Measurements and quality control of ammonia eddy covariance fluxes: A new strategy for high frequency attenuation correction

Abstract. Measurements of the surface-atmosphere exchange of ammonia (NH3) are necessary to study the emission and deposition processes of NH3 from managed and natural ecosystems. The eddy covariance technique, which is the most direct method for trace gas exchange measurements at the ecosystem level, requires trace gas detection at fast sample frequency and high precision. In the past, the major limitation for measuring NH3 eddy covariance fluxes has been the slow time response of NH3 measurements due to NH3 adsorption on instrument surfaces. While high frequency attenuation correction methods are used, large uncertainties in these corrections still exist mainly due to the lack of understanding of the processes that govern the time response. We measured NH3 fluxes over a corn crop field using a quantum cascade laser spectrometer (QCL) that enables measurements of NH3 at a 10 Hz measurement frequency. The 5 month measurement period covered a large range of environmental conditions that included both periods of NH3 emission and deposition and allowed us to investigate the time response controlling parameters in field conditions. Without high frequency loss correction, the median daytime NH3 flux was 8.59 ng m−2 s−1 during emission and −19.87 ng m−2 s−1 during deposition periods, with a median daytime random flux error of 1.61 ng m−2 s−1. The overall median flux detection limit was 2.15 ng m−2 s−1, leading to only 11.6 % of valid flux data below the detection limit. From the flux attenuation analysis we determined a median flux loss of 17 % using the ogive method. No correlations of the flux loss with environmental or analyser parameters, such as humidity or inlet aging were found, attributed to the uncertainties in the ogive method. Therefore, we propose a new method that simulates the flux loss by using the analyser time response that is determined frequently over the course of the measurement campaign. A correction that uses as a function of the horizontal wind speed and the time response is formulated which accounts for surface aging and contamination over the course of the experiment. Using this method, the median flux loss was calculated to be 34 %, substantially higher than with the ogive method.


deposition processes of NH3 from managed and natural ecosystems. The eddy covariance technique, which is the most direct method for trace gas exchange measurements at the ecosystem level, requires trace gas detection at fast sample frequency and high precision. In the past, the major limitation for measuring NH3 eddy covariance fluxes has been the slow time response of NH3 measurements due to NH3 adsorption on instrument surfaces. While high frequency attenuation correction methods are used, large uncertainties in these corrections still exist mainly due to the lack of understanding of the processes that govern the 15 time response. We measured NH3 fluxes over a corn crop field using a quantum cascade laser spectrometer (QCL) that enables measurements of NH3 at a 10 Hz measurement frequency. The 5 month measurement period covered a large range of environmental conditions that included both periods of NH3 emission and deposition and allowed us to investigate the time response controlling parameters in field conditions. Without high frequency loss correction, the median daytime NH3 flux was 8.59 ng m -2 s -1 during emission and -19.87 ng m -2 s -1 during deposition periods, with a median daytime random flux error of 20 1.61 ng m -2 s -1 . The overall median flux detection limit was 2.15 ng m -2 s -1 , leading to only 11.6 % of valid flux data below the detection limit. From the flux attenuation analysis we determined a median flux loss of 17 % using the ogive method. No correlations of the flux loss with environmental or analyser parameters, such as humidity or inlet aging were found, attributed to the uncertainties in the ogive method. Therefore, we propose a new method that simulates the flux loss by using the analyser time response that is determined frequently over the course of the measurement campaign. A correction that uses as a function 25

Introduction
Knowledge of ammonia (NH3) exchange processes between ecosystems and the atmosphere is essential for improving our understanding of its impact on air quality, global warming, and ecosystem health. As the most abundant base in the atmosphere, NH3 is responsible for the formation of ammonium aerosol, impairing air quality and affecting climate. For example, it is estimated that in Europe 34 % of fine particulate matter (PM) is directly linked to emission of ammonia that generates 5 secondary PM (Pozzer et al., 2017). Furthermore, deposition of ammonia has been shown to strongly impact low-nitrogen ecosystems thereby reducing biodiversity (Erisman et al., 2013).
Major sources of NH3 to the atmosphere are emissions from agriculture, with globally 24 Tg yr -1 being emitted from livestock and livestock waste and 9.4 Tg yr -1 from synthetic fertilizer applications on agricultural fields (Paulot et al., 2014). The understanding and quantification of these emission sources is critical to propose effective NH3 control and mitigation strategies. 10 However, to date large uncertainties on the magnitude of NH3 emissions exist. This is in part due to difficulties and challenges in measuring NH3 fluxes, especially in environments where NH3 fluxes are small and/or frequent changes between net emission and deposition are expected.
To measure NH3 exchange at the ecosystem level, micrometeorological methods including gradient, relaxed eddy accumulation (REA) and the eddy covariance (EC) have been used in the past. Which method is used largely depends on the 15 expected flux magnitude, the site layout and the available instrumentation. Systems based on the capture of ammonia by wet denuders or the thermal conversion to nitric oxide (NO) have been employed for flux gradient (e.g., Famulari et al., 2010;Flechard & Fowler, 1998;Sutton et al., 2000;Walker et al., 2006;Wolff et al., 2010) and REA-based flux measurements (Baum and Ham, 2009;Hansen et al., 2013Hansen et al., , 2015Hensen et al., 2009;Myles et al., 2007;Nelson et al., 2017;Zhu et al., 2000). The success of both methods often relies on the precise and accurate determination of a concentration difference 20 measured over a time period of typically 30 min up to several hours. Especially under conditions where the measured vertical concentration gradient is small, like for flux measurements above forests, the REA method is the much stronger approach. The flux gradient measurement heights and the size of the REA deadband largely impact the concentration difference, and their choice has to be considered carefully in the flux measurement setup and operation (Moravek et al., 2014).
The eddy covariance method is the most direct method to quantify ecosystem scale turbulent fluxes as it relies on the covariance 25 between the near ground turbulence and the scalar of interest. For this, the scalar has to be measured at a fast time response (≤ 0.1 s) with concurrent adequately high precision. In the last two decades, the development and improvement of mass spectrometry and infrared spectroscopy techniques for fast measurements of NH3 provided the opportunity to measure NH3 fluxes using eddy covariance. Shaw et al. (1998) presented the first eddy covariance measurements of NH3 using a tandem mass spectrometer and Sintermann et al. (2011) employed a proton transfer reaction mass spectrometer to measure NH3 30 emissions after slurry application. For laser-based eddy covariance measurements, Famulari et al. (2004) and Whitehead et al. (2008) used diode laser absorption spectroscopy using cryogenic cooled lasers. A greater laser stability and output power is given by quantum cascade lasers (QCL) which can be Peltier-cooled (McManus et al., 2010). Ferrara et al. (2012Ferrara et al. ( , 2016 and https://doi. org/10.5194/amt-2019-193 Preprint. Discussion started: 18 June 2019 c Author(s) 2019. CC BY 4.0 License. Whitehead et al. (2008) used pulsed QCL spectrometers for NH3 eddy covariance measurements on agricultural sites, while Zöll et al. (2016) employed a more sensitive continuous-wave QCL instrument to measure NH3 fluxes above a peatland.
The common limitation of the previous studies using fast-response NH3 analysers is the instrument time response due to the surface adsorption of NH3. It is well-known that adsorption and desorption of NH3 to and from surfaces, including inlet tubing, can be significant. This effect slows the time response of the measuring system leading to high frequency attenuation (HFA) 5 of the measured NH3 time series, which affects the NH3 flux measurements by eddy covariance. To avoid these effects, Sun et al. (2015) employed a custom-built open-path QCL (Miller et al., 2014) to measure eddy covariance fluxes above a cattle feedlot. While open-path QCL systems have the advantage to avoid HFA effects, they may introduce flow distortion due to their size when placed close to the sonic anemometer and require frequent cleaning of the exposed cell mirrors (Sun et al., 2015). Furthermore, to our knowledge they are not commercially available yet. As closed-path systems will still play a an 10 important role in the future, our study focuses on the performance and quality control of closed-path eddy covariance systems.
The magnitude of flux loss due to HFA in closed-path systems is highly variable depending on the instrumental setup and meteorological conditions. For past NH3 eddy covariance field measurements, the estimated flux loss ranged between 20 and 50 % (Ferrara et al., 2012(Ferrara et al., , 2016Sintermann et al., 2011;Whitehead et al., 2008;Zöll et al., 2016). Although the flux loss can be corrected for in post-processing with spectral correction techniques, the applied correction factor can vary significantly 15 depending on the chosen correction method (Ferrara et al., 2012). A lack of understanding of the factors impacting the time response of NH3 eddy covariance systems is responsible for this uncertainty.
From previous studies it is known that adsorption and desorption of NH3 is governed by the surface area and material of the inlets and internal instrument components, with stainless steel showing significantly slower time responses than polyethylene (PE) or polytetrafluoroethylene (PTFE) (Whitehead et al., 2008). Ellis et al. (2010) showed that heating of their 20 perfluoroalkoxy (PFA) inlet tubing to 40 °C for NH3 mixing ratios above 30 ppbv reduces the HFA significantly, while Sintermann et al. (2011) found that heating their drift tube inlet to 180 °C enabled a time resolution high enough for eddy covariance measurements. It is suspected that heating removes -at least partially -liquid and molecular water layers on the surface, which decreases the adsorption sites for the polar NH3 molecule (Sintermann et al., 2011), although NH3 can also interact directly with the surface material. Roscioli et al. (2015) showed that using active passivation by continuously adding 25 a fluorinated amine into the sample gas improves the time response significantly, as the polar amine group of the molecule occupies potential NH3 adsorption sites and NH3 does not react with its non-polar fluoro-chain. Still, there is a lack of comprehensive mechanistic understanding of NH3 sorption on surfaces (Sintermann et al., 2011), which is needed to reduce the uncertainties of the HFA correction for NH3 eddy covariance fluxes. There is evidence that adsorption and desorption processes act at different rates (Whitehead et al., 2008), which would skew the high frequency NH3 distribution and may 30 impact the flux covariance calculation. While Ellis et al. (2010) found the time response to degrade with the relative humidity of ambient air, the potential effect on NH3 fluxes is not accounted for in the currently used HFA flux correction methods (Ferrara et al., 2012;Zöll et al., 2016). Evidence that the time response is improved when the NH3 mixing ratio changes are larger (Ellis et al., 2010), can be interpreted either such that fluxes with higher magnitudes need to be corrected less than small https://doi. org/10.5194/amt-2019-193 Preprint. Discussion started: 18 June 2019 c Author(s) 2019. CC BY 4.0 License. fluxes or that fluxes at higher ambient concentrations are less attenuated due to a higher passivation of the surface. Finally, the time response effect of surface aging and surface deposition of particulate matter is poorly understood and accounted for in HFA correction methods (Roscioli et al., 2015;Sintermann et al., 2011;Whitehead et al., 2008). As our understanding of NH3 time response in changing environmental and instrumental conditions is limited, the analysis of flux datasets under a wider range of conditions than previously sampled are needed reduce the uncertainties in the HFA correction of NH3 fluxes. 5 In this study, we employed a NH3 eddy covariance system over an entire growing season from May to October 2017 at a corn field in Eastern Canada. The system used a closed-path continuous wave QCL spectrometer with a 5.5 m heated PFA inlet line. The objectives were to (1) limit adsorption/desorption of ammonia in the inlet of the QCL; (2) quantify its impact on the systems time response under a large range of environmental and instrumental conditions in order to obtain a deeper understanding of the processes that govern the time response and of how this is ultimately applied to the NH3 flux correction. 10 This includes, for example, the examination of the relationship between time response and humidity or the flux magnitude.
Due to the 5 month measurement period, we are able to examine the effect of inlet aging and the benefit of cleaning procedures on the NH3 flux measurement. Based on our findings, we present an approach to correct NH3 fluxes which uses our improved understanding of NH3 time response. The approach may be also used for flux correction of other species that show a strong surface adsorption, such as nitric or organic acids. 15 Next to the issue of HFA, NH3 measurement systems need to resolve small NH3 mixing ratios fluctuations at high time resolution. Especially under low flux conditions, a precise and stable operation of the NH3 measurement system is required.
For this reason, the paper also discusses the precision and flux detection limit of the QCL spectrometer depending on environmental and operational conditions. Currently, continuous wave QCL spectrometers are the most precise high time resolution NH3 measurement systems available; however, their operation under field conditions requires careful setup and 20 regular maintenance. With our findings we provide details on the setup and operation of the QCL which are helpful for investigators that aim to use it for eddy covariance NH3 flux measurements in the future.

NH3 detection with QCL 25
A Quantum Cascade Tunable Infrared Laser Differential Absorption Spectrometer (QC-TILDAS, Aerodyne Research Inc., USA) was used to measure NH3 mixing ratio at 10 Hz sampling frequency for eddy covariance flux measurements. The QC-TILDAS (referred to as QCL hereafter) retrieves the NH3 absorption spectrum at 967.3 cm −1 using a thermoelectrically cooled continuous wave quantum cascade laser (Alpes Lasers, Switzerland), which is scanned across the full NH3 transition within the spectral window. A continuous wave laser has an increased power output over a pulsed laser, which was used in the version 30 of the QCL described in detail by Ellis et al. (2010), and is therefore more suitable for measurements of high precision needed for the eddy covariance method. As illustrated in Fig. 1, the laser beam is directed into an astigmatic Herriot multiple pass absorption cell (0.5 L, 76 m effective pass length) coated with a hydrophobic coating to reduce the interaction of NH3 with cell walls. To minimize linewidth broadening of the absorption peak, the pressure in the absorption cell is kept at approximately 4.67 kPa. A reference cell containing ethylene (C2H4), a less surface reactive gas that contains an absorption line near that of NH3, is used for absorption-line lock. The signal and reference paths are focused on the same thermoelectrically cooled 5 Mercury Cadmium Telluride (HgCdTe) infrared detector (Vigo Systems, Poland).
The laser control, spectral retrieval and mixing ratio calculations are managed by the software package TDLWintel (Aerodyne Research Inc., USA) described in Nelson et al. (2004). The measured NH3 spectrum is fit at the 10 Hz sample frequency by convolving the laser line shape with a calculated absorption line shape based on the HITRAN (high-resolution transmission) molecular absorption database and the measured pressure, temperature and path length of the optical cell (Herndon et al., 10 2007). The software allows for automatic user-defined additions of zero air through the use of a solenoid valve.
Variations in pressure, temperature and other disturbances may significantly impact the instrument performance by influencing the absorption spectrum fringe pattern. Fringes are structures in the absorption spectrum which are caused by optical interferences within the laser beam path and can be responsible for signal drift if their pattern changes over time. Especially for species that are typically present in the atmosphere in the lower ppbv or pptv mixing ratio range, such as NH3, the impact 15 of fringes to the absorption peak range can be significant. For that reason, the operation of the QCL requires a stable environment to house the QCL and frequent background measurements with zero air to account for potential drifts in the background spectrum.

Setup and operation of NH3 flux measurements
Eddy covariance flux measurements of NH3 were performed from 28 May to 23 October 2017 on an agricultural corn field, 20 equipped with twin flux towers near Ottawa in Eastern Canada (see Fig. 7, Pattey et al., 2006). The experimental site is located on the premises of the Canadian Food Inspection Agency (CFIA) and managed by Agriculture and Agri-Food Canada (AAFC).
Prior to the measurements, the agricultural field was tilled and fertilized on 25 May using granular urea fertilizer (155 kg N ha -1 ). The corn crop was seeded on 28 May. The QCL was installed on the west eddy covariance flux tower, which had a fetch of 200 to 500 m depending on the wind direction. 25 To measure the 3D wind vector for the covariance calculation, a CSAT-3 (Campbell Scientific Inc, USA) sonic anemometer was installed on the tower at 2.5 m above ground level (a.g.l.). To accommodate the growing corn canopy, the tower was raised to a measurement height of 4.5 m a.g.l. on 6 July. Water vapour (H2O) and carbon dioxide (CO2) were measured using a closed-path infrared gas analyser (LI-7000, LI-COR, USA). The AAFC in-house data acquisition and control system, called "REAsampl" (Pattey et al., 1996), was used to record at 20 Hz the analogue channels of the various instruments. The 30 REAsampl software adjusts the predetermined lags between the various close-path analysers and the vertical wind velocity, and rotate the horizontally symmetrical sonic anemometer head in the mean horizontal wind direction every hour when the hourly mean horizontal wind velocity is greater than 1.5 m s -1 . During the few seconds of the rotation for aligning the anemometer, the raw data are not recorded. By aligning the anemometer, flow distortion and lateral loss of covariance are minimized.
As shown in Fig. 1 the setup of the QCL consisted of five major parts: (1) the inlet system, (2) the QCL and chiller unit, (3) the valve and heating control box as well as enclosure housing (4) the vacuum pump and (5) the zero air source. The inlet was mounted at mid-vertical distance of the sonic anemometer open-path and 25 cm behind the anemometer open-path to minimize 5 flow distortion on the wind velocity measurements by the inlet and lateral loss of covariance. The QCL uses a 10 cm quartz inlet which acts as virtual impactor to remove particulate matter from the sample air. As described in more detail in Ellis et al. (2010), about 90 % of the sample air makes a sharp turn (Fig. 1) and is pulled through the absorption cell, while 10 % of the flow, including particles larger than 300 nm due to their higher inertia, is pulled directly to the vacuum pump (TriScroll 600, Agilent, USA). To limit the condensation of water and the interaction of NH3 with inlet surfaces, the glass inlet is internally 10 coated with a hydrophobic fluorinated silane coating and was heated constantly to 40 °C. The glass inlet acts as a critical orifice which regulates the volume flow at the inlet. From 28 May to 27 July a glass inlet with a flow rate of 15.4 L min -1 and after that period a glass inlet with a flow rate of 13.4 L min -1 , which had a newly applied silane coating, was used. For closedpath eddy covariance measurements, a high volume flow rate is essential to keep a plug flow in the inlet system, in order to minimize HFA in the inlet system. Depending on the flow rate and the actual sample gas temperature, the Reynolds number 15 ranged between 3000 and 3700, indicating mainly transitional flow conditions with turbulent flow in the centre and laminar flow near the tubing walls. The glass inlet was connected to the QCL via a 5.5 m long 3/8" PFA sample tube, which was insulated and controlled to 40 °C. While raising the tower on 6 July, the QCL was also raised by 2 m using wood pallets to keep a constant 5.5 m sample tube length.
The QCL, located at the bottom of the flux tower, was housed in an insulated aluminium enclosure that was equipped with 20 two Peltier coolers to precisely control the internal temperature of the box to 28 (± 0.2) °C. Next to the QCL, the enclosure also housed the chiller (Oasis Three, Solid State Cooling Systems, USA) that is required for stable temperature control of the infrared laser and the optical and electronic parts of the QCL. To prevent the build-up of heat inside the enclosure, the intake and outtake vents of the chiller were connected to the outside of the enclosure. A dehumidifier was built in to prevent condensation inside the box. 25 The time series of the 10 Hz NH3 mixing ratios and analyser parameters were digitally recorded on the QCL computer. For precise time synchronization between the post-processed digital NH3 mixing ratios and the vertical wind velocity, an analogue signal of the NH3 mixing ratio was recorded through REAsampl. A remote monitor was placed in a nearby trailer for regular checks and other operations like data transfer and manual valve switching, to avoid opening the temperature-controlled enclosure. Data were collected from the QCL computer regularly (every 2-4 days) and plotted for routine quality checks. To 30 ensure optimal operation of the instrument of the QCL over the 5 month measurement period, the status of the QCL was checked regularly through remote access using a mobile hotspot.
The QCL glass inlet includes a port for the introduction of zero air for frequent background measurements, designed so that these flows encounter the inlet in the same way as ambient air (Fig. 1). For zero air, a heating catalyst (Aadco Instruments, USA) was used, which scrubs NH3 from ambient air by catalytic thermal conversion at 300 °C using palladium beads. The automatic background schedule was set to flush the inlet with zero air for 5 min at the end of every 30 min period from 28 May to 16 July. Due to an operation failure of the heating catalyst, on 17 July the zero air source was replaced by ultra high purity (UHP) compressed zero air (Praxair Canada Inc., Canada). To minimize the UHP compressed air consumption, the automated background interval was set to 1 hour with a reduced duration of 3 min from 17 July to 27 July. From 28 July to 31 5 August, the interval was set to 3 hours without significantly compromising on the data quality. In the final phase of measurement period, from 15 September to 23 October the background interval was set to 2 hours. The zero air gas flow was introduced into the inlet by triggering a solenoid valve installed in the valve and heating control box.
To test the effect of surface aging on the time response and to reduce the interaction of NH3 with surfaces, regular cleaning of the glass inlet, inlet line and the absorption cell was performed using a procedure of rinsing with deionized water and ethanol, 10 while the tubing was heated to about 80 °C during the cleaning process. The cleaning of the glass inlet was performed on 22 June, 27 July and 12 September. The inlet tubing was cleaned on 27 June, 27 July and 12 September. Along with cleaning of the glass inlet and the inlet tubing, the inner surface of the absorption cell was cleaned on 12 September.

Eddy covariance flux calculation 15
The processing of the NH3 data leading to the final calculated NH3 fluxes consisted of four major steps: (1) processing and quality control (QC) of the digital NH3 mixing ratio data, (2) time synchronization between the quality-controlled digital NH3 data and vertical wind velocity data recorded through REAsampl, (3) flux calculation and (4) flux random error calculation.
The processing of QCL NH3 data as well as all other processing was performed using the R software package (R Core Team, 2017). 20 The NH3 mixing ratio time series were first scanned for periods of instrument failures and maintenance, which were removed.
Spike detection and removal was conducted by using a running-mean low-pass filter. Spikes were identified best as data points that exceeded 3.5 times the standard deviation of a 21 point averaging window. To correct for a potential drift of the QCL between two automated background periods (varying from 30 min to 3 hr), the background mixing ratios were linearly interpolated between two consecutive background measurements and subtracted from the NH3 mixing ratios. Following 25 analogue acquisition using REAsampl, the 20-Hz CSAT-3 sonic anemometer and uncorrected ammonia data were extracted from the REAsampl raw data binary files, in which the data associated with tower rotation were already removed.
The time synchronization between the NH3 mixing ratios and vertical wind speed, which is essential for the eddy covariance flux calculation, was performed in two steps: (1) Time synchronization between the digital NH3 data and the REAsampl data and (2) the time synchronization between the digital NH3 data and the vertical wind velocity. For the former, a circular cross-30 correlation was performed between the digital and analogue NH3 signals. Accounting for the analogue output delay, the time lag between both systems was then determined as the position of the maximum correlation. In the second step, a circular crosscorrelation (using a ±5 s window) between the time synchronized sonic anemometer data and the digital NH3 data was used to https://doi.org/10.5194/amt-2019-193 Preprint. Discussion started: 18 June 2019 c Author(s) 2019. CC BY 4.0 License. account for delays caused by the inlet system and the horizontal displacement of the CSAT-3 and the glass inlet. As the crosscorrelation method between the vertical wind velocity and a scalar only works well when turbulent fluxes are large enough, a quality assessment was performed on the results of the cross-correlation. Only lag times which were less than ±2.5 s and had a cross-correlation value greater than 0.05 were used. Missing lag times were then replaced by the last previous valid lag time.
To detect further outliers, lag times that were offset by more than ±1.5 s were set to the preceding lag time if the difference 5 between the preceding and successive lag time was less than 0.5 s, indicating a spike and not a real shift in the lag time.
Background on the final eddy covariance flux calculation and the required correction methods is well-documented in literature Pattey et al., 2006). In brief, NH3 fluxes are calculated by the covariance of the NH3 mixing ratio ( 3 ) and the vertical wind velocity ( ) multiplied by the molar density of air ( ) as: where 3 ′ and ′ denote the fluctuations of NH3 mixing ratio and the vertical wind velocity from their 30 min mean value, respectively. The NH3 fluxes presented in this study are given in ng-NH3 m -2 s -1 . Prior to the eddy covariance flux calculation, the 3D wind vector coordinate is typically rotated to ensure zero vertical wind velocity over the averaging period (Finnigan et al., 2003;Wilczak et al., 2001). Due to the tower rotation mechanism used in this study, the wind vector was already rotated into the mean wind direction. Variations in the air density caused by temperature and air moisture fluctuations may impact the 15 eddy covariance flux and are typically corrected for by the WPL correction (Webb et al., 1980). The sensible heat flux induced fluctuations of ambient air temperature are expected to be efficiently damped by the heat exchange in the inlet system and the 5 m long heated inlet line. An effect of air moisture fluctuations caused by the latent heat flux on the NH3 flux is possible, however, it was found that the effect on NH3 fluxes is negligible (≤ 1 %) due to the relatively low concentrations of NH3 in ambient air (Ferrara et al., 2016;Pattey et al., 1992). For these reasons, the WPL correction was not applied for the NH3 fluxes 20 in this study. High frequency loss corrections, like for the flux loss due to the distance between the CSAT-3 and the QCL glass inlet (Moore, 1986), were not applied to the initially calculated NH3 fluxes as they were incorporated as part of the high frequency loss analysis discussed later in this paper.
The TK3 software package (Mauder & Foken, 2011) was used to calculate fluxes of momentum, sensible heat and latent heat and NH3 flux quality parameters. The quality flag scheme of Foken & Wichura (1996) was used to filter for periods of low 25 stationarity and low developed turbulence. Furthermore, the TK3 program derives the random flux errors of the NH3 flux. The random errors include (1) the errors due to stochastic nature of turbulence and (2) the random errors due to instrumental noise (Mauder et al., 2013). The former is calculated in TK3 after the method of Finkelstein & Sims (2001), which calculates the variance of the covariance function as a combination of the auto-covariance and cross-covariance terms with changing the lag time. The random flux error due to instrumental noise is calculated in TK3 by extrapolating the auto-correlation function of 30 the NH3 time series towards a zero lag time (Mauder et al., 2013). As the random error calculation in TK3 was not successful for all 30 min periods, we additionally determined the instrumental noise error ( ) by using the variance of the zero air source measurements (conducted every 30 min to 3 hr throughout the experiment) as the variance of the NH3 mixing ratio and by implementing that in the instrumental noise function used in Mauder et al. (2013). To be comparable to other NH3 flux studies, we applied in addition the approach used by Sintermann et al. (2011), where the random flux error is determined for each 30 min period by the standard deviation of the covariance function ( ℎ ) when using a time lag ranging between -120 and -70 s and +70 and +120 s. Using this approach, we defined the flux detection limit as 2• ℎ .

Time response determination of NH3 measurements 5
A fast time response of the NH3 measuring system is essential for performing eddy covariance measurements. To understand the processes that impact the adsorption and desorption of NH3 to the measurement system, the knowledge of the system's NH3 time response is important. The time response of the QCL NH3 measurements is determined mainly by two processes (Whitehead et al., 2008): (1) the exchange of the sample air volume in the inlet line and the sample cell and (2) the adsorption and desorption of NH3 at the inlet and sample cell walls. As a result, the time response can be described by a double exponential 10 function giving two time constants, τ1 and τ2, representing the time response towards the exchange of the sample air volume and wall interactions, respectively: where t0 is the start time and y0 the offset from zero (Fig. 2); A1 and A2 are proportionality coefficients which account for the contribution of each process to the overall time response. Accordingly, the percentage contribution of the wall interaction 15 processes can be described as (Ellis et al., 2010): Together with the time constants τ1 and τ2, can be used to evaluate the performance of the QCL system with respect to its time response. To determine the time response for our instrument setup, the double exponential function (Eq. 2) was fitted to the step change in NH3 mixing ratios when switching from ambient air to zero air measurements as part of the automated 20 background correction. Those fits were performed for each background period in order to obtain the temporal variation of the time response over the course of the entire measurement period.

2.4
Analysis of flux loss due to high frequency attenuation

Ogive method
The attenuation of the high frequency scalar time series due to a slow response leads to an underestimation of the calculated 25 eddy covariance flux. Both theoretical and experimental approaches are used to quantify the flux loss and ultimately correct for it; a summary is given in Foken et al. (2012). For NH3 the HFA needs to be determined experimentally, since to date no adequate description of the surface adsorption and desorption processes exists. Experimental approaches typically compare the co-spectrum of the vertical wind velocity and the attenuated scalar time series to the co-spectrum of a non-attenuated reference flux. As a reference flux the sensible heat flux is most often used, either from direct measurements or from parameterization available in literature (Kaimal and Finnigan, 1994). In this study we used the ogive method described in Ammann et al. (2006). The ogive of a scalar flux ( ) is calculated by the cumulative integral of the co-spectrum ( ) as over the observed frequency range, beginning with the lowest frequency (1/ , where is the averaging interval). If the ogive is normalized by the covariance, the ogive value at the highest frequency is 1. The flux loss is derived by scaling the normalized 5 ogive of the scalar flux to the normalized ogive of the sensible heat flux at a given limit frequency ( 0 ). The scalar ogive value at the highest frequency represents then the flux attenuation factor (between 0 and 1). The limit frequency is defined as the highest frequency at which no HFA occurs and can be determined from the comparing the scalar co-spectrum to the sensible heat flux co-spectrum.
Ideally, a reference scalar is used which shows the highest scalar similarity to the investigated flux. The water vapour flux is 10 expected to have a similar sink and source distribution than NH3, but the measured flux may experience high-frequency loss if measured with a closed-path analyser. Therefore, we used in this study the measured non-attenuated sensible heat flux as a reference.

Time response transfer function method
Another approach to correct for the HFA is to simulate the flux loss by knowing the flux loss transfer function. The transfer 15 function can be indirectly determined through the system's time response, which was determined over the entire measurement period as described in Sect. 2.3. The method applies a low-pass filter, which represents the HFA of the system, to the nonattenuated scalar time series and then compares the filtered and non-filtered co-spectra. As for the ogive method, we used the sensible heat flux as a reference flux. A low-pass filter can be characterized by the cut-off frequency ( ) using the time constant τ as 20 which is then used to define a transfer function T as The transfer function describes the low-pass filtering effect over the frequency range f and can be used to apply the HFA to the co-spectrum of the scalar and the vertical wind velocity. 25 In the case of the double exponential function as for NH3, which produces two time constants, we suggest using a transfer function combining both τ1 and τ2 by taking into account their contributions to the overall time response. The combined transfer function is then where 1 ( ) is the transfer function describing the exchange of sample air volume using τ1 in Eq. (5) then be determined by the ratio of the transfer function filtered co-spectrum and the non-filtered co-spectrum of the sensible heat flux: To investigate the possible NH3 flux loss over the measurement period, the flux loss was simulated for all 30 min periods, using τ1, τ2 and values that represent the range of observed time responses. 5

Measured NH3 fluxes, random flux error and flux detection limit
The QCL eddy covariance measuring system was operated over a period of 149 days between May and October 2017, covering a total of 7132 flux measurement periods. Times of system maintenance, quality control checks, and other system downtimes were discarded from the dataset, resulting in a flux data coverage of around 85 % (Table 1). 10 Figure 3 shows the flux statistics for each month from May to October, with distinct periods of NH3 emission and deposition.
Although fluxes in May only represent four days, the overall maximum flux of around 500 ng m -2 s -1 was observed in that period. While June and July were mainly dominated by small fluxes, typically less than ±10 ng m -2 s -1 , in August clear emission was observed with a daytime median value around 20 ng m -2 s -1 . September marked an emission to deposition transition period, with significant deposition fluxes reaching to -300 ng m -2 s -1 . In October, fluxes were dominated by deposition, with a median 15 daytime value around -30 ng m -2 s -1 . Over the entire period, the daytime median flux was 8.6 ng m -2 s -1 during periods of emission and -19.87 ng m -2 s -1 during periods of deposition (Table 1). Night time fluxes were significantly lower, with median values of 3.5 and -4.27 ng m -2 s -1 , respectively, showing clear diurnal cycles during periods of high emission and deposition.
The flux statistics are affected by the choice of the flux quality flags used to remove periods of weakly developed turbulence or non-stationarity (Foken and Wichura, 1996). A quality flag ≤ 3 is typically used for fundamental research, whereas fluxes 20 with quality flag ≤ 6 are used for long-term flux data sets. We only used fluxes with a quality flag ≤ 3, leaving 68 % of daytime and 46 % of night time fluxes (Table 1).
The random flux error due to instrumental noise ( ) is dependent on the precision of the QCL and variations in the friction velocity ( * ). Over the course of the field campaign, an average precision of 0.085 (±0.010) ppbv at the 10 Hz sample frequency was achieved. The resulting median flux error was 0.17 ng m -2 s -1 for daytime and 0.08 ng m -2 s -1 for night time fluxes; much 25 lower than the median observed fluxes (Table 1). In contrast, the random error derived from the lag time shift ( ℎ ) was significantly larger, with median values of 1.61 ng m -2 s -1 during daytime and 0.72 ng m -2 s -1 during night time. However, still only 11.6 % of the total flux data were below the detection limit (2.15 ng m -2 s -1 ), which shows the overall good performance of the QCL eddy covariance measuring system.

Ogive method
We applied the ogive method to the co-spectra of sensible heat and NH3 flux. Figure 4 shows an example of the normalized co-spectra. To illustrate the effect of HFA, the co-spectra are multiplied by the frequency; HFA is indicated if the slope of the NH3 co-spectrum in the inertial subrange (high frequency part of spectrum) is steeper than the slope of the sensible heat flux 5 co-spectrum. While this is not clearly visible in the co-spectra (by comparing against the expected -4/3 slope line in the inertial subrange), the ogives show a clear underestimation of the NH3 flux at higher frequencies, illustrated by the flattened NH3 ogive curve at high frequencies. Applying the ogive method to all 30-min periods yields a frequency distribution of the attenuation factor (Fig. 5). As the ogives may show significant noise or fluctuations, the choice of the limit frequency for the determination of the scaling factor is critical. For that reason three limit frequencies of 0.05 Hz, 0.067 and 0.10 Hz were chosen representing 10 time scales of 20, 15 and 10 s, respectively. Only periods were used where the standard deviation of the scaling factors determined at those three limit frequencies was smaller than 5 %. Furthermore, only data periods were used where both the NH3 and sensible heat flux were significant and of sufficient data quality (TK3 flag ≤ 3). After applying that filter, the median flux attenuation factor was calculated to be 0.83, corresponding to a flux loss of 17 %. The spread of the observed factors was significant with 50 % of the values lying between 0.75 and 0.91; the standard deviation was 16 %. Also, for some periods the 15 attenuation factor was above one, which is most likely caused by uncertainties in the ogive method.
To give insights into the causes of the NH3 flux loss, the attenuation factor was analysed for dependencies on different parameters. No clear change of attenuation factor was observed over time, which would have suggested an effect of surface aging or cleaning. Due to the uncertainties in the ogive method, the noise in the attenuation factors was larger than the effect of surface aging or cleaning. Also, no correlation of the attenuation factor with ambient air humidity was found. The highest 20 variation of attenuation factors occurred at conditions of neutral atmospheric stability, an indication of the limitations of the ogive method during these conditions. However, a slight decrease of the attenuation factor with increasing horizontal wind speed ( ) was found (Fig. 6), which is expected due to the shift of the turbulence spectra to higher frequencies with increasing wind speed. The linear regression yields a function of the flux attenuation factor using the ogive method ( ) with (in m s -1 ) as: 25 However, with an R 2 value of 0.04, the correlation is very weak. The correlation might be impacted by changes of the aerodynamic measurement height, due to tower raise and the growth of the corn canopy over the season. However, as no clear correlation of the attenuation factor with time was found, this effect is not accounted for in the presented relation with horizontal wind speed.

Time response
The double exponential function used to determine the system's time response is characterized by τ1, τ2 and the value. The median value for τ1 was 0.6 s. That value is slightly higher than the calculated time constant for the exchange of air volume in the 0.5 L absorption cell (~ 0.1 s). The time response needed to exchange the air volume in the inlet tubing is significantly less (< 0.03 s), due to its small volume and low pressure (< 15 kPa) in the inlet tubing. The time constant τ2, representing the time 5 scale for NH3 surface interactions, showed a median value of 27 s. Both τ1 and τ2 did not exhibit a clear trend over time, nor a correlation with humidity or other parameters. In contrast, the value revealed distinct differences over time. Figure 7 shows the evolution of the value over the course of the field campaign together with times of inlet, tube and cell cleaning. Initially, values were around 20 % and then increased steadily to more than 50 % at the end of June. While cleaning of the glass inlet and the inlet tube in June did not directly lead to a decrease in the value, the cleaning of the glass inlet and inlet tubing on 10 27 July led to a visible decrease of the value. On September 12, when the inlet, tubing and the surface of the absorption cell was cleaned, the largest decrease of the values after cleaning down to 10 % was observed. Additionally, a significant overall decrease of values was observed following July 17, after switching the zero air source from the heating catalyst to the UHP compressed air. Since the heating catalyst scrubs NH3 from ambient air, the moisture level of the heating catalyst zero air is more similar to ambient air than that of the dry UHP compressed air. The value differences for the two zero air sources 15 might be therefore caused by different moisture levels. However, no distinct correlation of the value with ambient air humidity was found for the heating catalyst period. A decrease of values with increasing ambient air NH3 mixing ratios, as discussed in Ellis et al. (2010), was detected (Fig. 8), suggesting a larger relative importance of adsorption and desorption processes at lower NH3 mixing ratios. The increase of values at lower NH3 mixing ratios may be also due to the variation of values caused by a larger uncertainty of the double exponential fit with small mixing ratio changes. While the relative 20 random errors of 1 and 2 from double exponential fit increase exponentially with decreasing NH3 mixing ratios, the propagated random error of was typically below 10 % for NH3 mixing ratios above 0.5 ppbv. For lower mixing ratios only values with a relative error of less than 50 % were used.

Time response transfer function method
To investigate the effect of the measured time response on the NH3 fluxes, the flux loss was simulated using the time response 25 transfer function method. As the observed time constants were fairly constant over time, τ1 and τ2 were fixed at 0.6 and 27 s, respectively. The flux loss was simulated for D values of 0, 20, 40 and 60 %, which represents the observed range of D.

Random flux error and flux detection limit
Closely connected to the issue of HFA is the requirement of the NH3 measurement system to resolve small NH3 mixing ratio fluctuations at high time resolution. Due to the challenges in measuring small NH3 fluxes, the quantification of the random flux error and flux detection limit is essential for the quality assessment and interpretation of NH3 fluxes. Distinguishing between two types of random errors, we found that the error due to instrumental noise ( ) was significantly lower than 15 the stochastic random error ( ℎ ). This can be attributed to the high precision of the QCL measurement achieved during the measurement period, resulting in median of 0.13 ng m -2 s -1 . To our knowledge, continuous wave QCL spectrometers are currently the most precise commercially available NH3 measurement systems. Since the instrumental noise affects also ℎ , ℎ (median 1.08 ng m -2 s -1 ) can be used as the total random flux error. Investigating the entire measurement period, we found increasing ℎ values with a higher (absolute) flux magnitude, although, due to variations, no clear relationship 20 could be formulated. Still, for (absolute) flux magnitudes above 20 ng m -2 s -1 , the median random flux error was 13 %, giving a general random error estimate for those higher observed flux magnitudes.
Defining the flux detection limit as 2• ℎ , the median value was 2.15 ng m -2 s -1 , which is about half of what was reported by Sintermann et al. (2011) for flux measurements using PTR-MS after slurry application (4.5 ng m -2 s -1 ) and 4.4 times lower than the detection limit given by Zöll et al. (2016) for measurement above a peatland (9.4 ng m -2 s -1 ), using the same QCL 25 analyser used in this study. While our analysis of ℎ covered the entire measurement period, Sintermann et al. (2011) determined their flux detection limit during a period when no significant NH3 fluxes were detected, most likely leading to a smaller flux detection limit calculation than with our approach. Since we observed that ℎ can vary significantly over time, when filtering the NH3 flux data the respective flux detection limit value of the relevant flux period seems to better reflect different turbulence conditions.

Parameters affecting time response
The QCL's time response for NH3 was determined over the 5 month measurement period, providing a large dataset of different operational and environmental conditions which may impact the time response. Known parameters affecting the NH3 time response are properties of the ambient air like humidity and magnitude of NH3 mixing ratios, surface material and surfaceadsorbing matter, surface temperature and sample flow conditions. Since the temperature of the inlet and the sample flow rate 5 were not significantly changed during the field campaign, we did not investigate the influence of surface temperature and sample flow conditions on time response. From the findings of our study, we discuss and summarize in the following the impact of humidity, ambient NH3 mixing ratios, surface contamination on time response, together with investigations on the availability of surface adsorption sites.

Ambient air humidity 10
Since we determined the time response from the decay of NH3 mixing ratios after switching from ambient to zero air, the humidity of the zero air source is important. While the UHP compressed air was dry, the humidity of the heating catalyst zero air is closer to ambient air humidity. There was a clear improvement in the time response when using the dry UHP compressed air. This indicates that humidity has an effect on the time response if dry air is compared with ambient air of average humidity.
However, no correlation was found between time responses and ambient air humidity when the heating catalyst was used as 15 zero air source. Since the amount of water vapour in the measured ambient air was about 10 6 times higher than the amount of NH3, a humidity effect on the NH3 time response might be not noticeable at ambient humidity levels. Since there are only limited adsorption sites for water molecules available on the inlet and cell surfaces, low humidity levels may already lead to a saturation of surface sites. Once all surface sites are occupied by water molecules, the time response is insensitive to variations in ambient air humidity, which might explain why no correlation of the time response with ambient air humidity was observed. 20 Ellis et al. (2010) found the time response to degrade with the relative ambient air humidity, however, their data was collected under significantly higher NH3 mixing ratios reaching up to 1000 ppbv and the correlation was much weaker when the inlet line was heated to 40 °C.

Ambient air NH3 mixing ratio
As shown in Fig. 8, the relative importance of the slow component of the system's time response was higher with lower NH3 25 mixing ratios. While our ambient air measurements were mainly below 10 ppbv, the laboratory study by Ellis et al. (2010) saw the same effect for NH3 mixing ratios between 30 and 1000 ppbv. The proposed hypothesis for this effect is that at lower NH3 mixing ratios a larger proportion of the ambient NH3 can interact with surface adsorption sites, which leads to a larger value.
As the time response experiments were performed by switching from an ambient mixing ratio to zero air, the time response could be improved if the start and end NH3 mixing ratio are offset by a fixed NH3 mixing ratio, leading to higher passivation 30 of the surface material. However, in a laboratory test (not shown), we found that the time response was not significantly improved when switching between two higher mixing ratios of the same mixing ratio difference. This may imply that the time response is also governed by the magnitude of the NH3 change instead of the NH3 mixing ratio alone. While for the former the NH3 high frequency loss would be related to the NH3 flux magnitude, for the latter it would be dependent on the NH3 mixing ratio, which has implications for the flux loss correction. For our dataset a mixing ratio decrease of about 10 ppbv increased the value by about 10 % (Fig 8.), which would result in an additional flux of about 10 % (Fig. 10), However, since more 5 tests are needed to distinguish whether the NH3 flux magnitude or NH3 mixing ratio are the determining factor, the correction factors developed here do not included this additional correction.

Inlet system surface material and contamination
Previous studies have shown the impact of surface material on time response, with PFA or PTFE being the most suitable material for inlet tubing (Whitehead et al., 2008;Zhu et al., 2012). Surface aging and contamination are known to be important 10 factors, thought it has been not satisfactorily quantified in the past. The presented 5 month dataset shows how the time response gets slower with time, most clearly visible in the increasing values in the first month of the measurements (Fig. 7). The fact that the time response is not improved after every cleaning of the inlet parts and absorption cell shows the complexity of NH3 time response and the importance of the other factors affecting it.

Availability of surface sites and active passivation 15
The discussion of parameters controlling time response shows that the mechanisms that govern the time response of NH3 measurements are still not well understood. One reason for this is the lack of knowledge on the amount of available surface sites for adsorption of NH3 under the different conditions. Also, the amount of NH3 adsorbed to the surface is uncertain. It can be estimated by introduction of 1H,1H-perfluorooctylamine into the sample inlet (Roscioli et al., 2015), where the NH3 originally adsorbed to the surface is replaced by the amine and measured by the QCL. By integrating the NH3 peak from the 20 desorption process, Roscioli et al. (2015) estimated a density of NH3 of 8•10 14 molecules cm -2 on the surface of their QCL system. We performed the same experiment for our system in the laboratory after the completion of the field campaign and found a surface coverage of about 4•10 13 molecules cm -2 , which is about a factor of 20 lower than the value observed by Roscioli et al. (2015). As we carried out the test using the same tubing and glass inlet (both not cleaned) that were used during the campaign, the desorbed NH3 amounts represents the amount of NH3 at the end of the measurement campaign, although it 25 cannot be excluded that some NH3 was desorbed from the surface after the experiment. Repeated application of the amine after exposure of the inlet to a calibration source of NH3 (8 ppbv) showed that the desorbed NH3 was proportional to the NH3 exposure time. This shows the direct link between NH3 exposure and adsorbed surface NH3. However, similar experiments would be needed for a wider range of conditions to obtain a quantitatively meaningful characterization of adsorbed surface NH3 and better understanding of available NH3 adsorption sites. Together with the NH3 adsorption capacity of the surface, a 30 determination of the adsorption and desorption rate constants is necessary to describe the interaction of NH3 with the surface material and to quantify the HFA during field conditions through adsorption and desorption mechanisms.

High frequency loss correction for NH3 fluxes
The knowledge about the parameters affecting the time response of NH3 measurements is important to make improvements on the NH3 measurements and to find the appropriate high frequency flux loss correction method. In past studies, both experimental and theoretical approaches have been applied for the correction of NH3 fluxes. Although experimental approaches were used in this study, we discuss in the following also theoretical approaches, based on the findings from the presented 5 5 months flux dataset.

Theoretical approaches
In the processing of eddy covariance fluxes, theoretical approaches to correct for the HFA are commonly applied. In this approach, transfer functions, which describe the HFA, are multiplied with the reference co-spectrum. The reference cospectrum is either derived from parameterizations or from direct measurements of a non-attenuated flux, such as the sensible 10 heat flux. For NH3 fluxes, the transfer function has to account for the NH3 surface interaction, as it has been formulated for fluxes of water vapour (Ibrom et al., 2007;Massman and Ibrom, 2008). Ferrara et al. (2012Ferrara et al. ( , 2016) used a reference co-spectral model combined with theoretical transfer functions for the effect of lateral/longitudinal separation between the sonic anemometer and the NH3 measurement and for the HFA along the inlet tubing and within the QCL absorption cell. The main parameters controlling these transfer functions are wind speed, inner diameter and length of tubing, sample flow rate and the 15 absorption cell time constant. The results from our study showed that the time response of the NH3 measurements over the 5 month period was governed by desorption and adsorption processes of NH3 on surfaces, which are not adequately considered in these transfer functions. Although short inlet tubing, a high sample flow rate, low tube pressure conditions and heating of the inlet line are factors that minimize the time response (Ferrara et al., 2012(Ferrara et al., , 2016, we found that changes in operational conditions, such as humidity and surface contamination, can have significant impacts on the time response and therefore the 20 flux attenuation. As argued also by Sintermann et al. (2011), a more mechanistic understanding of surface processes is necessary for a theoretical transfer function that adequately describes the time response of NH3. For that reason, approaches that evaluate the high frequency flux loss from measurements are necessary.

Experimental approaches 25
The high frequency flux loss in experimental approaches is determined by comparing the co-spectrum of the attenuated scalar flux to the co-spectrum of a non-attenuated flux. The ogive method (used for NH3 fluxes by Ferrara et al. (2012), Sintermann et al. (2011), Zöll et al. (2016 and in this study) uses the cumulative co-spectra, and is mathematically equal to approaches that use the ratio of the scalar co-spectrum and the reference co-spectrum directly (used for NH3 fluxes by Whitehead et al., 2008). Differences are mainly apparent in data pre-processing, averaging or fitting procedures used. Since variations of co-30 spectra from the ideal shape can be significant, the method which yields the most robust relation between the NH3 and reference ogives (or co-spectra) is preferred. In this study, we used the standard deviation of the attenuation factor from three different https: //doi.org/10.5194/amt-2019-193 Preprint. Discussion started: 18 June 2019 c Author(s) 2019. CC BY 4.0 License. limit frequencies, the flux magnitude and the flux quality flag to filter the dataset, yielding a median flux attenuation factor of 0.83. The flux loss was significantly less than for NH3 fluxes reported by Zöll et al. (2016), who found for their NH3 QCL a median flux attenuation factor of 0.67. Also using a NH3 QCL system, Ferrara et al. (2012) determined an average correction factor of 1.37 with the ogive method, which translates to a flux attenuation factor of 0.73. Reasons for the differences in flux attenuation factors may be linked to operational differences. In contrast to our study and Zöll et al. (2016), the system by 5 Ferrara et al. (2012) had a lower flow rate with laminar flow conditions, which would be expected to result in more significant attenuation factors. In contrast, their heated inlet line was shortest (2.5 m), followed by Zöll et al. (2016) (3 m) and our study (5.5 m). Regular cleaning of the QCL system over the 5 month period, likely contributed to the low flux attenuation in our study. Another factor affecting the flux loss is the flux magnitude. For example, the NH3 fluxes presented by Ferrara et al. (2012) are significantly larger, covering six days directly after urea application to the agricultural field, which may explain the 10 moderate flux attenuation despite the laminar flow conditions. Finally, differences in data processing and filtering of ogive results can be responsible for some of the discrepancies.
We observed a slight decrease of the flux attenuation factor with increasing horizontal wind speed, as was reported by Sintermann et al. (2011). However, uncertainties in the flux attenuation factor may be the reason why the correlation with wind speed was very weak (R 2 = 0.04). Zöll et al. (2016) did not find a correlation with wind speed, stating a random error of the 15 flux attenuation factor of 15 %. In our study, the random error of the attenuation factor was 19 %, while the flux attenuation factor in Ferrara et al. (2012) had a variation of 27 %. Again, the data processing and data filtering method largely impact the standard variation of the ogive method results. For example, as shown in Fig. 6, our data included attenuation factors greater than 1, which are not realistic, but were included for statistical reasons. Next to the random uncertainty, Zöll et al. (2016) states that there might be potential systematic deviations of the attenuation factor, caused by HFA that is not detected by the ogive 20 method; however, it is unclear which low-pass filtering processes might be responsible for that.
Due to the uncertainties in the ogive method and its failure to reflect the operational changes (such as inlet cleaning and tube aging) that we expected to affect the attenuation, we used the time response measurements to quantify the flux loss over the experiment period. We found that the time response transfer function method captured cleaning and surface aging overtime.
Also, from the change of the zero air source to dry UHP compressed air we found evidence that time response is sensitive to 25 humidity, although a distinct correlation between ambient humidity and the time response could not be found. Furthermore, the applicability of the time response transfer function method is not limited to the flux magnitude, and therefore can also be used to determine the flux attenuation of low NH3 fluxes where the ogive method shows larges uncertainties.
To correct the NH3 flux loss with the time response transfer function method, we determined the attenuation factor as a function of wind speed and the value from the time response measurements (Eq. 10). Since the observed values showed significant 30 fluctuations, which are attributed to the random uncertainty of the double exponential fitting procedure, we calculated a moving median value as illustrated in Fig. 7 effect has to be corrected for if the time response method is applied to the entire dataset. We therefore increased those values which were derived with the UHP compressed air by a fixed absolute value of 20 % (blue data points in Fig. 7), which was the approximate difference of values observed at the time of switching between the two sources. Applying the corrected value in Eq. 10 to the entire flux dataset, the median flux attenuation factor is 0.67, corresponding to a median flux underestimation of 34 % that has to be corrected for. This is significantly larger than the median flux loss from the ogive method, which was 5 17 % (Fig. 5). Using the relationship of the ogive flux attenuation factor with wind speed (Eq. 9) yields a median flux loss of only 8 % for the entire dataset. However, the relation might be skewed by flux attenuation factor values above 1, which were not rejected.
We showed that the time response transfer function method is a technique that can be used to correct for the HFA of NH3 fluxes. The method accounts for changes in operational and environmental conditions over time, with a significantly lower 10 random uncertainty than the ogive method. However, the applicability is largely dependent on the correct determination of the time response. The change of values after switching between zero air sources showed that it is crucial to perform time response experiments under the conditions of the flux measurements. This entails using similar air humidity conditions and NH3 mixing ratios as in ambient air, but also NH3 step changes of similar magnitude than the NH3 fluctuations during the eddy covariance measurements. For example, the time response can be determined between two different NH3 levels, as it was done 15 by Brodeur et al. (2009), by adding NH3 from a NH3 source to ambient air. However, challenges to use such a system in field conditions remain, as short-term fluctuations (< 5 min) of ambient air would have to be filtered out in order to obtain a reliable estimation of the double exponential fit and therefore the time response.

Conclusions
Challenges in measuring atmospheric NH3 at a fast sampling rate have limited the application of the eddy covariance technique 20 to investigate the surface-atmosphere exchange of NH3. While several studies have presented eddy covariance measurements, there is still a poor understanding on the drivers governing time response and how to account for HFA in the post-processing.
In the present study, we deployed a continuous-wave QCL from May to October 2017 to measure NH3 eddy covariance fluxes above a corn field in Eastern Canada.
Over the experimental period, the eddy covariance system was operated without major interruptions, while regular 25 maintenance of the QCL guaranteed a consistently high precision of the 10 Hz NH3 signal (90 pptv). The median random flux errors due to the instrumental noise was insignificant (0.1 ng m -2 s -1 , 4 % using absolute NH3 fluxes) compared to the stochastic error of the eddy covariance measurement (1.1 ng m -2 s -1 , 15 % using absolute NH3 fluxes), which is independent from the eddy covariance system performance. The median flux detection limit before applying the HFA correction was 2.15 ng m -2 s -Through the 5 month measurement period, we obtained flux measurements over a large range of environmental and operational conditions, which allowed us to study the parameters affecting the instruments time response and its effect on the NH3 flux measurements. While humidity is thought to be a factor affecting time response, we found no clear correlation between the ambient humidity and the time response. Instead we found that the time response was improved when dry UHP compressed air was used, which suggests the existence of a humidity effect. For that reason, the time response value had to be corrected 5 to account for the higher ambient humidity under flux measurement conditions. While we saw significant improvement of the analyser's time response after cleaning of the QCL sample cell, the cleaning of inlet tubing and the QCL glass inlet did not systematically lead to significant time response improvements. This shows the complexity of mechanisms governing NH3 time response and the need for the appropriate flux loss correction method.
From the flux attenuation analysis with the ogive method, we determined a median flux loss of only 17 % (±16 %), which 10 shows the overall good performance of the eddy covariance system. As for the time response, no correlation between the flux loss and ambient humidity was found, instead a slight increase of the flux loss with increasing horizontal wind speed, which was expected due to the shift to smaller turbulent scales at higher wind speeds. The ogive method did not detect the change of time response due to surface aging and instrument cleaning, which we attribute to noise and fluctuations in the co-spectra. Due to the uncertainties in the ogive method and complexity of NH3 time response, we introduce the use of the time response 15 transfer function method for NH3, which simulates the flux attenuation according to measured changes in the system's time response, and thereby accounts for changes of the time response, for example due to surface aging. We provide a correction factor as a function of the time response value and the horizontal wind speed. The obtained flux correction factors ranged from 1.2 to 2.11, with a median flux loss of 34 %, which is substantially higher than the values obtained through the ogive method. We argue that due to the complexity of NH3 adsorption and desorption processes to surfaces, it is important to 20 determine the time response over the course of a field experiment, also to make better informed decision on instrument operation and maintenance. In the future, improvements have to be made on how the time response is determined over the course of a field experiments. Also, a more in-depth understanding of NH3 surface adsorption and desorption processes is necessary to develop theoretical frameworks to correct the flux loss of NH3 eddy covariance measurements and give guidance for improved fast time response NH3 measurements. 25

Author contribution
AM and SS performed and quality controlled the NH3 measurements. LP and EP obtained the turbulence and H2O measurements and managed the experimental site. AM performed the NH3 flux calculation and quality control and did the high frequency attenuation analysis. JGM and EP initiated the project. AM wrote the manuscript with comments from coauthors.