Micrometeorological flux measurements were made during the period 7 August–19 December 2012 from a flux tower located on the roof of a building belonging to King's College, University of London (51.511667∘ N, 0.116667∘ W; ground altitude 30 m a.s.l.), on the Strand in central London. Although the site is within the London Congestion Charge Zone (an area encompassing central London requiring road tolls to be paid and hence an area with reduced traffic density), surrounding roads supported a medium to high traffic volume (annual average of 50 000–80 000 vehicles per day; Department for Transport, 2014) with the River Thames situated 200 m to the south. This site is classified as a local climate zone class 2 compact midrise according to Stewart and Oke (2012) (i.e. dense mix of midrise buildings; 3–9 stories; few or no trees; land cover mostly paved; stone, brick, tile and concrete construction materials). Land cover types (in %) were calculated based on the Ordnance Survey map for the surrounding 9 km2 area (Fig. 1) encompassing the site and are roads (37 %), buildings (31 %), other paved areas (14 %), unpaved/vegetation (11 %) and water bodies (7 %).

The sampling inlet and sonic anemometer were mounted on a triangular mast (Aluma T45-H) at approx. 60.9 m (2.3 times mean building height, zH) above ground level. The mean building height was around 25 m and the mast was located on an elevated area in the centre of the roof. A street canyon was located to the NW and an enclosed parking area to the SE, but generally surrounding buildings were of equal height. The sampling point (which we call KCL) is located 37 m west of a sampling point (KSS) that has been used for long-term energy and CO2 flux measurements (Kotthaus and Grimmond, 2012). Although the site is not optimal for micrometeorological flux measurements due to the heterogeneity of the urban canopy, its suitability has been assessed in detail by Kotthaus and Grimmond (2014a, b). This study describes in detail the measurement area and investigates the influence of source area characteristics on long-term radiation and turbulent heat fluxes for the KSS site. They conclude that the site can yield reasonable data on surface-to-atmosphere fluxes.

The weather in 2012 was somewhat cooler than the 1981 to 2010 long-term mean for London during summer and autumn, with several cold fronts bringing up to twice as much precipitation and associated winds as average, suppressing pollution levels. However, during the period of the Olympic and Paralympic games (27 July–12 August and 29 August–9 September 2012) the weather was hot and dry, causing sustained pollution peaks. Winter 2012/2013 was generally warmer and drier in London than the 1981–2010 mean (Met Office, 2013).

The CSAT3 sonic anemometer (Campbell Scientific Inc., Utah, USA) and inlet were faced toward the predominant wind direction (SW) to minimise flow distortion. Data from the sonic anemometer were logged at a frequency of 10 Hz and flux measurements were calculated using 25 min averaging periods. The rotation angle theta (θ), used to correct measurements of the vertical wind velocity for minor misalignment of the sonic anemometer, showed no significant disturbance of the turbulence from interactions with the building when plotted against wind direction. Data were recorded in UTC (universal time coordinated), which is 1 h earlier than local time in summer and coincident with Greenwich mean time in winter. However, all analyses used local time.

VOC concentrations were measured using a high-sensitivity proton transfer reaction (quadrupole) mass spectrometer (PTR-MS) (Ionicon Analytik GmbH, Innsbruck, Austria) with three Varian turbomolecular pumps (see for example de Gouw and Warneke, 2007; Hayward et al., 2002; Lindinger et al., 1998, for more detailed description of the instrument). Air was drawn through an inlet co-located with the sonic anemometer. Sample air was purged through a ∼ 30 m 12′′ OD (38′′ ID) PTFE tube at a flow rate of 81 L min-1 to the PTR-MS, which was housed in a utility room below. The high flow rate ensured turbulent flow was maintained and signal attenuation minimised (Reynolds number, Re =11177). During the campaign, PTR-MS operating parameters were maintained at 1.95 mbar, 510 V and 50 ∘C for drift tube pressure, voltage and temperature respectively to achieve an E/N (E is the electric field strength and N is the buffer gas number density) ratio of 123 Td (1 Td = 10-17 V cm2). This field strength forms a compromise between reagent ion clustering and fragmentation suppression (Hewitt et al., 2003). Further instrument parameters and meteorological conditions are summarised in Table 1. The inlet flow rate into the instrument was 0.25–0.3 L min-1.

The logging program was written in LabVIEW (National Instruments, Austin, Texas, USA) and operated the PTR-MS in multiple ion detection (MID) and SCAN modes for VOC concentrations of nine selected masses and a range of the protonated mass spectrum m/z 21–206 respectively. The sonic anemometer was not directly interfaced with the LabVIEW logging program, requiring the measurements to be synchronised during post-processing through the use of a cross-correlation function between the vertical wind velocity w and the VOC ion counts c. A valve system controlled the measurement cycle, which consisted of 5 min zero air (ZA), 25 min MID followed by 5 min SCAN of sample air and 25 min MID mode. During the ZA cycle, air was pumped through a custom-made gas calibration unit fitted with a platinum catalyst heated to 200 ∘C to provide instrument background values at ambient humidity. In MID mode the quadrupole scanned nine predetermined protonated masses with a dwell time of 0.5 s, to which each of the following compounds were ascribed: m/z 21 (indirectly quantified m/z 19 primary ion count via [H318O+]), m/z 33 (methanol), m/z 39 (indirectly quantified m/z 37 first cluster [H3O+ H2O+]), m/z 42 (acetonitrile, results not shown), m/z 45 (acetaldehyde), m/z 59 (acetone/propanal), m/z 69 (isoprene/furan), m/z 79 (benzene), m/z 93 (toluene), m/z 107 (C2-benzenes) and m/z 121 (C3-benzenes, results not shown). The total cycle time was 5.5 s. Secondary electron multiplier voltage, as well as O2+ (m/z 32) and photon “dark counts” (m/z 25) signals, were monitored weekly.

The PTR-MS cannot distinguish between different compounds with the same integer mass; therefore, isobaric interference can occur. For example, m/z 107 may result from several contributing C8 aromatics: ethyl benzene, (m+p)-xylene, o-xylene and some benzaldehyde (Warneke et al., 2003). Further interferences at measured m/z from additional compounds and fragmentation for this instrument in an urban environment are discussed in Valach et al. (2014). Although the O2+ and water cluster ions were kept < 2 % of the primary ion, interferences from 17O+ isotopes at m/z 33 were taken into account.

Single point calibrations were performed on-site once a month using a certified multiple component VOC gas standard (Ionimed, part of Ionicon Analytik GmbH, Austria), which was validated by cross-calibration with a second independent VOC standard (Apel Riemer Environmental Inc., CO, USA). Before and after the campaign, multistep calibrations were performed with both standards. Standards were diluted with catalytically converted zero air, since cylinder concentrations were approx. 1 ppm ± 5 % uncertainty (Ionimed Analytik) and 0.5 ppm ± 10 % (Apel Riemer). Error propagation resulted in a total calibration uncertainty of < 20 %. Measured normalised instrument sensitivities (SN, Table 1) based on Taipale et al. (2008) were used to convert normalised count rates (ncps) of protonated masses (RH+) to volume mixing ratios (Langford et al., 2010a). Only the o-xylene isomer was present in the Ionimed standard, which was used to determine instrument sensitivities for m/z 107, but sensitivities agreed well when compared with sensitivities for p-xylene present in the Apel Riemer standard. Any remaining humidity effects on calibrations were previously investigated for this instrument and were found to be within the overall calibration uncertainty (Valach et al., 2014). Detection limits of VOC concentrations (Table 2) were calculated according to Taipale et al. (2008).

Fluxes were calculated according to Karl et al. (2002) and Langford et al. (2009, 2010b) using F=1n∑i=1nw′i-tlagΔtw×c′i, where w′ and c′ are the instantaneous fluctuations around the mean vertical wind (w-w¯) and mean VOC concentration (c-c¯), n is the number of VOC concentration measurements per 25 min averaging period (n=273), tlag is the lag time between the wind and PTR-MS measurement due to the transit through the sampling line and Δtw is the sampling interval of the vertical wind speed measurements of the sonic anemometer (10 Hz = 0.1 s). Langford et al. (2015) recently demonstrated that the method used to determine the time lag becomes important where the signal-to-noise ratio of the analyser is poor, showing that methods that systematically search for a maximum in the cross-correlation function within a given window (MAX method) can bias the calculated fluxes towards more extreme (positive or negative) values. Their study recommends the use of a prescribed lag time determined either through the use of a monitored sample flow rate or by using the typical lag time derived by searching for a maximum. Here the prescribed lag times were determined by fitting a running mean to the time series of daytime lag times calculated using the MAX method for acetone, which had large fluxes and the clearest time lags. Prescribed lag times for all other compounds were set relative to that of acetone, accounting for the offset introduced by the sequential sampling of the PTR-MS.

Flux losses due to the attenuation of high and low frequency eddies were estimated for our measurement setup. High frequency flux attenuation was estimated to be on average 11 % using the method of Horst (1997), and a correction was applied. Attenuation from low frequency fluctuations for a 25 min flux period was investigated by reanalysing the sensible heat fluxes for longer averaging periods of 60, 90, 120 and 150 min. The coordinate rotation was applied to the joined files, which acted as a high pass filter to the three wind vectors, confirming that fluctuations of eddies with a longer time period than the averaging time did not contribute to the flux measurement (Moncrieff et al., 2004). The fluxes were compared to the 25 min average fluxes, which had the coordinate rotation applied before joining, again to ensure only turbulent fluctuations of ≤ 25 min contributed to the flux (Fig. A1 in the Supplement). Flux losses due to low frequency attenuation were estimated to be < 1.5 % and, therefore, no corrections were deemed necessary. The error due to the disjunct sampling was estimated by comparing the sensible heat fluxes calculated from the continuous data series with those calculated from a disjunct data series using a set sampling interval of 5.5 s. The continuous data were averaged to match the sampling frequency of the disjunct data (i.e. 2 Hz). The difference between the eddy covariance and DEC sensible heat fluxes was minimal (0.01 %) and thus no additional corrections were applied.

Many of the 25 min resolved flux measurements were close to the limit of detection (LoD), based on 1 standard deviation using the method of Spirig et al. (2005), with an average fail rate of 82 %. Various techniques to statistically analyse or replace values below the LoD have been developed (Clarke, 1998). However, they often result in significant bias, either high or low depending on the value substituted, because values tend to be below the LoD when fluxes are indeed small (Helsel and Hirsch, 1992). In this study, our analysis focused on diurnally averaged flux profiles and we decided not to filter out individual flux values on the basis of being < LoD in order to avoid this bias. When averaging the 25 min flux data it is appropriate to also average the LoD which, as shown by Langford et al. (2015), decreases with the square root of the number of samples averaged (N). Therefore, although the majority of the individual 25 min flux measurements were below the LoD, their diurnal average profiles may exceed the LoD for the average and thus still yield important data on the net exchange of VOCs above the city. LoD‾=1N∑i=1NLoD2. The following describes the additionally applied filter criteria. 25 min flux values with a friction velocity (u∗) < 0.15 m s-1 were rejected (3.4 % of total data) due to insufficient turbulence. The stationarity test and data quality rating methods of Foken and Wichura (1996) and Velasco et al. (2005) were used, and 47 % of the data files were rejected on this basis. The high number of files rejected in the stationarity test is to be expected for eddy covariance measurements over highly heterogeneous canopies, although horizontally averaged canopy morphology recovers some surface homogeneity. Furthermore, the low measurement height used can cause an increased sensitivity towards canopy roughness features resulting in non-stationarity. Since urban environments are inherently not ideal for micrometeorological flux measurements due to their heterogeneity, integral turbulence characteristics of this site were assessed by comparing the measured standard deviation of the vertical wind velocity (σw) normalised by u∗ to the parameters of a modelled ideal turbulence (Foken et al., 2004). Results showed that 99.6 % of all the data was rated category six or better and 0.4 % was rejected using the criteria of Foken et al. (2004). This large pass rate gives further confidence that the measurements were not unduly affected by wake turbulence generated from the structure of the building. Erroneous meteorological data (2.6 % of total) were removed around wind directions of 14–15 ∘ due to minor turbulence interferences from the presence of other sensors on the mast. Depending on the compound, between 40 and 61 % of flux data (N= 1934–2949) passed all of the above quality controls. Exactly 2014 h of concentration data (N=4834) was obtained. For consistency, regression coefficients (R2) were used throughout.

The traffic densities used for the analysis were obtained from a nearby site at Marylebone Road (approx. 3 km to the NW) and consisted of hourly vehicle counts covering the period 7–22 August 2012. The major roads of the Strand and the Thames Embankment surrounding the measurement site support a comparable traffic volume with an annual average of 50 000–80 000 vehicles per day (Department for Transport, 2014) and diurnal patterns in traffic are likely to be similar across central London.

Photosynthetically active radiation (PAR) and CO2 measurements used in the analysis were part of the long-term micrometeorological measurements at the same site and covered the period from August to September for PAR and from August to December for CO2. Average diurnal profiles were calculated for the boundary layer mixing height, which was measured using three LiDARs located on rooftops within central London during an approx. 2-week period in summer and winter 2012 (Bohnenstengel et al., 2015).

Average diurnal cycles of measured VOC fluxes and mixing ratios are shown in Fig. 2 with descriptive statistics for all the data summarised in Table 2. Largest median (interquartile range in parenthesis) fluxes per day were from C2-benzenes and toluene, with 7.86 (0.92–21.8) kg km-2 d-1 and 7.26 (1.83–15.3) kg km-2 d-1 respectively, followed by oxygenated compounds, i.e. methanol with 6.37 (2.99–10.0) kg km-2 d-1, acetaldehyde 3.29 (1.52–5.62) kg km-2 d-1 and acetone 5.24 (2.33–9.62) kg km-2 d-1. Isoprene and benzene showed the smallest median fluxes with 2.14 (0.56–4.85) kg km-2 d-1 and 1.78 (0.06–4.34) kg km-2 d-1 respectively. The highest median mixing ratios were of the oxygenated compounds methanol (7.3 (6.8–7.9) ppb), acetone (0.95 (< LoD–1.36) ppb) and acetaldehyde (0.82 (0.59–1.13) ppb), followed by aromatics (C2-benzenes, toluene and benzene), and isoprene.

Oxygenated compounds commonly have relatively long atmospheric lifetimes and widespread origin including anthropogenic and biogenic sources and photochemistry, resulting in elevated concentrations and less pronounced diurnal profiles (Atkinson, 2000). Most VOC fluxes and concentrations were comparable to or lower than those previously observed in London (Langford et al., 2010b) and other UK cities (Langford et al., 2009), although C2-benzene fluxes and concentrations, as well as isoprene and benzene concentrations, were slightly higher. The discrepancy in isoprene and benzene concentrations is consistent with photochemical loss during transport to the higher measurement height of the previous studies. Compared to other cities such as Houston Texas (Park et al., 2010) and Mexico City (Velasco et al., 2005), VOC fluxes and concentrations were lower with the exception of C2-benzenes, which were comparable or higher; however, it must be noted that C2-benzenes in this study represent the sum of multiple VOC species. Unlike the other studies cited, Park et al. (2010) use relaxed eddy accumulation to measure VOC fluxes and hence the data obtained are not directly comparable with measurements made by EC-based methods.

Diurnal profiles of aromatic fluxes and concentrations presented two clear rush hour peaks during the morning and evening (07:00–10:00 and 17:00–20:00 local time). Concentration peaks are thought to be linked to additional advection of traffic-related pollution from larger commuter roads outside of the city centre, as well as boundary layer effects and photochemistry. VOC concentration measurements at canopy height can be affected by boundary layer depth (Vilà-Guerau de Arellano et al., 2009). The rush hour emission peaks mostly coincide with the boundary layer expansion and collapse and therefore the effect of each factor cannot be separated. The morning concentration peak was slightly higher than the evening peak across traffic-related species even though fluxes tended to be larger during the evening rush hour. Morning emissions enter a shallow nocturnal boundary layer leading to relatively larger concentrations compared to higher afternoon emissions entering a developed boundary layer leading to relatively lower concentrations. This enhanced dilution effect is found more often during summer when the boundary layer mixing height is higher (Fig. 2). Therefore, the regression analyses below only refer to data from August (cf. Sect. 3.1.2 for comparisons with winter). Furthermore, increased photochemical degradation during the day removes VOCs, further contributing to the midday minimum in mixing ratios. The diurnal flux profiles of methanol, acetone, isoprene and to a smaller extent acetaldehyde showed one large peak just after midday (approx. 13:00 local time), which was only reflected in the concentration profiles of acetone and isoprene. Acetaldehyde concentrations presented a slight double peak similar to mixing ratios of aromatics. Methanol has a relatively long atmospheric lifetime and therefore high background concentrations, and hence mixing ratios showed no distinct diurnal profile.

Correlations of VOC / VOC fluxes (R2= 0.40–0.62, p < 0.001) indicated two groups of compounds with good correlations within each group, i.e. compounds related to traffic sources, such as aromatics, and oxygenated and biogenic compounds, such as methanol, acetone and isoprene (Fig. 6 top). Correlations of VOC / VOC concentrations (R2= 0.13–0.84, p < 0.001) showed the highest correlations between traffic-related compounds (R2= 0.45–0.84, p < 0.001) and good correlations between the oxygenated and biogenic compounds (R2= 0.55–0.69, p < 0.001) (Fig. 6 bottom). High correlations between oxygenated VOCs could indicate source commonality or formation mechanisms that depend on similar environmental factors. Scatter plots between aromatic compounds and isoprene/oxygenated compounds tend to show bimodal distributions indicating separate source contributions. Using temperature or, to a smaller extent, PAR as a third variable highlights a temperature or light dependency of the second source supporting the existence of additional biogenic and/or atmospheric sources. In the example of isoprene against benzene the relationship changes with temperature from 2 : 1 to 1 : 2.

Polar annulus and polar plots were constructed for VOC fluxes and mixing ratios respectively and representative compounds are shown (Fig. 9). Polar plots use a generalized additive model to interpolate between wind direction and wind speed averaged data points within the OpenAir package in R (see Carslaw and Ropkins, 2012; Hastie and Tibshirani, 1990; Wood, 2006). Polar annulus plots averaged by time of day instead of wind speed show diurnal variability with wind direction. The majority of the time (83 %), unstable and near neutral conditions prevailed (ζ < 0.2), although the frequency varied between months with 87, 89, 82, 84 and 69 % during August, September, October, November and December respectively. Wind directions with mostly unstable conditions were with W and S winds and near neutral with N or E winds. Mixing ratios were on average highest with low wind speeds (showing a negative correlation) when pollutants accumulate due to reduced mixing, indicating local emissions (Fig. 9, bottom).

Largest fluxes for all compounds were from the NW with either one daytime peak (e.g. isoprene) or two distinct rush hour peaks (e.g. benzene) (Fig. 9, top). On average, fluxes were largest from the W > E ≥ N > S (F statistic = 60.37–227.06, p < 0.001) because of increased emission rates of specific compound sources. Separated by month, fluxes were largest from W > N > E ≥ S in August and September, whereas during October, November and December fluxes followed the pattern W > E ≥ N > S. The flux footprint in this study was relatively small compared to that of measurements previously made at 190 m height from the BT Tower in central London (Langford et al., 2010b). Due to the relatively low measurement height in this study, flux measurements were always closely coupled with the surface layer, unlike measurements by Langford et al. (2010b) that were at times disconnected from the surface layer during stable night-time conditions.

The average length of the maximum flux footprint contribution (Xmax) was around 330 m and 90 % of all the fluxes (X90) originated from within 900 m. The median footprint area was 1.8 km2. This established that the majority of emission sources contributing to the measured fluxes must have been local. Additionally, the selected emission grid (cf. Sect. 2.3.1 above) encompassed 97 % of the footprint with S and W wind directions but only 80 and 84 % during E and N winds. Grid square 5 represented the maximum contribution area because it encompassed the measurement point. Average footprint contributions (mean ± SD) comprised of grid squares 1 (2 ± 4 %), 2 (5 ± 7 %), 4 (4 ± 5 %) and 5 (52 ± 31 %) during S and W wind conditions, squares 6 (4 ± 9 %) and 9 (4 ± 10 %) indicated E wind conditions and square 8 (18 ± 27 %) N wind conditions. During October contributions from square 9 increased to 10 % and were more frequent at 30 % in December. Squares 3 (0.6 ± 2 %) and 7 (0.9 ± 2 %) provided minimal average contributions.

The River Thames to the S may have caused the low fluxes associated with S winds (i.e. squares 1, 2 and 3). Contributions of traffic-related compound fluxes were statistically significant from the W (i.e. squares 4, 5, and 7), followed by the N (square 8) and E (squares 6 and 9) likely from the nearby heavily trafficked roads (Kingsway, Charing Cross, Strand and Blackfriars areas respectively). Biogenic compound fluxes were highest from the W and E, which coincides with significant nearby green areas within the flux footprint.

Correlations of fluxes with grid square contributions in the footprint can also give information on emission source strengths within the respective grid square (Fig. 1). Generally positive correlations with fluxes across most compounds were seen from the W (squares 4, 5 and 7), confirming that high emission rates from sources within these grid squares were driving the large fluxes. The strongest correlations of fluxes with contributions from squares 4, 5 and 7 were seen during October and November (R2= 0.40–0.46, p < 0.001), especially for masses associated with biogenic sources (m/z 33, 45, 59 and 69). Square 8 showed positive correlations for benzene and only in August for all compounds. Correlations of fluxes with contributions from squares 1, 2, 3, 6 and 9 were negative, indicating weaker emission sources in these squares or increased VOC deposition.

Highest mixing ratios with wind direction were from E > N ≥ W > S for traffic-related compounds, whereas oxygenated compounds/isoprene followed a similar pattern as the fluxes of W ≥ E > N ≥ S (F statistic = 47.49–86.95, p < 0.001). Easterly winds in London are often associated with synoptic conditions that bring European continental air masses to the UK, resulting in higher background concentrations. Furthermore, since the boundary layer was on average more stably stratified and mixing heights were lowest (640 ± 80 m) with E wind conditions, it is likely that pollutant concentrations were allowed to build up, resulting in the observed higher concentrations to the E for the more ubiquitous compounds, whereas concentrations of compounds with biogenic contributions additionally had strong sources to the W, such as several green areas (St. James' Park, Hyde Park and Regents Park; total 331 ha).

The LAEI and NAEI produce yearly emission estimates over the 1 km2 OS grid for a range of pollutants and emission sources. Total VOC emission estimates are provided, but only benzene and 1,3-butadiene are estimated separately. Measured emissions were compared with annual estimated emissions for the above OS grid area selection from 2012 for benzene using the LAEI and indirectly speciated VOCs of the NAEI. Using the average flux footprint, the grid square estimates were compared with the scaled flux measurements from the equivalent area (Fig. 10).

LAEI emission estimates included contributions from major (69 %) and minor roads (4 %) as well as evaporative emissions (27 %) (LAEI). No data were available on cold start emissions for benzene. The calculated standard errors provided some uncertainty approximation. Measured fluxes compared well with emission estimates, although the LAEI predicted slightly smaller benzene fluxes. Comparisons of fluxes with wind directions (Sect. 3.3) agreed well with the LAEI emission estimates for the respective grid squares with highest emissions from squares 4, 5, 7 and 8 (i.e. W and N directions). This comparison assumes that the benzene fluxes during the measurement period were representative of annual emissions with any significant seasonal variation in benzene emission rates captured in this 5-month period. Section 3.1.2 confirmed that there was little month-to-month variability in the benzene flux.

Using speciated VOC emission contributions (percent of total VOC emissions) for 2006 (Bush et al., 2006) and emission maps from 2012 for total non-methane VOC emissions, speciated estimates could be compared with observations (Fig. 10). The NAEI includes a wide range of emission sources divided into 11 SNAP (Selected Nomenclature for sources of Air Pollution) sectors including industrial, commercial and residential processes, transport, waste treatment, solvent use, point sources, agriculture and nature, although the latter two were unavailable for the London urban area. NAEI estimates for benzene exceed the LAEI due to the inclusion of a wider range of sources beyond traffic-related emissions. Total C2-benzene emission estimates consisted of ethyl benzene, (m+p)-xylene and o-xylene. Benzene and methanol emissions agreed very well; however, for all the other compounds, estimated emissions were significantly lower than the measured fluxes. Uncertainties related to the measurements, such as isobaric interferences within the PTR-MS could have contributed to measurement overestimation, whereas uncertainties within the modelled emissions and the use of older speciation values may have impacted the estimates. In the case of isoprene, only minimal emissions are assumed, which do not include the biogenic sources that contributed to the measured fluxes. It is also likely that some of the m/z 69 signal could be attributed to cyclic alkenes, but Sect. 3.1.3 showed that biogenic isoprene provided a significant contribution during August and September in central London.