Introduction
The Total Carbon Column Observing Network (TCCON) is a network of ground-based spectrometers that record near infrared (IR) direct solar spectra from
which column abundances of greenhouse gases are retrieved (Wunch et
al., 2011b, 2015). Column average dry-air mole fractions (DMFs, or
Xgas
where “gas” is the species of interest) measured by multiple TCCON sites
are used to evaluate Xgas retrievals from satellite measurements (for
example, Dils et al., 2014; Kulawik et al., 2016; Nguyen et al., 2014; Wunch
et al., 2011a). TCCON measurements are tied to the World Meteorological
Organization (WMO) in situ trace gas measurement scales through extensive
comparisons with in situ DMF profiles obtained by balloon and aircraft
measurements (Deutscher et al., 2010; Geibel et al., 2012; Messerschmidt et
al., 2011; Washenfelder et al., 2006; Wunch et al., 2010).
For the TCCON to meet the goals of satellite validation and carbon cycle
flux estimates, measurements need be precise and accurate. Currently, the
2σ single sounding uncertainties of the TCCON are estimated to be
0.8 ppm (0.2 %) XCO2 and 7 ppb (0.4 %) XCH4 (Wunch et al.,
2010). Systematic errors such as spectral ghosts (Messerschmidt et al.,
2010), pressure offsets, instrument misalignment, or improper fitting of the
continuum curvature (Kiel et al., 2016) can, however, produce systematic
biases between sites that will remain even after averaging many single
sounding measurements. An error analysis by Wunch et al. (2015) suggests
that biases of 0.2 % for XCO2 and 0.4 % for XCH4 could exist
in the network even though the retrieval algorithm (GGG) has undergone
continual improvements designed to reduce such biases.
In this study we quantify bias in XCO2 and XCH4 among the four
operational TCCON sites in the United States (US) in 2015. These sites
were at (1) the California Institute of Technology (Caltech), Pasadena,
California (CA); (2) Armstrong Flight Research Center (AFRC), Edwards, CA;
(3) Lamont, Oklahoma (OK); and (4) Park Falls, Wisconsin (WI). Bias quantification was
accomplished by comparisons with two mobile EM27/SUN spectrometers (Gisi et
al., 2012). A map of the US 2015 TCCON sites is shown in Fig. 1. The
campaign is described in Sect. 2; the data processing and some sensitivity
tests are described in Sect. 3. Comparisons between the sites are made in
Sect. 4.
Map of the United States with TCCON sites that were active in 2015
labeled. Normalized difference vegetation index (NDVI) from Terra MODIS
(Moderate Resolution Imaging Spectrometer; Didan, 2015) and nightlights from
VIIS (Visible Infrared Imaging Radiometer Suite) in red are shown for
September 2015.
Number of measurements prior to any filtering.
Site
Dates
No.
No. CIT
No. LANL
No. Co.*
TCCON
mFTS
mFTS
Caltech-1
10 Aug–15 Aug
708
22 338
18 119
145
AFRC
17 Aug–21 Aug
1831
31 980
22 402
283
Caltech-2
22 Aug–28 Aug
740
26 406
22 382
269
Lamont
31 Aug–4 Sep
1146
31 814
32 454
250
Park Falls-1
7 Sep–11 Sep
369
14 820
13 746
79
Park Falls-2
12 Sep
187
6018
6130
44
* Co. indicates 10 min averaged two-way coincident mFTS and TCCON data
points.
US TCCON 2015 intercomparability campaign
This campaign involved a comparison of simultaneous side-by-side measurements from two
EM27/SUN instruments with TCCON measurements. One EM27/SUN
instrument is operated by Caltech and one by Los Alamos National Laboratory
(LANL). These instruments have been described in detail elsewhere (Gisi et
al., 2012). Briefly, similar to the TCCON spectrometers, they measure direct
solar near IR spectra, albeit at a lower resolution (0.5 cm-1 versus
0.02 cm-1). They include an in-built solar tracker and are small and
stable enough to be easily transported. We also designate them as mFTSs for
mobile Fourier transform spectrometers (mFTSs) herein. For this study, both mFTSs
employed the standard InGaAs (indium gallium arsenide) detector. To reduce
the potential for drift between the mFTSs, the campaign was completed within
a 5-week period. Based on the lack of drift between the two mFTSs, we
conclude that the retrievals from their observations are internally precise
over this period so their Xgas measurements can be used as transferable
comparison products.
The general strategy of the campaign was to visit each of the four TCCON sites
shown in Fig. 1 and attempt at least 5 days of measurements. Two mFTSs were
used so any drift in their measurements would be noticed. In addition to the
spectrometers, a traveling Coastal Environment Weather Station with a
ZENO® data logger and Setra barometer was used for regular
meteorological surface measurements at the AFRC, Lamont, OK, and
Park Falls, WI, sites. At Caltech the on-site ZENO®
data logger and Setra barometer were used. This type of barometer is used at
each of the four US TCCON sites. The Setra sensor has a resolution of 0.1 hPa
and a stated accuracy of 0.3 hPa. A Paroscientific 765-16B Portable
Barometric Digiquartz® pressure standard with a stated
accuracy of ± 0.08 hPa or better was used as a traveling pressure
standard. The Digiquartz® was compared with each of the
on-site barometers. Surface pressure is important to the Xgas
retrievals because it is used to derive the pressure altitude for the site.
In Table 1 we present the dates of the campaign as well as the number of
coincident averaged measurements. Occasionally one mFTS recorded
significantly fewer spectra due to unexpected halts during acquisition. This
issue was mostly resolved by updating to the latest firmware provided by
Bruker™ while at AFRC, but it shows an advantage of having multiple mFTS
instruments. Our quality control filters were set after a preliminary look
at the data. For this study our filters included 392 ppm < XCO2 < 404,
1.79 ppm < XCH4 < 1.865 ppm,
and solar variation < 0.5 % within an interferogram. Prior to the
campaign several of the TCCON sites used a mercury manometer as an absolute
pressure reference. In the comparisons shown here, the current version of
the public TCCON data (R0 for Park Falls, R1 for all others) are used where
the surface pressure measurements at all sites are tied to the
Digiquartz® (Iraci et al., 2014; Wennberg et al., 2014a,
b, c). The mFTSs used the meteorological data from the Caltech on-site station or from the traveling Setra barometer with offsets applied to
match the Digiquartz®.
Site characteristics – Caltech
The Caltech site is located in Pasadena, CA (34.136∘ N,
118.127∘ W; 240 m a.s.l.), in the California South Coast Air Basin
(SoCAB). Pasadena is in an urban environment where there are large diurnal
variations of Xgas pollutants because of emissions and advection (Wunch
et al., 2009, 2016). Emissions from the basin are estimated to be 167 Tg CO2 yr-1 and 448 ± 91 Gg CH4 yr-1
(Wunch et al., 2016). Pasadena is located towards the northern end
of the basin, which is bounded by mountains. Two additional sides of the
basin are also bounded by mountains, and the other side is bounded by the
Pacific Ocean. General conditions during the August 2015 campaign were mostly
clear skies with some cirrus clouds. We treat 2 different weeks at Caltech
separately to estimate the limits of our methodology. The mean measured
daytime XH2O for both weeks was 3540 ± 840 ppm (1σ).
Site characteristics – AFRC
The AFRC (also called Dryden or Edwards)
is located in the Mojave desert at 34.960∘ N, 117.881∘ W (700 m a.s.l). It is approximately 100 km north of Caltech and 100 km east
of Bakersfield, CA. AFRC is on a military base, but the surrounding
area is much less densely populated than the SoCAB. The area is mostly flat
and devoid of vegetation. General conditions here during the campaign were
cloud free with daytime surface temperatures of 36.4-13.2+4.0 ∘C (95 % confidence intervals, or CI) and a mean measured daytime
XH2O of 2640 ± 250 ppm (1σ).
Site characteristics – Lamont
The Lamont, OK, site is located in an agricultural region that is
mostly flat with some rolling hills (36.604∘ N, 97.486∘ W; 320 m a.s.l.). It is situated on the Atmospheric Radiation Measurement
(ARM) Southern Great Plains (SGP) site. The surrounding area is sparsely
populated. During the campaign cumulus clouds were present covering from
less than 5 % to approximately 40 % of the sky. The mean measured
daytime XH2O for the campaign week was 5080 ± 890 ppm (1σ).
Site characteristics – Park Falls
The Park Falls, WI, TCCON site has been described in more detail
elsewhere (Washenfelder et al., 2006). Briefly, the site is in a sparsely
populated but heavily forested region with low topographic relief
(45.945∘ N, 90.273∘ W; 473 m a.s.l.). Conditions were
highly variable, ranging from nearly cloud free to full coverage by
stratocumulus clouds. Despite planning more days at this site, the often
cloudy conditions contributed to collecting the least amount of data. On 11
September 2015, the TCCON IFS 125HR instrument was realigned as part of
routine maintenance. We treat the days before and the day after alignment
separately. The mean measured daytime XH2O was 2480 ± 750 ppm
(1σ) for this period.
Comparisons
Because of different spectral resolutions between the TCCON instruments
(0.02 cm-1) and the traveling spectrometers (0.5 cm-1), we
anticipate systematic differences in their Xgas retrievals (Gisi et
al., 2012; Petri et al., 2012). Even in the absence of instrumental
problems, spectroscopic inadequacies can cause systematic differences that
correlate with T (temperature) errors, surface pressure errors, and solar
zenith angle (SZA; Wunch et al., 2011b). In addition, the instruments have
different averaging kernels (AKs) due to differences in spectral resolution. Thus,
even though we use the same a priori gas volume mixing ratio and
temperature profiles, errors therein will produce differences in the
retrieved Xgas products (e.g., compare Wunch et al., 2015, and
Hedelius et al., 2016). In this section we consider five reasons why the
Xgas products between the two instrument types (mFTSs and TCCON) may
differ.
First, we consider AM-dependent artifacts that arise due to the effect
of spectroscopic errors being resolution dependent. Second, we consider how
surface pressure bias could affect retrievals, noting that surface pressure
bias should be minimal amongst the current US TCCON data because
of standardization to the common traveling Digiquartz®
standard. Third, we consider effects of errors in the a priori temperature
profile on retrievals from higher- versus lower-resolution spectra. Fourth,
we consider the effects of differences in sensitivity from the AKs. Finally, we mention how a non-ideal ILS (instrument line shape) may
affect retrievals.
Unadjusted comparisons and AM dependence
The comparisons prior to accounting for differences in temperature
sensitivities and AKs are shown as box plots in Fig. 3 (Δ=TCCON-mFTS). The mFTS data were scaled to match the TCCON
product and center the difference about zero, by dividing by scaling factors
of 0.9987 for XCO2 and 1.0073 for XCH4. These factors were based
on the TCCON and mFTS data at all sites and were used in combination with
the TCCON to in situ profiles bias correction (Wunch et al., 2015). An
additional scaling factor is used because retrievals from lower-resolution
spectra are biased compared to higher-resolution spectra due to errors in a
priori profiles and spectroscopy (Gisi et al., 2012; Hedelius et
al., 2016; Petri et al., 2012). For the box plots, we use the convention that
the whiskers are 90 % CI.
AM- or SZA-dependent differences may arise due to spectroscopic
errors (Frey et al., 2015). At higher SZAs sunlight passes through a longer
atmospheric path, which increases the depth of the measured transmission
lines. Spectroscopic errors can lead to bias that varies with SZA, even in
clean air sites (Wunch et al., 2011b). Though adding in an AM-dependent
correction did not improve the long-term mFTS to TCCON comparison in
previous studies (Hedelius et al., 2016), here we noted significant
AM dependencies. Air-mass-dependent corrections are accounted for in TCCON
data, but these are developed for the high-resolution observations (Wunch et
al., 2011b). When we attempted to correct the Xgas from the mFTS
measurements as a function of SZA, we noted significant influences from
local sources and sinks, even at the non-Caltech sites. This complicated the
separation of the spurious effects with AM from true atmospheric variation.
Additional measurements in areas with little atmospheric variation could aid
in accounting for AM artifacts (Klappenbach et al., 2015). In this study, we
apply a symmetric basis function to the mFTS products following Eq. (A12) in
Wunch et al. (2011b), with coefficients determined empirically to reduce the
overall diurnally varying difference data between the mFTS and TCCON
retrievals. Further, for estimates of bias we only use data within ±2 h of local noon so that comparisons are over similar SZAs at all sites.
This constrains comparison data to have an AM between 1.05 and 1.85 (site
means between 1.10 and 1.46). Recent work has shown residual dependencies on
AM that could cause a high bias of ∼ 1 ppb XCH4 between
AM 1.10 and 1.46 (Matthaeus Kiel, personal communications, 2017).
Surface pressure and temperature considerations
Surface pressure is used in the calculation of the dry-air column in GGG. It
is an input to the retrievals to set the pressure altitudes of each site. A
+1 hPa bias in surface pressure leads to average biases of approximately
+0.036 % XCO2 and +0.039 % XCH4, respectively, for
10∘ < SZA < 20∘ and +0.034 %
XCO2 and +0.049 % XCH4, respectively, for
70∘ < SZA < 80∘ (Wunch et al., 2015). Because pressure
measurements are tied to the same Digiquartz® sensor
(accuracy of ±0.08 hPa), surface pressure errors are expected to
contribute less than 0.01 % to the XCO2 and XCH4 retrievals.
Histograms of differences in temperature from those used in the
retrievals at the surface (NCEP model) as opposed to the temperature
measured at the TCCON sites.
At different temperatures, the distribution of the molecular J states
differs, which can affect the relative strengths of overlapping lines from
different species. In GGG bands are chosen to be reasonably temperature
insensitive by including both high and low J lines to average out temperature
sensitivity. In the lower-resolution spectra, lines are less well resolved.
When the algorithm attempts to fit the lines, the overall fit may still be
good even if fits for individual species are incorrect, but in compensating
ways.
We define a temperature error as the a priori surface interpolated
temperature minus the measured site temperature. Histograms of the
temperature errors at the different sites are shown in Fig. 4. In general,
NCEP temperatures are typically cooler than those measured on site. At AFRC
the difference is particularly large: the NCEP reanalysis product
underestimates the surface temperatures by ∼ 10 K at times in
this desert region for this particular week. We also compared interpolated
surface temperatures from the European Centre for Medium-Range Weather
Forecasts (ECMWF; 0.125∘ × 0.125∘), MERRA-2
(Modern Era Retrospective-Analysis for Research and Applications), GEOS-5
(Goddard Earth Observing System Model), and NAM12 (North American Mesoscale
Forecast System, 12 km). Model surface temperature is lower than the AFRC TCCON
temperature in all cases, and three of the five models have noon differences of
∼ 10 K. Differences are ∼ 7 K for GEOS-5 and
∼ 5 K for NAM12. Though error in the measurement may
contribute to part of the T difference, the lower-resolution dynamical
models may have a difficult time reproducing surface T at AFRC.
To account for error in the a priori temperature profiles near the surface,
we apply two different tests separately. First, we define the temperature
error from the surface to 700 hPa as equal and apply the results described in
Sect. 3. Second, we apply corrections defining the temperature error
separately at each level. The error at each level k was defined as the
difference from the NCEP profile potential temperature θNCEP,k-θmeasured,s (where “s” stands for surface) when θmeasured,s >θNCEP,k. Thus potential temperatures
aloft are always greater than or equal to θmeasured,s. Both
corrections reduce the diurnal trend of the ΔXCH4 and ΔXCO2 during the middle
hours of the day but do not significantly
alter the comparisons in the late afternoon. True temperature profiles are
likely different from the NCEP noon profiles. Future releases of GGG will
apply a post facto temperature correction for the lowest 3 km based on
temperature-dependent water lines (Toon et al., 2016b). For future studies,
we recommend adding dedicated sondes as part of the instrument suite for
these field campaigns.
A comparison of the averaging kernels at three different SZAs for the
higher-resolution (HR) and lower-resolution (LR) instruments. The LR
instruments are more sensitive to changes at the surface but less sensitive
to changes in the stratosphere.
Averaging kernel differences
AKs (Fig. 5) are different for the 0.02 and 0.5 cm-1 instruments. We apply Eq. (A13) from Wunch et al. (2011a) to the
TCCON Xgas (c) product to reduce the smoothing error (the
contribution of different AKs). We denote the mFTS by
subscript 1, the TCCON by subscript 2, and the TCCON product adjusted to
reduce the smoothing error of the mFTS AKs (AKs) as
1←2.
c^1←2=ca+γ2-1∑jhja1jxaj
A ^ represents a retrieved quantity, the subscript “a”
denotes the prior, h is the pressure weighting function described
by Connor et al. (2008), a is the column AK, x is the
DMF a priori profile, and γ is the overall scaling factor applied to
the TCCON a priori profile to obtain the retrieved Xgas. Both the TCCON
and the mFTS use the same a priori profiles. In Eq. (1), the TCCON profile
γxa is treated as an approximation to the true
atmospheric DMF profile (compare Eq. 3 from Rodgers and Connor, 2003). This
is a better approximation in a sparsely populated location such as Lamont
than at Caltech where local anthropogenic emissions strongly influence the
atmosphere. However, overall the application of Eq. (1) only makes differences
of 0.00-0.04+0.04 ppm and 0.01-0.07+0.17 ppb (95 % CI)
for XCO2 and XCH4 in this dataset.
GGG2014 a priori profiles do not take into account local anthropogenic
emissions at the surface. In Fig. 6 we plot the in situ DMFs of CO2 and
CH4 measured near the surface throughout the day as well as those from
the a priori profiles used in the GGG2014 retrievals at the Caltech site.
The in situ measurements were recorded using a Picarro cavity ring down
spectrometer, with standardization by comparison to three NOAA (National
Oceanic and Atmospheric Administration) standards every 23 h. Given the
intense local emissions, the measured in situ DMFs are significantly larger
than the a priori near the surface. Using the same assumptions as Hedelius
et al. (2016), the Xgas retrievals for two instruments in a polluted
environment where the true and a priori profiles differ only at the surface
are related by
c^1=a1,sa2,sc^2-ca+ca.
Note the error term has been omitted. The subscript s represents
the surface. These assumptions are better for XCO2 than for XCH4
as changes in tropopause height can also make the a priori methane profile
significantly different from the true profile (Saad et al., 2014). Over this
time at Caltech, XHF averaged ∼ 50 ppt and γHF
averaged ∼ 0.87, suggesting an a priori tropopause height that
is too low. Using the β value from Saad et al. (2014) we estimate a
13 % difference in γHF due to tropopause height would cause
about a 0.24 % change in γCH4 (∼ 4 ppb), which is
large enough that Eq. (2) is not valid for XCH4. We apply Eq. (2) to
the XCO2 TCCON retrievals at the Caltech TCCON site, which leads to an
adjustment of 0.22-0.35+0.54 ppm (95% CI).
(a) Diurnal variation of in situ DMFs measured near the
surface at Caltech on the days of TCCON to mFTS comparisons. A priori
surface values are marked by an “x” at noon. (b) GGG2014 a
priori profiles used in the retrievals, with lower CO2 and CH4
than was measured near the surface. Surface pressure is indicated by the
dashed line.
Effects of a non-ideal ILS
Imperfections in the ILS due to misalignment of the TCCON
FTSs can also cause site biases. At the sites described in this study,
weekly internal lamp measurements of the internal, calibrated HCl cells
(Hase et al., 2013) are collected from the 125HR instruments. We use LINEFIT
14.5 (Hase et al., 1999) software on HCl lines from monthly-averaged spectra
to characterize the ILS. For Park Falls spectra were averaged before and
after realignment. In Fig. 7 are the ME and phase
error (PE) with OPD. An ME not equal to 1 can indicate instrument
misalignment, which may be from shear, angular, or defocus misalignment.
Effects of different types of misalignment on ME are not independent (Toon
et al., 2016a). However, parameterizing changes in ME with OPD can be used
to assess effects on Xgas retrievals (Griffith and Macatangay, 2010; Velazco et
al., 2016; Wunch et al., 2011, 2015). These previous studies have found that
each 1 % increase in ME at MOPD leads to a decrease on the order of 0.04 %
in XCO2, though the change does vary with SZA. For XCH4, there is
a decrease on the order of 0.03–0.05 % for a 1 % increase in ME at MOPD.
The cause of the change in ME with OPD can, however, also significantly
influence results. For example, Wunch et al. (2015) noted significantly
different results for the same change in ME when the cause is shear versus
angular misalignment.
Modulation efficiency and phase error for each of the 125HR
instruments describe the ILS. Results are calculated from HCl lines using
LINEFIT 14.5 on monthly averages of internal lamp spectra. For Caltech, 2
different months are shown and Park Falls-1 corresponds to August 2015 and
Park Falls-2 corresponds to October 2015.
We estimate biases based on ME at MOPD values alone, compared with AFRC.
Based on the LINEFIT analysis of the lamp spectra, we would expect a low
XCO2 bias of 0.02 % for Caltech, a high bias of 0.05 % for Lamont,
and a high bias of 0.09 % for Park Falls (prior to realignment). The
results of our study are not consistent with this expectation. Only Park
Falls is consistently in the right direction with a bias of ∼ 0.18 % before realignment. After realignment, Park Falls XCO2 was more
in line with the other spectrometers, although based on the ME at MOPD
results alone there should have been a change in the opposite direction. The
Park Falls ILS was much more symmetrical after realignment, as seen by the
PE curve in the lower panel of Fig. 7 being much closer to zero. For
XCH4, both Park Falls and Lamont are biased in the expected direction
from Armstrong, and the Park Falls-1 bias is ∼ 0.25 %.
However, the Lamont bias is greater than expected from the single value
parameterization. A more complex parameterization of the ILS effect on
Xgas (e.g., using the full function of ME with OPD, accounting for SZA
dependence) might reduce the expected versus observed mismatch.
The Xair parameter from GGG can be used as a diagnostic for large
misalignments, timing, and surface pressure errors. Xair is calculated
by dividing the sum of all non-water molecules based on the surface pressure
by the retrieved column of dry air based on column O2. Xair should
be close to 1.0 and not vary, though empirically it is approximately 2%
lower due to spectroscopic errors for oxygen (Washenfelder et al., 2006).
Wunch et al. (2015) showed an increase of about 0.3 % in Xair for a
1% increase in ME at MOPD due to shear misalignment, and the change due
to angular misalignment was < 0.03%. In Fig. 8 Xair is shown
for all the sites. At Park Falls Xair was approximately 0.979 before
and 0.983 after alignment, which could correspond to an ME increase of about
0.013 at MOPD from shear realignment. LINEFIT results actually show a
decrease in ME at MOPD after 11 September 2015, but XCO2 and XCH4
decreased. Based on Xair, XCO2 was expected to change by
∼ 0.2 ppm (compared with ∼ 0.08 ppm) and
XCH4 was expected to change by 0.7–1.2 ppb (compared with
∼ 1.5 ppb). Residual differences may indicate measurement
uncertainties.
TCCON Xair compared with mFTS Xair within
±2 h of local noon. The differences are scaled by 1.001 to be centered
about zero. Xair can be used as a diagnostic for misalignments, timing,
or surface pressure errors.
Medians and standard deviations of the TCCON data compared to the
mFTS product after various adjustments. Line style represents the
significance of the difference of the group median from the median of all
data by the Kruskal–Wallis test (p < 0.05 - , p < 0.2 - -,
otherwise …). Legend entries indicate
what adjustments were applied to the data to make measurements from the
different instrument types more comparable. Open symbols did not have a
scaling factor applied to center about zero. AM is air mass adjustment; T
is temperature error adjustment; AK is averaging kernel adjustment.
Truncated 125HR interferograms comparisons
Retrievals from the 125HR and mFTS instruments are inherently different due
to the differences in resolution. By truncating the longer 125HR
interferograms to the same length as those collected from the mFTS, similar-resolution spectra are obtained. This likely eliminates most discrepancies
between the different types of measurements, except for some residual
instrumental imperfections such as instrument misalignment or ghosts.
Truncation also reduces the effects of ME variations due to the smaller
MOPD. Truncation has been performed in past studies comparing retrieved
Xgas from different-resolution spectrometers (Gisi et al., 2012;
Hedelius et al., 2016; Petri et al., 2012). This test provides little new
information if truncation changed all retrieved DMFs in a uniform manner.
However, past studies showed truncation does not necessarily affect all
results the same way, which makes this test imperative in diagnosing
potential causes of differences. It helps in determining which biases likely
arise from instrumental issues and which arise from other issues such as
errors in the forward model (e.g., from temperature biases at different
locations).
The results of the truncation test are shown Fig. 9, and changes are most
easily seen from the unscaled (open) points. The sign of the change for
XCO2 is inconsistent for the different sites. Previous studies also
noted changes that were negative (Petri et al., 2012), positive (Gisi et
al., 2012), or both (but with a preference towards negative; Hedelius et
al., 2016) when using lower-resolution spectra. For lower-resolution spectra
XCH4 increases, in agreement with previous studies (Hedelius et al.,
2016; Petri et al., 2012).
Mean differences pre- and post-adjustment for ±2 h of
local noon.
XCO2
AM
AM+T
AM+T+
Trunc
(ppm)
AK
1n∑Md
0.17
0.18
0.11
0.14
1n∑σ
0.34
0.34
0.34
0.42
XCH4 (ppb)
1n∑Md
1.1
1.1
1.2
1.7
1n∑σ
1.9
1.8
1.8
1.8
Biases to overall median
The medians and standard deviations for data before and after considering
differences in AKs, and surface temperature are shown in Fig. 9. Though we
have attempted to reduce artificial diurnal variation between the different
instruments with the AM correction, there may still be some residual
dependence with SZA. To reduce this dependence, which is larger at higher
SZAs, only data within ±2 h of local noon are used. We use the
Kruskal–Wallis one-way analysis of variance test, which assumes ordinal but
not necessarily normally distributed data (Kruskal and Wallis, 1952), and
compare data from each site to the median of data from all sites. The null
hypothesis
of this test is the medians do not significantly differ. Line styles
indicate the degree of significance by the Kruskal–Wallis tests.
Pairwise 95 % CI of differences between sites. Differences for
data within ±2 h local noon. Comparisons are ranked in order of
decreasing mean difference. For each species, plots are shown for (1) corrections for air mass, differences in temperature sensitivity errors
defining temperature errors layer by layer, and a reduction of the smoothing
error from different averaging kernels; (2) differences by comparing results
from 125HR spectra with lowered resolutions. At the bottom are the site
orderings. Lines between indicate when the pairwise difference is first more
than 0.
Pooled differences are listed in Table 3 for different adjustments. These
are represented by the averages of the group median differences, the overall
median, and the average standard deviations. Park Falls TCCON data prior to
realignment of the spectrometer are omitted. The sum of the median
differences decreases for XCO2 after adjustments. However, this is not
true of XCH4, which increases in variability after adjustment. Despite
this overall increase for XCH4, these adjustments better reflect the
intercomparability of the sites rather than the intercomparability of
measurements from differing instruments. From Table 3, we estimate the
average biases of all sites compared to the median to be 0.03 % XCO2
and 0.08 % XCH4.
Confidence intervals of the pairwise differences
We use the Critchlow–Fligner method to estimate simultaneous CI for the differences between all pairs of sites (Hollander et
al., 2014). The Critchlow–Fligner test is nonparametric so it is less
sensitive to outliers and few assumptions are needed about the distribution
of the underlying population of data. We use α= 0.05 to obtain
95 % confidence intervals of the differences between sites. Results are
presented in Fig. 10 in order of decreasing median difference and separated
by gas and adjustments. At the bottom are the ordering of the sites.
This comparison suggests for XCO2 is lowest for Lamont and highest for
Park Falls-1 in both cases. There is a difference between the 2 different
weeks at Caltech for unknown reasons. The largest difference within a 95 %
CI is 0.6 ppm between Park Falls and Lamont; this difference is 1.0 ppm for
the truncation test. However, most mid-range values are ∼ 0.2
to 0.3 ppm.
For XCH4, there was more of a change in site order between the two cases.
For the truncation comparison the differences are even greater than
AM+T+AK comparison as indicated in Table 3. The largest difference
within a 95 % CI is 4 ppb between Lamont and Caltech. For the truncation
test the largest difference is between Armstrong and Caltech and is greater
than 5 ppb. Mid-range values are 2–3 ppb.
Conclusions
We estimate the range of statistically significant site-to-site bias amongst
the sites as < 0.3 ppm for XCO2 and < 3 ppb for
XCH4. These were determined by comparing TCCON data with simultaneously
collected data from co-located portable spectrometers, which we have assumed
to be internally precise over the duration of the campaign. This assumption
is supported by standard deviations of only 0.15 ppm for XCO2 and 1 ppb
for XCH4 for the 10 min averaged differences between the two mFTS
instruments over the campaign. Five reasons Xgas could differ among
instruments were considered: (1) differences in averaging kernels, (2) differences in spurious air mass dependence from spectroscopy errors, (3) the
a priori profile (e.g., temperature profile), (4) error in the measured
surface pressure, and (5) instrument misalignments. Of these, the last four
can cause site-to-site biases in the TCCON, and empirical adjustments to
make the mFTS and TCCON datasets more comparable were made to the first
three. When the 125HR interferograms were truncated so the spectra would be the
same resolution as the mFTSs, differences from the first three inherently go
away.
As the spectroscopy is improved, the data should have smaller
AM-dependent artifacts, though for now an empirical correction is used
for the TCCON (Wunch et al., 2011). Updates to the retrieval algorithm to
include line mixing may also make the AM dependence more predictable
(Hartmann et al., 2009). The corrections based on T errors described in
Sect. 4.2 are for the differences in sensitivity to T error between the mFTS
and TCCON instruments and not for the different T errors at each TCCON
site. Large temperature errors of +10 K from the surface through 850 hPa
could cause errors of 0.08 % in XCO2 and 0.11 % in XCH4 at
an air mass of 1.5. Biases due to a non-ideal ILS will be reduced in future
versions of the GGG retrieval algorithm. Biases in surface pressure data can
cause site biases but are expected be less than 0.01 % in the current
data revisions because surface pressure data were standardized to the same
traveling standard. We recommend regular (∼ annual, depending
on the pressure sensor accuracy) comparisons of meteorological pressure
measured by on-site barometers with a universal standard for those making
similar column measurements.
Remaining differences are most likely from a combination of other errors
mentioned by Wunch et al. (2015), such as instrumental misalignment and
Doppler shifting of solar lines with respect to telluric lines. Some of
these uncertainties will be reduced in the next version of GGG. Other
remaining differences may be due in part to noise. Sufficiently large sample
sizes should have helped reduce bias from noise, and the 15 min running
standard deviations for TCCON were 0.11 % XCO2 and 0.13 %
XCH4. Apparent differences between the weeks at Caltech suggest we are
near the precision limit of our current methodology. Though we reduced the
contributions of ΔXgas from different instruments, there may
remain additional contributions because of differences in resolution (Petri
et al., 2012).
United States TCCON site-to-site biases measured herein are within the
2σXCO2 and XCH4 uncertainties stated by Wunch et al. (2010).
We suggest repeating this study comparing results from traveling
spectrometers with those from the stationary TCCON sites, especially when
aircraft and air-core data are not available to check for bias. Ideally
repeat campaigns will include multiple traveling mFTS instruments. Others
may even consider taking three mFTS instruments so if there is a change from
one it would be noticeable by comparing with the other two. When
collocated, three or more EM27/SUN instruments can easily be operated by just one or
two people. Multiple instruments also provide backup in case problems arise
with one and can increase the signal to noise ratio. As a backup strategy,
one traveling mFTS can be taken in the field and compared with an mFTS
instrument left in a fixed location before and after the campaign. This
second strategy is acceptable when there are no instrumental issues, or if
it is known when and how issues affect Xgas measurements. This type of
campaign can be repeated every few years, or with different sites (e.g., Sha
et al., 2016), or with different gases that can be measured with an
extended-band InGaAs detector with spectral filters (Hase et al., 2016). Similar
studies should, however, also consider the current precision limits of these
comparisons on various timescales. We hope others will improve on our
methodology to estimate inter-site biases using portable spectrometers. A
sufficient number of aircraft profiles may also aide in determining
intercomparability. The NASA Atmospheric Tomography Mission (ATom), for
example, will conduct global flights summer 2016 through spring 2018 and
will include profile measurements of CO2, CH4, CO, and N2O
over many of the TCCON sites (https://espo.nasa.gov/home/atom). Data from
ATom can be used to reevaluate TCCON uncertainties in the next version of
GGG.