Use of networks of low cost air quality sensors to quantify air quality in urban settings

Low cost sensors are becoming increasingly available for studying urban air quality. Here we show how such sensors, deployed as a network, provide unprecedented insights into the patterns of pollutant emissions, in this case at London Heathrow Airport (LHR). Measurements from the sensor network were used to unequivocally distinguish airport emissions from long range transport, and then to infer emission indices from the various airport activities. These were used to constrain an air quality model (ADMS-Airport), creating a powerful predictive tool for modelling pollutant concentrations. For nitrogen dioxide (NO2), the results show that the nonairport component is the dominant fraction (∼75%) of annual NO2 around the airport and that despite a predicted increase in airport related NO2 with an additional runway, improvements in road traffic fleet emissions are likely to more than offset this increase. This work focusses on London Heathrow Airport, but the sensor network approach we demonstrate has general applicability for a wide range of environmental monitoring studies and air pollution interventions.


Introduction
Poor air quality is known to affect human health (Bernstein et al., 2004;Brunekreef and Holgate, 2002;McConnell et al., 2002;Parnia et al., 2002;Pope and Dockery, 2006;Ren et al., 2017;Samoli et al., 2008;WHO, 2009). Urban air quality monitoring, traditionally the remit of expensive and complex reference instruments (Kumar et al., 2015;National Audit Office, 2009) can now be achieved using readily deployable low cost instruments, revolutionising approaches to the study of urban pollution. Over the last few years, low cost sensors have been assessed for their viability in monitoring ambient air quality including measurements of gaseous and particulate matter (PM) pollutants. Whilst some studies show that there are still some challenges with using off the shelf devices (Borrego et al., 2016;Lewis et al., 2016), other studies have demonstrated that some of these limitations can be overcome with careful data processing and network design (Crilley et al., 2018;Kim et al., 2018;Penza et al., 2014;Popoola et al., 2013;Sun et al., 2017;Spinelle et al., 2015;De Vito et al., 2018;Heimann et al., 2015). There have also been attempts at deploying portable sensors as networks to better understand the spatial variability of air pollution Miskell et al., 2017;Mueller et al., 2017;Penza et al., 2014;Sun et al., 2016). Schneider et al., 2017 also used a data fusion technique to combine sensor network data with an air quality model which was then used to simulate the spatial pattern of pollutants.
Although there have been several studies on the impact of aviation on ambient air quality (Carslaw et al., 2006;Herndon et al., 2004Herndon et al., , 2008Hu et al., 2009;Masiol and Harrison, 2015;Schürmann et al., 2007), they mainly focused on using sparse array of fixed stations or short term mobile platforms which allowed source apportionment studies within and in the proximity of the airports.
Here we show how low cost air quality instruments, when deployed as a network rather than as individual sensors, provide unprecedented additional insights into the patterns and sources of air pollution.
The case study we have applied this methodology to is London Heathrow Airport and focusses in this paper mainly on NO X . The limit value for annual average NO 2 of 40 μg/m 3 is not currently being met at some of the areas around London Heathrow airport. The expansion of https://doi.org/10.1016/j.atmosenv.2018.09.030 Received 1 February 2018; Received in revised form 12 September 2018; Accepted 17 September 2018 the airport, including a third runway (UK Department for Transport, 2017), has raised concerns that meeting this limit value would be unachievable in the future. The sensor network we deployed at the airport consisted of up to 40 nodes sited across the airport and covering a range of emission environments, each measured a range of species including NO 2 , NO, CO and CO 2 . The sensor network, together with the novel analysis approach we describe, has allowed pollutant emissions attributable solely to the airport activities to be distinguished from other non-airport related sources. Using the CO 2 measurements made across the network this, in turn, has allowed the direct quantification of NO X and CO emission indices for a number of different airport related activities.
These measurements, and in particular the direct determination of emission indices, were used to optimize the ADMS-Airport air quality model, creating a powerful predictive tool for pollutant concentrations in and around Heathrow airport.
While this study focuses on the UK's Heathrow Airport, the techniques we describe, that exploit the emerging low-cost air quality sensor technologies and novel analysis approaches, have far wider applicability for environmental monitoring and air pollution interventions.

Materials and methods
In this section we briefly describe the low cost sensor network used for the study at LHR, including the state-of-the art modelling air quality model utilised in our study.

Sensor measurement: SNAQ boxes and network deployment
The sensor nodes were low cost portable air quality devices developed at the Department of Chemistry, the University of Cambridge, UK, as part of NERC (Natural Environmental Research Council) funded Sensor Network for Air Quality (SNAQ) project at London Heathrow airport. Each node measured the parameters summarised in Table 1, made at a 20 s time resolution. In addition, each node was equipped with GPRS and GPS for near real-time data transmission and location/ time data respectively . Both the particle size and gas species devices were calibrated under laboratory conditions, with the ambient performance of the NO, NO 2 and the CO 2 sensors presented in section 2.2.
Although a network of 40 sensor nodes was ultimately deployed, we are focussing here on a five-week period (4 October-11 November 2012) where only 17 of the sensor network nodes were fully operational, but where additional information on aircraft throttle settings was available (Fig. 1). The sensor network was designed to cover, as far as possible, the different activity zones in and around the airport including the terminals, runways and the main airside roads (Fig. 1). To contextualise the network density, only one routine air quality reference station (LHR2) is located within the deployment zone, with node S47 co-located with this reference station. Two nodes (S47 and S48) were also sited close to each other (∼5 m apart), and, were used to study the reproducibility of the SNAQ measurements. All the nodes were mounted at 3 m on lampposts within the airport perimeter. Fig. 2 (and Fig S1) illustrate the high temporal and spatial variability within the airport, influenced by both airport and non-airport related emissions captured by the network.

Sensor measurement validation
The performance of the SNAQ units for NO and NO 2 was evaluated by comparing the results from the node co-located with the reference air quality site within the airport (LHR2). Hourly averaged NO 2 data were derived from the hourly O X reading by correcting for O 3 interference using hourly modelled O 3 data. Comparison of the standard measured gas species (e.g. NO X ) showed that the SNAQ measurements captured variations at the airport with good agreement indicated by the slope (0.9379 ± 0.0081) and coefficient of determination R 2 ≈ 0.95 (Fig. 3A). This comparison was performed by taking hourly averages of the high temporal resolution (20 s) sensor measurements as the reference data were only reported as hourly averages. As there was no local CO 2 reference instrument available, validation of the SNAQ CO 2 measurements was performed post-study by comparing CO 2 observations made after the deployment with those of a Picarro G2201-I CO 2 / CH 4 analyser (Fig. 3B), giving results with a slope of (0.9553 ± 0.0021) and coefficient of determination R 2 ≈ 0.92. Although, this validation was performed using a single sensor node, we have demonstrated in this study -as with previous studies (Popoola et al., 2016;Mead et al., 2013) -the high level of sensor-sensor repeatability (slopes between 0.9 and 1.0 for all gas species), in this case for two nodes (S47 and S48) which were near co-located for the measurement period (Fig. 4). Errors are standard errors of the ordinary least square estimates.

Source apportionment
This section describes how we estimate the network baseline (referred to as the non-local signature) which is due to non-local emissions for the measurement period presented in this work (4 October to 11 November 2012).
The major novelty of the sensor network work described here is that local (airport) and non-local (non-local/airport related) fractions of the measured species could be determined independent of wind conditions directly from the network observations without the need for long-term (years) measurements, conditions that limit the extraction of similar information from the traditional sparse routine network measurements. We extracted both the airport fraction and non-airport fraction, as described below, using the local signature to validate the source apportionment of a model, the ADMS-Airport.
Measurements at each sensor node within the network comprise a contribution attributable to emissions local to that sensorthe local signal, and a component advected from outside the network area which will be essentially identical at all sitesthe non-local signal.
As emissions of NOx and CO local to each site will always lead to locally elevated concentrations of the pollutants, the concept is then to establish the minimum value of each pollutant across the entire network within a given time window and to equate that to the baseline or non-local signal for that species and period.
To illustrate the process, we will consider for simplicity a threesensor sub-network with measurements from three nodes (site29, site50 and site19 designated S 1 , S 2 and S 3 respectively). In this (arbitrarily chosen) period, S 1 is located at a high pollution site, while S 2 and S 3 are located at low pollution sites. The data at 20 s resolution, for three hours only in this illustration, are actual measurements used in this study and are shown in Fig. 5 (red, blue, green). It can be immediately seen that site S 1 has highly variable CO values, significantly elevated above any hemispheric background value (200-300 ppbv) (Lowry et al., 2016). In contrast, sites S 2 and S 3 both show lower variability, and lower absolute CO values. While sites S 2 and S 3 do both show some isolated pollution events, they together define a clear baseline, which in this period is around 400 ppbv and shows a slight upward trend over the three hour period. Fig. 5B, shows a probability histogram for a single hour of the ensemble of measurements shown in Fig. 5A. A clear maximum in probability is seen at ∼400 ppbv, associated with measurements from the low pollution sites where the baseline dominates across this sub-network. This reflects the low CO values associated with predominantly the absence of pollution events at sites S 2 and S 3 , while site S 1 mainly contributes to the less frequent but more elevated CO values (> 1000 ppbv). The baseline for this period is extracted by taking a percentile (in this case the 10th) of the measurements from the network (in this case just from the three sites) for the defined period, typically one hour. A percentile is chosen rather than the minimum to account for the measurement error for the sensors, with the choice of the 10th percentile reflecting the typical sensor error. The difference between the minimum CO value measured and the 10th percentile in this case is < 10 ppbv (see expanded probability histogram). These differences, which are broadly representative of the entire dataset for CO, are negligible compared to the pollution levels observed (see e.g. Fig. 6). The baseline shown in Fig. 5A is constructed from a linear interpolation between the baselines obtained for individual hour of data (and an additional hour before and after) and, as can be seen, reproduces both the average baseline and trend. A similar approach (Fig. S2) is applied to the other species (except for NO 2 , where hourly O 3 corrected data (see section 2 above) was used with the time window of three hours instead of one hour used for the 20 s resolution data). Although three sensors were used in this illustration, the entire network (17 sensor nodes) were used in the baseline determinations presented below.
Applying the above method to the sensor data for each gas species, we successfully extract the network baseline and use this to determine the local signal ( Fig. 6 shows results of the source attribution methodology for CO, and Fig. S3-S4 shows results for CO 2 and NO 2 ).

Modelling approach
The model used in this work is ADMS-Airport. This model is an  Note there are no comparisons for CO and CO 2 at LHR2 as these species are not part of the species routinely monitored at the reference station. The occasional missing data are due to data transmission issues. extension of the street scale resolution ADMS-Urban dispersion model (McHugh et al., 1997) for complex urban environments with additional capability for the explicit modelling of aircraft jet engine emissions as jet sources. ADMS-Airport has been recommended by the Project for the Sustainable Development of Heathrow (PSDH) used at Heathrow Airport (UK Department for Transport, 2006) and has been used by the Airports Commission (Module 6 (2015)) as well as the proponents of the three airport expansion schemes considered by the Airports Commission. The overall approach has been to set up an optimised modelling system for a base year of 2012 making extensive use of the sensors, and then to project the airport impact forward to 2030. Model and inventory optimisation for the airport sources was based on the detailed analysis of local fractions of model and monitored data for the sensor locations site 29 and site 30, located north and south of the 09R runway ( Fig. 7). Details of model configuration for the hourly and annual predictions are described below.

Hourly average model simulation
Airport emissions were based on a 2012 airport inventory complied by AEA Environment and Energy (AEA, Energy and Environment, 2010) which was then refined using emissions calculated from detailed aircraft activity data for the two-week period 4th November 2012   power units were also obtained from the AEA inventory.
Based on these data the following model/emission improvements were made: • The dimensions of stand sources were modified to be more representative of the area of emissions; • The location of take-off sources on runway 09R were moved to better represent the actual start of take-off roll points. For consistency with the new take-off locations, 09R taxi-out and initial climb sources were also moved.
• The treatment of taxiing was refined as follows: -a default aircraft speed of 15.4 m/s was previously used for taxiing. There is relatively low model sensitivity to taxiing speeds, but to ensure greater consistency with actual aircraft operations, a speed of 5 m/s was adopted for shorter taxi segments of less than 300 m in length. -In addition to engine emissions from the Landing and Take-off (LTO) cycle, taxiing emissions were also modelled as jet sources taking account of the momentum and buoyancy of the source, rather than as ground level volume sources with no initial momentum or buoyancy.
• NO X and CO taxiing emissions indices were recalculated assuming thrust levels of 4% for short taxi segments and 6.5% for longer segments, together with emission factors for NO and CO which were derived from measured concentration ratios with CO 2 from the sensor network data, rather than the ICAO (International Civil Aviation Organization) default emission factors for 7% thrust. When compared to the ICAO default thrust level of 7% for taxiing, these reduced thrust levels lead to higher CO emission indices and lower NO x emission indices. Fig. 7 shows scatter plots of the modelled airport contribution to CO, NO and NO 2 concentrations (ppbv) against the contribution to CO 2 concentrations (ppmv) at the site 29 sensor location, before and after the model refinements. The plots highlight the following effects of the refinements: the large reduction in modelled near ground level concentrations due to the replacement of passive volume sources with buoyant jet sources from taxiing emissions which results in significant plume rise; an increase in the average CO/CO 2 ratios and an increase in the average but greater spread of the NO x (NO + NO 2 )/CO 2 ratios, due to the higher CO emission indices and lower NO x emission ratios, used for the lower thrust settings for taxing, and the increased relative contribution of take-off NO x emissions to taxiing NO x emissions at the site.

Annual average model simulation
Non-airport emissions were obtained from the London Atmospheric Emissions Inventory (LAEI) and Slough Emissions Inventory (SEI). 2012 inventories projected from 2008 base year inventories were used; these were the latest versions of the LAEI and SEI available for the study. The road traffic emission factors used to compile these versions of the LAEI and SEI were found to underestimate real world emissions from diesel vehicles, therefore road emissions were recalculated using modified emission factors for Euro 2 to Euro 5 diesel vehicles. Non-airport emissions within 10 km of the airport were included within the model. Major roads and large industrial sources within 7 km of the airport were modelled explicitly, with all other non-airport emissions modelled as an aggregated source on a 1 km 2 basis.
Monitored concentrations from rural and suburban sites in the Defra's AURN network and the London Air Quality Network (LAQN) were used to estimate the background concentrations due to sources more than 10 km from the airport. Wind-direction dependent background concentrations for the modelled area were calculated using hourly concentrations from Harwell, Southall and Barnes Wetlands monitoring sites. This model set-up allowed use of the Generic Reaction Set chemistry scheme (Venkatram et al., 1994), considering NO x chemical reactions on an hour-by-hour basis.
Hourly sequential meteorological data from Heathrow Airport was used as input to the model. A surface roughness value of 0.2 m was used for the meteorological measurement site and a value of 0.5 m for the modelling domain (Air Quality Studies for Heathrow, 2007).
Apportionment of the contribution of airport and non-airport emissions to concentrations of NO x and NO 2 was also carried out. Included within airport emissions is the contribution from all airside emissions and the airport related traffic on surrounding major roads. It is assumed that all traffic on the airport perimeter road along with M4 and M25 spur roads, leading to the airport, is airport related traffic. The traffic flows on these roads is used to apportion an airport traffic contribution on adjacent major roads. To calculate the airport contribution to NO 2 the difference between two model runs was determined: the first including all emissions, the second excluding airport emissions.
The model performance was also verified against Defra and local authority automatic monitors around the airport in 2012, the results showed excellent agreement between modelled and monitored concentrations as summarised Table 2.

Year 2030 and third runway scenarios
The modelling for 2030 considered aircraft fleet changes and airport activity changes, with and without a third runway at Heathrow, equivalent to the 2030 North-West Runway (three runways) and 2030 Do-minimum (two runways) scenarios in the Airports Commission modelling. In addition, a 2012 North-West Runway scenario was modelled i.e. a third runway at Heathrow with the current aircraft emissions technology. Only the airside emissions were modelled for these scenarios due to the uncertainty in road layout with the extended airport.
Scenario aircraft emissions for 2030 were calculated by scaling the 2012 emissions by aircraft movements and fleet-weighted NO x emissions by aircraft category for 2030 used in the Airports Commission modelling. The location of the new runway and associated taxiways and stands are based on submissions by the scheme proposer to the Airports Commission (Independent report, 2014). Stand emissions are calculated by scaling the baseline 2012 emissions by annual passenger projections.
Runway activity assumed the ending of the Cranford Agreement (LHR Airports Limited, 2016). 1 The scheme design with the third runway proposes rotating runway use between, four modes of operation based on the northern and southern runways alternating between mixed and segregated mode operation, and the central runway operating in segregated mode only. For the modelling, it is assumed that there is, equal use of the four modes during the year, weekly alternation between departure and landing operation on the central runway and intra-day switch between mixed-mode and segregated mode operations on the northern and southern runways.
The airport contribution to total annual average NO 2 concentrations was calculated from modelled annual average NO x concentrations using primary NO 2 dependent NO x :NO 2 correlations based on the 2012 baseline modelling including airport and non-airport emissions and using the ADMS chemistry scheme. This was necessary since only sources of NO x related to the airport were modelled for 2030, which meant that the chemistry scheme within ADMS could not be employed for the 2030 calculations. These NO 2 concentrations, calculated from model derived NO x :NO 2 correlations, were found to be broadly similar to concentrations obtained using Defra's NO x Table 3 for the average traffic speed on Bath Road (37 km/h), along with total emissions for the typical traffic flow on the road; a decrease of over 80% in NO x emissions between 2013 and 2030 can be seen. The EFT v7.0 includes emissions factors extracted from COPERT 4 v 11.0; an update of the EFT taking into account revised emission factors for the majority of Euro 5 and Euro 6 light duty diesel vehicles in COPERT 4 v 11.4 is expected to be released by Defra shortly. Emissions of light duty petrol vehicles and all heavy duty vehicles are unchanged between COPERT 4 v 11.0 and COPERT 4 v 11.4. The emissions for all Euro 6 light duty diesel vehicles in EFT 7.0 have emission limit conformity factors in the range 2.6-2.8 (0.208 g/km to 0.224 g/km compared to the type approval limit of 0.08 g/km at 33.6 km/h, in accordance with Defra's definition of conformity factors). The updated emissions in COPERT 4 v 11.4 take into account the legislative stages of the Euro 6 standard, with the introduction of Real Driving Emissions (RDE) regulations from 2017 (Euro 6c and Euro 6d); for 2017-2019 vehicles (Euro 6c), the updated emissions have conformity factors in the range 3.94-5.05 (0.315 g/km to 0.404 g/km) and for post-2020 vehicles (Euro 6d) conformity factors in the range 1.97-2.45 (0.158 g/km to 0.196 g/km); this is expected to lead to a reduction in road transport emissions along Bath Road for 2030 when compared with emissions calculated using COPERT 4 v 11.0 (Updated Air Quality Re-analysis, 2017).

Air quality network measurements
To illustrate the type of information available using the low cost sensor technique, Fig. 8 shows time series of 20 s average carbon monoxide (CO) across the network on varying timescales: for the entire period in this study, for a week and for a single day. For clarity, the data are ordered from low to highly polluted sites in each panel. Several of the sensors have incomplete time series because of power supply and data transmission issues, however an overall pattern is clear. Firstly, there is a diurnal periodicity evident, with some sensor locations, generally those close to direct aircraft emissions (e.g. S10, S29, S30), showing high levels of CO in contrast to other locations at which much lower levels are seen (e.g. S17, S48). For the overall data set (Fig. 8A), the coarse pattern is that there are two periods of several days of highly polluted conditions separated by less polluted periods. These differences are associated with changing general meteorological conditions, mainly wind direction, which influences runway usage (direction of take-off) and background pollutant concentrations as well as the dispersion of pollutants. Viewing a week of data (Fig. 8B) the strong diurnal pattern associated with the daily airport operation becomes clearer, and finally, looking at a single day (Fig. 8C), individual pollution events can be seen associated with single aircraft movements. Similar patterns are observed across the network for the other species (see Fig. 2). The main novelty of this deployment was that there were multiple sensors positioned as components of a network of sensor nodes across the airport. Individual sensors in the network measure pollutant levels which are a combination of emissions local to that site and those from further afield. The critical point is that while individual sensors have pollutant signatures unique to the emission sources close to each site, sources which are external to the network itself (as a whole) produce a near identical response across every node of the network. The methods used to distinguish individual sensor responses (local signatures) from the network signature (non-local signature) are presented in section 2.3. The method was applied to all the sensor nodes in the network, and examples of such a separation for NO 2 and CO 2 are shown for a single site (S29) in Fig. 9.
For both NO 2 and CO 2 the observations at this site (which we term absolute) show periods of significant diurnal variation (in excess of 50 ppmv and 50 ppbv for CO 2 and NO 2 respectively), interspersed with periods where diurnal patterns are less evident. In contrast, the network measurements (which we term non-local or baseline) show less diurnal variation but do show longer periods of elevated pollutant concentrations, for example around October 9th and October 24th. Both types of  features are characteristic of meteorological impacts on dispersion of pollutants, runway usage, and, as will be seen below, longer range transport of pollutants. Finally, for both CO 2 and NO 2 the site specific (local) signatures retain the diurnal patterns, with the underlying longer-term variations removed. Polar bivariate plots (Carslaw and Ropkins, 2012), where pollutant concentrations for a time series are shown as functions of wind speed and wind direction, are powerful tools for source attribution studies, both for source location and for providing the basis for the determination of emission indices. Fig. 10A-D shows polar bivariate plots for the hourly average local signatures (i.e. absolute measurements with the network baseline signature removed) for CO, NO, NO 2 and CO 2 for the site discussed above (site 29). The meteorological data (wind speed and direction) for these analyses are taken from the anemometer on the sensor node itself (see panel (E)). The main features in Fig. 10A-D are two distinct regions of elevated concentrations to the NE and SE of the measurement site, both associated with aircraft activities near site 29. The NE lobe shows lower levels of both NO x (NO + NO 2 ) and the ratio of NO x to CO 2 and higher levels of CO than the SE lobe, signifying aircraft taxiing to the NE and aircraft taking off to the SE. This is consistent with the location of a taxiway and the runway relative to site 29 (Fig. 1C). NO 2 is comparable for both lobes, consistent with the lower NO but much higher ratio of NO 2 to NO x for taxiing emissions than for take-off emissions (Herndon et al., 2004). CO 2 in the NE lobe is also higher than in the SE lobe, which considering that the emissions rate of CO 2 is higher for take-off, suggests poorer initial dispersion of taxiing emissions relative to take-off emissions. The figures thus illustrate the additional benefits of multi-species measurements in characterising sources in detail as well as combining them with meteorological information. Note also that the higher concentrations in the local measurements are generally associated with higher wind speeds from the source direction reflecting the buoyant nature of ground level aircraft jet engine plumes (Bennett et al., 2010).
As is discussed in section 2.3, once the local pollutant signatures for each node are obtained (i.e. the background signature removed), emission indices (the amount of pollutant emitted per unit of CO 2 ) can now be calculated for the different sites. Fig. 4 shows how concentrations of local NO, NO 2 and CO vary with the local CO 2 at site 29. The ratio of the measurements (the graph gradient) is then used to estimate the emissions index (Table S1) for that pollutant which, in this case, correspond to 26.7 ± 0.56 g/kg for CO, 3.50 ± 0.16 g/kg for NO 2 , and 1.70 ± 0.04 g/kg for NO. Note that as these are ratios of coemitting species, they are not dependent on the degree to which any plume is intersected. Fig. 9. Time series depicting elements of the source separation for site 29 over the five-week study period. The upper panel (A) observed (absolute) values, together with local and non-local signatures (see text for definitions) for NO 2 and with the equivalent plot for CO 2 (B). Observations obtained at the site are shown in black (with offset of 100 ppbv for absolute and non-local NO 2 applied for clarity), with the network (non-local) signature in red. The derived local measurements are shown in each case in blue. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)   10F-I also show polar bivariate plots for CO, NO, NO 2 and CO 2 , except in this case for the non-local (non-airport) component, with the meteorological measurements now taken from the Met Office site at Heathrow airport (Met Office, 2012) which can be taken to represent the large-scale wind field (Fig. 10J). Note that as these plots are derived from the entire sensor network and not from any one sensor, unlike the local site-specific signatures there is only one per species for the entire network. Firstly, we observe that, in contrast to the local pollutant signatures (Fig. 10A-D), elevated concentrations are observed at lower wind speeds when dispersion of pollutants would be expected to be reduced, but secondly there are also elevated pollutant levels when the wind is from the easterly sector, consistent with high pollutant concentrations being advected across the airport from central London (Fig. 1B).
A key feature of the non-local NO 2 signature is that its diurnal pattern (Fig. 11A), with morning and evening peaks, matches that typically seen from road traffic emissions (Vardoulakis et al., 2007). This contrasts with the diurnal profiles from the local signatures at the airport ( Fig. 11B and C), which show flatter signatures during the day with a night time baseline associated with the more uniform level of daytime airport activities and the night time airport closure.

ADMS airport model and measurement comparison
This comparison uses data from the local pollution signatures extracted from the sensor network and modelled equivalent emissions for the ADMS-Airport model. The emissions and model set-up have been optimised using the sensors data, as described in section 2.4, and using known average thrust settings for British Airways (BA) aircraft at Heathrow for the period. The sensor data were averaged to hourly means to match the timescale of the ADMS-Airport model outputs. Overall, there was good agreement between the model output and the measurements for all the gas species of interest across the network, as shown in the example for site 29 close to the west end of the southern runway (Fig. 7) which compares, for each of CO, NO, NO 2 and CO 2 , the local signature from the sensor with the both the refined and unrefined model prediction taking account only of the airport emissions. Both measurement (blue) and refined model (red) are consistent in showing periods of high and low daily concentrations associated different modes of runway usage as meteorological conditions vary. The peak concentrations observed in early October and the week beginning 19 October are from the southern runway under the so-called easterly runway operational mode when aircraft take-off and land into the easterly wind. Heathrow airport most frequently operates under westerly mode (take-off and landing in this direction) consistent with the prevailing wind, however if the wind is from the east, an easterly operational mode is initiated (LHR, 2017). Fig. 12 shows a comparison of modelled and measured local NO 2 concentrations for all the active sensor sites, averaged over the fiveweek measurement period. Also shown are comparisons of the average NO 2 local and total mixing ratios at the one reference site at the airport (LHR2). Firstly, there is broadly good correspondence between modelled and measured local NO 2 across the entire network confirming the high level of ADMS model performance. Secondly, the co-located SNAQ and reference instrument show excellent agreement both in local and absolute terms confirming the performance of the low cost NO 2 sensor.
An average absolute NO 2 concentration of 25 ppbv (∼50 μg/m 3 ) was observed at the airport reference site (site 47), compared to local airport contribution of 9 ppbv (∼18 μg/m 3 ). This demonstrates that local airport emissions account for only ∼36% of the total NO 2 observed at that site and confirming that non-airport emissions dominate for the measurement period even at this airside location. A previous study (Carslaw et al., 2006), based on hourly NO x and NO 2 data for 8 measurement sites in the broad neighbourhood of the airport, reported a lower bound for the airport contribution of 27%. In contrast to the current study however, in that work the background contribution was estimated from one site only so that the effect of any emission sources between that site and the airport could not be accounted for.   12. Comparison of measured local NO 2 amounts for all the active sensor sites with the ADMS model averaged over the five-week measurement period (solid red circles). Also shown are the average local and total NO 2 mixing ratios at the one airport reference site (green and blue solid stars respectively) as well as the measured and modelled local NO 2 at the reference site (black). The error bars on y and x axes are ± 3σ. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Model annual average predictions using the ADMS-Airport model
In order to identify regions where the NO 2 annual mean limit value is exceeded, we used the ADMS-Airport model, as refined by the sensor network measurements, to predict annual mean NO 2 concentration in and around Heathrow Airport. A contour plot of modelled annual average NO 2 concentrations for 2012 obtained by averaging the model output for each hour of the year is shown in Fig. 13A. This shows exceedence (yellow, orange and red) of the NO 2 limit value within and close to the airport and close to major roads. Source apportioned model calculations are depicted in Fig. 13B and C, detailing the annual average contributions of airport and non-airport emissions; the airport emissions include road traffic emissions associated with the airport.

Discussion and conclusion
We have used the network of sensor nodes to distinguish between and quantify local and non-local contributions to the different pollutant measurements. From this we have been able to independently derive emission ratios for the airport activities at the different node sites.
The non-airport mean NO 2 concentration contribution derived from the observations of ∼16 ppbv (∼32 μg/m 3 ) shown in Fig. 11A indicates that most of the NO 2 observed in the proximity of the airport arises from emissions which are unrelated to the airport activities. The polar bivariate plot analysis of this non-airport contribution also unequivocally indicates that the predominant non-airport emission sources are in the ENE/E direction, i.e. from the Greater London region. Importantly, the diurnal profile of this signal is characteristic of road traffic related emissions as the morning and evening rush hour peak events are clearly evident in the figure and thus it can be concluded that these sources are largely associated with road traffic emissions from Greater London.
The ADMS-Airport model, with emission indices validated and refined by comparison with the equivalent airport emissions derived directly from the sensor network measurements, shows that within the airport perimeter annual average NO 2 levels, at typically 50 μg/m 3 (Fig. 13A), are above the EU annual average limit of 40 μg/m 3 . However, beyond the airport perimeter, annual average NO 2 levels only marginally exceed this limit except near major roads (40-44μg/m 3 up to 1 km to the north of the airport and lower further away). However, the model results show that in the area immediately outside the airport where the NO 2 annual average limit value is exceeded, only 12-16 μg/ m 3 (∼6-8 ppbv) of NO 2 (Fig. 13C) can be attributed to airport activities, ∼30-36% of the total NO 2 . Model evaluations of the potential impact of an airport expansion (the addition of a third runway with its increased aircraft movements) suggest that the airport contribution to the north east of the airport will rise to ∼20 μg/m 3 (∼10 ppbv) with the addition of a third runway (Fig. 13D-E).
This work demonstrates a technique for source apportionment in complex environments, in this case a major international airport, using a low cost air quality sensor network. The wider impact of such work is that is can lead to an improved understanding of the impacts of policy decisions and interventions. For Heathrow airport, there is an expectation (Airports Commission July 2015) that on the timescale of the airport expansion, changes to the road traffic fleet, particularly the introduction of cleaner (Euro 6) and zero emission vehicles, NOx levels in the London plume reaching the airport will fall significantly. As an example of the likely magnitude of this effect, projections from 2013 to 2030 using the Emission Factor Toolkit (EFT) 7 published by Defra (2016) suggest a reduction of more than 80% in road NO x emissions just to the north of the airport (see section 2.4.4). If this projection is correct, then even with a third runway and the associated roadside activities, NO 2 levels would be expected to fall below the current NO 2 annual average limit, meaning that the area would then be compliant for NO 2 . Clearly crucial to achieving compliance with the NO 2 annual average limit value by 2030 would be the extent to which vehicle engine technologies do indeed reduce NO x emissions across the vehicle fleet.
The critical components of this study are the explicit separation of local and non-local emissions and the direct determination of emission indices using the sensor network coupled to appropriate analysis methods, and together are important demonstrations of the potential of this emerging low-cost air quality sensor technology.
Heathrow airport represents in some senses a special case, particularly in that on the spatial scale of the airport the pollutant baselines at any time can be represented by single values. Extending the technique across wider domains such as across an entire mega-city, while conceptually very similar to the approach we have taken, would be expected to require baselines which could vary spatially.
The measurement and analysis methodology we have demonstrated in this case study has wide applicability in many complex air quality environments including developing megacities, and has the clear potential to inform in a cost-effective manner policies and interventions which would mitigate potential health impacts of air quality and other environmental issues including the monitoring of greenhouse gas emissions.