Comparison of two closed-path cavity-based spectrometers for measuring air – water CO 2 and CH 4 fluxes by eddy covariance

In recent years several commercialised closedpath cavity-based spectroscopic instruments designed for eddy covariance flux measurements of carbon dioxide (CO2), methane (CH4), and water vapour (H2O) have become available. Here we compare the performance of two leading models – the Picarro G2311-f and the Los Gatos Research (LGR) Fast Greenhouse Gas Analyzer (FGGA) at a coastal site. Both instruments can compute dry mixing ratios of CO2 and CH4 based on concurrently measured H2O, temperature, and pressure. Additionally, we used a high throughput Nafion dryer to physically remove H2O from the Picarro airstream. Observed air–sea CO2 and CH4 fluxes from these two analysers, averaging about 12 and 0.12 mmol m−2 day−1 respectively, agree within the measurement uncertainties. For the purpose of quantifying dry CO2 and CH4 fluxes downstream of a long inlet, the numerical H2O corrections appear to be reasonably effective and lead to results that are comparable to physical removal of H2O with a Nafion dryer in the mean. We estimate the high-frequency attenuation of fluxes in our closed-path set-up, which was relatively small (≤ 10 %) for CO2 and CH4 but very large for the more polar H2O. The Picarro showed significantly lower noise and flux detection limits than the LGR. The hourly flux detection limit for the Picarro was about 2 mmol m−2 day−1 for CO2 and 0.02 mmol m−2 day−1 for CH4. For the LGR these detection limits were about 8 and 0.05 mmol m−2 day−1. Using global maps of monthly mean air–sea CO2 flux as reference, we estimate that the Picarro and LGR can resolve hourly CO2 fluxes from roughly 40 and 4 % of the world’s oceans respectively. Averaging over longer timescales would be required in regions with smaller fluxes. Hourly flux detection limits of CH4 from both instruments are generally higher than the expected emissions from the open ocean, though the signal to noise of this measurement may improve closer to the coast.


Introduction
Eddy covariance is a direct and non-intrusive method for quantifying the vertical transport of carbon dioxide (CO 2 ) and methane (CH 4 ).This micro-meteorological technique derives fluxes from rapid fluctuations in the atmospheric CO 2 and CH 4 and thus requires a high-frequency (typically 10 Hz) chemical sensor.Over the last couple of decades, open path infrared gas analysers (IRGAs), such as the LI7500 (LI-COR Biosciences, Lincoln, Nebraska, USA), have been widely used to measure the atmosphere-biosphere as well as atmosphere-ocean exchange of CO 2 .A similar instrument (LI7700) is available for measurements of CH 4 flux.The advantages of open-path sensors include small size, low power consumption, minimal lag time between chemical and wind sensors, and no high-frequency signal attenuation from the inlet tube.Miller et al. (2010), Blomquist et al. (2014), and Landwehr et al. (2014) demonstrated that air-sea CO 2 fluxes measured by LI7500s are subject to substantial biases due to variations in water vapour (H 2 O).When measuring air-sea CO 2 fluxes, which are typically of the order of a few mmol m −2 day −1 , bias in the measured CO 2 fluxes due to H 2 O can be 1 to 2 orders of magnitude greater than the actual CO 2 flux signal.This measurement bias is not related to the application of the air density correction (i.e.Webb et al., 1980); rather, it seems to be caused by cross-sensitivities between CO 2 and H 2 O in the forms of spectral interference and pressure broadening (Kondo et al., 2014).The low-flux magnitude of CO 2 over the ocean and high humidity exacerbate this measurement bias.Furthermore, there appears to be little consistency even within the LI-COR family of products, with the LI7500 and the more recent LI7200 showing opposite signs in the H 2 O bias (Landwehr et al., 2014).Uncertainties and nonlinearity in these cross-sensitivities complicate any mathematical corrections that try to address them (e.g.Prytherch et al., 2010;Edson et al., 2011).Miller et al. (2010) first showed that by converting the open path LI7500 to a closedpath configuration and physically drying the sampled air with a Nafion dryer, the accuracy and precision of the covariance CO 2 flux is significantly improved.The benefit of removing H 2 O from the sampled air when measuring air-sea CO 2 fluxes with IRGAs was rigorously and quantitatively confirmed by Blomquist et al. (2014) and Landwehr et al. (2014).
In more recent years, cavity ring-down spectroscopy (CRDS; O' Keefe and Deacon, 1988) as well as off-axis integrated-cavity output spectroscopy (OA-ICOS; O' Keefe et al., 1999) have been developed and commercialised to measure CO 2 and CH 4 .Two of the leading manufacturers are Picarro Inc., Santa Clara, California, USA (CRDS) and Los Gatos Research (LGR), Mountain View, California, USA (OA-ICOS).These instruments regulate the cavity (i.e.measurement cell) temperature and pressure, obviating the parts of the Webb et al. (1980) correction that are caused by fluctuations in air temperature and pressure.Blomquist et al. (2014) coupled a prototype Picarro CRDS instrument (G1301-f) to a Nafion dryer and measured CO 2 fluxes with an order of magnitude better precision (and free from H 2 O bias) compared to measurements with open-path IRGAs (see more details in Sect.3.2).
The latest models from Picarro (G2311-f) and LGR (enhanced performance Fast Greenhouse Gas Analyzer, FGGA) measure CO 2 , CH 4 , H 2 O, as well as temperature and pressure simultaneously at 10 Hz.Observed H 2 O mixing ratio is used to numerically correct on a point-by-point basis for the effect of humidity on the CO 2 and CH 4 signals by the instruments' internal software.This in theory enables the derivation of dry CO 2 and CH 4 fluxes without the need for a physical dryer (e.g.Nafion).The LGR FGGA is a fairly recent instrument and we are not aware of any published eddy co-variance CO 2 or CH 4 fluxes using this instrument at sea.To the best of our knowledge, measurements of air-water fluxes Picarro and the LGR analysers have never been compared side by side.
Previous tests over terrestrial regions of very high fluxes are briefly reviewed here but do not provide sufficient information with respect to the suitability (or preference) of the Picarro and LGR instruments for air-sea flux measurements.Peltola et al. (2014) compared the performance of eight different instruments for eddy covariance CH 4 flux measurements over grassland, including the Picarro G2311-f and the LGR FGGA.They relied on the numerical H 2 O corrections from the Picarro and the LGR instruments for the CH 4 flux calculations and did not look at the effect of using a Nafion dryer.The authors found that, in this environment with a typical CH 4 emission of the order of 2 mmol m −2 day −1 , cumulative measured fluxes from the G2311-f and the FGGA over 14 days agreed within 3 % (∼ 0.8 mmol m −2 ).Tuzson et al. (2010) found good agreement between an earliergeneration LGR instrument (with numerical H 2 O correction) and a quantum cascade laser-based absorption spectrometer (physically dried) when measuring artificially generated CH 4 fluxes (> 1 mmol m −2 day −1 ) by eddy covariance.In contrast to these areas of high fluxes, estimated CH 4 emissions over the open ocean are of the order of 0.01 mmol m −2 day −1 (e.g.Bange, 2006;Forster et al., 2009), placing greater demand on instrument accuracy and precision.
In this work, we compare the performance of the Picarro G2311-f and the LGR enhanced performance FGGA at measuring air-water fluxes of CO 2 and CH 4 by eddy covariance.Measurements were made from a coastal site, free from platform motion that might interfere with the flux signal (e.g. on ships or buoys, Prytherch et al., 2015).We compare the precision and flux detection limits of the two analysers, quantify the effects of H 2 O fluctuations on the measured CO 2 and CH 4 fluxes, and examine the necessity of drying sample air when using these cavity-based instruments downstream of a long inlet.The gas flow rate required by the Picarro is fairly low (owing to the small cavity size), enabling efficient physical removal of water vapour from the sampled air with a dryer.The LGR has a much larger cavity cell that requires a higher gas flow to minimise high-frequency flux attenuation.The high flow rate results in decreased dryer performance and the LGR was thus configured without a dryer.

Instrumental set-up
The Picarro G2311-f and the LGR FGGA were deployed side by side between 25 September and 2 October 2015 at the Penlee Point Atmospheric Observatory (PPAO; Fig. 1) on the south-west coast of the United Kingdom.Yang et al. (2016) provided detailed descriptions of this site and reported air-Atmos.Meas. Tech., 9, 5509-5522, 2016 www.atmos-meas-tech.net/9/5509/2016/sea CO 2 and CH 4 flux measurements from the open water wind sector (i.e.south-west) during spring/summer 2014-2015 using the Picarro G2311-f.During the week-long intercomparison described here, wind direction was predominantly from the north-east (Fig. A1) over the Plymouth Sound with a wind fetch over water of 5-6 km.Within this wind sector, measured momentum and sensible heat flux suggest that the flux footprint was also over the water only (Appendix, Figs.A2 and A3).
Both analysers sampled from the same ∼ 18 m long PFA inlet tube (1/4 inner diameter), which led from inside the observatory to a sonic anemometer (Gill Windmaster Pro) on a retractable mast above the rooftop (∼ 18 m, a.s.l., above mean sea level).The gas inlet tip was about 30 cm below the sonic anemometer centre volume.A small, stainless steel particle filter (2 µm pore size, Swagelok SS-6F-05) was installed inline to protect the gas sensors from sea salt.An external scroll pump (BOC Edwards XDS-35i) was adjusted to pull ∼ 20 SLPM of air through the inlet tubing (measured by a mass flow meter).About ∼ 15 SLPM went through the cavity of the LGR analyser, which has a volume of 408 mL.The LGR cavity pressure was held at ∼ 140 Torr (0.184 atmosphere), resulting in a volumetric flow through the cavity of ∼ 80 L min −1 .At this rate, the flushing time through the LGR cavity was 0.3 s.
The Picarro analyser subsampled from the main tubing ahead of the LGR at a flow rate of approximately 5 L min −1 (at atmospheric pressure).Maintained at a pressure of 153 Torr (0.20 atmosphere), the cavity volume of the Picarro (35 mL) is about 10 times smaller than that of the LGR.The flushing time in the cavity is thus less than 0.1 s and the manufacturer's stated response time is 0.2 s or less.Between 25 September and 1 October, a high throughput dryer (Nafion PD-200T-24M) was installed immediately upstream of the Picarro instrument (i.e.not affecting the LGR).The dryer was set up in the reflux configuration, utilising the reduced pressure of the Picarro exhaust air to dry the sample air.Using the LGR as reference, the dryer eliminated ∼ 80 % of H 2 O (and ∼ 95 % of the rapid fluctuations in H 2 O) in the sample stream of the Picarro instrument.We will refer to this sample period as Picarro (dry) vs. LGR (wet).Between 1 and 2 October, the Nafion dryer was removed in order to briefly examine the humidity dependence in the Picarro analyser.We will refer to this period as Picarro (wet) vs. LGR (wet).

Numerical corrections for water vapour
Both analysers report ambient mixing ratios of CO 2 and CH 4 (C CO 2 and C CH 4 ) in parts per million (ppm).They compute the "dry mixing ratios" (C CO 2 _d and C CH 4 _d , in ppm) based on the measured volume fraction of water vapour (C H 2 O , in %).Water vapour affects CO 2 and CH 4 measurements in cavity-based instruments in at least two ways: volumetric dilution and spectroscopic line broadening.Rella (2010) proposed the following corrections to account for both effects together: For the Picarro G-2311f analyser, Chen et al. (2010) found a = −0.01200,b = −2.674× 10 −4 , c = −0.00982,d = −2.393× 10 −4 .When air is dried with a Nafion dryer, the reliance of the Picarro on these numerical corrections is reduced by about an order of magnitude.Different coefficients have been estimated for the LGR FGGA (Hiller et al., 2012) 2) would imply that these mathematical corrections are inaccurate or insufficient at describing the cross-sensitivities between the trace gases and H 2 O.
The greatest relative biases are expected to occur when the magnitude of the H 2 O fluctuations in the measurement cavity is large and when the trace gas fluxes are small.The slope in mixing ratios (intercept, r 2 ) during Picarro (dry) vs. LGR (wet) was 1.014 (−15.5 ppm, 0.99) for C CO 2 _d and 0.976 (0.0113 ppm, 0.99) for C CH 4 _d (Picarro sampling behind a dryer).Uncertainties (relative standard deviations) of the slopes and intercepts in the fits above were ∼ 0.1 %.The calibrations of these two instruments were also crudely checked with a concentrated calibration gas mixture (8000 ppm CO 2 and 40 ppm CH 4 , BOC gas).C CO 2 _d and C CH 4 _d from both instruments were within the certified uncertainty of the gas standards (5 %) so no calibration curve was applied.For the Picarro, the use of the Nafion dryer reduced the observed CO 2 mixing ratio by ∼ 0.3 % and reduced the CH 4 mixing ratio by ∼ 0.06 % when measuring the concentrated calibration gas mixture.This suggests that small amounts of CO 2 and CH 4 permeate through the Nafion dryer, qualitatively consistent with results from Welp et al. (2013).

Flux processing
Fluxes of CO 2 , CH 4 , and H 2 O are computed using the Picarro and LGR data along with the streamline-corrected vertical wind velocity (w) from the Windmaster Pro sonic anemometer (see Yang et al., 2016 for further details about flux processing).Lag correlations between C CO 2 _d , C CH 4 _d from the Picarro and LGR with w generally showed a consistent maximum covariance lag time of about 3 s.The maximum covariance lag time between H 2 O and w (without a dryer) tended to be much longer and more variable (∼ 20 s) due to severe attenuation of the high-frequency water vapour signal in the inlet tube (see Sect. 3.3).
Covariance from both instruments are computed in the same, non-overlapping 10 min intervals between lag-shifted gas-mixing ratios and w (e.g.w C CO 2 _d , w C CH 4 _d ), which are multiplied by the ambient air density to yield mass concentration fluxes.Here the primes indicate fluctuations from the means while the overbar denotes temporal averaging.Because fluxes are computed from mixing ratios, an air density correction due to fluctuating temperature and pressure on the fluxes (e.g.Webb et al., 1980) should be unnecessary.The 10 min fluxes are further averaged to hourly intervals in order to reduce random noise.The relatively short averaging time of 10 min is chosen here to more easily satisfy the stationarity flux criteria in this dynamic environment.As demonstrated by Yang et al. (2016), low-frequency flux not captured by the 10 min averaging window is usually small at this site (within a few percent) and in any case the same for both analysers.
The LGR showed a positive offset in C CO 2 _d of ∼ 15 ppm relative to the Picarro when sampling ambient air.When measuring high-purity nitrogen, the LGR CO 2 signal was also around 15 ppm initially and decayed towards zero after more than an hour (the Picarro reported CO 2 , CH 4 , and H 2 O levels very close to zero immediately).Post-campaign tests suggest that the offset in the LGR was likely due to inaccuracy in the instrument calibration.Since the covariance flux is computed from deviations from linearly detrended mean in 10 min intervals, the positive bias in the LGR CO 2 mixing ratio should not significantly affect our flux comparison.

Results and discussion
Time series of CO 2 and CH 4 flux measurements during the Picarro (dry) vs. LGR (wet) period are shown in Fig. 2. For the LGR, both ambient (e.g. from C CO 2 ) and numerically dry (e.g. from C CO 2 _d ) fluxes are shown.Considering all wind directions (137 flux hours), there is good agreement between the two instruments for dry CO 2 flux (slope = 0.99) and dry CH 4 flux (slope = 1.00).The mean (±standard error) dry CO 2 and CH 4 fluxes from the Picarro were 27.9 ± 6.4 and 0.167 ± 0.021 mmol m −2 day −1 respectively.These values agree within uncertainties with dry CO 2 and CH 4 fluxes from the LGR, which were 28.3 ± 6.7 and 0.176 ± 0.022 mmol m −2 day −1 respectively.Likewise, during the Picarro (wet) vs. LGR (wet) period, there was no sta- tistically significant difference between the dry CO 2 and CH 4 fluxes from the Picarro and the LGR.
We examine the impact of the numerical H 2 O correction on the CO 2 and CH 4 fluxes within a given instrument.This "intra-instrument" difference in flux, computed from numerically dry and ambient mixing ratios (i.e.w C CH 4 _d − w C CH 4 ), is plotted against the apparent H 2 O flux (i.e. w C H 2 O at the optimal lag time of CH 4 / CO 2 ) for both the LGR and Picarro in Fig. 3.For the Picarro, these data are taken from the Picarro (wet) vs. LGR (wet) period.The slopes between the "intra-instrument" flux differences and H 2 O fluxes are steeper for the Picarro than for the LGR.This is consistent with the fact that the Picarro G2311-f applies both a dilution and a line-broadening correction (Eqs. 1 and 2), while the LGR FGGA applies only the dilution correction.At typical background levels of these greenhouse gases, the expected slope of [C CO 2 _d − C CO 2 ] vs. H 2 O mixing ratio is ∼ 0.0004 for the LGR and ∼ 0.0006 for Picarro; the analogous slope in the case of CH 4 is ∼ 1.9e −6 for the LGR and 2.5e −6 for the Picarro.We see that the slopes between fluxes in Fig. 3 are essentially equal to the slopes expected between the gas-mixing ratios as a result of the numerical H 2 O corrections.This is because the H 2 O corrections are largely linear (i.e.coefficients b and d are close to zero).

Comparisons of air-water fluxes of CO 2 and CH 4
As indicated in Fig. 2 by the gray shading, air-water CO 2 and CH 4 fluxes tend to be much lower than terrestrial fluxes.To further evaluate the performance of these analysers at airsea flux measurements, we limit our comparison to the sector when the wind was coming from over water only.Within the north-east quadrant, we find that the fluxes of momentum and sensible heat are reasonably consistent with air-water transfer when the wind directions were between 45 and 80 • (Appendix, Figs.A2 and A3).We compare fluxes of CO 2 and CH 4 from the Picarro and LGR instruments within this wind sector in Fig. 4 (73 flux hours).Here CO 2 and CH 4 fluxes have further been filtered for stationarity following Yang et al. (2016).This removed occasional periods when the atmospheric CO 2 and CH 4 mixing ratios were highly variable (e.g.due to large horizontal transport or emissions from passing ships in the Plymouth Sound).
For both gases, fluxes from the two instruments scatter around the 1 : 1 line and differences in the mean fluxes between the two instruments are within the measurement uncertainties.During the Picarro (dry) vs. LGR (wet) period, the mean (±standard error) dry CO 2 flux was 10.7 ± 1.8 mmol m −2 day −1 from the Picarro and 12.5 ± 2.9 mmol m −2 day −1 from the LGR.The dry CH 4 flux was 0.110 ± 0.009 mmol m −2 day −1 from the Picarro and 0.123 ± 0.011 mmol m −2 day −1 from the LGR.The root mean square (rms) error between the physically dry Picarro  LGR (numerically dry and ambient) vs. Picarro (physically dry) CO 2 (left), and CH 4 (right) fluxes for the air-water wind sector of 45 to 80 • .For CO 2 , rms error between the two instruments is 16.1 mmol m −2 day −1 using numerically dry LGR data and 16.3 mmol m −2 day −1 using ambient LGR data.For CH 4 , rms error between the two instruments is 0.063 mmol m −2 day −1 using numerically dry LGR data and 0.064 mmol m −2 day −1 using ambient LGR data.
flux and the numerically dry LGR flux (hourly average) is 16.1 mmol m −2 day −1 for CO 2 and 0.063 mmol m −2 day −1 for CH 4 .The rms errors increase by only ∼ 1 % when comparing the ambient LGR fluxes to the dry Picarro fluxes.Since the gas fluxes were computed using the same wind data and the two analysers were sampling the same gas stream, hourly differences in the fluxes are primarily due to noise in the two instruments, rather than by the presence of water vapour.
The effects of H 2 O on the LGR fluxes were detectable but relatively small compared to the magnitude of the fluxes.This is illustrated in Fig. 5, where the differences in CO 2 and CH 4 fluxes between the LGR (numerically dry as well as ambient) and Picarro (physically dry) are plotted against the bulk air-water latent heat flux predicted from the Coupled Ocean-Atmosphere Response Experiment (COARE) model (Fairall et al., 2003).Discrepancies between the two sets of the LGR trace gas fluxes are most apparent (e.g. up to 0.04 mmol m −2 day −1 in either direction for CH 4 ) under conditions of high latent heat flux (i.e.highest fluctuations in H 2 O mixing ratio).Qualitatively similar trends are seen when the differences in trace gas fluxes are plotted against the measured H 2 O flux by the LGR (see Supplement).These results are consistent with the findings from Tuzson et al. (2010), who showed that with an earliergeneration LGR instrument, not accounting for water vapour (either physically through drying or numerically) results in a bias in the eddy covariance CH 4 flux.In our set-up, additional application of a spectral line-broadening correction (using b and d from Hiller et al., 2012) has a negligible impact on the LGR CO 2 and CH 4 fluxes.
Figure 6 shows the mean cospectra of CO 2 , CH 4 (Picarro physically dry, LGR ambient, LGR numerically dry), as well as momentum when winds were from the air-water wind sec- tor.Each cospectrum is normalised by the respective mean flux.The gas cospectra from the two instruments have similar spectral shape to the momentum cospectrum.Numerical drying of the LGR only has a visible impact on the cospectra at low frequencies, which is likely due to the severe dampening of the H 2 O flux in the main inlet tubing (Sect.3.3).The LGR shows much more noise at frequencies above ∼ 0.1 Hz, which is partly caused by the less than optimal ring-down time of the LGR analyser during our testing period (∼ 8 µs instead of nominal value of ∼ 12 µs).Dust and/or sea salt might have entered the LGR cavity during instrument installation, thereby reducing the ring-down time and instrument sensitivity.

Instrument noise and eddy covariance flux detection limit
Scatter (i.e.random uncertainty) in the hourly LGR flux was about 50 and 20 % higher than in the Picarro for CO 2 and CH 4 respectively.Variance spectra of C CO 2 _d and C CH 4 _d from the Picarro and LGR show much higher noise in the latter instrument (Fig. 7).These were averaged from the Picarro (dry) vs. LGR (wet) period for the air-water wind sector.Excluding the small spikes near 3 Hz (likely instrument artifacts), the mean CO 2 variance above 1 Hz (i.e.bandlimited noise) was 0.0023 ppm 2 Hz −1 for the Picarro and 0.3 ppm 2 Hz −1 for the LGR.In the case of CH 4 , mean variance above 1 Hz was about 0.23 ppb 2 Hz −1 for the Picarro and 5 ppb 2 Hz −1 for the LGR.Interestingly, at low frequencies the LGR variance was also greater than the Picarro (by ∼ 2 times for CO 2 and ∼ 60 % for CH 4 ).This may be partly because the LGR does not regulate the cavity temperature and pressure (1σ of about 0.005 • C and 0.16 Torr within an hour) as precisely as the Picarro (1σ of about 0.0007 • C and 0.06 Torr).Low-frequency fluctuations in H 2 O, if not fully accounted for by Eqs. ( 1) and ( 2), could also cause some apparent variance in CO 2 and CH 4 dry mixing ratios.Uncertainty, as well as detection limit, in eddy covariance gas flux depends upon variance in both the vertical wind ve-  LGR (wet) period.The LGR shows significantly greater variance, especially at high frequencies.
locity as well as in the gas-mixing ratio.The latter is the sum of the ambient variance of the gas (estimated as the second point of the autocovariance of the mixing ratio; see Blomquist et al., 2010) and instrumental noise (estimated as the difference between the first and second points of the autocovariance).Both the Picarro and the LGR instruments are able to resolve rapid ambient fluctuations in atmospheric CO 2 and CH 4 .For example, ambient variance in C CH 4 _d (i.e.without instrument noise) from the two instruments agree almost exactly (slope = 1.00; r 2 = 0.96).Based on earlier measurements at PPAO, Yang et al. (2016) estimated a CH 4 flux detection limit of 0.02 mmol m −2 day −1 (hourly average) for the Picarro G2311-f at a wind speed of 10 m s −1 .This was quantified using two different methods that yielded largely comparable results: a theoretically based method on instrument noise and ambient variance (Blomquist et al., 2014) and an empirically based method on scatter in w C CH 4 at an implausible lag time (e.g.300 s).Applying the same methods to CO 2 here results in a flux detection limit between 2.2 mmol m −2 day −1 (theoretical) and 4 mmol m −2 day −1 (empirical) for the G2311-f.At a wind speed of 10 m s −1 and in a neutral atmosphere, Blomquist et al. (2014) computed a CO 2 flux detection limit of 1.3 mmol m −2 day −1 (hourly average) for a prototype Picarro analyser G1301-f.The instrument used by Blomquist et al. (2014) only measures CO 2 and thus has a lower noise level of 700 ppb 2 Hz −1 (see their Fig. 6) than the Picarro we deployed at PPAO.The high-frequency noise of dry CO 2 and CH 4 in the Picarro are comparable with and without the Nafion dryer.Thus, we do not expect the dryer to cause a significant difference in the Picarro detection limits.
Following Blomquist et al. (2014), we theoretically estimate the flux detection limits for the LGR FGGA using the band-limited noise shown in Fig. 7 and ambient variance from periods of very low variability (100 and 0.04 ppb 2 for CO 2 and CH 4 respectively).At a wind speed of 10 m s −1 and in a neutral atmosphere, we estimate a flux detection limit of 13.5 mmol m −2 day −1 for CO 2 and 0.060 mmol m −2 day −1 for CH 4 (hourly average).We also examined an earlier measurement period when the LGR had the optimal sensitivity (ring-down time of 12 µs), with bandlimited noise of 0.11 ppm 2 Hz −1 for CO 2 and 3.7 ppb 2 Hz −1 for CH 4 .Coupling these noise levels to the ambient variability above, at the same environmental conditions, we estimate a best case hourly flux detection limit for the LGR of 8.2 mmol m −2 day −1 for CO 2 and 0.053 mmol m −2 day −1 for CH 4 (see Table 1 for summary).In the case of CO 2 , the LGR flux detections limits are of the same magnitude as those from the LI-COR LI7200 with a dryer (as estimated by Blomquist et al., 2014).
It is useful to put these flux detection limits into the context of regions of the ocean and times of year that are favourable for future process-level studies with direct flux measurements (e.g. research cruises that aim to improve our understanding of the gas transfer velocity).For illustrative purposes, we compare the CO 2 flux detection limits with the estimated global air-sea CO 2 flux maps from the National Oceanographic and Atmospheric Administration's Pacific Marine Environmental Laboratory (http://www.pmel.noaa.gov/co2/story/Surface+_CO2+_Flux+_maps;Feely et al., 2006;Sabine et al., 2008Sabine et al., , 2009;;Park et al., 2010).Four monthly maps from 2015 to 2016 (4 • resolution) are chosen to capture seasonal variations: March, June, September, and December.Globally, the absolute value of the estimated monthly mean CO 2 flux exceeds 2 mmol m −2 day −1 (hourly flux detection limit of the Picarro) in roughly 39-44 % of the oceans (highest in March) and exceeds 8 mmol m −2 day −1 (detection limit of the LGR) in 2-6 % of the oceans (highest in December).The North Atlantic and North Pacific, southeastern Pacific, and the Southern Ocean are not surprisingly amongst regions favourable for eddy covariance CO 2 flux measurements.When estimated fluxes are comparable or below the hourly flux detection limit, further averaging of the eddy covariance measurements (temporally or in bins of wind speed, etc.) will likely be necessary to extract statistically meaningful results (as random uncertainty in flux decreases with increasing number of independent hourly mea-Table 1. Performance of the Picarro G2311-f and LGR FGGA at measuring CO 2 and CH 4 fluxes by eddy covariance.Precision refers to observed noise at 10 Hz.Band-limited noise is averaged between 1 and 5 Hz from the mean variance spectra.Flux detection limit is estimated at a wind speed of 10 m s −1 in a neutral atmosphere (hourly average).The FGGA band-limited noise and detection limit correspond to when the ring-down time was optimal and the instrument sensitivity at its highest.

CO 2
Precision (10 Hz) Band-limited noise Flux detection limit Precision (10 Hz) Band-limited noise Flux detection limit surements, N, by the relationship of N −0.5 ).For example, if the hourly fluxes are binned to 6 h intervals the percentage of exceedance as described above may improve to 65 and 24 % for the Picarro and LGR respectively (averaged across all seasons).The predicted open ocean emission of CH 4 is of the order of 0.01 mmol m −2 day −1 (Forster et al., 2009).Thus, further binning of hourly eddy covariance measurements (e.g. to 6 h or daily intervals) would likely be needed to better resolve CH 4 fluxes.The signal-to-noise ratio in the eddy covariance CH 4 flux measurement should be improved near the coast (e.g.Bange, 2006) and in the Arctic (e.g.Shakhova et al., 2010), where surface saturations and hence emissions of CH 4 are likely greater.Combining data from a floating chamber with a turbulent diffusivity model, Kitidis et al. (2007) estimated a CH 4 emission of 0.06 to 0.17 mmol m −2 day −1 in a large coastal embayment.CH 4 flux of a similar magnitude has been measured here at PPAO (this paper and Yang et al., 2016).Aircraft observations suggest CH 4 emissions over 0.1 mmol m −2 day −1 from the partially ice-covered Arctic (Kort et al., 2012), which are above the flux detection limits of both analysers.
We did not evaluate the effect of motion on these analysers.During the High Wind Gas Exchange Study (HiWinGS) cruise in 2013 (Yang et al., 2014b), a Picarro G1301-f was deployed with a Nafion dryer along with a second generation model from Picarro (G2301-f) which was not sampling after a dryer.In moderate seas, the two instruments yielded similar CO 2 fluxes, implying that the numerical H 2 O correction in G2301-f is reasonable.However, the different sensitivities of these two models to the ship's motion in high seas complicated these flux comparisons (B.Blomquist, personal communication, 2016).Side-by-side shipboard deployments of the Picarro and LGR are required to provide a more conclusive verdict on the optimal flux analyser for air-sea flux measurements.

High-frequency flux loss
Based on measurements of dimethylsulfide (DMS) flux, Blomquist et al. (2010) and Yang et al. (2011) reported highfrequency flux attenuations to be of the order of 5 % using a long inlet and the same type of Nafion dryer as used here for the Picarro.CO 2 and CH 4 are less polar (i.e. less "sticky") gases than DMS so we expect their flux losses in the dryer to be comparable or less.CO 2 and CH 4 flux attenuation by the tubing itself should be a few percent at most given the fully turbulent flow used in our study (Lenschow and Raupach, 1991;Ibrom et al., 2007).Below we quantify the highfrequency attenuations in the CO 2 and CH 4 fluxes from the Picarro (with and without a dryer) and the LGR (without a dryer).
The effect of the Nafion dryer on the Picarro instrument response is clearly illustrated in Fig. 8.Here we made rapid, stepwise reductions in the flow rates of the concentrated CO 2 and CH 4 gas standard (here controlled by a digital mass flow controller, EL-FLOW, Bronkhorst), which was teed into the gas inlet with and without the dryer.For ease of comparison, we normalise the dry CO 2 and CH 4 such that the signals decrease from one to zero.Without the dryer, the observed response time (defined as the time needed for the signal to fall to 1/e of the initial value) was about 0.2 s.This is 0.1 s longer than the expected instrumental response time, likely in part due to the finite time response of the mass flow controller (settling time of 0.1-0.2s).With the dryer, the signal drop-off was noticeably slower and the observed response time increased to about 0.5 s.The response time for CO 2 appears to be very slightly longer than for CH 4 in the presence of the dryer, as might be expected from the greater permeation (i.e.breakthrough) of CO 2 through the Nafion membranes compared to CH 4 .Knowing the instrument response time allows us to estimate the high-frequency flux loss using a filter function (e.g.Bariteau et al., 2010;Yang et al., 2014a;Blomquist et al., 2014).At a response time of 0.3 s with the dryer (accounting for the settling time of the mass flow con- troller), the predicted flux loss in the Picarro averaged over the Picarro (dry) vs. LGR (wet) period (mean wind speed of 10 m s −1 ) is 10 %.Without the dryer the predicted flux loss is reduced to only ∼ 4 % at this wind speed.
Compared to the momentum cospectrum, the Picarro CO 2 and CH 4 cospectra are noticeably attenuated above the frequency of about 1 Hz (Fig. 6).High-frequency loss in the gas fluxes can also be estimated by similarity scaling with respect to another variable (often sensible heat).Air-water sensible heat flux varied in sign between day and night during this period, resulting in a very noisy mean cospectrum.Thus, we use momentum cospectrum for comparison instead.We convert the cospectra of gases and momentum to ogive (Oncley, 1989), or the cumulative sum of the cospectrum from low to high frequency.Following the method outlined by Spirig et al. (2005), from the ogive, we estimate that the mean loss of flux is about 6 % for CO 2 and 4 % for CH 4 , somewhat less than the filter function estimates above.
For the LGR, the large volume of the cavity cell was the principal limitation to the instrument response time, so it was likely the main cause for flux attenuation.Judging from the cospectra, the CO 2 and CH 4 flux attenuation in the LGR is comparable to the Picarro with a Nafion dryer.We note that the flow rate through the LGR can be increased substantially (by approximately a factor of seven) relative to the set-up here by adjusting the external scroll pump and using an inlet tube with a larger inner diameter.A faster airflow through the LGR cavity (e.g.flushing time of 0.1 s) would reduce the estimated flux loss in the LGR analyser to only a few percent; it likely would not significantly improve the instrument noise or flux detection limit, however.
Unlike CO 2 and CH 4 , fluctuations in water vapour are severely attenuated in our main inlet tube (similar to Ibrom et al., 2007).This is because the polar H 2 O is a much "stickier" gas than CO 2 and CH 4 and tends to absorb onto the wall of the tubing.Figure 9 shows the mean cospectrum of LGR H 2 O, which was computed at the maximum covariance lag time of CO 2 / CH 4 (3 s) as well as at the optimal lag time for H 2 O (∼ 20 s).H 2 O cospectrum computed at a lag time of 3 s is positive at the lowest frequencies and negative between about 0.01 and 0.1 Hz (resulting in mean apparent H 2 O flux near zero).In reality, the calculated bulk latent heat flux was always positive during this period, averaging 120 W m −2 (range of 60-230 W m −2 ).
Computed at the optimal lag time of ∼ 20 s, H 2 O cospectrum from the LGR is positive at frequencies below ∼ 0.1 Hz.Above 0.1 Hz the measured H 2 O flux is essentially zero (i.e.flux completely attenuated).Comparison with the bulk latent heat flux suggests that only up to 20 % of the H 2 O flux is still detectable after going through ∼ 18 m of inlet tubing (see Supplement).H 2 O flux from the Picarro without the Nafion dryer yielded similar results, confirming the importance of the sample tubing at attenuating the H 2 O flux.In our set-up, H 2 O flux attenuation might be particularly severe because the tubing we used had been constantly exposed to marine air for 1.5 years, resulting in accumulation of hygroscopic sea salt particles on the tubing wall.In a case of a shorter, cleaner inlet tubing, dampening of H 2 O fluctuations in the tubing should be somewhat reduced and the impact of H 2 O on CO 2 and CH 4 fluxes may be greater.Forgoing a physical drier would increase the reliance on the numerical H 2 O correction for accurate determination of CO 2 and CH 4 fluxes.

Conclusions
In this paper we compared the performance of the Picarro G2311-f (with and without a Nafion dryer) and the Los Gatos Research FGGA (without a dryer) at eddy covariance measurements of air-water CO 2 and CH 4 fluxes.Within measurement uncertainties, dry CO 2 and CH 4 fluxes from these two analysers agree.For winds over the Plymouth Sound, air-water CO 2 and CH 4 fluxes averaged about 12 and 0.12 mmol m −2 day −1 respectively.In our closed-path set-up with a long inlet tube, the numerical corrections of CO 2 and CH 4 mixing ratios based on concurrently measured H 2 O appeared to be reasonably effective for both instruments.Numerically dry CO 2 and CH 4 fluxes from the LGR agree within uncertainties in the mean with results from the Picarro after physical removal of H 2 O using a Nafion dryer.Coupling the LGR to a dryer, which we did not test, may have slightly improved the flux precision.The addition of a dryer would reduce the reliance on the instrumental numerical correction for H 2 O, but could increase flux attenuation and potentially cause a small bias in the mean CO 2 and CH 4 mixing ratios.Flux attenuation in the Picarro (likely occurring primarily within the dryer) was within 10 % for both CO 2 and CH 4 .Similar flux losses were observed for the LGR without a dryer in our set-up (likely occurring primarily within the large cavity), which could be reduced by increasing the airflow through the instrument.In contrast to CO 2 and CH 4 , H 2 O fluctuations were severely dampened in our inlet tubing, resulting in ≥ 80 % loss of the latent heat flux.Compared to the LGR, the Picarro demonstrated significantly lower noise levels, so flux detection limits were also lower (Table 1).These estimates help to advise us on the choices of location and timing for future process-level studies on air-sea CO 2 and CH 4 transfer.

Data availability
Processed hourly eddy covariance fluxes can be found in the Supplement of this paper.Raw data files at 10 Hz are very large (tens of gigabytes) and currently not archived in an online database.Please contact us directly if you are interested in the raw data and we would be very happy to share them.

Appendix A: Selection of wind sector for air-water transfer
The PPAO site is exposed to marine air over a wide wind sector from about 30 to 250 • (Fig. 1).Wind was mostly from the north-east during this period of instrument intercomparison (Fig. A1).Within this quadrant, we find that the fluxes of momentum (Fig. A2) as well as sensible heat (Fig. A3) are reasonably consistent with air-water transfer when the wind directions were from 45-80 • .As shown by the inset in Fig. 1, the distance between PPAO and the water's edge is about 20-30 m towards the north-east.North of a wind direction of 45 • , the flux footprint begins to overlap with land (Mount Edgcumbe Country Park).Immediately south of 80 • , the longer foreshore in front of PPAO likely affects the flux footprint and increases the effect of airflow distortion on the measured wind speed.The 10 m neutral drag coefficient is computed as C D10N = u * /U 10N .Here u * is the friction velocity measured by eddy covariance and the 10 m neutral wind speed U 10N is determined using Businger-Dyer relationships (Businger, 1988) from the wind speed and air temperature at PPAO, tidal-dependent sampling height, and sea-surface temperature (SST) from the L4 mooring station (∼ 6 km south of PPAO).Within 45-80 • , C D10N increases with wind speed, as expected for air-water transfer.Values of C D10N are, however, about ∼ 20 % lower than parameterisations for open water in the mean (e.g.COARE model version 3.5, Edson et al., 2013;Smith, 1980), possibly due to flow distortion of the mean wind speed or a residual bias in the Windmaster Pro wind measurement (see Yang et al., 2016 for more details).
Sensible heat flux was computed from the sonic temperature and further corrected for the bulk latent heat contribution (Schotanus et al., 1983).It shows a surprisingly good agreement (slope = 0.98 and r 2 = 0.87) with sensible heat flux estimated with the COARE model (Fairall et al., 2003) using air temperature and wind speed at PPAO as well as SST from L4.This close agreement despite the relatively long distance between the flux footprint and the mooring station was possibly due to fair weather in September 2015 and minimal riverine influence in the flux footprint (i.e.negligible spatial gradient in SST).Sensible heat flux was on average positive at night and negative during the day, primarily due to diel variability in air temperature.In spite of this variation, the atmosphere was near neutral during this measurement period as a result of the relatively strong winds (mean of 10 m s −1 ).The predicted atmospheric stability parameter z/L varied from about −0.20 at night to 0.05 during the day.Thus, we do not expect the flux footprint to extend beyond the 5-6 km fetch of water to the opposite side of the Plymouth Sound (as could happen during periods of extreme stability).Note that a flux footprint extending over land would cause momentum and sensible heat fluxes to grossly deviate from the predicted bulk air-sea fluxes, which was not observed.The Supplement related to this article is available online at doi:10.5194/amt-9-5509-2016-supplement.

Figure 1 .
Figure 1.Location of Penlee Point Atmospheric Observatory (orange 4-point star).The yellow lines mark the air-water wind sector of 45 to 80 • .The inset shows the area around the observatory (small hut in the left-centre of the image).

Figure 2 .
Figure 2. Time series of CO 2 (top) and CH 4 (bottom) fluxes.Picarro sampled after a dryer, while LGR sampled ambient (moist) air."LGR ambient" indicates flux prior to the numerical H 2 O correction.The shaded region indicates the wind sector between 45 and 80 • .

Figure 3 .
Figure 3. "Intra-instrument" difference in CO 2 (left) / CH 4 (right) fluxes computed from numerically dry and ambient CO 2 / CH 4 mixing ratios from the same instrument vs. apparent H 2 O flux (latter computed at the lag time of CO 2 / CH 4 ).Both the LGR (red) and Picarro (blue) were directly sampling ambient air.Slopes (b) and r 2 values from the linear regressions are also shown.

Figure 4 .
Figure 4.LGR (numerically dry and ambient) vs. Picarro (physically dry) CO 2 (left), and CH 4 (right) fluxes for the air-water wind sector of 45 to 80 • .For CO 2 , rms error between the two instruments is 16.1 mmol m −2 day −1 using numerically dry LGR data and 16.3 mmol m −2 day −1 using ambient LGR data.For CH 4 , rms error between the two instruments is 0.063 mmol m −2 day −1 using numerically dry LGR data and 0.064 mmol m −2 day −1 using ambient LGR data.

Figure 5 .
Figure 5. Differences in CO 2 (top), and CH 4 (bottom) fluxes between LGR (numerically dry as well as ambient) and Picarro (physically dry) vs. the bulk latent heat flux for the air-water wind sector of 45 to 80 • .The numerical H 2 O correction has the largest effect on the trace gas fluxes under conditions of high latent heat flux.

Figure 6 .
Figure 6.Mean cospectra of CO 2 (top), CH 4 (bottom) with momentum for the air-water wind sector of 45 to 80 • during the Picarro (dry) vs. LGR (wet) period, normalised by the respective mean flux.Gas cospectra from the two instruments are similar in the mean and show reasonable agreement with the momentum cospectrum.Spectral similarity suggests high-frequency flux losses of ∼ 6 % for CO 2 and ∼ 4 % for CH 4 during this period.

Figure 7 .
Figure 7. Mean variance spectra of dry CO 2 (top) and CH 4 (bottom) from the air-water wind sector during the Picarro (dry) vs.LGR (wet) period.The LGR shows significantly greater variance, especially at high frequencies.

Figure 8 .
Figure 8. Fall offs in the Picarro dry CO 2 and CH 4 signals shortly after reductions in the calibration gas flow rate.The use of a Nafion dryer increases the observed response time by about 0.3 s.

Figure 9 .
Figure 9. Mean cospectra of LGR H 2 O flux for the air-water wind sector, computed at the optimal lag time of CO 2 / CH 4 (3 s) and the optimal lag time of H 2 O (∼ 20 s).The much longer lag time of H 2 O is due to severe high-frequency attenuation of H 2 O in the inlet tube (e.g.no flux signal above ∼ 0.1 Hz).Even at the optimal lag time, measured H 2 O flux is only about 10 % of the predicted latent heat flux (mean of 120 W m −2 ).

Figure A1 .
Figure A1.Time series of wind speed and direction during the Picarro (dry) vs. LGR (wet) period.The shaded region indicates the wind sector between 45 and 80 • , which we consider to be representative of air-water transfer.

Figure A2 .
Figure A2. 10 m neutral drag coefficient vs. 10 m neutral wind speed for the air-water wind sector.

Figure A3 .
Figure A3.Measured vs. predicted sensible heat flux for the airwater wind sector.