Replacing the AMOR with the miniDOAS in the ammonia monitoring network in the Netherlands

In this paper we present the continued development of the miniDOAS, an active differential optical absorption spectroscopy (DOAS) instrument used to measure ammonia concentrations in ambient air. The miniDOAS has been adapted for use in the Dutch National Air Quality Monitoring Network. The miniDOAS replaces the life-expired continuous-flow denuder ammonia monitor (AMOR). From September 2014 to December 2015, both instruments measured in parallel before the change from AMOR to miniDOAS was made. The instruments were deployed at six monitoring stations throughout the Netherlands. We report on the results of this intercomparison. Both instruments show a good uptime of ca. 90 %, adequate for an automatic monitoring network. Although both instruments produce 1 min values of ammonia concentrations, a direct comparison on short timescales such as minutes or hours does not give meaningful results because the AMOR response to changing ammonia concentrations is slow. Comparisons between daily and monthly values show good agreement. For monthly averages, we find a small average offset of 0.65± 0.28 μg m−3 and a slope of 1.034± 0.028, with the miniDOAS measuring slightly higher than the AMOR. The fast time resolution of the miniDOAS makes the instrument suitable not only for monitoring but also for process studies.

The table shows that instrument performance has increased over the years: detection limit and integration time have decreased simultaneously. The other entries in the table show how the groups tackled various challenges posed by the DOAS technique. The line "set-up" in the table describes the instrument lay-out. The Edner instrument is bistatic, i.e. light source and detector are in two separate locations. This makes alignment difficult, as both parts of the set-up must be precisely aligned. It also requires power to be available at both ends of the path. For this reason, all other instruments have light source 5 and detector combined in a single instrument. Such a monostatic set-up folds the path back with a retroreflector. Losses due to the extra reflections at the reflector and to the optical geometry (part of the light beam gets reflected back into the light source, missing the detector) makes this set-up less light efficient. This is reflected in the shorter path lengths used.
In 1990, Edner and co-workers selected a scanning slit monochromator with a photomultiplier tube as detector, noting that a spectrograph with an array detector had many advantages but was financially out of reach. Improvements in semiconductor 10 technology made diode arrays and CCDs accessible to the makers of the other instruments. The miniDOAS uses an uncooled CCD rather than a cooled detector as the other instruments do, sacrificing some performance for lower costs. To help tackling stray light, the Edner, RIVM and miniDOAS instruments use an interference filter to block out visible light. The Mount system instead employs a double spectrograph, eliminating the need for an interference filter. The Sintermann system uses a deuterium arc lamp rather than a xenon arc lamp, as the other systems do. A deuterium arc lamp emits hardly any 15 visible light, which is the reason this system can do without an interference filter. Its lack of visible light also makes this system less obtrusive, which may facilitate its placement and avoids attracting vandalism. A disadvantage of a deuterium arc lamp is its shorter lifetime when compared to a xenon arc lamp. This tipped the balance towards xenon arc lamps for the other systems.
Another consideration is that its copious amounts of visible light makes a xenon arc lamp inherently more eye-safe than a 20 deuterium arc lamp, as the natural reaction of people to the bright visible glare of the xenon arc lamp is to look away. The pale purple glow of the deuterium arc lamp offers no such reflex, so onlookers may inadvertently be exposed to ultraviolet radiation. For an instrument in a monitoring network, that is to operate unattended 24 hours per day, this extra safety offered by the xenon arc lamp is an important advantage.
When comparing the miniDOAS with the other systems, we note that the detection limit is slightly higher than of the two 25 contemporary systems, but lower than of the systems built 10 and 22 years earlier. The path length of the miniDOAS is shorter than that of all other systems, reflecting the low-power lamp used. The use of an uncooled CCD does not affect the performance unduly.
The miniDOAS shows adequate performance for a monitoring network: a detection limit of 0.25 µg m -3 , an accuracy of 0.25 µg m -3 , a true time resolution of 1 minute and an instrument uptime exceeding 90%. When compared to the AMOR, 30 purchasing price and maintenance requirements are much lower, leading to an attractive cost reduction while increasing measurement quality. Before the transition from AMOR to miniDOAS was made on 1 January 2016, an extensive comparison period was conducted, from September 2014 to December 2015, in which both instruments were operated in Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License.
parallel. This paper describes the implementation and performance of the miniDOAS, the intercomparison with the AMOR and some issues associated with the transition.

2
Measurement methods

The Dutch National Air Quality Monitoring Network
The Dutch National Air Quality Monitoring Network (LML) was established more than 50 years ago, to monitor air 5 pollution. Starting in 1992, ammonia has been measured at eight of these stations (since 2014 on six stations). The instrument used from 1992 to 2016 was the AMOR (Ammonia Monitor), see Sect. 2.2.1. A map of the network is shown in Fig. 1. More detailed maps of the monitoring sites are shown in Appendix A.

2.2
The AMOR

Description of the instrument 10
The AMOR, Ammonia MonitoR, is an automatically operating continuous-flow denuder system. It was developed at ECN in the early 1990s from the AMANDA (Wyers et al., 1993). The AMOR differs from the AMANDA mainly in its remote control options and its ability to operate unattended for prolonged periods of time (up to 4 weeks). It is described in detail in Erisman et al. (2001), Wyers et al. (1993) and in Buijsman et al. (1998). The procedure in which this instrument was selected for use in the monitoring network LML, as well as tests of its performance and the results of the first years of measurements 15 are described in Buijsman et al. (1998) and in Mennen et al. (1996). The published specifications for the AMOR were a detection limit of 0.01 µg m -3 , an accuracy of 2% and a time resolution of 3 min (Erisman et al., 2001).

Implementation in the monitoring network LML
The instrument was installed inside the climate-controlled housing of the monitoring stations. Air was sampled from an inlet on the roof of the housing, at 3.5 m above ground level. The air flow through this inlet was 250 m 3 h -1 . From this air flow, a 20 small fraction (25 L min -1 , or 0.6%) was sampled, just after a 90° turn in the inlet tube, and fed into the AMOR. This arrangement served to minimise the amount of particulate matter being sampled by the AMOR.

Calibration
Calibration of any trace gas monitor is preferably done by offering the instrument a gas stream with a high but realistic concentration of the gas to be measured. For ammonia, this means a concentration of e.g. 400 µg m -3 . A gas bottle with a 25 mixture of such an ammonia concentration in e.g. nitrogen is not stable. For the AMOR in the monitoring network, calibration was therefore carried out by offering the instrument a solution of NH 4 + of a known concentration, corresponding Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License.
to a realistic ammonia concentration in the ambient air. Calibration of the system took place automatically, every 80 hours, with NH 4 + solutions in two concentrations.

AMOR dataset used in this study
The AMOR data were corrected for a small offset compared to data published in national and international databases. The offset is caused by a digital to analogue conversion of the data. See Appendix B for more information. 5

Description of the instrument
The miniDOAS is extensively described in Volten et al. (2012a). The instrument uses a xenon lamp as ultraviolet light source and a retroreflector to measure optical absorption spectra along an open path, typically 42 m long. It uses the DOAS technique, Differential Optical Absorption Spectroscopy, to retrieve concentrations of several atmospheric trace gases along 10 this path. See Fig. 2 for a schematic representation of the optical set-up of the instrument.

Retrieval of concentrations
The spectral window used is from about 203.6 to about 230.9 nm, the precise window differs slightly between instruments.
In this region, three gases commonly present in the atmosphere show specific absorption patterns: NH 3 , SO 2 and NO. These three gases are retrieved together. Other atmospheric constituents either do not absorb in this region (e.g. NO 2 ) or have 15 absorption features that change only slowly with the wavelength (e.g. O 3 ), those are filtered out by the fitting algorithm.
Central to the DOAS technique is the Lambert-Beer law. Eq. (1) is one form to write it (CEN, 2013): Here, I meas () is the measured spectrum, I 0 () is the intensity spectrum as emitted by the instrument, a() is the specific absorption coefficient of the component through which the light passes, c is the concentration of that component and l is the 20 optical path length. In the open atmosphere, light is attenuated not just by absorption but also by Rayleigh and Mie scattering. The wavelength dependent attenuation by the optical system must also be taken into account. The key to the DOAS technique is to separate narrow-band absorption features in the specific absorption spectrum a() from broadband features due to interfering compounds, atmospheric scattering and the intensity spectrum of the light source. This leads to the following equation: 25 Here, I bgc () is the background corrected measured spectrum, see Sect. 2.3.4 for the determination of this background.
is the differential initial intensity, all broadband features are collected in this term.
is, for component i to be measured, the part of the specific absorption spectrum containing the narrow-band absorption features. In logarithmic form: To approximate , we use a moving average of the measured spectrum, denoted as [I bgc ()] moving average in Eq. (4) below.
We found two consecutive passes with averaging over 41 channels each time to work well. We define the DOAS curve To retrieve the concentrations, we use a three component least-squares fit (Kendall and Stuart, 1976;Volten et al., 2012a): Here,  is the standard deviation of the fit. The wavelength is denoted by j, n is the total number of wavelengths, X() j , Y() j and Z() j are reference spectra for NH 3 , SO 2 and NO, respectively. The parameters ,  and  are proportional to the concentrations of NH 3 , SO 2 and NO to be retrieved. The retrieval algorithm minimises  2 to find the best fit. The 10 concentration of NH 3 is calculated according to Eq. (6): Concentrations for SO 2 and NO are calculated analogously.

Calibration: span measurements
To determine the reference spectra X(), Y() and Z() used in Eq. (5), two methods exist (CEN, 2013): calibration with 15 complete spectral modelling using reference spectra, and gas cell calibration with and without including the atmosphere.
Both methods have their advantages and difficulties.
The spectral modelling method involves modelling the complete system. This requires knowledge of the instrument line shape function (ILS) of the spectrometer, of the presence of straylight in the spectrum, of the linearity and dark current of the detector, and of the differential absorption coefficient of each component at the wavelengths used. The last parameter can be 20 obtained by measuring a spectrum with the instrument itself, or by convolving a high-resolution absorption spectrum from the literature with the ILS.
For gas cell calibration, a cell with the gas for which the reference spectrum is to be determined is placed in the light path.
The concentration of the gas should be known. Care should be taken that the gas is stable in the cell, or a flow-through cell should be used. When applying this method with inclusion of the atmosphere, only an incremental calibration can be 25 performed, i.e. the system will measure an increase due to the gas in the gas cell with respect to the atmospheric background concentration. This requires stable atmospheric conditions. For the gas cell method with exclusion of the atmosphere, the light path should be routed directly from the light source through the gas cell into the detector. This eliminates all atmospheric influence. It also allows a calibration under zero gas conditions. 30 Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License.
The systems discussed above all use the gas cell method to calibrate. Edner et al. (1993) do not specify whether they included the atmosphere or not. Mount et al. (2002) discuss the possibility of using the spectral modelling method but prefer the gas cell method, because it automatically convolves the ILS with the gas cross section spectrum. They did exclude the atmosphere. Sintermann et al. (2016) do not specify whether they included the atmosphere or not. They did apply the modelling method, as a check on their calibration spectrum, for NH 3 only. For the RIVM DOAS, we used the gas cell 5 method, with exclusion of the atmosphere (Volten et al., 2012a).
For the miniDOAS, we use this method as well. We place a 75 mm path length quartz flow cell in the instrument, see Fig. 2.
Pressure and temperature of the gas are continuously measured so that the amount of gas in the cell is known. Because the path in the optical cell is, at 75 mm, only 1/560 st of the full 2 x 21 m path that is used in the open air, the concentration in the cell must be 560 times the ambient concentration. This is an important advantage, as it enables DOAS systems to be 10 calibrated with high-concentration gas mixtures. These are much more stable than the mixtures at ambient concentrations that an air-sampling instrument would use. The gas mixtures used are listed in Table 2.
To exclude the atmosphere, we originally used a shortened optical path of 1 m rather than 42 m. However, we found that, for the miniDOAS, this short path yielded distorted spectra compared to spectra measured over a full length optical path. We attribute this distortion to the difference in Rayleigh scattering over a longer versus a shorter path. At the short wavelengths 15 we use for ammonia DOAS, this effect is much more pronounced than in DOAS applications at longer wavelengths, since the intensity of Rayleigh scattering is proportional to the inverse fourth power of the wavelength of the light. In addition, the tail from the Schumann-Runge absorption bands of O 2 affects the spectrum at the low wavelength side (Yoshino et al., 1984). The combined distortion negatively affected the fitting procedure. Therefore, we decided to measure the reference spectra with the full optical path. A disadvantage of this long path is that any gas present in the atmosphere will leave its 20 spectral signature on the reference spectra, whereas those spectra are assumed to contain only the known concentration of the target gas. To address this issue, we set up the miniDOAS that has its reference spectra measured (the miniDOAS under test) in the laboratory next to another DOAS (the reference DOAS), with the optical paths of both instruments running parallel, so that they measure the same parcel of ambient air. Both instruments measure a full-length outdoor atmospheric path.
The reference DOAS can be any previously calibrated DOAS, or indeed, any instrument capable of measuring ammonia. It 25 turned out to be most convenient to use an RIVM DOAS (as described in Volten et al. (2012a); see also Table 1) that had been calibrated with exclusion of the atmosphere. This DOAS system reports values at 5 minute intervals, it has a detection limit of 0.15 µg m -3 for NH 3 and it receives regular maintenance.
We use the reference DOAS to determine the concentrations of NH 3 , SO 2 and NO in the ambient air during the reference spectra measurements. If those ambient concentrations are too high (>10 µg m -3 for NH 3 , >5 µg m -3 for SO 2 or >20 µg m -3 30 for NO) the resulting reference spectra are rejected and measured anew. In this way, we make sure any remaining effects are small: Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License.
-The concentrations in the gas cell are chosen so that the equivalent concentrations in the atmosphere are realistic, but high. As an example, for NH 3 , an equivalent concentration of 500 µg m -3 is used. If NH 3 is present in the outdoor optical path at the typical ambient concentration in the Netherlands of 5 µg m -3 , the resulting error is 1%.
-The effects of one gas being present in the reference spectrum of another gas is also small. As an example, suppose some ambient NO is present while the reference spectrum for NH 3 is being measured. The resulting spectrum will have the 5 spectral features of both NH 3 and NO, but since NH 3 is present in a much higher concentration, the features of NO will be negligible.
As measuring NH 3 concentrations is the prime aim of this instrument, the calibration for NH 3 is checked with a Primary Reference Material (PRM), obtained from VSL in the Netherlands. This check is done by measuring the apparent concentration when a PRM gas mixture of NH 3 in N 2 is passed through the flow-through calibration cell, using the same 10 procedure as outlined above for the measurement of the reference spectra. If the concentration reported by the instrument is within the tolerance of the PRM gas, as indicated by its manufacturer (this amounted to 2%, see Table 2), the instrument is considered to have passed the test. If the reported concentration is outside the tolerance, the instrument is considered to have failed the test. In the latter case, its reference spectra are discarded and measured anew.
Reference spectra are measured in the following cases: 15 -Before a new system is deployed for the first time.
-After a system has its xenon lamp replaced. This usually takes place after operating a system for one year.
-After any repairs to the spectrograph.
Because replacing the xenon lamp makes measuring new reference spectra necessary, such a replacement takes place by exchanging a system with a life-expired lamp with a system with a new lamp installed and newly measured reference spectra 20 present. This minimises the instrument downtime.

Calibration: zero measurements and dark current corrections
The background subtracted spectrum referred to in Eq. (7) is calculated as follows: , Here, I bgc () is the background corrected spectrum, I meas () is the spectrum measured by the spectrograph, I dark (,T) is the 25 temperature-dependent dark spectrum and I background () is a reference spectrum containing no spectral features from atmospheric gases.
The temperature-dependent dark spectrum I dark (,T) is determined by placing the spectrograph in a dark temperaturecontrolled room and measuring the spectral response at a number of temperatures and integration times.
Dividing by the reference spectrum I background () improves the signal-to-noise ratio of the measurement. To measure 30 I background (), we need an optical path that is free of NH 3 , SO 2 and NO. For the RIVM DOAS reported on in Volten et al. concentrations. Other groups handled this in a similar way, e.g. Mount et al. (2002), who used a shortened path away from local sources.
An alternative is to somehow deal with the background concentrations of NH 3 , SO 2 and NO along the full length optical path. Sintermann et al. (2016) identify three ways to do this: 1) Monitor for an extended period of time. During this time, some episodes with near-zero ambient concentrations are likely 5 to occur.
2) Measure I background () at a remote location where ambient concentrations are assumed to be very low. Alternatively, create an artificial low-concentration environment in the laboratory, over the full optical path.
3) Remove traces of NH 3 , SO 2 and NO by excluding narrowband absorption from I meas (). This may be done by applying a low-pass filter. 10 They tested method 3) and found it to work well; however, the method also introduced extra uncertainty to the results.
We used method 1) for the reference DOAS mentioned above. A disadvantage of this method is that I bgc () is only available after operating the instrument for a long time, of many weeks or even months. This is fine if measurement results are analysed after a monitoring period, but it is not acceptable in a monitoring network, when measurements are to be reported in near real-time. For the miniDOAS, we adopted a different approach. 15 We use the set-up with a reference DOAS and the miniDOAS under test running parallel that we described in Sect. 2.3.3.
The I background () of this reference DOAS was measured using method 1), as outlined above. The reference DOAS indicates when the interfering gases reach low concentrations in the ambient air. Spectra measured with the miniDOAS under test during those episodes are used as its I background (). This I background () still contains the signature of low ambient concentrations of NH 3 , SO 2 and/or NO, typically below 5 µg m -3 for NH 3 . We determine the average differences in concentration between 20 miniDOAS and reference DOAS and correct for these in the retrieval procedure. For an example, see Fig. 3.

Error sources, detection limit and precision
We distinguish between random and systematic error sources. The random error sources, e.g. the correction for the dark current, end up in the residual spectrum. Their combined magnitude is estimated by the standard error that is reported by the fitting algorithm (Stutz and Platt, 1996). To use the standard error to estimate upper limits for the detection limit and the 25 precision we select episodes at the monitoring stations with very low ambient concentrations. An example of such an episode is shown in Fig. 4. Note that this episode still contains both a background concentration and some natural variability in the ammonia concentration. From such episodes, we estimate the upper limit of the precision to be 0.25 µg m -3 and the detection limit to be 0.25 µg m -3 , both at the instrument time resolution of 1 minute (Table 3). The precision of hourly averaged data we estimate to be 0.1 µg m -3 . 30 Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License.
The systematic error sources are listed in Table 4. As an example, the calibration gas concentration is known with a precision of 2%. This concentration is used only once, in the preparation of the reference spectra, and the associated error is the same for as long as this spectrum is used. These will never be estimated by the fitting algorithm.
We use a tape measure to determine the path length, we estimate that we do this with a precision of 5 cm, or 0.25% of 20 m.
The calibration gas concentration is given by the manufacturer with a precision of 2%. The zero concentration measured by 5 the reference DOAS we determine with a precision of 0.45 µg m -3 , i.e. 3 times its precision of 0.15 µg m -3 (Volten et al., 2012a). Combined, we estimate the detection limit to be 0.45 µg m -3 and the precision to be 2.25%, with a minimum of 0.25 µg m -3 .

Instrument intercomparison campaign
To fully characterise the differences between the AMOR and the miniDOAS, an instrument intercomparison campaign was 10 conducted. From 2 September 2014 to 31 December 2015, on each of the six operational stations, both an AMOR and a miniDOAS were operated in parallel. Both instruments were operated in the regular monitoring network mode. For the miniDOAS, operating procedures and the data transfer set-up were updated and refined during the campaign. This did not influence the measurement data, as these refinements dealt with issues not directly affecting the measurements or the retrieval. These changes in the procedures did improve the instrument uptime. 15 The AMOR measurements were conducted under ISO 17025 accreditation.

2.4.1
Height difference assessment AMOR measurements have always been conducted with an air inlet at 3.5 m above the local ground level. This is the standard air inlet height for the Dutch Air Quality Monitoring Network. For the miniDOAS this height was considered unpractical, as it would mean mounting the instrument outside the station housing, or using a complex optical setup. Instead, 20 a measuring altitude was chosen of 2.2 m. This corresponds to the highest practical mounting position inside the station housing.
The effect of the difference in measurement height was studied using passive samplers. At all stations, three sets of three Gradko passive sampler tubes (Lolkema et al., 2015) were installed: -Set 1: attached to the AMOR air inlet at 3.5 m height, above the station housing roof. 25 -Set 2: at 3.5 m height (the AMOR air inlet height) in a separate mast, halfway the miniDOAS optical path.
-Set 3: in the same mast, but at 2.2 m height (the height of the miniDOAS optical path). See Fig. 5 for an overview of the situation.
The passive sampler sets were exchanged and analysed monthly. Measurements took place between January and December 2014. Table 5 lists in which months and on which monitoring stations the samplers were deployed. 30 Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License. Results

Dataset and uptimes of the miniDOAS and AMOR
In Fig. 6 uptimes of the AMOR and miniDOAS systems are given over the period from 1 September 2014 to 1 September 2015. We excluded periods when the miniDOAS systems were in the laboratory for instrument characterisation, when station housings were renewed, et cetera. 5 Uptimes for both the AMOR and the miniDOAS instruments were mostly between 80-90%, as is required for instruments in a monitoring network. Note that the miniDOAS systems were during this period not formally part of the monitoring network and therefore not under continuous surveillance, in contrast to the AMOR instruments.
Comparing any two instruments can only be done with data gathered on a timescale that permits both instruments to produce meaningful data. Both AMOR and miniDOAS generate a data point every minute. However, it takes time -about half an 10 hour -for ammonia to be processed by the AMOR system. This causes a delay and a smoothing effect in the AMOR values.
In addition, instruments that employ inlet lines are known to suffer from a memory effect due to ammonia sticking to walls of the inlet line (Parrish and Fehsenfeld, 2000). This ammonia may be released later, depending on temperature and relative humidity. The miniDOAS has no inlet lines and shows an instant response to ammonia in the air. Therefore, comparison between AMOR and miniDOAS minute value is not feasible, as will be shown below. Comparison between hourly values is 15 complicated but possible, comparison between daily and monthly values works fine.
We will briefly illustrate the smoothing and delaying effects of the internal works and inlet lines of the AMOR on its data by applying a similar effect to the miniDOAS data using the simple formula in Eq. (8) (Volten et al., 2012a;Von Bobrutzki et al., 2010): where c'(t) is the delayed smoothed concentration, c(t) is the measured miniDOAS concentration data and f is a smoothing factor which would be unity for an instrument equally fast as the miniDOAS. We use an e-folding time  1/e of 1 hour, where  1/e = 1/f. The value of 1 hour was adopted for illustration purposes, similar to what was reported earlier in Volten et al. (2012a). After applying the smoothing and delaying effect the miniDOAS data is remarkably similar to the AMOR data, as illustrated in Fig. 7. It is not our aim to find the perfect smoothing and delaying curve for the miniDOAS data to reproduce 25 the AMOR data. We just wish to illustrate that for comparisons of the miniDOAS and the AMOR data it is more meaningful to compare averages over longer time intervals. Below we give some examples of the AMOR and miniDOAS data compared for different, increasing time intervals, starting with hourly values, to daily values to monthly values.

Hourly values compared
When comparing hourly values of the miniDOAS and the AMOR the delay effect is less pronounced than for the 30 comparison of the minute values, but a smoothing effect and a delay is still clearly visible as illustrated in Fig. 8, containing hourly data for a selected period at the monitoring station in Vredepeel. This station has strongly varying concentrations with Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License. relatively high ammonia peaks, up to several hundred micrograms per cubic meter. Since the miniDOAS suffers no smoothing effect, it captures the concentration peaks more effectively than the AMOR.
The delay effect is clearly visible during 17 March. In a period when the concentrations are less variable and less extreme, e.g. from 19 to 22 March, the AMOR and miniDOAS data agree much better. This is reflected in the scatter plot shown in Fig. 9. Here, all the largest deviations from the y = x line are cases of the miniDOAS value being larger than the AMOR 5 value.
On a monitoring station where on average the ammonia concentrations are much lower, such as De Zilk, the delay and smoothing effects are visible as well. As is apparent from the scatter plot (Fig. 10), deviations from the y = x line still occur, again the deviations are larger on the side where the miniDOAS values are higher than the AMOR values.

3.1.2
Daily values compared 10 This is also apparent when we compare the scatter plots of the hourly values at Vredepeel (Fig. 9) with the scatter plot of the daily values at the same station (Fig. 12). The latter shows more clustering around the y = x line, although some extremes remain.

Monthly values compared
In Fig. 13 we give monthly averages for both miniDOAS and AMOR. The monthly averages are based on hourly values, but 20 only on those where both AMOR and miniDOAS values were simultaneously available. A monthly average is included when it is based on at least 100 hourly values. The series of monthly values tend to agree well, showing very similar patterns. In many cases, though not all, the AMOR values tend to be slightly below the miniDOAS monthly values. Some episodes, e.g. September 2014 to March 2015 for Wekerom, agree really well. Some other episodes, e.g. July to December 2015 for the same station, show less agreement. The reasons are so far unknown. There is no correlation between high 25 AMOR-miniDOAS differences and episodes with high or low values. Neither is there any seasonal influence.
To evaluate the comparability of the AMOR and miniDOAS data the monthly values in Fig. 13 have been used for a orthogonal regression plot presented in Fig. 14. The number of data pairs included is 89 and the R 2 = 0.94. This orthogonal regression yields a relation between the values of miniDOAS and AMOR given by Eq. (9) 1.034 28 • 0.65 28 (9)  30 where the uncertainty (1) in the last two digits of the slope and the offset are given between brackets.

3.2
Results of the height difference assessment 5 Table 7 lists the annual averaged differences in ammonia concentrations measured with passive samplers at different measurement heights averaged over all sites, and standard 2-sigma errors. Given the small concentration differences and the relatively large statistical variance associated with these passive samplers, analysis per station or per season is not feasible with this dataset.
We see no significant difference between the set at the AMOR inlet (at 3.5 m) and the set at the miniDOAS path (at 2.2 m). 10 Results do show a difference between the two measurement points at 3.5 m, i.e. those at the AMOR air inlet and at the separate mast. The concentrations at the AMOR air inlet are lower. This may be explained by the station housing influencing the air flow: the air sampled by the AMOR is not pure air from 3.5 m height, it is mixed with air from lower heights forced upwards by the station housing. In both cases the statistical error is substantial.

4
Discussion 15 Analysing the results obtained in the comparison we see that the uptime of both instruments is comparable. At 80-90%, the miniDOAS uptime is adequate for an instrument in an automated monitoring network. We expect that the uptime of the miniDOAS will further improve in 2016, as from then on the instrument will benefit from the regular monitoring of the network performance.

Timescale of the intercomparison 20
Both instruments provide minute values of ambient ammonia concentrations. When looking at short timescales (minutes, hours) we see relatively large differences between the datasets. The differences get smaller as the timescale gets longer.
When we look at the fits in the scatterplots of hourly, daily and monthly averages (Fig. 9, Fig. 10 and Fig. 14, respectively) we see that the slopes approach unity: 1.54 for hourly averages, 1.27 for daily and 1.03 for monthly averages. The offsets approach zero, from -7.34 µg m -3 for hourly averages, via -3.06 µg m -3 for daily to 0.65 µg m -3 for monthly averages. 25 For this reason, we focused our comparison on longer timescales: daily and monthly values. Daily value pairs showed good agreement in a direct comparison, i.e. when the concentration values are plotted in the same graph (see e.g. Fig. 11). The smoothing and delay effects that are apparent in the minute and hourly values have largely disappeared. However, scatter plots (see e.g. Fig. 12) show still some deviations from y = x, indicating that some delay effects are still not smoothed out. This is to be expected, a high peak just before the transition to a new day will cause differences in two consecutive days. In

Intercomparison at longer timescales
Monthly averaged concentrations show a linear relationship, as indicated in the previous paragraph. We conclude that for monthly averages the instruments compare well. Over the whole comparison period there is an average offset of 0.65 ± 0.28 5 µg m -3 and a slope of 1.034 ± 0.028 between the techniques. Thus, the miniDOAS measures on average slightly higher than the AMOR, over all concentration ranges.
From a scientific point of view this correspondence is excellent, especially since two completely different measurement techniques are used. As a reference, we refer to a study by Von Bobrutzki et al. (2010) that shows much larger discrepancies between different techniques. The systematic difference found between AMOR and miniDOAS amounts to roughly 10% of 10 the typical ammonia background concentrations in the Netherlands of around 5 µg m -3 . Fortunately, not so much the absolute difference between the techniques is politically relevant but any jumps in ammonia trends, see e.g. Wichink Kruit et al., 2017) for two studies in which this data is being used. We will discuss some possible explanations for the difference between the techniques.

Possible difference due to height difference
The effect of the difference in measurement altitude (the AMOR measured at 3.5 m, the miniDOAS at 2.2 m) was studied using passive samplers.
The results reported in this paper show no significant difference between AMOR inlet and miniDOAS path, so they offer no explanation for the observed positive bias between miniDOAS and AMOR. The results do show a difference between both 20 measurement points at 3.5 m, indicating that the AMOR measurement may be influenced by the station housing resulting in lower measured values. In both cases the statistical error is substantial, and consequently the results are inconclusive. Further research with more precise equipment would be needed to reduce the statistical error and study the effects of the altitude difference between 2.2 and 3.5 m, and also the possible influence of the station housing. be from a scientific point of view) would however improve the comparison only slightly. Therefore, the validation procedure can be ruled out as a major source of the offset.

Possible difference due to ammonia loss in the AMOR inlet line
It is conceivable that ammonia is lost in the AMOR air inlet system, as this is a known effect in ammonia inlet lines (Yokelson et al., 2003). However, the AMOR air inlet system has been designed to minimise such effects. Especially the 5 relatively high airflow through the instrument, of 25 L min -1 rather than the mL min -1 flows found in other instruments, should be effective in minimising these effects. As discussed in Sect. 3, no indication for ammonia loss was found in the measurement data. It seems therefore unlikely that ammonia loss is a major contributor to the bias found.

Possible difference due to AMOR calibration procedure
It should be noted that AMOR calibrations are performed using calibration fluids, and thus only pertain to the 'liquid' part of 10 the instrument, after ammonia has been absorbed in the denuder. Any losses in the airborne phase, e.g. in the inlet system, are not included in the calibration procedure. As stated previously, the reason for omitting this part in the calibration procedure is that it is virtually impossible to generate an adequate calibrated gas flow, as the AMOR tries to minimize inlet effects by using a very high airflow of 250 m 3 h -1 , from which a further 25 L min -1 is sampled by the instrument. We have not been able to study this aspect further in the framework of this comparison. 15

Possible difference due to miniDOAS zero calibration
The miniDOAS zero is determined by comparison to a DOAS reference instrument. Any offset in the reference instrument will show up as a similar offset in the reported miniDOAS values. The zero of the reference instrument is determined by study of a long time series, looking for periods of lowest values and assuming these occur at constant zero ammonia levels.
If this assumption is incorrect, it results in the reference instrument underestimating the real concentrations. This would 20 therefore lead to a negative bias in the concentrations reported by the miniDOAS, never to a positive one. There is no evidence for this in the dataset.

Intercomparison at shorter timescales
On a timescale of minutes or even hours, the instruments do not compare well. This is caused by a distinct difference in temporal resolution: the typical integration time of the miniDOAS is 1 minute, its minute-measurements are delay-free and 25 mutually independent. The AMOR has a much larger response time, despite its claimed temporal resolution of 3 minutes. Its response to abrupt changes shows a delay (order of 30-60 minutes) and a spread-out and flattening of short peaks. In general, the integral over time of the AMOR-observed ammonia seems to remain conserved, as is reflected by the good comparison of longer timescale averages discussed above. This means that (virtually) no ammonia is lost in the AMOR, but it will be recorded at a different moment in time than its actual appearance at the AMOR inlet. 30 Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License.
On timescales of hours, e.g. when looking at daily cycles, we consider the miniDOAS concentrations to be more representative for the actual ambient ammonia concentrations than the AMOR measurements.

Conclusions
The Dutch National Air Quality Monitoring Network has been monitoring ambient ammonia concentrations since 1992, using automated AMOR instruments. Over a period of 22 years, an hourly dataset was obtained at eight stations in the 5 Netherlands. In 2014 the number of stations was reduced to six. On 1 January 2016, six miniDOAS instruments have replaced the AMOR instruments. The DOAS technique is an open-path remote sensing technique that does not require bringing ammonia inside an instrument. This technique avoids all adverse effects typical for most commercial ammonia measurements: adsorption to tubing, filters and instrument interior, and interference from aerosols generating ammonia. In addition, a substantial reduction of operating costs is obtained. 10 Prior to the transition, both instruments ran in parallel at six stations for a period of 16 months. The comparison during this period shows that both instruments have a similar uptime, obtaining 80 to 90% of the possible hourly values. This is adequate for network operations.
The introduction of the miniDOAS in the Dutch Air Quality Monitoring Network results in a substantial reduction of the instrument response time and thus in a gain in temporal resolution. Consequently, miniDOAS minute values and hourly 15 values will be more representative for ambient ammonia concentrations. The resulting dataset will be better suited for the study of daily cycles and processes than the dataset based on AMOR data. Compass analysis, i.e. sorting concentration data by episodes of a single wind direction to investigate in which direction ammonia sources are located, also becomes possible with the high temporal resolution of the miniDOAS.
Daily-averaged and especially monthly-averaged values of both instruments compare well. The miniDOAS dataset shows a 20 small positive offset of 1.0 ± 0.6 µg m -3 to the AMOR dataset. The origin of this offset is presently unknown. As a potential cause for this offset, we cannot rule out possible losses in the AMOR inlet, as this part of the instrument is not included in the calibration process. Other possibilities are: the height difference between AMOR inlet and DOAS path in combination with a deposition gradient, and the possible influence of the monitoring station housing on ammonia concentrations at the AMOR inlet. 25

Outlook
Over recent years, we received several requests to make a miniDOAS available to other parties. So far, we worked together with Agroscope in Switzerland (Sintermann et al., 2016) and the Flanders Environment Agency (VMM) in Belgium. As a result, some 10 miniDOAS systems are currently operational in these countries. We are now exploring the possibilities to make the miniDOAS instrument available as a commercial instrument, through collaboration with one or more partners. 30 Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License.
To further improve the calibration, especially the zero measurements (Sect. 2.3.4), we intend to construct a laboratory facility with zero concentration over the full path length of the instrument.
We anticipate being able to measure ammonia deposition with miniDOAS soon. Using the gradient technique, we aim for hourly deposition measurements. This development will be subject of a forthcoming paper.

7
Data availability 5 The full dataset of hourly data, daily averaged data and monthly averaged data is provided in a supplement to this paper. Interactive 360° views of the monitoring stations may be found on the following website:

Implementation of the miniDOAS in the monitoring network
General remarks: at all sites, the optical path is at about 2.20 m above the ground. The ground is level at every site, so the path stays at this height over its entire length. Path lengths are given as 2 x the distance between the miniDOAS and the 15 retroreflector. Path directions are given in degrees east of north, i.e. 270° is due west. Fro each station, its ID that is used in national and international databases is given, its place name and street name. Remarks: In November 2014, this station was re-sited, it was moved 110 m to the west. Also, the optical path was turned 90°. As a consequence, in March and October, the instrument looks straight into the rising sun. To avoid damage to the sensitive optics, the instrument is switched off and its optics are covered during those months. Installation of an automatic shutter that only covers the instrument when the sun is actually in the field of view is envisaged. Remarks: Because of a restricted site, the optical path is shorter than on other locations.

Appendix B: AMOR data offset 30
As indicated in the main text, the AMOR data were corrected for a small offset. This offset is stable in time but varies between stations. It is produced in the digital to analogue converter (DAC) that transforms the digital AMOR signal to an Atmos. Meas. Tech. Discuss., doi:10.5194/amt-2016-348, 2017 Manuscript under review for journal Atmos. Meas. Tech. Discussion started: 3 March 2017 c Author(s) 2017. CC-BY 3.0 License. analogue signal which in turn is digitised by the data acquisition system of the station. Table 8 shows the resulting offsets per station. AMOR data used in this study have been corrected for these offsets, in contrast to the original AMOR dataset that is now present in national and international databases. A correction of the data in the official database is planned and will be documented in a separate publication.          Mast 3.5 m -Mast 2.2 m 0.5 ± 0.3 (n = 50) Possible gradient due to deposition

Author contribution
Mast 3.5 m -AMOR inlet (3.5 m) 0.5 ± 0.6 (n = 47) Expected to be zero, unless e.g. influence of station housing on AMOR measurement