
Different experimental methods were applied for spectral correction, e.g.
according to Mammarella et al. (2009) in EddyUH and Fratini
et al. (2012) in EddyPro. Moreover, additional correction for water vapour
cross-sensitivity (henceforth point-by-point spectroscopic correction) of
closed-path CH
Flux data were quality-screened prior to analysis. CH
Fluxes measured by the same system were analysed and compared as estimated
by the two software packages in the reference run. In terms of regression
statistics, a very good agreement between EddyUH and EddyPro was obtained
for LE and
Finally, good agreement resulted from the LI-7700
EC data processing scheme for open- and closed-path gas analyser
data. Relative magnitude of each processing step is also
reported, according to this and other studies. Note the different
ways with different sign conventions (see footnotes) in which the relative
magnitudes are calculated in different studies, where
Comparison of the reference run fluxes estimated by EddyUH and EddyPro. CH
Median diurnal variation of flux ratio (left side) and bias (right
side) for the studied instruments and variables between the two software.
Light blue shows the uncorrected fluxes after despiking, coordinate
rotation, detrending and time lag compensation; blue the WPL corrected
fluxes (for G1301-f and LI-7700
Scatter plots of
Effect of different calculation procedure of the estimated flux at
the Siikaneva site, presented as deviation (in %) from the reference run
(ref). Deviation is defined as (run-ref)/ref, where run refers to the run
performed with no spectral correction (no spec), theoretical spectral
correction (theor spec), no WPL (or dilution) and spectroscopic correction
(no WPL) and using a constant time lag (const lag). Bars indicate the
median values, and error bars denote 25th and 75th percentiles. Note the
different scale on
As in Fig. 5, but for Erottaja. Note the different scale on
In order to further evaluate the discrepancies between the two software
packages, diel patterns of flux ratio (left side in Fig. 3) and bias (right
side in Fig. 3) were plotted for each flux at different processing levels.
In general, the uncorrected (raw) fluxes do not show significant deviations
from unity (left side of Fig. 3) or from the zero line (right side of Fig. 3), which suggests that the preparations done at the raw data level
(despiking, coordinate rotation, time lag compensation) did not make a
significant systematic difference to the fluxes. For LE calculated from
LI-7500 data, the WPL correction tends to be slightly larger in EddyUH than
in EddyPro, meaning that it increases the daytime positive fluxes more (cf.
WPL curve in Fig. 3r). The daytime WPL correction in EddyUH is approximately
2 W m
Fluxes measured with different gas analysers are compared as estimated using
the reference combination in EddyUH and EddyPro (Fig. 4). At Erottaja, very
good agreement was found between
The impact of calculation steps on the final flux estimates in EddyUH and
EddyPro is presented as deviation (in %) from the reference run for
Siikaneva (Fig. 5) and Erottaja (Fig. 6) datasets. For closed-path systems in
Siikaneva, the run without spectral correction had the largest effect on LE,
being (in the EddyUH run) 14 and 9 % lower for G1301-f, and 16 and
12 % lower for the LI-7000 system, during night-time and daytime,
respectively (Fig. 5c and d). When compared to the EddyUH run, slightly smaller
deviations (10 and 8 % for G1301-f and 12 and 11 % for the
LI-7000) were found in the EddyPro run, but the trend in the deviations was
consistent between the software packages. A similar range of deviations was
found in Erottaja LE measured by LI-7200 (Fig. 6b). However, performing the
theoretical spectral correction on the Siikaneva LE had a minimal effect
(Fig. 5c and d), while the deviations found in LE measured by LI-7200 at
Erottaja ranged between 6 and 3 % (Fig. 6b). The use of a constant time
lag produces a small effect on LE, except during night-time when the higher
RH increases the sorption of H
For open-path systems the critical step is represented by the WPL (and
spectroscopic) correction. In Erottaja, the net CO
Software set-ups for the reference combination.
Cumulative sums of non-gap-filled flux time series are mostly within
The absolute differences between the cumulative CH
We have performed an inter-comparison between EddyUH and EddyPro, two public software packages for EC flux calculation. Both software packages feature up-to-date methods for EC raw data processing steps and corrections. Flux data as estimated by the reference combinations (Table 1) were in good agreement. In general, there were not significant systematic differences in the uncorrected fluxes calculated by EddyUH and EddyPro (Fig. 3). This suggests that an optimal choice for the raw data preparation and processing schemes leads systematic biases to be avoided in the fluxes between the two software packages. The most significant differences between the software packages only occurred after the flux corrections, and the impacts of these steps are further discussed below for fluxes measured by closed- and open-path systems.
Relative humidity (RH) dependence of low-pass filter time constant
as estimated with the two software packages. Note the different scale on
EC measurements from three closed-path systems (LI-7000, LI-7200 and
G1301-f) were processed, and the impact of different processing step
combinations was analysed using runs performed with EddyUH and EddyPro.
Among the different calculation procedures analysed, the spectral correction
was the most relevant for the closed-path LE measurements at the two sites,
the median values being between 6 and 16 %. On average, the use of
theoretical spectral correction gave up to 6 % lower LE in Erottaja, while
in Siikaneva the deviation respect to the reference run was generally below
3 %. We determined a stronger RH dependence of low-pass filter time
constant in Erottaja than in Siikaneva (Fig. 7) caused by the non-heated
sampling line there (Nordbo et al., 2013). Moreover, because of
different approaches in the spectral correction methods, the low-pass filter
time constants estimated by EddyPro in Erottaja were larger than those
estimated by EddyUH for RH values lower than 80 % (Fig. 7c). This may
explain the 2 % difference between LI-7200 LE as estimated by EddyPro and
EddyUH (Fig. 2g). Although the relative magnitudes of spectral corrections
are not directly comparable with other studies, they are in the same range as previously reported (Fig. 1). Surprisingly, at Siikaneva the theoretical
spectral correction gave on average 7 % higher
Cumulative sums estimated with EddyUH and EddyPro. Relative differences and data coverage are also shown. Data were not gap-filled prior to calculation of the cumulative sums.
In addition, we found that in Siikaneva the LI-7000
Finally, for our sites and datasets, the use of nominal constant time lag
was only an issue for nocturnal LE, when the absorption effect on H
For open-path gas analysers the WPL correction is the most critical step,
and although it depends on ecosystem type, season and target gas, the WPL
correction terms can often surpass the magnitude of the flux itself and also
change the sign of the measured target gas flux (e.g. Peltola et al., 2013).
Thus it is critical to perform this correction accurately, especially when
small target gas fluxes co-occur with large
If the response time which characterises the measurement system ability to
measure the flux contribution of small eddies, i.e. high frequencies, is
determined using power spectra, as done in EddyPro (Fratini et al., 2012),
then the high-frequency dampening caused by spatial sensor separation needs
to be estimated separately. In EddyPro it was done using the method proposed
by Horst and Lenschow (2009), while EddyUH uses cospectra to
estimate the measurement system's high-frequency response, and thus no
additional correction for sensor separation is needed. The Horst and
Lenschow (2009) method is based on co-spectral peak frequency
(
For LI-7700 at the Siikaneva site, it was shown that this correction resulted in
systematic differences between the software packages (Sect. 3.4) and between
the two co-located CH
We have estimated and analysed the flux uncertainty due to the use of two
software packages, using datasets 2 and 3 months long including CO
Data are freely available upon request from the authors.
EddyUH (
Post-field EC data processing with EddyUH is done through user-defined projects. A project in this context means a certain time period of data from a certain site that are processed with certain user-defined processing methods. These methods are determined by the user using the GUI and are saved in a set-up file where the site specifics and measurement system characteristics among other things are also defined. Therefore, all the processed data are always related to the saved project. The same project may include up to five different gas analysers combined with the same ultra-sonic anemometer, giving the possibility for the user to process several raw datasets at the same time. The software includes
Flow chart of EddyUH.
a number of modules, which operate at different levels of
post-processing (Fig. A1). Preliminary fluxes are calculated in the
preprocessor, where the first level of processing is done to the raw dataset.
Then several corrections are applied in the flux-calculation module, and the
final fluxes are calculated (Table A1). In order to optimise the processing
and properly apply all needed corrections, several software tools are
available (Fig. A1). Co-spectral data are used in the high-frequency spectral
transfer function estimator, where the low-pass filter time constant is
experimentally estimated for each gas according to Mammarella et al. (2009).
This approach is particularly relevant in the case of closed-path systems.
Further, the time lag optimiser is a useful tool to verify the correctness of
the chosen time lag window (and eventually refine it) for each gas, as well
as to determine the varying window boundaries for H
List of implemented methods for data processing in EddyUH.
In the first level of data processing several operations are done to the raw dataset in order to calculate uncorrected covariances of interest. Several methods related to these processing steps are available (see Table A1), and they are briefly presented below.
The raw data are quality-flagged according to physical plausible ranges of high-frequency values of each variable, diagnostic parameters (if available) and several tests, as described in Vickers and Mahrt (1997). Further spikes are then detected, and commonly, this is done applying the Vickers and Mahrt (1997) method. However, other methods also exist in EddyUH, e.g. the difference limits method, which compare the difference between consecutive data points to a given threshold for each raw data time series (see Rebmann et al., 2012 for more details). If the time series contains too many spikes, the data might be useless and a flux should not be calculated for the averaging period of interest (commonly 30 min). Foken (2008) suggests excluding time series with more than 1 % spikes from further analysis.
Current closed-path gas analysers measure H
In addition, H
A coordinate rotation is applied to the wind
velocity components, in order to align the
In order to extract the turbulent fluctuations from the measured time series, the mean values are subtracted from the time series. There are three methods available in EddyUH, i.e. block averaging, linear detrending and autoregressive filtering (Rebmann et al., 2012). Of these three methods, only block averaging fulfils the Reynolds averaging rules. All methods attenuate the low-frequency part of the cospectra. Block averaging has the smallest effect on the cospectra, whereas autoregressive filtering attenuates the cospectra the most (Rannik and Vesala, 1999). Linear detrending and autoregressive filtering are methods used to remove unwanted low-frequency variation (trend) in the signal (e.g. Mammarella et al., 2010). However, often block averaging is recommended.
Sonic anemometers calculate sonic temperature
The gas signal measured by closed-path analysers usually lags behind the wind
speed measurement made with the sonic anemometer. The time lag can be
estimated theoretically if sampling tube length and diameter are known, in
addition to the flow rate in the tube. However for H
Finally, covariances are calculated as a final step of the first processing level, which is performed by the preprocessor in EddyUH. Besides covariances and time lag estimates, the EddyUH preprocessor outputs include wind and gas signals statistics (mean, standard deviation, skewness, kurtosis), power spectra and cospectra for each averaging time period. In addition, quality statistics parameters are also calculated, e.g. flux steady-state and integral turbulence characteristics (Foken and Wichura, 1996), instrumental noise (Lenschow et al., 2000) and random flux error (Finkelstein and Sims, 2001). All these data are saved in monthly binary files, and then used by other modules (Fig. A1).
In the second level of processing, several corrections must be applied to
the 30 min covariances, and the set of corrections are different for closed- and open-path systems (see Fig. 1). At this stage the estimated covariances
are used to calculate the stability parameter defined as
Flux loss at high frequency is due to the
incapability of the measurement system to detect small-scale variation. The
inadequate frequency response, sensor separation and line averaging, and, in
closed-path systems, the air sampling trough tubes and filters are the main
reasons causing co-spectral attenuation. On the other hand, flux loss at low
frequency is due to limited averaging period (30 min) and trend removal. The
frequency response correction is usually performed based on a priori
knowledge of the system transfer function and the unattenuated cospectrum,
e.g.
The high-frequency transfer function TF
For measurements done with an open-path gas analyser, fluctuations in air
density cause apparent variations in measured scalar concentration, and this
needs to be corrected according to Webb et al. (1980). The correction is
performed to 30 min fluxes of the target gas, and the H
Gas molar concentration measurements carried out with instruments based on
laser spectroscopy (like LI-7700 and G1301-f analysers) also require
corrections for spectroscopic effects that affect measured values, in
addition to the above-mentioned WPL or dilution corrections. As these
spectroscopic effects are related to the changes in shape of the absorption
line, due to the changes in gas
temperature, H
When estimating the CH
For closed-path systems measuring H
The correction is based
on the transformation of sonic temperature (
Corrections to the covariances are repeated in an iteration loop until the flux change is smaller than 0.01 % (see Fig. 1). In EddyUH these steps are performed in the “flux calculation” module, including the estimates of flux density according to Eqs. (1)–(4).
The study was supported by Väisälä Foundation, EU projects InGOS and GHG-LAKE (project 612642), Nordic Centre of Excellence DEFROST and National Centre of Excellence (272041), ICOS-FINLAND (281255), CarLAC (281196), funded by Academy of Finland. We would also like to thank Gerardo Fratini for the useful discussion related to this study. Edited by: C. Ammann Reviewed by: M. Aubinet and two anonymous referees