Due to their major role in global change and the climate system, time series of atmospheric greenhouse gases (GHGs) are a cornerstone of earth system science. Since the atmosphere is generally well mixed, the gradients and spatial variability of GHGs and other trace species can be quite small, necessitating high-quality, accurate, and precise measurements.

The Southern Hemisphere is underrepresented relative to the Northern Hemisphere in the various sampling programs and networks of ground-based stations that make continuous measurements of GHGs . Coverage is particularly poor in Africa, which complicates efforts to monitor the variations and emissions of GHGs from developing economies, biomass burning, and the terrestrial biosphere .

In this paper, we discuss the configuration and performance of a new station constructed in 2012 at a near-coastal background site at Gobabeb (GAW station ID “NMB”), Namibia, here referred to as the Namib Desert Atmospheric Observatory (NDAO). Surface flask samples have been taken near Gobabeb since 1997 as part of the US National Oceanic and Atmospheric Administration Earth System Research Laboratory Global Monitoring Division (NOAA ESRL GMD) Carbon Cycle Cooperative Global Air Sampling Network. The NMB site is ∼ 3 km from NDAO. The goals of this project are to expand the ground-based station network in Africa and to use the time series as a top-down perspective on regional biogeochemical cycling and surface–atmosphere exchange of GHGs. The particular focus of the project at a regional level is on the influence of biomass burning and coastal upwelling on GHG budgets. As the site receives air largely free from anthropogenic influences, it is representative of both the terrestrial and the marine background (Fig. ). The main quantities measured at the observatory are carbon dioxide, methane, nitrous oxide, carbon monoxide, and atmospheric oxygen.

Absorption spectrometers featuring high-finesse optical cavities are proving to be a highly sensitive and user-friendly approach to measuring GHGs and are being widely adopted by the GHG measurement community . Commercial analyzers using the cavity ring-down spectroscopy (CRDS) and off-axis integrated cavity output spectroscopy techniques can now be purchased to measure all of the major non-synthetic greenhouse gases. While there are a fair number of reports on using such analyzers to measure CO2, CH4, and the tracer CO , there are few assessments of such analyzers for N2O, particularly from the field. Given that the natural variability of this gas in the atmosphere is quite small – consider that the seasonal cycle of N2O is typically less than 1 ppb in the lower troposphere – the potential for off-axis integrated cavity output spectrometer (OA-ICOS) instruments to greatly increase our understanding of this powerful GHG is quite high.

Atmospheric oxygen can be used as a top-down constraint on the carbon cycle, since the major biogeochemical processes that consume or produce carbon dioxide on the global scale also consume or produce oxygen, such as respiration, photosynthesis, or combustion . Since both CO2 and O2 are conservative tracers, the composition of air masses affected by these processes conform to the stoichiometry of these reactions. The stoichiometry of marine respiration and photosynthesis is not preserved after surface–atmosphere exchange, however, as the ratio of O2 to CO2 is altered by the slower equilibration of CO2 with the atmosphere due to the carbonate equilibrium system. Furthermore, as the world's oceans are not a long-term sink or source of O2, but are a sink for CO2, measurements of atmospheric oxygen can be used to constrain the marine uptake of CO2 . The success of efforts to use atmospheric oxygen to quantify the global carbon cycle is dependent primarily on the availability of data, since coverage is lower for O2 than for CO2 . As oxygen is a major constituent of the atmosphere, measuring small changes against the large background is technically challenging but feasible with several methods, including mass spectrometry, paramagnetic analyzers, gas chromatography, vacuum ultraviolet absorption, and fuel cell techniques .

The Namib Desert is one of the driest areas on Earth, with annual precipitation usually less than 25 mm yr-1, and consecutive years without rain are not unusual . A distinctive feature of the area is the common occurrence of fog (60–200 days yr-1), a result of the nearby cold waters of the Benguela current; fog is a larger source of moisture than rainfall to the region . The Köppen–Geiger climate classification is BWh (sometimes BWn, “mild desert”) or BWk (“cold desert”) . The Namib Desert extends ∼ 2000 km along the coast and some 150 km inland where it terminates at the Great Escarpment. The biomes bordering the Namib are classified as xeric shrublands, drylands, or savannas .

NDAO is located at Gobabeb (23.563∘ S; 15.046∘ E), in the Erongo region, ∼ 600 m from Gobabeb Research and Training Centre (GRTC), a desert research station. The station is situated at an elevation of 408 m above sea level, ∼ 50 km from the coast, along the Kuiseb River, an ephemeral river which demarcates the northern margin of the Namib Sand Sea from a gravel plain region to the north. After years with heavier rains, sparse vegetation appears in the Sand Sea. Likewise periodic greening of the gravel plains region is not uncommon. There are some trees and vegetation along the riverbed, though much of it is dead or dormant; otherwise the region is mostly barren. All of the facilities at Gobabeb are solar powered, although there is a backup generator which runs occasionally when there is fog or the load on the system is too high. Day-to-day maintenance is performed by GRTC personnel.

The top of an aluminum 21 m mast serves as the basis for all atmospheric observations at the observatory. Each of the three main instruments and the flask sampler has its own dedicated sample line. The intakes of the sample lines are each protected from suspended material and precipitation with a custom inlet. The inlet is mushroom shaped and allows for the flow-through of air, such that sample air is drawn from underneath the cap of the inlet, through a 5 µm polyester filter, and then a portion is siphoned into the main sampling line. From the top of the mast to the outer wall of the container the air sampling lines are 12 mm SERTOflex tubing.

Large volume (5 L min-1) pumps run continuously to pull sample air down from the tower. Inside the container, sample lines (stainless steel tubing with 1/8" outer diameter) pull a smaller volume of air with a lower flow rate from the larger diameter tubing running up the tower via a T-junction. Buffer volumes of 7 L are located upstream of the large pumps to minimize pressure fluctuations. Three lines of sample air pass through cool (4 ∘C) water traps filled with glass beads and cryotraps (-80 ∘C) also containing glass beads. All water traps are made from electropolished stainless steel. Sample air then is routed to the main instruments: an Oxzilla FC-II differential fuel cell analyzer (DFCA), measuring atmospheric oxygen (Sable Systems International, Las Vegas, NV, USA), a Picarro ESP-1000 cavity ring-down spectrometer measuring CH4 and CO2, (Picarro Inc, Santa Clara, CA, USA), and a Los Gatos N2O/CO-23d cavity-enhanced absorption spectrometer measuring N2O and CO (Los Gatos Research Inc, Mountain View, CA, USA). Downstream of the T-junctions, the flow rates are slightly different between instruments, reflecting their different requirements for gas supply: flows are between 90 and 95 mL min-1 for the DFCA, between 150 and 210 mL min-1 for the CRDS, and between 110 and 220 mL min-1 for the OA-ICOS. The larger range for the latter two instruments is given because flow and sample line pressures gradually decline in between filter replacements due to the very fine dust present at the site.

A complete diagram of the instrument system can be seen in Fig. , and a part list is presented in Table . Four-port two-position valves, determined to have no cross-port leakage, are used to switch between reference gas streams and the air sample stream. Small membrane pumps are used to maintain sample flow during measurements of reference gases, so that sample air does not sit stagnate in the lines and so that no additional purge is necessary upon the completion of a reference cycle. The entire system is under automatic control, using a custom program written in LabVIEW. The LabVIEW system also serves as the data recorder for the meteorological instrumentation and the diagnostic sensors inside the laboratory. Data are logged at an interval of 1 s; data generated at a higher frequency than this are averaged by the software. In addition to the control and data acquisition computer running the LabVIEW software, a second computer is used for off-site data transfer and remote access to the laboratory network.

As both the CRDS and OA-ICOS instruments are upstream of their pumps, these sample lines are under-pressurized (100–200 mbar under ambient), and the instruments actively regulate the pressure inside the measurement cavity. The CRDS has been fitted with an external diaphragm pump for better ease of maintenance. Since the DFCA does not have its own pump, air is forced through the analyzer via a small membrane pump located upstream of the fridge traps and cryotraps (Fig. ); consequently in this region the sample lines are over-pressurized. The residence time of sample air in the entire system was determined to be between 15 and 200 s, depending on which analyzer air was routed to.

The Picarro ESP-1000 (Serial No. CFADS-06) instrument is an early-model CRDS. This optical technique is a type of laser absorption spectroscopy, whereby a continuous wave laser emits a pulse of light into an optical cavity. The laser is tuned to several specific wavelengths where the measurand has strong absorption and to regions without absorption in order to determine the ring-down time with and without absorption by the measurand. The difference in decay time of the light intensity (as measured by a photo-detector) between the two modes is the signal of the analyzer, which is converted to a mole fraction . Since the CRDS software calculates the mole fractions of CO2, CH4, and H2O, this output can be used directly, after the application of an instrument-specific water correction and a calibration (see Sects.  and , respectively).

As the CRDS is tuned to specific absorption features, it measures only the main isotopologues of each species, namely 12C16O2, 12C1H4, and 1H216O . Therefore, the isotopic composition of calibration standards and other reference gases must be close to that of the sample to avoid artifacts . Likewise, standard gases should have the same buffer matrix as the sample, meaning that synthetic air standards should not be used . As the shapes of these spectral lines are quite sensitive to temperature and pressure, the levels of both must be carefully maintained by the instrument. The cavity is held at a constant pressure of 186.7 ± 0.1 mbar and a constant temperature of 45 ± 0.01 ∘C.

The Los Gatos N2O/CO-23d Analyzer employs cavity enhanced absorption off-axis integrated cavity output spectroscopy (OA-ICOS) . In this technique, light from a laser is emitted into a cavity and the temporally integrated intensity of the transmission of the light out of the cavity is the raw measurement signal . Like the CRDS, the OA-ICOS software calculates the mole fractions of N2O, CO, and H2O; this output is used directly after the application of an instrument-specific water correction, calibrations, and a drift correction (see Sects. , , and , respectively).

As the instrument is fairly new, few published studies on the analyzer are available, particularly those that evaluate both N2O and CO. Initial assessments indicate that the instrument has the potential for use in making atmospheric measurements, eddy covariance methods, and aqueous dissolved gas concentrations measured in equilibrated headspace, provided the drift of the analyzer is accounted for and care is taken with respect to gas handling, since the mole fraction ranges are usually small .

The DFCA measures the O2 mole fraction directly in dry air through an electrochemical method . There are two sensors in the device, each containing an anode, cathode, and weak acid electrolyte. Atmospheric oxygen is reduced on the cathode, generating a current which is the analytical signal. Differences in pressure on the sensor will necessarily result in a higher signal; therefore pressures and flows must be kept completely uniform. In practice this is quite challenging, so to minimize artifacts measurements of reference gases or sample air are always made differentially against a dedicated reference cylinder that flows continuously while the instrument is operational. Since it is not necessary to know the exact δ(O2/ N2) content of this reference gas, it is not measured; such a cylinder is called a “working tank”. A four-port two-way valve switches between these two streams every 1.5 min to avoid artifacts from the drift of either fuel cell, and computation of the final δ(O2/ N2) is thus made relative to both the difference between the two fuel cells and between the sample air and reference gas within one fuel cell: Δraw=[(CWT1-CX1)-(CX2-CWT2)],

δ(O2/N2)=Δcal+[(CO2)M-(CO2)ref]⋅SO2(1-SO2)⋅SO2.

Here C is the raw measurement from the DFCA fuel cells in percent. The subscript refers to a discreet measurement of the working tank (CWT) or the gas stream one wishes to quantify (CX), be it a reference gas or sample air. The numbered superscripts refer to fuel cell/channel 1 or 2. Each value of C represents the average of 1 s measurements for the duration of the period between switches. The first 30 s of this period are discarded, as the fuel cell has to stabilize after a valve switch. Solving Eq. () gives the uncalibrated differential signal, Δraw. After a calibration is applied (see Eq. ), Δcal is used to calculate the final O2/N2 in per meg, as per Eq. (), after and . (CO2)M is the final dry mole fraction in parts per million as measured by the CRDS, and (CO2)ref is an arbitrary reference value of the Scripps scale, 363.29 ppm. SO2 is the average mole fraction of O2 in standard dry air, 0.209392 .

The working tank gas flow is controlled precisely by a mass-flow controller (MKS Instruments, Andover, MA, USA). A pressure transducer (also MKS) measures the pressure difference between the two gas streams and a proportional valve equalizes the pressures so that they do not vary by more than ± 0.2 mbar.

We also present some atmospheric potential oxygen (APO) data in this paper as a further evaluation of data quality since they are a synthesis of both atmospheric oxygen and carbon dioxide measurements. APO is a data-derived tracer for the portion of atmospheric oxygen that is influenced by marine processes and is defined here via the more common, simplified formulation as

APO=δ(O2/N2)+1.1XO2(CO2-350).

Here XO2 is the mole fraction of O2 in standard air, and 350 is the dry mole fraction of CO2 in ppm in the same. CO2 is the dry mole fraction in ppm, which must be measured simultaneously with δ(O2/N2). Both APO and δ(O2/N2) are in per meg. The factor 1.1 is the approximate global stoichiometric ratio of O2 to CO2 surface–atmosphere exchange of the change in both species due to the terrestrial biosphere .

In order to provide an independent corroboration of the in situ data, flask samples were taken on a (nominally) weekly basis, with variations in timing and small gaps due to personnel availability. Since sampling is done manually, the time of day is always between sunrise and sunset, but typically in the afternoon. This preferentially samples the coastal wind sector, since a strong sea breeze is common (but not assured) during the afternoon hours at NDAO. Meteorological conditions are not taken into account during sampling, however, as there are no strong local sources of contamination, provided the generator is not running. All species that are measured continuously at the station are also measured in flasks; additionally the flasks are also analyzed for δ13C–CO2, δ13C–CH4, δ18O–CO2, δD–CH4, H2, Ar/N2, and SF6.

The flasks used by the Max Planck Institute for Biogeochemistry (MPI-BGC) are borosilicate glass and 1 L in volume, with two valves equipped with Kel-F^® (polychlorotrifluoroethylene, PCTFE) seals. PCTFE has been shown to have a low permeability to most gases compared to other common sealing materials, as long storage times can change the composition of sample air in flasks . Flasks are shielded from light exposure with a black sheath. Before sampling the flasks are pre-conditioned by evacuation at 60 ∘C for 72 h and then filled with dry air at a pressure of 1.6 bars.

Samples are taken in triplicate and connected in series upstream of a pump. A dedicated line (identical to the continuous sample lines) is used exclusively for the flasks, although the portion that is downstream of the main pump is not flushed or purged when not in use. The pump body and valve plates are aluminum, and the structured diaphragms are made of PCTFE. When in use the flow rate (3.2 Lmin-1) is higher than the in situ analyzer flow rates (100–200 mLmin-1). Air is dried with a cryotrap identical to the one used for the oxygen sampling line. During sampling, the line is flushed for 5 min before any air is directed to the flasks, then a bypass is opened and the flasks are flushed for an additional 15 min before they are sealed again. After closure, the pressure of the flask is about 1.6 bars.

Due to the lengthy shipping route and remote location of the station, storage time of flasks can be lengthy – the mean number of storage days is 100, the maximum thus far is 226 – and for reactive species like CO this can result in storage-related artifacts and/or permeation.

Laboratory demands and a large volume of samples from this and other stations in the MPI-BGC network mean that not all flasks are analyzed for all species, and in the case of methane isotopic composition, only two out of three flasks are analyzed. δ(O2/N2) is measured with mass spectrometry, CH4, CO2, and CO with GC-FID, and N2O with GC-ECD. Isotopic ratios are determined with isotope ratio mass spectrometry.

A small meteorological sensor array is located at the inlet height on the mast. Wind direction and wind speed are measured with a Thies Clima 2-D sonic anemometer (model 4.3810.30.310), temperature and relative humidity with a combined Galltec-Mela instrument (model C 2.4), barometric pressure with a Young pressure sensor (model 61202V), and solar irradiance with a Kipp and Zonen ISO 9060 first class pyranometer (model CMP11). Data are passed to the LabVIEW system via a controller area network (CAN bus).

In order to characterize the stability of the main instruments, a stream of dry air with a constant composition was supplied to each device for a 24 h period during the final stages of the laboratory installation on site. From this the Allan variance was computed. The Allan variance, a method for computing the frequency stability of clocks, can also be used to analyze the noise of a sensor . Allan variance is given as σy2(τ)=∑i=1N-2n+1(Ti+2n-2Ti+n+Ti)22τ2⋅(N-2n+1), where τ is the sampling interval, N is the number of measurements of some quantity Ti, and n is the number of adjacent values of Ti in τ . Akin to the standard deviation, the Allan deviation is the square root of the variance, σy(τ)=σy2(τ).

The advantage of the Allan deviation is that it is determined in the same units as the measurand is reported. In some cases, the minimum value of the deviation is the ideal averaging interval for the sensor, being the integration time for which sensor noise is the lowest. In practice, however, the integration time that yields the lowest deviation may be too large to be practical.

The absolute abundance of water vapor varies widely throughout the atmosphere; even under well-mixed conditions in the troposphere, the mixing ratio of H2O can span several percent. This dilutes the mixing ratio of trace gases and can obscure their “true” variability if mole fractions are not determined against moles of dry air.

As both the CRDS and OA-ICOS instruments measure water vapor, they afford the opportunity to dispense with sample drying and apply a water correction instead. There are drawbacks to any water correction or drying method. Sample drying introduces a physical alteration of the sample matrix and increases the total number of seals and connections of the gas handling system which must be rendered gas tight. Drying also generally requires additional consumables. Finally, depending on the method used there is also contingent maintenance required to remove the collected water, which can be problematic at remote sites.

Water corrections, however, require continuous, extremely precise, stable measurements of water vapor, which is analytically challenging. It is also a newer approach and as such represents a major break in methodological continuity, a concern for long-term time series. Switching between humid (sample) and dry (reference) air streams is also not ideal. Nevertheless, the method has been shown to be robust for measurements of CO, CO2, and CH4 with the type of instrumentation at NDAO .

It was decided to forgo sample drying for the CRDS and OA-ICOS instruments at NDAO for several reasons: first, the station is in a desert environment, where water vapor mole fractions are typically less than 2 % and water corrections perform best; second, since the site is remote and there are no formally trained technicians available, simple maintenance procedures are greatly preferred; third, the precise, continuous measurements of H2O made by the CRDS can be corrected to yield accurate observations of water vapor, data which otherwise would be unavailable ; and fourth, this obviates the need for drying greatly simplifies the plumbing system for these two instruments, both of which operate under ambient pressure, and reduces the potential for leaks inside the laboratory.

A version of the “water droplet” method was used to humidify the air stream of a target gas cylinder, using a slight variant of one of the methods in (Method 2, “Empa variation”). The apparatus allows the operator to easily switch between dry and humidified gas streams with the aid of the manual valves. The gas stream passes over the surface of a water droplet (< 200 µL) placed in a small well. After the water droplet is completely evaporated and the residual humidity in the tubing is fully removed, the entire shape of the correction function is delimited. The “tuning” parameters used to achieve a sufficient initial water vapor concentration are the height of the tubing over the droplet, the absolute pressure in the trap, and the temperature of the lab (J. Winderlich, personal communication, 2013).

Due to broadening effects on the spectral lines, both the CRDS and the OA-ICOS must be water corrected for each species, X, with a second-order function, in the form of : XwetXdry=1+a⋅H2Orep+b⋅H2Orep2.

H2Orep is the mole fraction of water vapor measured by the instrument in question without correction, usually in percent, and a and b are empirically determined factors determined by the experiment described above. The results of the water corrections are discussed in Sect. .

The recommendations for measurement compatibility between and within different laboratories and measurement sites made by the Global Atmosphere Watch Programme of the World Meteorological Organization (WMO/GAW) are given in Table . We used these compatibility goals as guidelines for evaluating target measurements and instrument error.

All reference gases are stored horizontally in an insulated box to minimize stratification and temperature fluctuations. Reference gases are calibrated against primary standards at the Max Planck Institute for Biogeochemistry (MPI-BGC) GASLAB for CO2, N2O, CO, and CH4. The O2 / N2 ratio of reference cylinders is measured at MPI-BGC by mass spectrometry. All measurements were tied to primary standards on the following scales: WMO X2007 for CO2, NOAA 2004 for CH4, NOAA 2006a for N2O, WMO X2004 for CO, and the Scripps Institute for Oceanography scale for δ(O2 / N2). All reference gases are comprised of dry ambient air and stored in 50 L aluminum cylinders. The initial pressure of each cylinder is 200 bar. Calibration of the instruments is done through four working secondary standards and instrument performance is periodically checked with “target” cylinders (i.e., tanks of known mole fraction which are regularly remeasured). Atmospheric abundances of CO2, CH4, N2O, and CO all are reported as dry air mole fractions, e.g., 1 µmolmol-1 = 1 ppm.

Instrument calibrations are performed automatically by the LabVIEW program. The set points of these tanks were selected to bracket the natural variability which was expected for this site. Mole fraction ranges were 370–410 ppm for CO2, 1600–1920 ppb for CH4, 300–340 ppb for N2O, 40–250 ppb for CO, and between -300 and -700 per meg for δ(O2 / N2). The interval between calibrations was 123 h for the CRDS and OA-ICOS and 71 h for the DFCA. Target measurements were made every 49 h for the CRDS and every 35 h for the OA-ICOS and DFCA. Various schemes were implemented, but the ones used represent the best compromise between the need to save reference gases and the drift of the sensors. As preparation, shipment, customs clearance, and local delivery of a pallet of cylinders to NDAO from MPI-BGC in Germany usually takes about 1 year, we implemented the most conservative (in terms of gas usage) calibration frequency possible without unduly compromising measurement accuracy.

Reference gases (working secondary standards or targets) were measured for a total of 12 min after a 2 min high-flow-rate (250 mL min-1) purge of the sample line. During purges the reference gas flow is not directed to the instrument but vented at the junction closest to the instrument. A stable signal is generally reached after 6 min of measurements for all measurands. As non-linearity was not observed for any instrument in the set point ranges of tanks used at NDAO, the instrument response functions for all species were taken as a linear fit of the average of the last 5 min of working secondary standards measurements and the mole fraction determined by the MPI-BGC facilities: y=mx+b, where x is the “true” value of the measurand, supplied by the analytical facilities at MPI-BGC, and y is the dry air mole fraction or O2 to N2 ratio measured by the instrument during a specific calibration. The slope (m) and intercept (b) are linearly interpolated between calibrations.

A dedicated reference cylinder was used as a target tank to assess long-term repeatability and instrument precision. Target measurements are an imperfect descriptor of system performance, or at least not a complete guarantor of data quality, since such measurements would not reveal a leak in air sampling lines upstream of the common tee. In the case of our measurement system, the pressure of sample line of the CRDS and OA-ICOS while measuring target gas is slightly over ambient and dry, meaning it is measured under different conditions from sample air. For some species, like CO2, there can be small adsorption-related artifacts during gas storage or gas handling which might appear to be due to poor instrument performance but would not be shared by the in situ time series. Nevertheless, regular target measurements can reveal long-term problems or biases and can give an estimate of measurement uncertainty and repeatability.

Following in part the recommendations of , we report the measurement uncertainty as the target-derived measurement uncertainty see Eqs. 9a–d in. This quantity, uTAR, was calculated as the 67th percentile of the absolute value of the difference between the laboratory assigned value or “set point” and the individual mean target determination in a moving window of 1 week.

During laboratory tests before the deployment of the OA-ICOS instrument, it became apparent that it is susceptible to temperature and pressure-related drift. Our Los Gatos analyzer does not have the improved temperature control that later models do (the “enhanced performance” feature). We opted to attempt to correct this drift empirically by measuring a working tank at 2 h intervals, a determination arrived at through consideration of the rate and magnitude of the sensor drift, the loss of in situ data, and the usage rate of the working tank.

Data are corrected for drift after the water correction but before applying the calibration. Each working tank measurement lasts 8 min, without a line purge, and only the last 2 min are used. A spline is fit through the average of all working tank measurements made on the same tank, and the difference of these interpolated measurements from the mean of the tank is then applied to each time step as a correction factor.

High temperatures are a major challenge of working at this site, and maintaining a constant temperature in the laboratory and for the devices is crucial. The CRDS and DFCA are both housed in an insulated rack, which reduces the range of temperatures the devices are exposed to compared to the variation in the laboratory itself. The standard deviation of the air temperature laboratory was 1.5 ∘C over the operational lifetime of the station and 0.77 ∘C for the instrument rack; for comparison the value for the in situ air temperature was 6.8 ∘C. The CRDS cavity temperature was stable within 0.02 ∘C. The OA-ICOS fared less well since it was too large to fit inside the rack (1σ=0.90 ∘C). An insulated box was built for it instead, with active ventilation control. The insulated box performed better than no protection at all but is still not as efficient as the rack, which is reflected in the OA-ICOS instrument temperature; it should be noted, however, that the CRDS has superior temperature control for its measurement cell. For the DFCA, which also has active temperature control, the standard deviation was 0.05 ∘C.

There is one large gap in the record during July 2013 when the hard drive of the control and data acquisition computer failed. This was a solid-state drive (SSD), and there is some evidence to suggest that these drives are more prone to failure than traditional hard drives . Given that the SSD of the communication computer also failed in January 2014, we have forsworn the use of SSDs at NDAO in favor of traditional hard drives.

The second large gap in the records of δ(O2 / N2), CO2, CH4, and H2O is due to a dust storm during a berg-wind event which clogged the CRDS sample line shortly after the communication computer failed and remote connection with the site was no longer possible. The DFCA sample line was not as affected but δ(O2 / N2) requires in situ CO2 data. Due to the large amount of fine dust at the site, filters have to be exchanged frequently and flow rates generally slowly decline as they are clogged; sudden events, however, can cause a rapid decline in flow and pressure in the sample lines and necessitate shutdown until the filters can be replaced.

The data stream also has smaller gaps, due to calibrations, target measurements, LabVIEW software issues, or maintenance. Software and computer issues, rather than instrument failure, is probably the more common cause of data gaps related to technical problems; the exception to this was from March to July 2013, when the station was plagued by frequent (sometimes daily) power outages. Nevertheless, absolute data coverage since October 2012 is approximately 75 % or higher for most measurands.

The time series for N2O and CO start 8 months after the beginning of the station operation due to a leak in the tower line developed a few weeks after installation and contaminated the line with air conditioner exhaust. Long (> 24 h) leak tests were performed during site installation and again after the OA-ICOS leak was discovered and fixed. The leak test was conducted by capping the top of the sample line on the mast and capping the tubing just upstream of the device. The line was then evacuated with a small hand pump to the typical operating pressure of the sample line. After accounting for temperature fluctuations, leak rates determined after this time were lower than 3.0×10-4 mbarL-1s-1, which we estimate, under worst-case scenarios, would result in artifacts smaller than 0.014 ppm for CO2, 0.0017 ppb for CH4, 7.6×10-5 ppb for N2O, and 0.0052 ppb for CO. Fractionation tests were performed for the DFCA sample line, which resulted in no detectable fractionation.

The Allan deviation of both CO2 and CH4 exhibit minima around the 1 to 1.5 h window (Fig. ), after which the drift of the sensor begins to dominate. In the case of CO2, this drift was clearly (inversely) dependent on the air temperature surrounding the working tank, as measured by a temperature sensor inside the insulated box housing all cylinders. Correlations were not seen with rack temperature, regulator temperature (housed outside of the insulated box for cylinders), or room temperature. While it is not atypical for Picarro CRDS analyzers to drift with temperature to this degree , the evidence points to an artifact related to the tank. The effect is about -0.06 ppm∘C-1 over a 24 h period. A weaker relationship with opposite sign was seen for methane. Repetitions of this test during maintenance visits have shown similar results.

Using the raw (i.e., uncalibrated) target measurements as an assessment of the long-term drift of the instrument, we note that long-term drift was not linear over the entire measurement period but wandering due to a slight seasonality in the temperature of the laboratory. The absolute range of raw target measurements on one tank was 0.4 ppm for CO2 and 2.3 ppb for CH4.

While the OA-ICOS is very precise over short intervals, it is much more prone to drift than the CRDS. The optimal averaging interval for this instrument is around 100 s for both species, after which drift begins to dominate the sensor error. The OA-ICOS is prone both to sudden spikes and rapid linear drift when the instrument heats or cools the cavity to maintain a constant temperature. Raw target measurements over 2 years showed a maximal spread of 2.9 ppb for N2O and 0.8 ppb for CO, although the long-term drift was not linear but governed by laboratory temperature.

Due to the differential nature of the DFCA, characterizing the noise of the individual sensors would serve little purpose, since the operator expects that the absolute signal of both fuel cells will drift. The simplest and most meaningful test of sensor stability in the case of this instrument was to provide both fuel cells with flow from a working tank and calculate the Allan deviation of the differential signal (Fig. , top panel), regardless of the low frequency of measurements.

From the time series of the DFCA output (Fig. , bottom panel), it seems that the analyzer took roughly 5 h to stabilize during the stability test. This apparent start-up drift is fairly small and could represent a gradual flushing of the regulator. Despite some small, sudden spikes, the DFCA shows the least dependence of the main instruments on the temperature and pressure of its environment, not surprising since the nature of the dual differential measurement should cancel out temperature effects. WMO/GAW compatibility goals for δ(O2 / N2) can be reached with an averaging interval greater than 7 min.

While the fuel cells have been monitored for degradation, no detectable loss in sensitivity has been noticed over the 2 years of operation.

The aggregated results of several water correction tests can be seen in Fig. . The final fit parameters applied to the time series were determined as a quadratic fit to all tests, and were consistent between tests and across a range of mole fractions. The coefficients show small but significant differences from other values reported in the literature for similar instruments (Table ), highlighting the importance of performing instrument-specific water correction tests with this class of CRDS. For instance, using the values from to correct a NDAO CO2 wet value of 390 ppm at 1.5 % water vapor content would result in a difference of 0.23 ppm. It should be noted that the reference values cited from the literature were obtained using different models of the same CRDS instrument. If one wishes to use the CRDS to measure water vapor, the values produced by the instrument (here referred to as H2Orep) must be corrected to acquire values for H2Oact, the true absolute mole fraction of water vapor. H2Oact can be determined as H2Oact=0.0292+0.7718⋅H2Orep+0.0197⋅H2Orep2.

This relationship is thought to hold between Picarro CRDS instruments, since they are all calibrated on the same scale during manufacture. A major assumption here is that the absolute drift of the water vapor signal is negligible. It should also be noted that the OA-ICOS instrument was not used for determining H2Oact, since tests showed that it was both less accurate and less sensitive than the CRDS; the water vapor data from the OA-ICOS is used only for correction.

Table  gives the values of the coefficients for a second-order water correction for both N2O and CO. Given the only recent availability of the instrument, and the greater prevalence of drying, there are no coefficients to compare to. performed a thorough comparison of the response of several different analyzers to water vapor, including the OA-ICOS N2O/CO-23d, but only considered carbon monoxide.

We also note that the absolute mole fraction of water vapor reported by the OA-ICOS disagrees considerably with that of the Picarro (which in part accounts for the lower water vapor mole fractions displayed in Fig.  for the OA-ICOS), the OA-ICOS values being 2329 ppm (1σ = 342 ppm) lower on average. We verified the accuracy of the CRDS water vapor measurements by comparing them to water vapor mole fractions derived from in situ NDAO relative humidity, temperature, and barometric pressure. Effectively, this means that the OA-ICOS instrument cannot detect values below ∼ 2000 ppm, which results in artifacts during reference gas cycles (targets and calibration runs), as the air stream continues to dry after the instrument returns a 0 value for water vapor and the water correction cannot be applied. In fact, much of the scatter in the N2O time series is related to periods of higher relative humidity and water vapor mole fraction in the in situ measurements.

Finally it should be mentioned that the OA-ICOS's internal water correction is not sufficient to remove the full influence of water vapor and does not account for the quadratic nature of the response curve to water vapor for both N2O and CO. When one compares the dry air mole fraction using the OA-ICOS linear internal water correction to the second-order water correction function used in this work, the two diverge with increasing water vapor mole fraction. For instance, an error of approximately 1.8 ppb at an ambient water vapor mole fraction of H2Orep= 20 000 ppm or ∼ 3.5 ppb at a mole fraction of 30 000 ppm would be produced for N2O using the linear function relative to the quadratic. For CO the effect was less pronounced but would result in errors of 0.7 and 1.5 ppb, respectively.

Calibrations were subjected to quality control (QC) assessments and discarded when they failed to meet specific QC targets for individual species.

For the CRDS, 4 % of calibrations were excluded based on the following QC criteria: a coefficient of determination greater than 0.999, a deviation from the mean slope greater than 0.002 (unitless; ppm/ppm) or a deviation from the mean intercept larger than 1 ppm. For CH4, the QC flags were 0.999, 0.02, and 10 ppb, respectively. This approach ensures that anomalous calibration coefficients were not used in producing the final time series. In most cases the poor calibrations were related to recent power outages or large variations in temperature in the laboratory, and the species causing the flag was CO2. The mean R2 was 0.9999987 for CO2 and 0.999995 for CH4.

Due in almost all instances to N2O not meeting QC standards, 20 % of the OA-ICOS calibrations were rejected. The QC standards were: greater 0.999 coefficient of determination, slope deviation from the mean not greater than 0.015, and an intercept within ± 5 ppb of the mean of all intercepts. For CO the corresponding values were 0.99996, 0.02, and 1 ppb. The mean R2 was 0.99997 for N2O and 0.999989 for CO.

In the case of atmospheric oxygen, 33 % of calibrations were discarded, most of these during a period of frequent power outages and hence poor temperature control in the instrument rack. The QC criteria were an R2>0.99, a slope not deviating more than 0.01 from the mean, and an intercept within 25 per meg of the average. The mean R2 was 0.996.

Target measurements (tanks where the mole fraction all measurands are known but are treated as unknowns) for the time series are presented in Fig. . Periods of poorer performance are associated with inadequate electricity supply, dust events, and/or degraded temperature control of the laboratory or a specific instrument. The target tank had to be changed several times during the first 6 months of operation due twice to the loss of target gas during power outages before a more robust power-down procedure was implemented. In the new procedure, after power loss the control computer returns all valves to a configuration that prevents the loss of gas while running on an uninterrupted power supply (UPS) system. Other target tank changes were due to the necessity of switching which tanks served as target and which as working secondary standards as the calibration scheme was adjusted to conserve gas before more cylinders could be delivered. Bias (mean offset) between the assigned values obtained from MPI-BGC laboratories of reference tanks and the average of all target measurements at NDAO are given in Table .

For oxygen this bias is large enough to warrant some speculation. As the target gas shares the same plumbing lines as the calibration gases, a leak or fractionation effect is unlikely, since it would effect the calibration gases as well. It is always possible that the target was contaminated during installation due to blowback when installing the regulator, which might account for the slightly enhanced CO2 and CO and the depleted δ(O2/N2), while N2O and CH4 are unaffected as this is the sort of profile one would expect in the laboratory environment. The slope of the linear fit to the calibration data is also quite sensitive to the absolute value of the cylinder with the lowest mole fractions, so these small biases could also be due to inter-laboratory differences at the low end of the scale for those three species.

Measurement uncertainty was estimated from the target tank time series (uTAR) to give a time-varying quantity (Fig. ). The average bias of each target tank was removed to calculate the measurement uncertainty since this varied from tank to tank and would not necessarily impinge on sample uncertainty of the in situ data. Note that this bias was not removed from the target tank time series in Fig. .

The mean uncertainty for each species was 0.027 ppm for CO2, 0.16 ppb for CH4, 0.22 ppb for N2O, 0.15 ppb for CO, and 6.7 per meg for δ(O2/N2).

The average drift correction over each 2 h period for the N2O data was ± 0.47 ppb (or 0.2 ppb h-1), with a maximum of ± 1.9 ppb. The corrections for CO were a little lower, with an average of ± 0.31 ppb (or 0.16 ppb h-1) and a maximum of ± 1.1 ppb. The application of this empirical drift correction significantly improved the data quality for both species as measured through the target cylinders (Fig. ). In the case of N2O, this allowed for the performance goal for this species to be reached on average.

A small inter-laboratory comparison was performed with the global GAW station run by the South African Weather Service at Cape Point, South Africa. A cylinder of dry background air from Cape Point was measured at both sites for CO2, CH4, N2O, and CO. The absolute value of the difference between the final determinations, i.e., the compatibility, was 0.1, 1.2, 1.3, and 0.2 ppb, respectively. More comparisons are planned.

Flask samples also offer a kind of external validation of the in situ measurements through assessing the compatibility of the flask and continuous data. The flask time series begins in mid-2013 and is ongoing. The average standard deviation of the flask triplicates were 0.04 ppm for CO2, 1.6 ppb for CH4, 0.06 ppb for N2O, 1.1 ppb for CO, 4.0 per meg for δ(O2/N2), and 3.9 per meg for APO.

Measurement compatibility between flask and in situ measurements is given in Table . After accounting for differences in flow rates, the final 10 min of the 15 min filling procedure was averaged to compare to the flasks. Most of the bias seen between flask and in situ measurements can be explained for carbon monoxide. CO mole fractions are known to increase in flasks during storage . At MPI-BGC, tests showed rates between 0.02 and 0.03 ppbday-1, in some cases as high as 0.05 ppbday-1 (A. Jordan, personal communication, 2014). Additionally, the offset showed some dependency on storage time (Fig. ); the specific mechanism is unknown.

Additionally, the NDAO in situ time series was compared to the available data set from the nearby NOAA ESRL GMD sampling site NMB (Fig. ). It should be noted that this is not a direct assessment of the compatibility between the two measurement programs, since the two stations do not really measure the same air masses at exactly the same time; the sites are about 2 km apart and have a height differential of 32 m. This can be of consequence since the boundary layer oscillations in the local wind field often create considerable heterogeneity in flow, resulting in spatial and temporal gradients of trace gases . While small differences between concurrent measurements at the two sites would be expected, the synoptic variability, seasonality, and long-term trend should be the same, and this is what is observed (for a discussion of these broader trends, see Sect. ). After excluding outliers smaller than 3σ, the average absolute difference of all concurrent measurements (using 1 h means for the in situ data) was 0.43 ppm for CO2, 2.4 ppb for CH4, 0.25 ppb for N2O, and 3.4 ppb for CO.

The full station time series consists of the five main measurands, plus atmospheric potential oxygen (APO), presented in Fig. . Six meteorological parameters are also measured continuously, presented in Fig. . The time series has been filtered based on diagnostic data and for extreme outliers.

Since the aim of this paper is to present and evaluate measurement techniques from the field, we do not provide a detailed scientific analysis of the data but instead briefly characterize the variability of the data with respect to other measurements from the Southern Hemisphere as a final assessment of measurement accuracy.

Annual growth rates for measurands were determined for NDAO, the nearby NOAA ESRL flask site NMB, the observatory run by the South African Weather Service at Cape Point (CPT), and NOAA ESRL flask data for Cape Grim (CGO). The growth rate for NDAO CO2 between October 2012 and October 2013 was 2.9 ppmyr-1 and 1.6 ppmyr-1 for the period October 2013–October 2014. Since the data sets were of different lengths, comparisons were made between sites for identical time periods. For the period October 2012–December 2013, the growth rate of CO2 was 3.0 ppmyr-1 for NMB and 2.9 ppmyr-1 for NDAO, CGO, and CPT. The latter value was the average global growth rate of CO2 for the calendar year 2013 .

For the same period (October 2012–December 2013) the growth rate of methane was determined to be 12.9 ppbyr-1 for NDAO, 11.1 for NMB, 12.3 for CPT, and 10.3 for CGO. During October 2013–October 2014, the growth rate was 16.9 ppbyr-1 at NDAO. Some of this discrepancy is due to the short time span and low frequency of sampling, which makes the calculation sensitive to outliers and intraseasonal variability. The average global increase during 2013 was 6 ppbyr-1 . Growth rates of methane vary with location and latitude , so these higher growth rates could be a result of a larger source of CH4 from the Southern Hemisphere for this year.

Atmospheric growth rates of N2O were only obtained for the two sites with sufficiently overlapping time series at the time of writing, NMB and NDAO. The period considered was July 2013–October 2014. At NMB the growth rate was 0.9 ppbyr-1 and at NDAO the value was 1.0 ppbyr-1.

Seasonality is apparent in the meteorological time series, although in the case of air temperature, the diurnal and synoptic variability masks much of the small seasonal cycle. Wind direction also exhibits some distinct seasonality, with predominately westerly winds (from the ocean) during austral summer and easterly winds during austral winter.

All measurands also display the expected seasonality and generally agree well in phasing and amplitude to the other Southern Hemisphere sites. For CO2 and CO, the seasonal cycle is somewhat distorted, likely due to seasonal changes in atmospheric transport (evident in the seasonality of atmospheric pressure, water vapor, and wind direction) at this site which bring continental air masses to the station during austral winter. This variability is also seen in the NMB data.