VOC emission rates over London and South East England obtained by airborne eddy covariance †

Volatile organic compounds (VOCs) originate from a variety of sources, and play an intrinsic role in in ﬂ uencing air quality. Some VOCs, including benzene, are carcinogens and so directly a ﬀ ect human health, while others, such as isoprene, are very reactive in the atmosphere and play an important role in the formation of secondary pollutants such as ozone and particles. Here we report spatially-resolved measurements of the surface-to-atmosphere ﬂ uxes of VOCs across London and SE England made in 2013 and 2014. High-frequency 3-D wind velocities and VOC volume mixing ratios (made by proton transfer reaction – mass spectrometry) were obtained from a low- ﬂ ying aircraft and used to calculate ﬂ uxes using the technique of eddy covariance. A footprint model was then used to quantify the ﬂ ux contribution from the ground surface at spatial resolution of 100 m, averaged to 1 km. Measured ﬂ uxes of benzene over Greater London showed positive agreement with the UK ’ s National Atmospheric Emissions Inventory, with the highest ﬂ uxes originating from central London. Comparison of MTBE and toluene ﬂ uxes suggest that petroleum evaporation is an important emission source of toluene in central London. Outside London, increased isoprene emissions were observed over wooded areas, at rates greater than those predicted by a UK regional application of the European Monitoring and Evaluation Programme model (EMEP4UK). This work demonstrates the applicability of the airborne eddy covariance method to the


Introduction
Non-methane volatile organic compounds (VOCs) are a class of organic molecules that are sufficiently volatile to enter the atmosphere. Some, including benzene, are directly toxic to humans, while others are reactive in the atmosphere producing secondary pollutants such as ozone and particles, and hence impact air quality. 1 Globally, most VOCs originate from the terrestrial biosphere, but in cities and other areas of high anthropogenic emissions, pollution sources usually dominate.
Emissions of VOCs in cities such as London have been well studied, with vehicles recognised as a signicant source, 2-4 either via emission in the exhaust gas or by the evaporation of unburnt or partially burnt fuel. 5-7 Na et al. (2005) found that 58% of aromatic VOC emissions in Seoul, Korea, originated from vehicle exhausts. 6 Measurements made during summer 2001 in Sacramento by Rubin et al. (2006) found fuel evaporative emissions could contribute up to 29% of total vehicular VOC emissions. 7 Quantifying the emission rates of individual VOCs is a prerequisite to their cost-effective and successful control and this is routinely attempted by the construction of bottom-up emission inventories such as the UK's National Atmospheric Emissions Inventory (NAEI). 8 In the case of VOCs of biogenic origin, the most widely used models of emissions are derived from that of Guenther et al. (1995). 9 The validation of emission inventory estimates is difficult as they are built up from many emission factors and activity rates. Measurements of atmospheric volume mixing ratios do not allow their direct validation but require a model of atmospheric chemistry and transport to infer emission rates. Micrometeorologically-based surface-to-atmosphere ux measurements can allow direct validation, but normally only at one point on the surface, which is not representative of an entire city. For example, Langford et al. (2010) assessed VOC emission uxes from central London by making ux measurements at the BT Tower. They found good agreement for benzene, toluene and C 2 alkyl-benzenes uxes to NAEI emission estimates generated using a spatial footprint model 2 but the study was limited by conned spatiality.
Measurements from aircra allow for larger spatial assessment compared to tower measurements. Due to the high speed at which aircra move, coupling aircra measurements to the surface has different challenges compared to measuring at a tower site. Small aircra which can y low and slow have previously been used to study atmospheric turbulent structure through airborne eddycovariance. 10,11 One such approach employs disjunct eddy-covariance (DEC), as rst described by Karl et al. (2002). 12 DEC allows for sample processing time to be slower than the air sampling rate (<1 s) causing data to become discontinuous, while still being able to fully capture the turbulent statistics required for eddycovariance. [12][13][14] Karl successfully implemented DEC on-board an aircra, to assess city-wide emissions of toluene and benzene from Mexico City, and found that sampling rates above $2 Hz captured the majority of eddy ux contributions. 13 Misztal et al. (2014) also investigated VOC emissions from an aircra via DEC.
High spatiality isoprene uxes were quantied over a large region of California. The study gave direct assessment of a statewide emission inventory used in predicting ozone concentrations, which previously had not been possible on such a large scale. 15 Shaw et al. 2015 assessed the temporo-spatial distributions of benzene, toluene NO x mixing ratios over London using an aircra in 2013. The strong correlation between aircra and ground site measurements highlighted the applicability for using aircra to assess ground level pollution sources from London, with vehicles predicted to be a major source for measured VOCs over central London. 16 During the same ights, Vaughan et al. (2016) measured uxes of nitrogen oxides (NO x ) over London and found that the NAEI greatly underestimated NO x uxes, mainly because road vehicle sources were underpredicted by the Inventory. 17 This was the rst attempt to validate the NAEI at the city scale, and the present study continues this by directly assessing the NAEI, and an inventory of biogenic VOC emissions, to provide a direct validation of the current ability of emissions inventories to predict VOC emissions in SE England. The work has important policy implications as it highlights current weaknesses in the ability to estimate VOC emission rates to the atmosphere.

Methodology
The Ozone Precursor Fluxes in an Urban Environment (OPFUE) campaigns ran during July 2013 and July 2014. The strategy for both campaigns was to determine the highly spatially resolved emissions of a range of ozone precursor species from both anthropogenic and biogenic sources, over Greater London and SE England. 16

Flight strategy
The OPFUE campaign (24 th June-9 th July 2013) consisted of 12 research ights aboard the Natural Environment Research Council's Airborne Research and Survey Facility (ARSF) Dornier-228 aircra, based at Gloucestershire Airport. Each ight involved replicated legs across Greater London and the rural area to its south (described as South Sussex below). Fig. 1a shows all ight transects across Greater London ($50 km length). Fig. 1c shows all ight transects across the rural South Sussex region ($50 km length). At the end of every ight a prole ascent was conducted to high-altitude (up to 2400 m) allowing an assessment of the boundary layer height and for instrument calibrations to be conducted in clean free tropospheric air.
In July 2014 (9 th -16 th July), similar ight paths were used over London, with extra legs added to form an incomplete gure of 8 ( Fig. 1b) with each transect $120 km long. The expanded ight path allowed for greater spatial coverage, passing over important emission areas, including the industrialised area around the Dartford crossing over the River Thames. Flights over the South Sussex region followed the same ight path as used in July 2013, shown in Fig. 1d.
In summary, total spatial coverage over Greater London during 2013 was $400 km 2 and $700 km 2 during 2014. Over rural South Sussex, total spatial coverage of $170 km 2 was achieved during 2013 and $180 km 2 during 2014.  18 Accurate aircra position data were obtained using an Inertial-Position and Altitude System (IPAS 20), which generates aircra coordinates and altitude data at the same rate as the AIMMS-20.
2.2.2. Volatile organic compound measurements. Measurements of VOC mixing ratios were made using a high sensitivity proton transfer reactionmass spectrometer (PTR-MS, Ionicon, Innsbruck, Austria). In-depth theory and design of PTR-MS have been well documented in previous studies. 13,15,[19][20][21] The instrument used in this study has been described in detail elsewhere de Gouw and Warneke 2007;Hewitt et al. 2003;Lindinger and Hansel 1997;Shaw et al. 2015) therefore only instrument set up, operation and ight modications are outlined here. A pressure controller (Bronkhorst) was added to the instrument to regulate the inlet ow (50-500 STP sccm), such that pressure upstream of the controller was maintained at a constant value. Thus, the PTR-MS dri tube pressure was independent of uctuations in ambient pressure caused by varying  ight altitude. Ambient sample air was only exposed to heated (70 C) Teon and stainless steel tubing, minimizing memory effects, inlet losses and the build-up of impurities in the inlet system. Considerable efforts were made to prevent VOC contamination of the PTR-MS inlet during operation on the ground and during take-off. On the ground, the PTR-MS inlet remained closed and all sample tubing capped. VOC measurements were obtained at a sampling rate of 5 Hz and a repetition rate of $2 Hz. The target protonated masses and likely contributing compounds were m/z 57 (methyl tert-butyl ether, MTBE), m/z 69 (isoprene), m/z 71 (methyl vinyl ketone/methacrolein, MVK/MACR), m/z 79 (benzene), m/z 93 (toluene), m/z 107 (C 2 alkyl-benzenes ethylbenzene/benzaldehyde/xylene isomers) and m/z 121 (C 3 alkyl-benzene isomers). Additionally, both the primary ion m/z 21 (H 3  Td. For ights at $360 m a.g.l, the m/z 21 primary ion count ranged between (4-7) Â 10 7 ion counts per second (cps) with an average of 6 Â 10 7 . Ion counts of m/z 32 ranged between (0.8-3) Â 10 6 cps, with an average of 2 Â 10 6 cps, which represented 3% of the primary ion signal. Ion counts of m/z 39 ranged between (1-5) Â 10 6 cps with an average of 3 Â 10 6 cps, which represented 6% of the primary ion signal. Due to O 2 + and NO + impurities in PTR-MS ($1-5%), elevated concentrations of butane and higher alkanes are going to be detected at the most abundant ion fragments such as at m/z 57 and 71, which will contribute to the ion signals of MTBE and MVK/MACR. 22 In PTR-MS several major vehicle exhaust emission components are expected to contribute to the ion signal at m/z 57, including MTBE and C 4 H 8 (butene) isomers. The butenes react to form protonated molecular ions at this mass while greater than 95% of the MTBE proton transfer reaction products fragment to the mass 57 ion. 23 It is now recognized that the concentrations derived from ion intensity measurements at m/z 57 involving petroleum vehicular emission predominately reect the sum of the butene isomers and MTBE. However, Rogers et al. (2006) concluded that m/z 57 ion receives signicant intensity from a wide variety of neutral components including the isomeric butenes, acrolein, higher order alkenes and alkanes particularly abundant in vehicular diesel exhaust. 24 The ion signal at m/z 57 (MTBE ‡) was therefore assumed to represent the vehicular emission source across Greater London with mixing ratios calculated assuming k c MTBE ‡ of 2 Â 10 À9 cm 3 s À1 .

Paper Faraday Discussions
This per ppbv (16-18 ncps) for C 2 alkyl-benzenes, C 3 alkyl-benzenes, benzene, toluene, isoprene, total monoterpenes and MVK/MACR respectively. Instrument uncertainties were 19 AE 5%, 17 AE 5%, 16 AE 5%, 21 AE 9%, 6 AE 8%, 16 AE 4% and 7 AE 9% for C 2 alkyl-benzenes, C 3 alkyl benzenes, benzene, toluene, isoprene, MVK/MACR and total monoterpenes respectively, calculated using the standard deviation of linear regression (S m ) of pre-ight calibrations. Instrument limits of detection (LoDs) were determined by the method outlined by Taipale et al. (2008) and were 23 AE 7 pptv, 20 AE 9 pptv, 13 AE 8 pptv, 18 AE 11 pptv, 18 AE 5 pptv, 19 AE 7 pptv and 39 AE 6 pptv for C 2 alkyl-benzenes, C 3 alkyl-benzenes, benzene and toluene, isoprene, MVK/MACR and total monoterpenes respectively. 25 During ights, ambient air was sampled from the forward facing isokinetic inlet along a heated (70 C) 5 m 1/4 00 Teon (PFA) tube pumped by a stainless-steel diaphragm pump (Millipore) at a ow-rate of 22 L min À1 . A portion of this ambient air ($300 sccm) was diverted into the pressure controlled inlet of the PTR-MS instrument such that the overall delay time was <3 s. To determine blank VOC mixing ratios, the remaining ambient air was purged into a custom built zero air generator, which consisted of a 3/8 00 stainless steel tube packed with 1 g of platinum coated quartz wool (Elemental Microanalysis) which efficiently removes VOCs. 26 The zero-air generator was operated at 350 C and 30 psi for the duration of the ights to maintain optimal operating conditions. The catalytic converter does not remove water vapour from the sample stream, which is of importance as background impurities may depend upon sample air humidity. Zero air was periodically back-ushed through the inlet system to determine the instrument background signal.

Disjunct eddy covariance
Eddy covariance is a well-dened technique for quantifying trace gases emissions from the surface to the atmosphere. Here we use a moving aircra to gain greater spatial coverage than possible from a tower, hence continuous wavelet transform (CWT) was chosen to generate uxes with high time (and hence spatial) resolution. CWT gives both frequency and time-resolved ux information, compared to frequency resolved information only from the more conventional Fast Fourier Transform (FFT) approach. CWT gives advantages over conventional FFT, with stationarity not needing to be conserved and time resolved spectral contributions quantied. All calculated mixing ratios from the PTR-MS are dry mixing ratios, with uxes calculated using this methodology not requiring water vapour content correction. Prior to calculating the ux, lag time correction via cumulative covariance was carried out.
Using identical methodology as proven successful in previous studies, 15,17,27,28 we conduct CWT via eqn (1), which denes the transformation of the discrete sequence of data x(n) (being either the instantaneous change of concentration or vertical wind speed data from its mean), using the daughter scaled Morlet wavelet: 29 j denotes the mother wavelet, with a and b being parameters scaling and localising the wavelet in frequency and time. p is the wavelet normalized factor. Due to differences in sample location and time taken for sample to reach the PTR-MS, lag-time correction was conducted. Through cross-correlation analysis, the correlation coefficient between instantaneous vertical wind and concentration was quantied as a function of lag time. Aer correction, maximum correlation between the two would be at time ¼ 0. Each VOC species showed different lagtimes, requiring individual RF transects to be corrected independently. Any transect not showing a clear coefficient peak was dismissed. CWT calculates eddy contributions over all frequency periods with respect to the distance along the ight path, which is shown as a global cross-spectrum. Spatial scales used were long enough to capture all ux frequencies (0-128 s), with no loss of low frequency ux contributions. Average ux is then calculated across all frequency periods. Fig. S1 in the ESI † shows the global cross-spectrum for a C 2 alkyl-benzenes ight leg over Greater London, with the "hotter" coloured regions showing positive uxes and "cooler" colours showing deposition (negative uxes).
For ux data quality control, conventional FFT ux analysis was run at the same time as the CWT analysis. A comparison was then made between the two. By analysing the co-and cumulative-co-spectra, we are able assess the need for any frequency correction. Fig. 2 shows the CWT and FFT ux spectra across all frequencies for heat and C 2 alkyl-benzenes uxes for one run over London. For all analysed runs, we found spectra followed the general trend shown, with all major ux contributions below 0.1 Hz, inferring that high frequency correction for ux loss was not needed. Any major deviations found between FFT and CWT spectras would result in data being discarded. Good uxes were deemed to display similar spectra to that of the standard FFT, showing no high frequency loss, and a ratio of FFT : CWT of 0.7-1.3. The calculated FFT ux data would still be affected by nonstationarity and inhomogeneity. 15 However, this comparison still points to whether the CWT ux analysis had been successful.
A nal data quality process involved the removal of any ux contributions which sat outside of the cone of inuence (COI). The COI is the area of the wavelet cross-spectrum which is free from edge effects, with the area outside it being of lower ux quality. 30 Edge effects were found to be greatest at the beginning and end of each ight run. To help reduce this, the beginning and end of each run was padded with instrument zeros before conducting CWT analysis; however, uxes calculated at the beginning and end of each run will still be prone to greater uncertainties than uxes from the more central parts of each run. 2.3.1. Eddy covariance ux errors. To reduce disjunct errors for the calculated uxes, ight tracks were kept relatively long ($50 km) and the frequency of VOC sampling greater than $2 Hz. Due to ight restrictions over London, stacked altitude ight legs were not possible, hence we were not able to quantify chemical losses occurring between the surface and the point of measurement by directly observing the ux gradient. The vertical divergence of NO x uxes over London due to reaction of NO 2 with the hydroxyl radical occurring between the surface and the average ight altitude of 360 m has previously been assessed as 1-2%. 17 Our reported uxes for the more reactive VOCs are therefore underestimates by at least this amount. 28 The calculated error associated with each ux transect considers random, systematic and disjunct errors as an average across the transect, using the methods of Lenschow et al. (1994) and Karl et al. (2013). 28,31 It should be noted that the error associated with instantaneous uxes was much greater than the average error across each ight track, with calculated errors increasing to 100-300%. A full summary of calculated ux errors as an average over each transect is shown in Table 1. 2.3.2. Isoprene loss correction. Biogenic emissions of isoprene were measured during both campaigns over SE England. Isoprene reacts about ten times faster with hydroxyl (OH) radicals than benzene, hence reactive losses of isoprene during the time taken to reach the ight level from the surface require correction.  quantied the vertical ux loss of isoprene through oxidation with the hydroxyl radical. 32 For the 2013 ight data, we used the same methodology as Karl et al., with known or estimated OH concentrations (2.0 Â 10 6 molecule per cm 3 ) and boundary layer height during each ight used to estimate ux loss correction factor estimates. Vertical ux divergence was calculated to be 7.15 Â 10 À5 mg m À2 h À1 per vertical metre. Overall, the average correction for isoprene during 2013 ights was $33% at an altitude of 360 m. For the 2014 ights, loss of isoprene due to oxidation was corrected using the measured methyl vinyl ketone/methacrolein (MVK/MACR) ux. MVK and MACR are direct oxidation products of isoprene, and hence this ux ratio gives a direct estimate of isoprene ux loss. At measured NO concentrations of <1 ppbv, MVK/MACR formation from isoprene oxidation was calculated to be 2.4 at 0.1 ppbv NO using the Master Chemical Mechanism v 3.3.1, 33 which was then used to correct the isoprene uxes.

Footprint calculation
A footprint model was used to calculate the origin of the emissions as measured by the aircra: the scalars of interest are mixed and transported downwind during the interval from the time of emission from the surface to their registration at the aircra position. The footprint model permits calculating an ensemble trajectory for each aircra observation, and thus to spatially attribute the measured emissions to surface sources in a probabilistic framework. Here, a computationally efficient parameterization for along-wind dispersion (Kljun et al. 2004) is coupled with a cross-wind dispersion function to permit 2-dimensional source area denition. Metzger et al. (2012) provide an in-depth description of this coupled parameterisation alongside a comparison with alternative footprint models. 34,35 The parameterisation depends upon friction velocity u*, measurement height z, standard deviation of the vertical wind s w , standard deviation of the crosswind wind s v , aerodynamic roughness length z 0 , and the boundary layer depth z i , and is valid in the range À200 # z/L # 1, u* $ 0.2 m s À1 , and 1 m # z # z i . As previously used in Vaughan et al. For the aircra ux observations, each 1 km along the ight track, the model calculates a surface weighting matrix at 1 km resolution, which is the same as the emission inventory grid resolution. The surface weights for each observation sum to unity, which permits cell-wise multiplication of the surface weights with the emission inventory, and subsequent aggregation over the entire grid extent. This results in the estimate of the emission strength that should be detected by the aircra, based on the emission inventory. 35,39 Additional methodological detail is provided in Metzger et al. (2013).

NAEI comparison
Measured anthropogenic uxes were compared to emission estimates of benzene and total non-methane hydrocarbons (NMHCs) from 11 source sectors made by the UK's National Atmospheric Emissions Inventory at 1 km 2 spatial resolution. 40 The source sectors used in the NAEI include solvent use, road transport, agriculture and industrial processes. NAEI road transport estimates follow the COPERT 4 emission factor model, which can be found in the European Monitoring and Evaluation Programme/European Economic Area (EMEP/EEA) air pollutant emission inventory guidebook. 41 A full description of NAEI model methodology can be found in Bush et al. (2008). 8 Comparison to measured uxes using the NAEI was done using year-specic inventory data for either 2013 or 2014.
Emission estimates at 1 km 2 resolution from each source sector were quan-tied using the footprint model described above. As NAEI emission estimates are annual averages, month-of-the-year, day-of-the-week and time-of-day scaling is required to produce hourly average emission estimates for the specic days of the year on which ux measurements were made. Each source sector was scaled separately, with each scaling factor accounting for temporal variation in emissions for each month, day and hour. 40 Scaled estimates for each source sector were then summed up for every 1 km along the ight track, giving time-of-day emission estimates along the ight track. Measured uxes of toluene and C 2 alkyl-benzenes were compared to NMHCs estimates from the NAEI. As before, annual estimates were scaled to time-of-day using the described scaling factors, giving temporally-resolved hour of the day estimates. To extract specic VOC emission estimates from the inventory, the Passant (2002) study was used. 42 This gives detailed characterisation for each NAEI source sector, with percentage contribution of each VOC quantied. Each source sector was scaled using the relevant percentage contribution, with all sectors summed up every 1 km along the ight track, giving time-of-day emission estimates for toluene and C 2 alkyl-benzenes.

EMEP4UK comparison
Emission measurements of biogenically-derived isoprene were compared with emission estimates made using the European Monitoring and Evaluation Programme model (EMEP MSc-W) for the United Kingdom (EMEP4UK). The model gives 5 km 2 emission estimates of isoprene for the whole of the UK with hourly resolution using near-surface air temperature and photosynthetically active radiation (PAR) ux. [43][44][45][46] A detailed description of EMEP4UK biogenic isoprene emissions can be found in Simpson et al. (2012). 47 Using the described footprint model, time-of-day estimates of isoprene emission rates were quantied using interpolated 1 km 2 EMEP4UK estimates. Comparison using the EMEP4UK model to measured isoprene uxes was done using year-specic inventory data either 2013 or 2014.

Anthropogenic VOC uxes over London
Measurements conducted during the 2013 ights over London focused on two anthropogenic VOCs, benzene and toluene. Both compounds are highly volatile with a wide range of previously identied emission sources in London. 2,16 Meteorological conditions during the ights are summarised in Table S1, † with prevailing wind directions predominately from the SW and temperatures above 20 C. Inverse distance weighing (obtained using ArcGIS) at 500 m, was used to interpolate all measured uxes from all RF transects, generating spatially-average benzene and toluene uxes, as shown in Fig. 3a. Approximately 4 hours of benzene uxes were obtained from 5 RFs, covering a total area of $400 km 2 over London. These uxes highlight the spatial heterogeneity of VOC emission over central London, with maximum emission rates of 0.20 mg m À2 h À1 . Higher benzene uxes were also observed around the M25 ring road around Greater London and industrial areas of London, as shown in Fig. 3a. The mean benzene ux from all RFs was 0.051 AE 1 mg m À2 h À1 .
Toluene uxes were quantied during from 33 ight ux transects in 6 RFs, giving $5.5 hours of ux data with total coverage over London of $400 km 2 . The observed structure in toluene emissions was comparable to that of benzene, with highest averaged emissions of 0.30 AE 1 mg m À2 h À1 found over central London.
Toluene emissions across all of London were always higher than those of benzene, with a mean toluene ux from all RFs of 0.18 AE 1 mg m À2 h À1 . Emission rates exceeding 1.0 mg m À2 h À1 of toluene were observed in central London ( Fig. 3b). In 2014, toluene uxes were obtained from 5 RFs, with >75% of transects generating uxes. A total of $3 hours of toluene uxes with coverage of $700 km 2 over London were obtained. The emissions prole was comparable to the previous year's data, with the highest measure toluene ux of 0.30 mg m À2 h À1 . Fig. 3d shows interpolated toluene uxes from all transects across London. No benzene ux data were available from the 2014 ights due to reduced instrument performance.
As Suburban London showed an average b/t ratio of 0.43, indicative of vehicular exhaust emissions as the main VOC source. Within the suburban region of London is the M25 motorway ring road, which encompasses the entire Greater London area. The highest b/t ratio observed within this area was 0.82, indicating direct VOC emissions coming from the motorway. As discussed by Rogers, higher b/t ratios can also be attributed to correct operation of vehicle emission control systems, i.e. the catalytic converter. Ratios will however depend on the performance of the converter and the fuel used. 23 For Southwest Greater London, the average b/t ratio was lower at 0.28. Ratios below 0.41 are indicative of other sources being dominant. 13,23 The average b/t ratio for central London was also 0.28, suggesting similar sources are responsible for these two regions' VOC emissions. Northeast Greater London showed the lowest average b/t ratio of 0.20, which as discussed by Karl Table S1. † Flux measurements were not possible during RFs 1 and 2 due to adverse weather conditions and instrument problems. Prevailing wind directions during many of the RFs over London were from the SW, with high summer air temperatures. The higher order benzenoid compounds, C 2 alkylbenzenes and C 3 alkyl-benzenes, were measured during the 2014 ights. Fluxes of C 2 alkyl-benzenes were measured during 5 RFs, giving 1.5 hours of ux data and a total spatial coverage of $450 km 2 . High emissions were observed over central London and industrial areas of the Greater London region (Fig. 3c), with a maximum ux of 1.00 mg m À2 h À1 . Measurements of C 3 alkyl-benzenes uxes were obtained during 2 RFs, giving $30 min of data, over a total area of 150 km 2 . petroleum fuels. 49,50 Maximum observed emissions of C 3 alkyl-benzenes were found to be $1.00 mg m À2 h À1 , observed over central London. Fig. 3e shows interpolated instantaneous C 3 alkyl-benzenes uxes from all RF transects. MTBE ‡ uxes were measured during two (2014) RFs, giving total spatial coverage of $100 km 2 . Methyl tert-butyl ether (MTBE) is a key additive in petrol fuel, acting as an anti-knocking agent. 50 The highest measured emissions of MTBE ‡ were observed over areas with similarly high toluene emissions, with up to 1.00 mg m À2 h À1 of MTBE ‡ observed. Fig. 4 shows latitude averages of all ight transects (1 km resolved) of MTBE and toluene emission, with both in good agreement. The average MTBE ‡ ux for Central London was calculated to be 0.21 mg m À2 h À1 , compared to 0.20 mg m À2 h À1 for toluene, suggesting an approximate 1 : 1 ratio. In 2008 in the UK, the average MTBE concentration in petroleum was 3.4% (v/v). 51 As MTBE ‡ is a good indicator of the relative importance of vehicular emissions, the observed toluene emissions from central London can be attributed mainly to vehicle emissions.

NAEI comparison
Benzene uxes quantied for all RFs in 2013 were compared to 1 km 2 footprint estimates of benzene from the NAEI, scaled to time-of-day as described above. The lateral displacement of the footprint from the ight track due to the horizontal boundary layer wind extended 4-12 km upwind and is shown in Fig. 5. Fig. 6a shows 1 km latitude averages of measured benzene uxes and NAEI benzene emission estimates with the standard deviation shaded. Good agreement between measurements and NAEI emission estimates was found for across most of London. Due to the signicant uncertainty of benzene ux at 1 km 2 resolution (100-300%), benzene ux and NAEI estimates were averaged across the entire ight path, reducing uncertainty. The average measured benzene ux (0.051 mg m À2 h À1 ) and the average NAEI estimate (0.037 mg m À2 h À1 ) agree within the bounds of measurement uncertainty, with a ratio of 1.40. To investigate the spatial emission distribution over London, the ight path was split into four zones as previously discussed, reducing the overall error associated with 1 km averages. The suburban London region was found to have the largest difference between averaged benzene ux and NAEI. The average ux was found to be 3 times higher than predicted NAEI estimates. Greater and Central London areas showed better agreement, with ratios of average ux/NAEI below 1.70. Toluene uxes measured in 2013 were also compared to the NAEI emission estimates. Fig. 6c shows 1 km 2 latitude averages of toluene ux and NAEI estimates with standard deviation shaded. In suburban areas of London, measured uxes are a factor of 1.4 higher than NAEI estimates. In the other zones, measured uxes are 30 to 50% lower than estimates. When averaged across the whole of London, the NAEI emission estimates for toluene were a factor of two higher than the average measured ux (0.39 mg m À2 h À1 from the NAEI compared with the observed ux of 0.18 mg m À2 h À1 ).   Comparison of 1 km 2 latitude average toluene ux (2014) and NAEI estimates are shown in Fig. 6d, with good agreement for outer London. Again, higher NAEI emission estimates are observed for most of London. Average NAEI toluene estimates are a factor of 2 higher compared to measured uxes, with average NAEI estimate of 0.28 mg m À2 h À1 , compared to the average measured value of 0.14 mg m À2 h À1 . With stricter legislation now restricting emissions from sources such as road transport, solvent emissions are now predicted to be the main source for NMHCs in London. 52 As observed in 2013, Greater and Central London areas show measured uxes to be 50% less than predicted by the NAEI, further implying the need for renement of emission sources for toluene within the NAEI. C 2 alkyl-benzenes compared to 1 km 2 NAEI estimates (Fig. 6b) show good agreement across London. Average NAEI estimates for C 2 alkyl-benzenes are higher than the average measured C 2 alkyl-benzenes ux. This overestimation is likely due to errors in the source sector factors for C 2 alkyl-benzenes emissions. Average suburban measured C 2 alkyl-benzenes ux agreed well with the NAEI estimates, compared to the average Southwest Greater London average ux which was 70% lower than NAEI estimates. Table 2 gives an in-depth review of measured uxes and NAEI footprint estimates for both years' ights, giving averages and ratios for each of the four dened London zones.

Biogenic VOC uxes
Measurements of several VOCs of predominantly biogenic origin were made over the South Sussex region of Southern England. Fig. 1c and d show all ight transects conducted during the two campaign periods. Fig. 7 shows the calculated footprint extent from the ight track overlaid onto the UK's National Land Classication map at 25 m 2 resolution. This shows that for a typical ight over the region, areas of mixed broadleaved woodland containing Quercus and other tree species known to emit isoprene occur within the footprint. [53][54][55] Measured isoprene uxes varied from 0.20 to 2.00 mg m À2 h À1 (Fig. 8a and b). In general, the temperature and the ux of photosynthetically active radiation (PAR) were higher during the 2013 ights than during the 2014 ights. Fig. 9 shows mean isoprene emissions from identical ight transects conducted on three consecutive days (7-9 th July 2013), each ight consisting of 1.5 hours and $40 km repeated ight legs. Mean measured isoprene emission rates in areas of broadleaved woodland displayed a dependence on both PAR and ground temperature. Maximum isoprene emissions measured during both campaigns were found to be 2.0 mg m À2 h À1 and 1.0 mg m À2 h À1 for July 2013 and July 2014, respectively. This difference can be explained by differences in weather conditions (temperature and cloudiness) during the two measurement periods.
MVK/MACR uxes were quantied during two RFs in 2014. Overall, one hour of ux data was obtained, giving spatial coverage of $180 km 2 . Maximum uxes of $0.20 mg m À2 h À1 were observed between latitudes 51.10-51.20 , with high isoprene uxes observed over the same area. NO mixing ratios measured during all ights were found to be #1 ppbv. Under these conditions, the MCM 3.1.1 predicts that 70% of the measured ux will be MVK and 30% MACR.
Emissions of monoterpenes were also quantied during both campaigns. PTR-MS intensities at m/z 137 were assumed to give a measure of the total monoterpene ux, since no separation of individual species was possible. In 2013, 1 hour of monoterpene ux data was obtained from 2 RFs and $160 km 2 of spatial coverage. Monoterpenes measured during 2014 generated 1.25 hours of ux data, with overall spatial cover of $180 km 2 . Both campaigns showed similar emission characteristics, with highest emissions found at latitudes 51.00-50.90 . Maximum emissions were found to be 1.00 mg m À2 h À1 during 2013 and 0.80 mg m À2 h À1 during 2014. Fig. 8c and d show interpolated monoterpene uxes during both years. Table S2 † gives a full statistical breakdown of all measured biogenic VOCs during 2013 and 2014.

EMEP model comparison
The measured isoprene uxes from 2013 and 2014 were compared to hourlyaverage EMEP4UK isoprene estimates at 5 km 2 resolution. Because EMEP4UK estimates only account for biogenic sources of isoprene, the NMHCs NAEI was used to account for any anthropogenic contribution. Isoprene contribution for each of the 11 source sectors within the NAEI was accounted for in remaining unspeciated contributions, with no separation. 42 Total anthropogenic   contributions of isoprene for the region were found to be less than 1% of EME-P4UK estimates, highlighting biogenic sources as dominant.
The ux footprint represented an area 5-10 km upwind from the ight tracks (Fig. 7 le). Latitude averages of isoprene ux and EMEP4UK time-of-day estimates for 2013 (Fig. 10 le) show good agreement from latitudes of 51.15-51.30 . The EMEP4UK model however fails to capture the high isoprene uxes observed at latitudes between 51.05-51.15 . The model predicts isoprene emissions of <0.20 mg m À2 h À1 , with measured isoprene uxes ranging from 1.00-2.00 mg m À2 h À1 . This discrepancy may be due to the relatively coarse (5 km 2 ) land classication used by the emission model (Fig. 7 le) or too low normalized ('base') emission rates used in the model. Fig. 10 (right) shows latitude-averaged isoprene uxes compared to the EME-P4UK estimates for 2014. The degree of agreement between the measurement and the inventory is poor, with measured ux consistently higher than EMEP4UK estimates. Measured isoprene uxes range from 0.10-1.00 mg m À2 h À1 , compared to the EMEP4UK estimates of #0.40 mg m À2 h À1 . Again, the discrepancy may be due to land classication resolution or unrealistic normalized emission rates used in the emissions model.
The highly spatially resolved (1 km 2 ) ux measurements have signicant uncertainty (100-300%). To overcome this, overall average uxes obtained from all the measurement data were compared with the corresponding EMEP4UK   emissions estimate. A full statistical description is given in Table 3 for average comparison between measured isoprene ux and EMEP4UK estimates. In 2013, the average measured isoprene ux was 60% higher than the EMEP4UK estimate. In 2014, the average measured isoprene ux was 140% higher than the EMEP4UK estimate.

Conclusions
This study has allowed the rst direct measurements of VOC emission rates over Greater London and SE England. A total of 11 ights during July 2013 and 2014 gave 16 hours of high spatial resolution ux measurements over London for a range of anthropogenic NMHCs, corresponding to emission measurements from $4500 grid cells of 1 km 2 each. Measured benzene uxes over London displayed high spatial heterogeneity and compared relatively well to NAEI estimates. Measured toluene and C 2 alkyl-benzenes uxes showed high emissions emanating from central and industrial regions of London, although these were overestimated by the NAEI. It is clear that further renement of the NAEI, both in terms of the speciation of its NMHC inventory and in terms of individual VOC emission rates, is required. Isoprene and monoterpene uxes, presumably predominantly from biogenic sources, were measured over rural SE England. A total of 5 ights over SE England gave 6.5 hours of highly spatially resolved uxes covering $800 km 2 . Measured isoprene uxes showed relatively higher emissions over areas containing Quercus and other broadleaf tree species known to emit isoprene. The EMEP4UK inventory signicantly underestimated the measured isoprene uxes. This may be due in part to the failure of the model to capture regions of elevated isoprene ux due to its use of land classication at 5 km 2 resolution when much of the woodland in SE England is smaller than this. Improvements in the spatial resolution of land use classication could help improve model performance with respect to isoprene and monoterpene emissions. This is important as it is possible that biogenic VOCs (especially isoprene) could become of increasing importance to secondary pollutant formation in the UK and Northern Europe in a future warming climate, with increased prevalence of high temperature events, and as anthropogenic VOC emissions continue to decline with improved regulation and control technologies. Overall this work demonstrates the suitability of using a low-ying aircra to determine anthropogenic and biogenic VOC uxes by eddy covariance and the possibility of validating emission inventories with these measurements. Further evaluation and improvements to the emission inventories of anthropogenic and biogenic VOCs used for regulatory, policy and research purposes is clearly an urgent requirement, especially as some current air quality policies are based on a awed understanding of VOC emission rates.
Author contribution statement CNH, ACL, JL and RP conceived the study and obtained funding. AV, JL and MS wrote the paper. AV, SM, PM, AG, TK and LC carried out the ux data analysis, footprint modelling and interpretation the results. JL, MS, ACL, CNH, RP and BD made the measurements. MV provided the EMEP4UK isoprene emissions inventory model, as well as help interpret the results. All authors contributed to the discussion and commented on the manuscript.