Measurements of Ozone Deposition to a Coastal Sea by Eddy Covariance

A fast response (10 Hz) chemiluminescence detector for ozone (O3) was used to determine O3 fluxes using the eddy covariance technique at the Penlee Point Atmospheric Observatory on the south coast of the UK during April and May 2018. The median O3 flux was -0.132 mg m−2 h−1 (0.018 ppbv m s−1) corresponding to a deposition velocity of 0.037 cm s−1 15 (interquartile range 0.017–0.065 cm s−1) – similar to the higher values previously reported for open ocean flux measurements, but not as high as some other coastal results. Eddy covariance footprint analysis of the site indicates that the flux footprint was predominantly over water (> 96%), varying slightly with tide. At moderate-to-high wind speeds, ozone deposition increased with wind speed, and showed a linear dependence with friction velocity of comparable magnitude to predictions from the one-layer model of (Fairall et al., 2007). Deposition was also elevated at very low wind speeds, most 20 likely because the footprint contracted to include a greater land contribution in these conditions.

such as turbulence at the air/sea interface. More recent flux measurements have been made with the eddy covariance method, which is the best way of observing fluxes in a system without perturbing it. Eddy covariance measurements have been made from coastal towers (Gallagher et al., 2001;Whitehead et al., 2009;McVeigh et al., 2010), aircraft (Lenschow et al., 1982;Kawa and Pearson, 1989), and ships (Bariteau et al., 2010;Helmig et al., 2012). The reported deposition velocities (vd) over 35 saltwater vary greatly: 0.01-0.15 cm s⁻¹, with windspeed dependencies evident in some measurements and absent from others.
The reported eddy covariance measurements use two different techniques to measure ozone at high frequency, both utilising chemiluminescent reactions of ozone. In the instruments used for tower-based measurements (Gallagher et al., 2001;McVeigh et al., 2010;Whitehead et al., 2009), ozone is reacted with a coumarin-based dye on the surface of a silica gel disk. 40 Aircraft (Kawa and Pearson, 1989;Lenschow et al., 1982) and ship-borne (Bariteau et al., 2010;Helmig et al., 2012) instruments have instead reacted ozone with gas phase nitric oxide.
Ozone deposition is likely to depend both upon physical exchange, facilitated by diffusion and turbulence, and chemical reaction at the water's surface (Chang et al., 2004;Fairall et al., 2007;Luhar et al., 2018). Iodide in sea water has been identified as a key reactant (Garland et al., 1980), and there has been considerable recent progress in understanding its global 45 distribution Macdonald et al., 2014;Sherwen et al., 2019). However, there has only been one report of the dependence of the iodideozone rate constant with temperature (Magi et al., 1997), and this remains a considerable uncertainty in global models. Dissolved organic material (DOM) has been suggested to be of similar importance to ozone deposition as iodide (Martino et al., 2012;Shaw and Carpenter, 2013), especially given its enrichment in the sea surface microlayer (SML) (Zhou and Mopper, 1997). The complex and variable composition of DOM makes assessing its global 50 reactivity towards ozone a challenge.
A better characterised ozone deposition sink to the oceans would significantly improve our understanding of global tropospheric O3 cycling. Here we present coastal ozone flux measurements made at Penlee Point Atmospheric Observatory (PPAO; https://www.westernchannelobservatory.org.uk/penlee/) on the southwest coast of the UK using a fast response gas phase chemiluminescence detector (CLD). Causes of natural variability and uncertainty in the observed deposition velocity 55 are discussed, including the effects of the changing relative contributions from sea and land within the flux footprint.

Measurement location
The PPAO is on a headland just south-west of Plymouth, UK, located 50° 19.08' N, 4° 11.35' W. The observatory is located 11 m a.m.s.l. with an extendable mast. It lies 30-60 m away from the sea, depending on tide, with the intervening land 60 predominantly bare rock with some grass immediately surrounding the tower. For the work presented here, the top of the tower was extended to 19 m a.m.s.l. The dominant wind directions are from the south-west, followed by the north-east https://doi.org/10.5194/amt-2020-65 Preprint. Discussion started: 27 May 2020 c Author(s) 2020. CC BY 4.0 License.
( Figure 1). The focus of this work is the south-west (180-240°) wind sector, which brings in air from the Atlantic Ocean and English Channel to the site (Yang et al., 2016).

Experimental set-up 65
The ozone chemiluminescence detector was adapted from an Eco Physics® CLD 886 NOx detector, working on the same principle as the instrument used by Helmig et al. (2012). A supply of excess NO is introduced to the sample, which reacts with O3 to generate NO2 in an excited state. The relaxation process leads to emission of a photon that is amplified and detected using a photomultiplier tube (PMT). In order to maintain a low number of dark counts, the PMT is cooled to -5°C by a Peltier cooler. Clean dry air is continuously pumped over the PMT to avoid the build-up of water ( Figure 2). 70 Sample air was drawn from the top of the tower through ~10 m of 3/8'' PFA tubing by a vacuum pump at 13.5 L min⁻¹. This maintained a turbulent flow in the main sampling line (Reynolds number = 3000). A flow of 300 mL min⁻¹ was drawn from this sample manifold through 1/8'' PFA tubing and into the analyser using an internal vacuum pump ( Figure 2 part 11), limited by a critical orifice (part 5). Before entering the analyser, the sample air was first passed through a dryer (part 3) consisting of 60 cm of Nafion tubing coiled in a container of desiccant (indicating Drierite) to reduce humidity. A three-way 75 solenoid valve (part 2) allowed for a sample of indoor air passed through a charcoal filter (part 1) to remove O3 to record an instrument zero. A 50 mL min⁻¹ flow of 2% NO in N2 was supplied separately to the analyser at a pressure of 4 bar through approximately 1.5 m of 1/8'' PFA tubing. The NO and O3 were then mixed immediately before the reaction chamber (part 9, at ~26 mbar pressure) and the resulting chemiluminescence was detected by the PMT.
The CLD counts were logged at 10 Hz and converted into ozone mixing ratios using the signal from a co-located, recently 80 calibrated 2B model 205 dual beam ozone monitor. The CLD sensitivity was determined to be 240 counts s⁻¹ ppbv⁻¹ and showed no obvious dependence on humidity ( Figure 3) providing evidence for the efficacy of the dryer. Instrument dark counts were 480±40 count s⁻¹, leading to a 10 Hz signal-to-noise ratio of 33 for the average 46 ppbv O3 measured during this work.
Three-dimensional wind data were obtained from a Gill WindMaster Pro 3D sonic anemometer at 10 Hz. Humidity, air 85 pressure and temperature data were logged at 0.25 Hz from a Gill MetPak Pro. Vertical wind data were adjusted by +16.6% and +28.9% in magnitude for positive and negative values, respectively, in line with the corrections recommended for a reported firmware bug in the Gill WindMaster instruments: (http://gillinstruments.com/data/manuals/KN1509_WindMaster_WBug_info.pdf).

Pre-flux processing 90
The eddy covariance method (EC) relies on the simultaneous measurement of vertical wind speed (w) and the relevant scalar (in this case, ozone concentration). These values were determined at 10 Hz in order to resolve the full range of eddies responsible for vertical ozone transport. It is necessary to average data over a suitable period to reduce random noise and https://doi.org/10.5194/amt-2020-65 Preprint. Discussion started: 27 May 2020 c Author(s) 2020. CC BY 4.0 License. capture transport from large eddies, whilst avoiding too long a period such that non-turbulent transport and non-stationarity become issues. An averaging time of around 30 minutes is often recommended (Foken, 2008). Previous measurements of O₃ 95 flux have used averaging periods from 10 minutes (Helmig et al., 2012) to 1 hour (Gallagher et al., 2001), and a 20-minute period was chosen for this work. Prior to the flux calculation, data were despiked using a median filter despiking method (Brock, 1986;Starkenburg et al., 2016) using an order of N = 4 (9 points in a window). This involves binning the differences from the normalised data into exponentially more bins until bins exist within the range of the histogram that have zero values. Difference values beyond these empty bins are then identified as spikes and removed. For the flux calculation, data 100 were linearly detrended to determine deviation from the mean within the averaging period. A double rotation was applied to the wind data in each averaging period to align the u axis with the mean wind and remove any tilt in the wind vector, resulting in a mean vertical wind of zero.
Due to the Nafion dryer and the fixed temperature and pressure of the reaction chamber, density corrections known as WPL corrections (Webb et al., 1980) were unnecessary for However, the presence of water vapour was taken into account for the 105 determination of ancillary parameters such as the Obukhov length used in footprint modelling. It should be noted that beyond its effect on mixing ratio, water vapour also quenches the chemiluminescence of the reaction of NO with O3. This can be dealt with either by determining the instrument sensitivity over a range of water vapor conditions (at the cost of some sensitivity) and applying a correction, or by sufficiently drying the sample air. The latter approach was taken here. Despite a range of humidity (2.8 × 10⁻⁵-1.8 × 10⁻² mol/mol, Figure 3) over the 42-day observation period, the two instruments 110 compare well when using a fixed sensitivity for the CLD. The sensitivity value of 240 ppbv s⁻¹ also compares favourably to 213 ppbv s⁻¹, which was estimated using a supply of dry ozone during lab tests prior to deployment. These results suggest that the dryer removed any major effect on the detection of ozone concentration and flux.
The sample air must travel from the inlet to the detector through the inlet, which introduces a time lag relative to the instantaneously measured wind data. The two datasets must therefore be realigned in order to calculate the covariance. A 115 cross-correlation function (CCF) was calculated at different time lags, with a high-pass Butterworth filter applied to the input values. The presence of a negative peak in the resulting CCF spectrum indicated a strong anticorrelation between ozone concentration and vertical wind, characteristic of deposition. Individual CCF plots were noisy, and gave scattered lag values, with a high density around 4 seconds. Daily average CCF plots indicated clear peaks in all but one case and drifted from 3.9 to 4.1 seconds over the course of the experiment (e.g. Figure 4). This is likely a consequence of slight particulate build-up in 120 the sample line filters over the course of the measurements. Individual 20-minute flux interval lags were accepted if they fell between 3.5 and 4.5 seconds to allow for some variability in conditions, vacuum pump strength etc. Lags that fell outside of these boundaries were then set to a value determined by a linear fit of the accepted data ( Figure 5). Simply setting the lag to 4 seconds in all instances was found to decrease the flux by 5% relative to the method used here (CCF lag determination maximises the flux magnitude). The expected lag was also estimated from the inlet setup: a 13.5 L min⁻¹ flow rate through 125 10 m of 3/8'' tubing plus a 300 mL min⁻¹ sample flow through 2 m of 1/8'' tubing yields a 4.2 second lag, similar to the CCF-determined values.
Following these steps, the ozone flux was calculated on a 20-minute basis using eddy4R (Metzger et al., 2017) with a modified workflow. Flux values were then used to determine the deposition velocity according to Eq. (1): (1) 130 where vd is deposition velocity in cm s⁻¹, F is flux in mol cm⁻² s⁻¹, and [O₃] is ozone concentration in mol cm⁻³. Molar flux was calculated using the instantaneous vertical wind, ozone mixing ratio and density of dry air. Similarly, the ozone concentration used in Eq. (1) was calculated for dry air using the mean ozone mixing ratio for the averaging period to avoid introducing a dependence on water vapour to the deposition velocity.

Data selection 135
A series of selection criteria were applied to the calculated 20-minute flux data. Firstly, periods with more than 10% missing data were excluded. Missing data were most commonly caused by periods of maintenance, or when heavy rain disrupted the sonic anemometer readings. Data were also selected by wind directiononly data between the true wind direction of 180° and 240° were accepted to avoid observing deposition on the headland to the north-west.
A selection criterion based on ozone variation, as used by Bariteau et al. (2010), was introduced to avoid periods of non-140 stationarity i.e. significantly different conditions within an averaging period (such as a sudden change in the air mass passing by the sensor, or a change in wind direction). Data were excluded if the ozone concentration drifted significantly (> 6 ppbv in 20 minutes) or if the standard deviation in ozone was above 2 ppbv. Data with a standard deviation in wind direction of > 10° were also removed to avoid non-stationarity of wind, as performed by Yang et al., (2016) for the same site. We note that the discontinuity in wind direction at for northerly winds (360°-0°) can incorrectly increase the standard deviation measured 145 near to north. However, this issue does not arise as we consider only winds from the south-westerly sector.
Periods of low wind speed were also excluded because of suspected land influence, as indicated by elevated deposition velocities (see Sect. 5). This is contrary to the trend of increasing deposition with wind speed proposed by Chang et al. (2004) and observed during open ocean cruises by Helmig et al. (2012). Yang et al. (2016Yang et al. ( , 2019) observed a similar enhancement in CO₂ transfer at wind speeds, and chose to filter out data when wind speeds were < 5 m s⁻¹. Footprint analysis 150 was used to investigate the potential for land influence within the footprint area. Land influence may increase as the footprint contracts at low wind speeds. Using the flux footprint parameterisation of Kljun et al. (2015), footprints were calculated for each averaging period using observed wind and stability conditions, and aggregated into 1 m s⁻¹ wind speed bins. Using these aggregated footprints, the percentage of land area contribution in the footprint area was estimated to increase from 1-2% at high wind speeds to 15% at winds below 2 m s⁻¹ ( Figure 6). It should be noted that the footprint model is designed for 155 flat homogeneous terrainnot a heterogeneous coastal site. For instance, land influence may be higher than estimated at low wind speeds as a consequence of the elevation of the headland relative to sea level.
Where 0 is roughness length in m, is measurement height in m, is the von Kármán constant, is wind speed in m s⁻¹, * 160 is friction velocity in m s⁻¹, and ( ) is the integral of the universal function, defined as (Businger et al., 1971;Högström, 1988): Roughness lengths at high wind speeds are scattered approximately around 0.0002 m, which is expected for an open sea fetch (World Meteorological Organisation, 2008), but a large increase can be seen at wind speeds < 3 m s⁻¹. The increase is indicative of a surface with more roughness elements, such as the rocks and grass found on the headland. Inaccuracies in the double rotation method at low wind speeds can mean that the removal off horizontal wind from the rotated vertical 170 component is incomplete, further contributing to the elevated surface values. This indicates the need for a filter to exclude land-influenced flux data. A wind speed filter of > 3 m s⁻¹ was used in this work, though filters on the basis of z₀ could also be used to similar effect. While it could further decrease the possibility of land influence, a more stringent filter has not been applied to avoid excessive data removal.
Previous eddy covariance work on CO2 has applied filters on the basis of friction velocity ( * )(e.g. Barr et al., (2013)) to 175 avoid underestimation of flux during periods of poorly developed turbulence, especially at night (Aubinet, 2008). However past measurements of oceanic ozone deposition have not reported using such a filter (Gallagher et al., 2001;Helmig et al., 2012;McVeigh et al., 2010), likely because very low wind speeds and * are uncommon over the ocean. For our data, removing data with * < 0.15 cm s⁻¹ made no difference to the observed median deposition velocity. Therefore, given that a wind speed filter was already applied, no friction velocity filter was included. 180 Longer averaging periods than 20-minutes were also considered, but 60-minute averaging caused a large loss of data to the selection criteria. Missing data, as well as stationarity of wind and ozone especially contributed to an overall 23% reduction in total data accepted compared with 20-minute averaging. This shorter averaging time was therefore used to avoid loss of data to stationarity requirements while still observing reasonable lag times and cospectral shape.  (Whitehead et al., 2009) and 0.13 cm s⁻¹ (Gallagher et al., 2001). These measurements were carried out at Mace Head (west Ireland), Weybourne (east UK) and Roscoff (north-west France) respectively. Our median vd of 0.037 cm s⁻¹ is towards the upper end of previous work, though much lower than Gallagher et al. (2001). 200

Wind Speed Dependence
The dependence of vd on wind speed and friction velocity ( * ) is examined in Figure 11A  Kármán constant (0.4), and * is the water-side friction velocity in m s⁻¹. The fit shown in blue in Figure 11B was determined using the relevant parameters during the experiment at the PPAO, with * derived from * using: * = √ * (7) 220 where and are the densities of air and water respectively. , , and were determined empirically according to Eq. (8) (Morris, 1988), Eq. (9) (Magi et al., 1997), and Eq. (10) (Johnson and Davis, 1996): where is the sea surface temperature (in K) and [ − ] is the aqueous iodide concentration in mol dm⁻³. We note that Eq. (9) only accounts for the reactivity of ozone with iodide in the sea surface. Other species present in the SML have also been shown to react with ozone (Martino et al., 2009;Shaw and Carpenter, 2013), but given the uncertainty surrounding their rate constants and any temperature dependence, they have been omitted here. , and thus vd (cm s⁻¹) using Eq. (6) (shown in blue on Figure 11B). This can be simplified to: = 0.01324 + 0.09378 * In comparison, the linear fit of our experimental 20-minute vd values against * is: = 0.02017 + 0.07537 * Our results therefore show comparable, but slightly lower dependence on friction velocity (and therefore also wind speed) 235 than predicted by the parameterisation of Fairall et al. (2007). Given the assumptions of the simplified model (Eq. (6)) and the uncertainties in various parameters, not least the rate constant for the reaction of O₃ with I⁻ (e.g. Moreno & Baeza-Romero, 2019), this agreement is remarkable. The two-layer model of Luhar et al. (2018) for the same data is shown in black in Figure 11B. Considering only iodide reactivity, this model appears to under-predict deposition compared with the onelayer model of Fairall et al. (2007), and lacks any major dependence on wind speed except during very calm conditions (see 240 Sect. 6 for further discussion).

Land and Tidal Influence
Footprint analysis of the PPAO site (as discussed in Sect. 4) suggests that the spatial contribution of land surfaces to our observed deposition is 3.9%. However, deposition to land is typically greater than to the ocean, amplifying the potential https://doi.org/10.5194/amt-2020-65 Preprint. Discussion started: 27 May 2020 c Author(s) 2020. CC BY 4.0 License.
influence on our data. If our observations were adjusted for 3.9% spatial contribution of grassland (vd ≈ 0.25 cm s⁻¹, 245 Hardacre et al., 2015), then our calculated coastal water vd would be 0.028 cm s⁻¹ (23% lower than we measured). In reality the terrain is a mixture of grassland and rocky shoreline, varying in extent with the tide, so the land vd discussed above may be an overestimate. Although there are insufficient data over the land to the north-west to reliably determine a vd value for the land around the PPAO, an estimate can be made by obtaining a least square solution using the land cover determined in Figure 6 and the observed vd values in Figure 11A. Data from wind speeds > 14 m s⁻¹ were not used (only 4 data points). 250 Using all data from 2-13 m s⁻¹ yielded values of 0.167 cm s⁻¹ and 0.034 cm s⁻¹ for land and sea respectively, suggesting a lesser effect from land than using the fixed value from Hardacre et al. (2015). Given that the land contribution in Figure 6 doesn't stabilise until 9 m s -1 , it is possible that constant vd between 4 and 10 m s⁻¹ wind speeds ( Figure 11A) may be a consequence of land influence and wind speed enhancement counteracting one another. Estimated water-only vd values, calculated by subtracting the product of the land fraction and the land vd value from the measured vd, are shown in Figure 13. 255 It is worth reiterating that this footprint model is designed for use in homogenous environments, which is not true of our site.
Furthermore, the double rotation applied to the wind data will result in varying pitch angles relative to the water surface, introducing a dependence of the footprint extent on this pitch angle. These limitations may be important for work relying on direct interpretations of the flux footprint, such as comparisons to emissions inventories (Squires et al., 2020;Vaughan et al., 2017). In contrast, we use aggregates of these individual footprints only to develop a strategy for robust data selection. 260 The PPAO site flux footprint also experiences periodic variations associated with the tide, which alters the effective measurement height and changes the land type in the footprint when the shoreline is exposed. Whitehead et al. (2009) provide an extreme example of this, reporting vd increasing from 0.030 cm s⁻¹ at high tide to 0.21 cm s⁻¹ at low tide during the day at a site with a tidal range of 9 m. The tide also causes periodic movement of the river plume around the Penlee headland, altering the salinity and composition of the surface water (Yang et al., 2016). 265 Measurement height was adjusted for tide height using tidal data from the British Oceanographic Data Centre (BODC), measured approximately 6 km upstream. Periodograms were also used to look for periodic deposition variation from exposed shoreline or riverine water, but none could be identified above the variability in the data. Gallagher et al. (2001) report a tentative (though statistically insignificant) diurnal cycle for coastal water during observations made at Weybourne in East Anglia, UK. However, no such trend was observed in the PPAO flux data. 270

Measurement uncertainty
To understand the variability in our vd observations, a flux limit of detection was obtained empirically according to the method of Langford et al. (2015). For each averaging period, cross-correlation functions (discussed in Sect. 3) were calculated at a series of improbable lag times (150-180 seconds), and the root mean squared deviation of these values was taken to be representative of the random error of the flux measurement. Limits of detection were calculated for each 275 averaging period due to its dependence on wind speed and atmospheric stability, giving a median 2σ flux limit of detection https://doi.org/10.5194/amt-2020-65 Preprint. Discussion started: 27 May 2020 c Author(s) 2020. CC BY 4.0 License. of 0.113 mg m⁻² h⁻¹. At the average ozone concentration of 48 ppbv, this equates to a deposition velocity of 0.033 cm s⁻¹, with 305 of the 491 averaging periods exceeding their individually determined 2σ limit of detection.
Alternatively, a theoretical estimation of flux uncertainty can be made according to the expression given by Fairall et al. (2000): 280 where ∆ is flux uncertainty, w' is instantaneous vertical wind velocity fluctuation, X' is instantaneous ozone fluctuation, σw is the standard deviation in vertical wind velocity, σX is the standard deviation in ozone concentration, T is length of the averaging period in seconds, and τwca is the integral timescale for vertical fluctuations. A factor with a value of 1-2 is sometimes also included to reflect uncertainty in this relationship (Blomquist et al., 2010). The integral timescale τwca can 285 either be determined from a flux cospectrum peak frequency: or empirically according to: where z is measurement height in meters, U is mean wind speed, and a is a value that varies with atmospheric stability. The 290 value of a has been reported variably as 0.3-3 for near neutral conditions (Blomquist et al., 2010;Lenschow and Kristensen, 1985) and on the order of 10-12 for convective/unstable conditions (Blomquist et al., 2010;Fairall, 1984). Using the peak frequency of the cospectrum shown in Figure 14 (0.07 Hz), τwca was determined to be approximately 2.2 s during nearneutral conditions and wind speeds of 12.1 m s⁻¹. This corresponds to a value for a of 1.5, similar to the literature. Since individual 20-minute cospectra were too noisy, this a value was used with Eq. (13) to determine for each 20-minute 295 period. It should be noted that the value of a is stability dependent. However, since stability was near neutral for most periods (z/L = -0.39 to 0.15, 20 th -80 th percentile), the effects of varying stability on a are expected to be small.
Using these integral timescales, a theoretical flux uncertainty can be calculated for each averaging period using Eq. (11). The theoretical values obtained were much higher than those found empirically -the median theoretical 2σ limit of detection was 0.241 mg m⁻² h⁻¹. We note however that this is an approximation, derived from the work of Lenschow & Kristensen (1985) 300 who defined twice the right-hand side of Eq. (11) to be an upper limit on flux uncertainty. Equation (11) demonstrates how the variability of ozone and vertical wind are directly related to uncertainty in the measured flux. White noise in the wind measurement is expected to be very small, such that random instrument noise likely represents a significant contribution to the total variance of ozone observed at 10 Hz. Given the relatively low sensitivity of the instrument used in this work (240 counts ppbv ⁻¹ s⁻¹ compared to 2800 counts ppbv⁻¹ s⁻¹ reported by Helmig et al. (2012)), 305 autocovariances were calculated for each averaging period using the 10 Hz ozone data to examine the extent to which variance in ozone concentration is caused by instrument white noise. White noise only correlates with itself at zero lag time, https://doi.org/10.5194/amt-2020-65 Preprint. Discussion started: 27 May 2020 c Author(s) 2020. CC BY 4.0 License. so it can be estimated from the difference between the first and second points in an autocovariance plot (Blomquist et al., 2010). Instrument white noise derived using this approach was found to contribute 45-98% to the total ozone variance (10 th -90 th percentile), with a median σnoise of 1.4 ppbv. A more sensitive ozone instrument could therefore significantly improve 310 the flux uncertainty at a 20-minute averaging period.
Besides the random uncertainty discussed above, systematic errors are also worthy of some consideration. Specifically, whether the highest and lowest frequencies of turbulence have been adequately observed. High frequency information can be lost if measurements are made too infrequently, or if the sample is attenuated significantly in the sample line. Measurements at 10 Hz, as performed here, are widely considered sufficient to observe this high frequency structure. Laminar flow was also 315 avoided through the length of the sample line (Reynolds number = 3000). As a result, the cospectrum in Figure 14 shows no major loss of high frequency information compared to theory. Since fluxes were calculated over 20-minute averaging periods using linear detrending, there is also a chance that low frequency information may not be fully observed. Firstly, using a simple block average in place of linear detrending had little effect on the median flux observed (+1.7%). Using an averaging period of 1 hour instead of 20 minutes gave slightly larger magnitude flux (+4.1%) as well. However, the longer 320 period lead to much greater data loss (22%) to the selection criteria in Sect. 4, hence the 20-minute average was used for this work. This suggests that any low frequency loss is approximately 5% the total fluxa small amount relative to the calculated random uncertainty.

Discussion
For the average meteorological conditions observed during this work, the one-layer model of Fairall et al. (2007) predicts a 325 deposition of 0.037 cm s⁻¹. Here, one-layer refers to considering the surface water to have uniform reactivity to ozone with depth, rather than a thin sublayer at the surface where reactivity is enhanced (a two-layer model). By contrast, the revised 2layer model of Luhar et al. (2018) predicts a deposition of 0.016 cm ⁻¹ for the same conditions using a fixed reactiondiffusion sublayer (δm) of 3 μm. An iodide concentration of ~600 nmol dm -3 would be necessary to yield the observed deposition -much higher than a typical oceanic value of 77 nmol dm⁻³ . However, DOM (Shaw and 330 Carpenter, 2013), chlorophyll (Clifford et al., 2008) and surfactants (McKay et al., 1992) have also been shown to enhance ozone deposition. Therefore the effective pseudo-first order rate constant for the reaction of ozone with water, A, is likely to be significantly higher than accounted for by iodide alone in Eq. (9). Chang et al. (2004) defined this total reactivity as: Where A is the effective pseudo-first order rate constant for the reaction of ozone with water, and are the second order 335 rate constant and concentration of species respectively. We can therefore include an estimate of the effects of DOM reactivity using a typical oceanic DOM concentration of 52 μmol dm⁻³ (Massicotte et al., 2017) and a rate constant of 3.7 × 10⁻⁶ dm³ mol⁻¹ s⁻¹ (average of the values reported by Sarwar et al. (2016) and Coleman et al. (2010)). Doing so increases A https://doi.org/10.5194/amt-2020-65 Preprint. Discussion started: 27 May 2020 c Author(s) 2020. CC BY 4.0 License.
from 544 s⁻¹ to 737 s⁻¹ and leads to increased deposition values of 0.048 cm s⁻¹ and 0.028 cm s⁻¹ for the models of Fairall and Luhar, respectively. 340 The magnitude of the effect of DOM on O₃ deposition remains highly uncertain due to the uncertainties in how O₃ interacts with DOM and surfactants, variability in the sea-surface microlayer (SML) composition, and the effect of temperature. The coastal waters near the PPAO experience large phytoplankton growth during the 'spring bloom' (Cushing, 1959;Smayda, 1998), and the organic content and composition of the SML could be very different compared to the open ocean. The seasonal and spatial variations in these O₃-reactive substances could, in turn drive differences in ozone deposition. For 345 example, Bariteau et al. (2010) reported vd increasing from 0.034 cm s⁻¹ to 0.065 cm s⁻¹ as the waters changed from open ocean into coastal during the TexAQS-2006 cruise. It is unclear how much of the observed gradient is a result of SML composition or of terrestrial influence. Similarly, Ganzeveld et al. (2009) encountered underestimation of coastal ozone deposition in their modelling work when DOM reactivity was omitted, suggesting that this may be a particularly important factor in coastal environments. While the model of Fairall et al. (2007) appears to match our observed vd well, it is possible 350 that this is a consequence of some missing reactivity. Inclusion of DOM causes the one-layer model to overestimate vd, as reported by Luhar et al. (2018).
If the two-layer model provides more accurate deposition velocities with adequate reactivity information, then it shows little dependence upon wind speed in all but the calmest conditions. This would stand in contrast to the one-layer model, and a number of experimental observations including those presented here. 355

Summary and conclusions
An ozone chemiluminescence detector adapted from an Eco Physics® CLD 886 NOx detector was used to measure the ozone deposition velocity to the sea surface at a coastal site near Plymouth, on the southwest coast of the UK. The median observed deposition velocity was 0.037 cm s⁻¹, comparable with past work, but at the upper end of the values obtained by Helmig et al. (2012) during ship-based, open-ocean measurements (0.009-0.034 cm⁻¹). 360 Using observed meteorology with the model of Luhar et al. (2018) yields a predicted vd of 0.018 cm s⁻¹ in the absence of DOM reactivity, or 0.026 cm s⁻¹ with estimated DOM concentration of 52 μmol dm⁻³ and a O₃ + DOM rate constant of 3.7 × 10⁻⁶ dm³ mol⁻¹ s⁻¹. We suspect that the difference from our measured vd is due to the uncertainty surrounding the reaction between O₃ and DOM, and the timing of our measurements, which coincide with the spring bloom and potential enhancements in surface microlayer reactive organics. 365 Elevated deposition was observed at low wind speeds, contrary to predictions (Chang et al., 2004) and to previous observations (Helmig et al., 2012). We attribute this observation to a contribution to vd from land within the footprint during periods of low wind. Periods with wind speeds > 3 m s⁻¹ (corresponding to approximately < 10% land cover in the footprint) were used to evaluate vd. However, the possibility of land influence could not be completely removed, with our oceanic vd estimates potentially overestimated by 8%, even after wind speed filtering. Deposition velocity showed a linear dependence 370 https://doi.org/10.5194/amt-2020-65 Preprint. Discussion started: 27 May 2020 c Author(s) 2020. CC BY 4.0 License. on friction velocity comparable to that predicted by the parameterisation of Fairall et al. (2007), though with considerable scatter. The potential for tidal effects on vd (exposing shoreline and input of river water with different chemical composition) were also examined, though no clear periodicity could be observed, either at the tidal frequency or on a diurnal timescale.
Cross-covariance was used to empirically determine a 2σ limit of detection for each averaging period. This limit of detection was exceeded in 305 out of 491 periods. Auto-covariance of high-frequency ozone data indicated that instrument noise was a 375 significant component in the observed ozone variability, and lowering the noise level would reduce the flux uncertainty.
Future work will link the properties of the sea-surface microlayer in the footprint area to observed O₃ fluxes. A larger dataset may help to elucidate the influence of biogeochemical parameters, seasonal variation and wind speed dependence, which have not been definitively characterised to date.
Code and data availability: the eddy4R software packages used in these analyses are maintained at 380 https://github.com/NEONScience/NEON-FIU-algorithm. 20-minute data have been submitted to the Centre for Environmental Data Analysis (CEDA), awaiting DOI. The corresponding author can be contacted directly for the full highfrequency data.
Author contribution: Experimental work was carried out by DCL, TGB and MY. DCL also conducted the formal analysis and visualisation of the results, with relevant supervision from TGB and MY. SM developed the eddy4R codebase, with 385 ARV providing modification for its use here. RJP provided software for instrumentation and validation of model applications to the data. JDL and LJC supervised the interpretation of the results. The work was proposed by LJC, who also acquired the necessary funding. DCL Prepared the manuscript with all authors contributing to the editing process.
Competing interests: The authors declare that they have no conflict of interest.