Accuracy and precision of eddy covariance flux measurements depend among other factors
on the method of high-pass filtering (HPF). Even if the time series are not filtered by linear
detrending or recursive filtering, measurements are high-pass filtered as a consequence of the
limited sampling interval. This makes correction for HPF necessary. Although some
recommendations exist, different methods are used by different groups and affect the flux
estimates in a different way.
We investigated annual CO2 flux data sets from 8 different European flux sites comprising
a wide range of surface properties, climates and technical setups for the effects of linear
detrending, recursive filtering and block averaging on systematic and random errors of the
flux estimates. To correct for HPF effects, we used the spectral transfer function approach
based on a common model spectra parameterisation.
Results showed comparably large absolute effects of HPF correction (Tab. 1), especially
above rough surfaces (4 to 15% of the annual C uptake). Above smooth surfaces, like the
cropland the grassland and the wetland sites, HPF effects were small. The higher the
attenuation of the filter, i.e. linear detrending and recursive filtering with small time constants,
the higher was the reduction of both the random error and the absolute flux signal. The HPF
effects were the larger the higher measurement height and the smaller wind velocity which
were two main effects leading to different accuracies at the different sites in this
comparison.
Tab. 1 Effects of high-pass filtering correction on annual flux estimates and their standard
deviation (g C m-2 yr-1). Treatments: BA, block averaged, LD, linear detrended, RM,
recursive filter with time constants (Ï) equal to different fractions of sampling time
(Ts)
surface type BA LD RM Ï=1/8 TsRM Ï=1/4 TsRM Ï=1/2 Ts
smooth (n=3)-1(±0.2)-1(±0.5) -2(±0.8) -1(±0.4) -1(±0.2)
rough (n=5) -8(±2.1)-16(±4.0)-27(±6.6) -15(±3.6) -7(±1.8)
However, like also found earlier, the systematic effect was largely reduced, but not fully
compensated by the correction (Tab. 2). This effect introduced systematic bias to the flux
estimates caused by the choice of the HPF filter method. Block averaging was used as a
reference, but it is not clear how much the corrected flux after block averaging is biased by an
incomplete correction. With one exception the residual differences were much smaller than
not correcting the flux for HPF.
Tab. 2: Differences between high-pass filtering corrected annual flux estimates after linear
detrending or recursive filtering compared to block averaging and their standard
deviation.
surface type BA LD RM Ï=1/8 TsRM Ï=1/4 TsRM Ï=1/2 Ts
smooth (n=3)reference2(±1.3)3(±1.0) 1(±0.4) -1(±0.4)
rough (n=5) reference3(±0.6)5(±1.4) 5(±0.6) 7(±1.5)
The investigation raises the question, if our understanding of turbulent fluxes is yet
complete, i.e. whether low-frequency flux contributions should be accounted for in plot scale
flux studies or not and how these flux contributions can be determined and realistically
represented by model spectra. The quantitative results indicate that this question is
practically more relevant for forest studies than for cropland or grassland studies.
With our presentation we aim at raising a discussion about the usefulness of the
suggested spectral correction approach and want to raise awareness on accuracy and
precision issues with regard to high pass filtering among users of the eddy covariance
technique.
Acknowledgements:
We gratefully acknowledge funding by the EU-IMECC (Infrastructure for Measurements
of the European Carbon Cycle) project. |