Effectiveness of Green Infrastructure for Improvement of Air Quality in Urban Street Canyons

Street-level concentrations of nitrogen dioxide (NO2) and particulate matter (PM) exceed public health standards in many cities, causing increased mortality and morbidity. Concentrations can be reduced by controlling emissions, increasing dispersion, or increasing deposition rates, but little attention has been paid to the latter as a pollution control method. Both NO2 and PM are deposited onto surfaces at rates that vary according to the nature of the surface; deposition rates to vegetation are much higher than those to hard, built surfaces. Previously, city-scale studies have suggested that deposition to vegetation


Introduction
Outdoor air pollution causes 35 000-50 000 premature deaths per year in the UK 1 , and more than 1 million worldwide 2 , in addition to increased morbidity 3 . The pollutants mostly harmful in cities in the developed world are nitrogen dioxide (NO 2 ), ozone, sulfur dioxide and particulate matter with aerodynamic diameter less than 10 μm (PM 10 ), all of which cause or exacerbate pulmonary and cardiac diseases 4,5 . Attempts to reduce concentrations of these air pollutants have been ongoing for several decades, with much progress being made 3 . Methods usually center on the reduction of pollutant emissions, an increase in atmospheric dispersion, or the locating of high emitters away from existing pollution hotspots or areas of high population. Yet concentrations of air pollutants in many urban areas still consistently exceed public health standards, with mean concentration trends that are near-zero or even increasing 6 . Furthermore, there is a growing body of evidence that there is no safe threshold for exposure to air pollutants, especially PM 7,8 , so strategies are required to continue to drive concentrations down. Air quality management is particularly needed in poorly ventilated street canyons 61 .
This study focuses on NO 2 and PM 10 , which are the dominant pollutants in most urban areas, where they are largely derived from vehicle emissions (e.g. ~50% of NO 2 and ~80% of PM 10 in Central London, U. K. 9 ). Although vehicular pollutants are reduced by dispersion, this is limited at the streetlevel by in-canyon air recirculation and low wind speeds. Pollutant concentrations can also be reduced by increasing dry deposition to surfaces. Compared to controlling emissions or enhancing dispersion, relatively little attention has been paid to deposition as a pollution control measure. An effective and accessible means of achieving an enhancement in pollutant deposition is to plant additional vegetation.
Dry deposition reduces pollutant concentration (C i ) through a first-order process, where V d,i is the deposition velocity and z is the height through which the pollutant is well-mixed. V d,i depends on the pollutant species, i, and the nature of the surface, and is generally higher to vegetation than to other urban surfaces due to metabolic uptake by the plant, the 'stickiness' of the leaf surface, the large surface area of plants and their aerodynamic properties 10 .
Previous estimates of the effect of dry deposition to urban vegetation suggest that it makes small reductions in NO 2 and PM 10 concentrations on the city-scale [11][12][13][14][15]57,58,59 . For example, in Chicago, reductions of less than 1% are estimated based on current vegetation cover and less than 5% if the urban area was totally covered by trees 11 . These studies are based on relatively large domains of 12-3350 km 2 , and use aggregate variables to describe the city or sub-regions of a city. Thus these estimates fail to take account of how the complex geometry of the urban surface affects street-level concentrations, where people are primarily exposed, in particular through the occurrence of street canyons Street canyons are virtually ubiquitous in dense urban areas such as central London, Paris, Rome or Manhatten. Within street canyons, overturning eddy circulations are largely isolated from the urban boundary layer (UBL) above, leading to greatly increased residence times of air within the canyon [16][17][18][19][20] (Fig. 1a). Residence times increase substantially as the aspect ratio (height/width; h/w) of the canyon increases, and as the above-roof wind speed decreases 16,17 . Where street canyons contain a pollutant source (e.g. traffic), the increased residence time within the canyon acts to increase street-level pollutant concentrations.
Deposition in street canyons acts to reduce pollutant concentrations, and is more effective than deposition from the UBL, because of (i) the increased surface-to-volume ratio in the canyon as compared with the UBL, (ii) the decreased volume into which the pollutant is initially mixed, and (iii) the higher concentrations found within the street canyon, especially at low wind speeds (Fig. 1b). All these effects could be exploited for pollution control by enhancing deposition velocities to in-canyon surfaces. "Green walls" and street trees offer two means to achieve this enhancement.
Here, a model of street-canyon chemistry and deposition was used to show that judicious use of enhanced-deposition surfaces in concert with the urban form can very substantially reduce pollutant concentrations in one of the parts of the atmosphere where people are most likely to be exposed, i.e. at street level in street canyons. The model results were evaluated against available measurements. They demonstrate that vegetation can be an important component of pollution control strategies in dense urban areas, but only if it is applied with due regard to in-canyon air recirculation and the spatial distribution of emission sources. Urban greening initiatives whose focus is purely to increase urban tree coverage will fail to achieve their maximum air quality potential, and may even worsen air quality in street canyons. By taking into account the particular characteristics of street canyons the potential for air quality improvements could be greatly enhanced.

Model formulation
The tortuous flows in street canyons can be simulated using computational fluid dynamic (CFD) models 16,[18][19][20] or deduced from measurement studies 17 . Although some of the modeling studies have used simple chemical schemes to study reacting pollutants, such simulations are very expensive computationally, limiting the scope for sensitivity studies. Therefore, the atmospheric chemistry model CiTTyCAT 21 has been enhanced to simulate mixing and dry deposition within street canyons.
Conceptually, the urban form consists of two compartments, the lower of which can represent either a single street canyon or every street canyon within a city. In this model, CiTTy-Street (Fig. 1c), deposition velocities for roofs, canyon walls and floors can be assigned separately. Emissions of NO x , PM 10 and volatile organic compounds (VOCs), including, if necessary, biogenic VOCs, are input to the lower compartment. Mixing, M, between the two compartments is parameterized using dimensionless air exchange rates (E) for different canyon aspect ratios 16 for the case of a perpendicular wind and modified by h and above-roof wind speed, u: Based on the available literature, it is currently not possible to draw a firm conclusion as to how E varies for above-roof wind directions that are not perpendicular to the canyon axis, or indeed the effects of junctions (see supporting information). It was assumed here that for a large area of street canyons the air exchange rates of Liu et al. 16 are valid for all wind directions, although the exact value of E for a particular street or set of streets remains a significant uncertainty, and one that may be sensitive to small features of canyon geometry or downwind fetch. For the case of single street canyons under a nonperpendicular wind horizontal ventilation is highly uncertain, but may be substantial (see supporting information). Thus the exchange rates applied in this study must be taken as a first approximation to the generic street canyon situation, and extrapolations to specific canyons must be carried out with caution.CiTTy-Street was evaluated against measurements made in and above street canyons in Hanover, Berlin and Copenhagen 22,23 . It was able to simulate successfully the magnitude of in-canyon NO 2 concentrations in a non-vegetated canyon given information on NO 2 and O 3 concentrations in the UBL, traffic emission rates, photolytic flux and above-roof wind speed (Fig. S2). The simulated change in anomaly between canyon and UBL concentrations with above-roof wind speed fell between that measured at leeward and windward walls in these canyons ( Fig. 1b and Fig. S1). This indicates that the lower compartment represented the canyon-average anomaly well. Full details of the model formulation and evaluation are given in the supporting information.
CiTTy-Street, was used to calculate the effects of urban vegetation on pollutant concentrations, taking central London, UK, as a case study. One control and three green wall/green roof scenarios were considered (Table 1), and each was evaluated for both a single canyon (f = 40 m) and for a large area of street canyons (f = 10 000 m). Note that the latter study does not represent a true simulation of central London, but rather a scaling-up of the single canyon run to represent a large area of generic street canyons. In the following, green walls are considered as a proxy for any in-canyon vegetation which minimally affects in-canyon residence time. Street trees are addressed later, as they have the potential to substantially lengthen canyon air residence times, and so increase street-level pollution concentrations 24,56 . The model results and conclusions we present below are sensitive to h/w, but not to canyon volume or cloud cover (see supporting information).

Deposition velocities
In this study, green walls can refer to any type of wall greening, from Hedera spp.  40 have reported deposition velocities exceeding 30 cm s -1 for particles less than 1 μm in diameter, although to the knowledge of the authors these measurements have not yet been replicated elsewhere. Particle deposition rates are strongly dependent on the surface properties and orientation of the surface, and the wind speed, with higher wind speeds producing greater impaction rates, and hence yielding higher deposition velocities. As leaves typically present favorable surfaces for particle capture, but wind speeds in canyons and immediately above roofs are likely to be low relative to the above-roof mean wind speed, the relatively conservative value of 0.64 cm s -1 used by Nowak 11,12 is adopted. Use of this commonly-used deposition velocity aids comparison of the results herein with those of previous studies. However, this V d of 0.64 cm s -1 is considered suitable for this study because it is comparable with that predicted by the process-based model of Petroff and Zhang 41 for deposition to grass (LAI 1-2; the same as the assumed LAI for a green wall) for PM 10 mass distributions measured in polluted street canyons (mass distribution peaking at 3-10 μm) 42,43 , and broadly comparable (although smaller) than the ~1 cm s -1 measured for particles of comparable size over moorland by Nemitz et al. 44 . The considerable uncertainty in this deposition velocity should be emphasized, and the deposition velocity may need to be re-evaluated if a different mass size distribution is assumed (see supporting information for further discussion). Secondary processes such as resuspension and possible deposition limitation due to leaf PM 10 loading are not explicitly modelled. Both these processes would act to decrease overall deposition rates, but the uncertainties are much less than those in the initial choice of V d,PM10 . Modeled PM 10 deposition fluxes are compared to measurements below.

Results
Simulating adoption of green walls across large areas of street canyons in CiTTy-Street reduced incanyon concentrations of NO 2 and PM 10 by as much as 15% and 23% respectively at u=1 m s -1 and h/w=1 ( Fig. 2; for results in terms of absolute concentrations, please refer to the supporting information). These reductions were strongly dependent on residence time (i.e. wind speed and canyon geometry) and fraction of canyon wall greening (Table 1), but not on the initial pollutant concentration (see supporting information). The net pollutant flux out of the canyon was itself reduced by 2-11% for NO x (Fig. S6 left) and became inward for PM 10 , leading to small concentration reductions in the UBL above the canyon. As cities are a major regional source of air pollutants, for cities with large areal coverage of street canyons (e.g. London, Paris) this is expected to make an important difference to pollutant transport and regional-scale photochemistry, but this aspect is not pursued further here.
Release of VOCs from urban trees can influence regional-scale photochemistry 14 , but did not significantly alter the NO 2 and PM 10 budgets in the canyon in this study.  Fig. S7, see supporting information).
Modeled PM 10 deposition rates were evaluated by comparing the modeled deposition flux per unit leaf area, F L,i , with measurements, where F L,i was calculated from the deposition velocity and the modeled concentration: Assuming a single-sided LAI of 2, the modeled F L,PM10 varied from 6 -9 mg m -2 (leaf area) day -1 for wind speeds from 0.5 to 5 m s -1 . This is well within the range of available measurements. For roadside trees F L,PM10 =11-119 mg m -2 (leaf area) day -1 has recently been measured 46 . Another study measured Currently it is believed that large-scale tree planting across the city is required for vegetation to make discernible improvements to street-level air quality. Contrary to this, the results presented herein show that, because the air within a street canyon is, to a degree, isolated from the air in the UBL and all the other street canyons [16][17][18][19][20] , greening in one canyon may have a profound effect on air quality in that canyon, and will have a small effect elsewhere through reductions in UBL concentrations (Table 1 and Fig. 2). The street-level reductions were slightly smaller than for greening of large areas of street canyons, because actions in a single street canyon did not significantly reduce UBL pollutant concentrations. Using the 2008 Kew Gardens wind speed climatology produced reductions over a year for action in a single canyon (h/w=1) of 7% and 11% for NO 2 and PM 10 concentrations respectively.
This increased to 20% and 31% respectively when h/w=2. Note that when considering single-canyons, along-street ventilation may also be important when the above-roof wind is not near-perpendicular to the along-canyon axis (see supporting information). Counter-intuitively, increasing h/w for green street canyons reduced absolute concentrations at low wind speeds (Fig. S4), as the increased overall deposition rate more than compensated for the greater pollution-trapping effect at high h/w. This implies that there may be a case for artificially increasing the aspect ratio of some streets in conjunction with greening activities, perhaps by the addition of living vegetation (green) "billboards" on top of existing buildings.
At low wind speeds, when the effect of in-canyon vegetation was enhanced, the greening of canyon walls offered considerable potential of reductions in the frequency of exceedence of air quality limit values. In these circumstances, reductions in NO 2 and PM 10 concentration of as much as ~40% and ~60% respectively were predicted by the model (Table 1 and Fig. S4). This indicates that street canyon vegetation not only results in a substantial overall reduction of in-canyon pollutant concentrations, but it also forms a natural buffer against high-pollution episodes (which are often associated with low wind speeds), and associated acute impacts on human health 4 .
Like green walls, street trees increase deposition, but in addition they reduce mixing, M, between street canyon air and the UBL 24,56 . Because of this potential to alter M, it remains difficult to quantify the effect of street trees on in-canyon deposition fluxes. In order to assess whether trees have a beneficial or negative effect on in-canyon pollutant concentrations, the sensitivity of CiTTy-Street to deposition velocity and to canyon residence time was explored using a bi-variate sensitivity study (Figs. each pollutant species, i. Above V d,i (P), in-canyon concentrations decreased as residence times increased at a given deposition velocity. Therefore, if trees increase in-canyon deposition velocity sufficiently, they will improve, rather than worsen, in-canyon air pollution. The position of V d,i (P) for each pollutant depends on the in-canyon emission rate of that pollutant. Higher emission rates require a higher canyonaverage deposition velocity to prevent concentration build-up. This could be achieved through using different or greater amounts of street tree vegetation. For the high emission scenario used here (central London), V d,PM10 (P) corresponded to a LAI of 1.3 averaged across the canyon width, whereas V d,NO2 (P) was beyond the maximum of the sensitivity study. Hence, it is expected that street trees will act to reduce street-level PM 10 but increase NO 2 concentrations in highly polluted canyons in most circumstances. However, for streets with moderate or low emissions, trees will have an unambiguously beneficial effect. In this case the situation is analogous to air in the centre of a large wooded stand, where measurements have shown substantial concentration reductions 60 .
Note that the effect of trees on deposition rates and residence time is unlikely to be constant.
Residence time in street canyons will instead varying according to wind direction and speed. Deposition velocities will vary with aerodynamic factors, and tree species/size and health, and season. No sufficiently detailed dataset on urban tree health and net primary productivity exists to enable timevarying deposition velocities to be built into CiTTy-Street.

Discussion
These results show that in-canyon vegetation offers a method to improve urban air quality substantially. Urban greening can be effectively enacted on the local scale, providing a complement to top-down policy and regulation that encourages local ownership of pollution mitigation strategy, and helping to focus intervention on problem areas. Even if in-canyon pollutant sources are removed, incanyon vegetation continues to offer substantial pollutant removal benefits (very close to the single canyon values in Table 1 and Fig. 2), with lower concentrations in the canyon than in the UBL above, in effect creating 'filtered avenues'. This effect is particularly important for pollutants with atmospheric lifetimes long enough to be transported long distances, such as PM 10 and ozone. Hence, greened urban canyons may ultimately experience better air quality than in surrounding rural areas. The use of street trees must be considered on a case-by-case basis. In streets with low street-level emissions (i.e. light traffic), the filtered avenue effect will apply. Where street-level emissions are high, however, tree planting must be used with the utmost caution. The specific combination of tree species, canopy volume, canyon geometry and wind speed and direction must be modeled on a case-by-case basis.
Unlike tailpipe-based emission reduction strategies, greening also offers wider benefits, including reduced surface temperature and noise pollution and increased biodiversity and amenity value 45 . But it also offers challenges in ensuring vegetation health and minimizing damage to non-green infrastructure (e.g. underground water infrastructure). There are potential feedbacks between urban climate and tree health which cannot as yet be captured by the model. In reality, the existence of suitable plant species and the ongoing costs of maintenance will determine the viability of green infrastructure. The results presented here must be considered as part of the wide-ranging inter-disciplinary discussion on the merits and implementation of urban greening 50,51 . In particular, we expect there to be strong interdependencies between urban vegetation cover and urban water resources. It is not yet possible to treat such interdependencies in CiTTy-Street or, to the authors' knowledge, in any other urban land-atmosphere model.
Many key uncertainties remain, which should be addressed as a matter of urgency. These are the residence times of pollutants under different canyon geometries and vegetation type/coverage (especially trees), the relationship between residence time and wind speed, deposition velocities of air pollutants to canyon walls and vegetation which take account of life-cycle and seasonality, and the behavior of vegetation in the street canyon environment. An alternative to green infrastructure for air quality benefits would be to increase deposition using e.g. titanium oxide or activated carbon surface coatings 54,62 , although research suggests that these should also be applied with care as studies have indicated the re-volatisation of adsorbed NO x as HONO 53,55 .
Of the green infrastructure options available in a densely-populated urban area, in-canyon vegetation offers by far the biggest benefits for street-level air quality, much greater than, for example, green roofs.
The results of this analysis show that street-level reductions of as much as 40% for NO 2 and 60% for PM 10 are achievable using green walls. This suggests that the potential benefits of green infrastructure for air quality have been substantially undervalued [11][12][13][14][15] . These results are consistent with field measurements of deposition to vegetation and point to the utility of innovative urban greeninge.g.
increasing canyon aspect ratios with green billboardsfor air quality control. Such changes may be retrofitted to existing developments or designed into new ones, with potential implications for how urban areas are structured. Green infrastructure in street canyons maximizes the ability of vegetation to remove pollutants, and offers the potential for large and sustained improvements in urban air quality in both single canyons, and across large areas of street canyons. It is therefore essential that the potential pollution mitigation effects of in-canyon greening inform the future development of urban areas.By not considering the adverse effects of tree planting on canyon ventilation, urban greening initiatives that concentrate on increasing the number of urban trees, without consideration of location risk actively worsening street-level air quality, whilst missing a considerable opportunity for air quality amelioration

Model description
Previous studies of the effect of deposition to urban vegetation on air pollutant concentrations have used a zero-order method based upon average urban boundary layer (UBL) pollutant concentrations, the height of the UBL, and the deposition velocity at the surface 1-6, 43, 44 . Below, this method is referred to as the "constant boundary-layer concentration (CBLC) method". The CBLC method makes two assumptions that limit its utility for estimating the true deposition efficiency of urban surfaces. Firstly, the assumption of constant air concentration in the UBL implies that the depositional loss is too small to significantly impact the concentration; when that is not the case, the CBLC method will over-estimate the deposition. Secondly, as described in the main paper, the CBLC method assumes a well-mixed urban boundary layer (UBL), and so fails to account for the urban geometry which can inhibit mixing of air in street canyons with air in the overlying UBL.
The CiTTyCAT model of atmospheric chemistry 7 has been modified to represent an idealized urban area containing street canyons. The new model variant, CiTTy-Street, is run with two compartments ( Fig. 1c of main text), the lower compartment represents a street canyon of height h and width w.
Depending on the treatment of the upper compartment, the canyon compartment can be considered to represent a single street canyon or be representative of a large area of street canyons (in some cities this may be practically city-wide). In this work, unless otherwise stated, h=w=20 m, which are common street canyon dimensions (e.g. Table S1). The upper compartment represents the UBL, and was parameterized with a depth, z UBL =H(t d )-h, where H(t d ) is the UBL depth and t d is time-of-day. Although H is strictly a function of time of day, for the simulations here H was set to a typical daytime value of 1000 m. Mixing between the compartments was parameterized using the dimensionless air exchange rates (E) calculated by Liu et al. 9 for above-roof wind perpendicular to the along-canyon axis: where Q is the air ventilation rate (m 3 s -1 ), V is the volume of the street canyon (m 3 ) and T is a characteristic timescale (s), where u is the above roof wind speed (m s -1 ). In this work Q and V always appears as the ratio Q/V=M, where M is the fraction of canyon air replenished by mixing with the UBL per second, and therefore the length of the modeled canyon is arbitrary (The implications of finite street lengths are discussed below).
Eqs. S1 and S2 are used to define the mixing parameter, M: For a fixed canyon aspect (h/w) ratio, an increase in h will result in a decrease in M and hence an increase in the canyon air residence time. Conversely an increase in wind speed, u, will bring about a decrease in the canyon air residence time. Liu et al. 9 and in the upper compartment, where z UBL is the depth of the upper compartment in meters, r is the width of the roof between two adjacent street canyons, and subscript U and L denote the upper and lower compartments respectively.
The term (h·w)/(z UBL ·(w+r)) in Eq. S5 accounts for the different areas of the UCL and the UBL, including the fact that the upper compartment is wider than the lower compartment due to the space occupied by the buildings between street canyons (Fig. S1). In order to prevent the build-up of unrealistic concentrations, the upper compartment was ventilated by predefined background air with concentrations (C i,B ) at a rate determined by the above roof wind speed: where f is the length scale (m) that the model is being used to represent, i.e. the model footprint. In this work f=10 000 m for modeling large areas of street canyons, and f=40 m for single canyon modeling The modeled mixing rate out of the street canyon was based on a parameterization developed for wind flow perpendicular to the street canyon axis 9 , this was a reasonable approximation for the purposes of this work as parallel flow is less effective at ventilating the street canyon vertically due to the lower roughness generated 10,11 . Under perpendicular or near-perpendicular flow, v l , the along-canyon component of the in-canyon wind speed vector v, will be close to zero and horizontal mixing can be is expected to be broadly representative of M under all wind directions across an urban area. When considering mixing within a single canyon, however, horizontal mixing rates will need to be considered for cases where the above-roof wind is not near-perpendicular. These cases are discussed further below.
Traffic-induced turbulence was not explicitly considered in this study. Traffic-induced turbulence may be expected to dominate over wind-induced turbulence when the above-roof wind speed is less than ~1.2 m s -1 16 . Therefore for very low wind speeds in high-traffic canyons the Liu et al. 9 parameterization may underestimate M. However, the parameterizations of Liu et al. 9 were not modified here for these low wind speeds as the results presented in this study also apply to canyons with low amounts of, or no, traffic, in which cases traffic-induced turbulence will be absent or minimal. Under these conditions the air tends to stagnate in the canyon, as predicted by the model of Liu et al. 9 .
The geometry of street canyons means that, in sunny conditions, a proportion of the canyon will be in shadow. This proportion will change according to the azimuth and zenith angles of the sun relative to the street, and according to the aspect ratio of the street canyon. In overcast conditions, nearly all solar radiation will be diffuse and the shadowing effect may be ignored. Koepke et al. 17 have calculated the reduction in photolysis rates of NO 2 within street canyons for a range of aspect ratios and solar zenith angles. For the conceptual modeling studies reported here, the simplifying assumption of Koepke et al. 17 was adopted, i.e. that street orientation within the urban area is random, and therefore the reduction in photolysis rate is averaged across all possible street orientations. Koepke et al. 17 show that the principal control on the average photolysis rate reduction is the aspect ratio, rather than the solar zenith angle, for those solar zenith angles found in the middle-latitudes. As a result, they suggest an average NO 2 photolysis rate reduction of 60% for a canyon of h/w=1. This reduction was applied in CiTTy-Street for all photolysis rates in the lower compartment. Differences in the spectral albedo of the canyon surface may result in some variation of the reduction for different wavelengths. However, Koepke et al. 17 determined the surface albedo to have a relatively minor effect compared to h/w and solar zenith angle.
All runs described in the study described herein were carried out under clear sky conditions, unless stated otherwise, because these are conditions under which stomatal deposition to vegetation is expected to be greatest. The implications of this are discussed in Section S5.
Emission fluxes of NO, NO 2, volatile organic compounds (VOCs) and PM 10 were made into the lower (street canyon) compartment. Aerosol particle concentration was treated as a passive scalar, in common with other approaches 18 , implying that coagulation and growth/evaporation do not significantly affect the overall particle deposition mass flux on the time and space scales characteristic of the canyon.
No emissions were made into the upper compartment. Liu et al. 9 calculate that for a canyon of h/w=1, the timescale of the primary circulation is about one quarter of the pollutant retention time within the street canyon. Therefore the air within a street canyon may be considered well-mixed relative to the timescale for exchange with the UBL; the assumption of instantaneous mixing is discussed further below. For the sensitivity runs, emissions were supplied by the National Atmospheric Emissions Inventory (NAEI) 19  In the lower compartment, deposition was assumed to occur to both canyon walls and floor. It is assumed that horizontal and vertical mixing rates are equal, which is consistent with the formation of the rotational circulation typically seen in computational fluid dynamic (CFD) modeling studies 21 and shown schematically in Fig. 1a where length scales w and h are as defined in Fig 1(c) of the main text. In the upper compartment deposition occurs to roofs only, and was parameterized as where V d,i r is the deposition velocity to that roof. All roofs are assumed to be similar. Fractional coverage of green roofs, can be accommodated by scaling the deposition velocity accordingly. In principle, V d,i is also a function of the type and health of plant used to provide the green wall or roof, and the density of planting. Since in this study no particular kind of green infrastructure was being modeled, constant values of V d,i were chosen from the literature that are characteristic of gas and particle deposition to urban plant canopies (see Deposition Velocities section below). Unless otherwise stated, r=20 m. Studies have shown that there is variation in pollutant concentration within a street canyon due to the circulation driven by the above-roof wind, and the existence of point emissions sources 22,23 . In particular pollutant concentrations have been shown to be relatively higher at the base of the leeward wall when the above-roof wind is perpendicular to the canyon, and to show a generally negative gradient with height (assuming a traffic emission source at the canyon base) 22 . CiTTy-Street cannot capture these inhomogeneities but, as deposition rates vary linearly with concentration, the average deposition rate within the canyon is unaffected as long as the timescale for mixing within the canyon is shorter than the timescale for a pollutant concentration to be depleted via deposition. For an in-canyon NO 2 deposition rate equal to the 100% green wall scenario described in Table 1 of the main paper, a 10% depletion of in-canyon NO 2 occurs on a timescale of 340 s. Liu et al. 9 found that the rotation timescale for in-canyon air is ~15T. Therefore, for the longest canyon residence time investigated here, T=40 s (equivalent to u=0.5 m s -1 when h=20 m), the rotation timescale is ~600 s. The error in deposition rate introduced by the well-mixed approximation is therefore less than ~12% and decreases with residence time (~3% when u=2.0 m s -1 and h=20 m).
PM 10 particles are substantially larger than trace gas molecules and therefore their distribution within the canyon will be affected to a greater extent by gravitational settling. However, the terminal settling velocity of a 10μm particle in air at 293K and 1013 hPa is 3 mm s -1 ; significantly smaller than likely vertical velocities generated by turbulence and recirculation within the canyon.
To summarize, the model is characterized by five length scales (z

Deposition velocities
The following details are given in addition to those in the "Deposition velocities" section in the main text. For PM 10 deposition no differentiation was made between horizontal and vertical vegetation surfaces since leaf angle varies considerably relative to the surface, and thus is implicitly incorporated in the deposition velocity used. PM 10 deposition velocities to horizontal and vertical brick and concrete surfaces were calculated following the model for rough surfaces described by Piskunov 24 . Median particle size was assumed to be 3 μm. Roughness lengths for the particle calculations were taken to be 3 mm for brick and 1 mm for concrete, and the friction velocity was set to 0.2 m s -1 . The mean particle size was selected for consistency with particle volume distributions measured from traffic emissions 25 , and to avoid the submicrometer size range for which deposition is particularly uncertain. Theoretical models of dry deposition (e.g. Petroff and Zhang 26 ) predict that deposition velocity continues to reduce for particles below 1 μm in aerodynamic diameter, before increasing again for particles less than ~100 nm in aerodynamic diameter. Yet field measurements suggest that deposition velocities may be relatively constant in the range 0.3 -10 μm 27 , with one study predicting deposition velocities that increase with decreasing particle diameter, yielding very large sub-micron deposition velocities (~10-30 cm s -1 ) 28 .
Thus, based on measurements, assuming a particle diameter of 3 μm appears representative of deposition velocities for most of the sub 10 μm particle mass, and may even underestimate sub-micron deposition. Note that these assumptions are less likely to hold for the ultra-fine size fraction, of ~100 nm or less in diameter. There has been some discussion of a possible "saturation effect" which leads to reduced PM deposition on leaves as deposited material accumulates 27 . However the authors know of no systematic study which tests this hypothesis. In this work leaf surfaces were assumed to be washed clean by rainfall with sufficient regularity that any saturation effect could be disregarded.
Nitric acid was assigned a deposition velocity of 8.0 cm s -1 following measurements by Aikawa et al. 29 above a concrete roof in Kobe, Japan. Although this value is high, the modeled scenarios in this paper were not sensitive to the nitric acid deposition velocity. Brick and vegetation surfaces were used for the walls and/or roofs as specified in Table 1

Model mixing evaluation
The authors are aware of no available dataset against which to compare the modeled concentrations in vegetated canyons. This is probably because, until now, a measurable effect was not expected. Some field measurements were available, however, to compare the mixing performance of the model, upon which the results of this study are dependent. The CFD model of Liu et al. 9 , from which the canyon mixing parameterization is taken, has also been evaluated against wind tunnel studies.
From Eq. 2 in the main text, a decrease in above-roof wind speed is expected to bring about an increase in canyon retention time, and hence an increase in pollutant concentration is expected.
However a decrease in above-roof wind speed will also tend to lessen dispersion of pollutants from the urban area as a whole, leading to higher boundary layer pollutant concentrations, and therefore higher canyon concentrations. In order to control for this effect it is necessary to have simultaneous measurements both within and above the street canyon. Measurement data meeting this criterion recorded at three different street canyon sites was used (Table S1) to control for the effect of above-roof concentrations and hence illustrate the wind speed effect on pollutant retention. The street canyon model was then compared against this measurement data.
At each of these sites NO x concentration measurements were made at street level on one side of the street canyon, and at a nearby rooftop measurement site. Wind speed and direction measurements were collected from a mast 10 m above the rooftop level. For the following analysis only measurements recorded when the wind direction was within a 12.5° arc of being perpendicular to the canyon axis were used. These measurements were sorted into windward and leeward groups, according to the wind direction relative to the position of the street-level measurement site. Larger concentration anomalies are typically found at the leeward wall where pollutants tend to accumulate 30  .
In order to correct for the effect of varying traffic emissions within the street canyon, ΔNO x was divided by a correction factor C E following Ketzel et al. 31 , where , E where e L and e H are the emission factors for light vehicles (e.g. cars, vans) and heavy vehicles (e.g. trucks, buses) respectively, and N L and N H are the number of light and heavy vehicles. Emission factors were taken from Schädler et al. 32 . E was the mean emission rate over the entire dataset. Following this correction c x ΔNO was binned according to wind speed to aid comparison between the normalised measurements and the model. Finally a normalised concentration anomaly for the street canyon (A NOx ) was calculated by (S13) Using A NOx allowed the mixing response of the model to changes in wind speed to be evaluated independent of other factors. Figure S1 shows A NOx as a function of wind speed for the three different canyons observed in the TRAPOS experiment (http://www2.dmu.dk/atmosphericenvironment /Trapos/datadoc.htm). More concentration variation with wind speed were seen on the windward side of the canyon, as concentrations here tended to be suppressed under all but very low (less than ~2 m s -1 ) wind speeds.
The model output plotted in these runs was produced by running the model for each wind speed bin using the same setup as described in Section S1, with f=40 m. It was then processed according to Eqs.
S9 and S13 (the model uses a constant emission and therefore has no need for emissions normalization).
The model output fell between the normalized concentration anomaly for the windward and leeward walls, as would be expected for a canyon-average statistic, for all three street canyons. This indicates that the CiTTy-Street model is able to represent well the change in the canyon pollutant retention effect with changes in above-roof wind speed.

Evaluation against absolute measurements
The model setup was adjusted for an explicit comparison with measurements from Göttinger Strasse (Table S2) Figure S2 shows good agreement between the model and mean street-level measurements for NO 2 , indicating that this model is able to simulate well the concentration differential between street canyon and UBL.

Setup for sensitivity runs.
Initial and background concentrations for the model sensitivity runs are summarized in Table S2. For the numerous canyon runs background concentrations are taken from measurements made at Writtle (51.52ºN, 0.13ºE) during the TORCH measurement campaign (25/07/03-31/08/03). These measurements were chosen to give as wide a range of background species as possible. PM 10 concentrations were taken from an urban background site in London Bloomsbury (51.73ºN, 0.41ºE). For single canyon runs, upwind NO, NO 2 , O 3 and CO concentrations were also specified using London Bloomsbury measurements.

Sensitivity of results to canyon volume
The absolute reductions in modeled street canyon concentrations were virtually insensitive to the size of a canyon of a given aspect ratio. Increasing h and w decreased the deposition rate per unit volume in the canyon. However, following Eq. S2, as h is increased, T must also increase proportionally, for a constant wind speed. Therefore, for larger h and w, the deposition flux now acts for a proportionally longer time due to the increased canyon residence time. As a result, the integrated loss per unit volume of a pollutant in the street canyon is invariant with canyon size (assuming h/w=1). When h=w and h For a species X, which is not emitted (e.g. O 3 ), Assuming P X ≈L X and integrating over time until the characteristic timescale, T, That is, for a constant wind speed   X Δ is independent of canyon size.

Sensitivity to cloud cover
The runs described in this work were carried out under clear sky conditions. The reduced photolysis rate of NO 2 , J(NO 2 ), under cloudy skies would be expected to lead to higher ambient NO 2 (S17) concentrations. Note, however, that the 50% reduction of peak J(NO 2 ) calculated by CiTTy-Street for 100% cloud cover is similar to the 60% reduction of all photolysis rates applied inside the canyon during clear skies to account for shading. This canyon shading reduction was not applied under cloudy skies where all incident radiation is diffuse. Hence, the modeled reductions in NO 2 due to addition of incanyon vegetation under cloudy skies were virtually identical to those under clear skies.
The reduction in direct radiation due to increased cloud cover may also reduce stomatal opening 33 , and consequently the NO 2 and O 3 deposition velocities. The treatment of deposition in CiTTy-Street does not directly calculate this effect; however, the deposition velocities employed (Table S1) were collected under a variety of conditions and are therefore believed to be broadly representative of a wide variety of conditions. To assess the importance of stomatal closure due to reduced solar irradiance the bulk canopy stomatal resistance equation, r s , of Wesely 34 was used, where r i is the minimum bulk canopy resistance for water vapor, and F G and F T are modification factors for solar irradiance and surface air temperature respectively.
where G is the solar irradiance (W m -2

Sensitivity to emission magnitude
The modeled relative reductions in street canyon mixing ratios of NO 2 were not strongly sensitive to a doubling or halving of pollutant emissions. For instance a halving of NO x , VOC and PM emissions led to a change in the NO 2 reduction caused by moving from the control scenario to the 100% green wall scenario from -11.6% to -11.9% (for the single canyon run

Sensitivity to canyon residence time
The sensitivity of pollutant concentrations to canyon residence time was tested by varying u from 0.5 m s -1 to 20 m s -1 , equivalent to residence times of 840 s and 21 s respectively for the modeled canyon. the effect is much more modest due to the lower PM 10 emissions. Figure S1 shows that the greatest anomaly in canyon pollutant concentrations compared to those in the UBL was found at low above-roof wind speed, when the canyon residence times were greatest. In addition, low wind speeds were effective in dispersing pollution in the UBL, increasing both UBL and canyon pollutant concentrations. Therefore, exceedences of limit values for pollutant concentrations are more likely to occur at low wind speeds. These calculations show that the efficacy of canyon vegetation for pollutant removal is greatest at low wind speeds (i.e. long canyon residence times). Figure S4 illustrates

Along-canyon mixing
The importance of along-canyon ventilation can be estimated using the horizontal mixing parameter, where l is the length of the canyon, v l is the along-canyon component of the in-canyon wind speed vector, v, and β is a dimensionless factor accounting for aerodynamic resistance to mixing at the concentration reductions due to urban greening would be smaller, but still close to those for a perpendicular wind (Fig. S4). However, the maximum single canyon pollutant concentration reductions reported in the main text will only be realized under the case of a near-perpendicular wind. It is expected that v l will decrease with increasing aspect ratio, and thus M H will reduce along with M as aspect ratio increases. More work is required to deduce the form of this relationship.

Street Trees
Trees represent an obvious way of increasing in-canyon deposition rates, as they may have a very high leaf area index (defined as m 2 leaf per m 2 ground). Two rows of large trees can have a leaf area equal to or exceeding that of a green wall 27 [37][38][39][40][41][42] . Only one of these studies however, has included a consideration of dry deposition 42 . This study suggested a small role for deposition in reduction pollutant concentrations, but importantly only investigated a very small portion of the parameter space.
Although CiTTy-Street is unable to simulate the complex fluid-dynamical flows needed to calculate the exchange efficiency for a particular canyon with trees, it can be used to map the parameter space due to changes in exchange efficiency (or residence time) and deposition velocity. Using the model setup for a single canyon, as described above, a bi-variate sensitivity study was carried out, varying the canyon where, , is the canyon deposition velocity for pollutant i. For NO 2 in this scenario, an increase in canyon residence time always led to an increased NO 2 mixing ratio, although the mixing ratio increase became less marked as V d increased. , 2 = 1.0 cm s -1 is a factor of 1.5 greater than that estimated for 100% green walls. Note, however, that the model response was strongly dependent upon the scenario, specifically the in-canyon emission flux of NO x . The emission fluxes used here are typical of central London. At lower NO x emission fluxes a compensation deposition velocity could be reached at which the canyon deposition velocity effect balanced the residence time effect, leading to a decrease, rather than an increase, in in-canyon NO 2 concentrations. An equivalent compensation deposition velocity is clearly seen for PM 10 at , 10 =0.4 cm s -1 in the lower panel of Figure S5. This canyon deposition velocity for PM 10 is only 27% of that for the 100% green wall scenario. The efficacy of in-canyon vegetation as opposed to green roofs for air pollutant removal is particularly demonstrated by Figure S6 (right), where the NO 2 deposition fluxes to a vegetated canyon were several times larger than those to green roofs. Indeed, under long residence times, the brick walled canyon was more effective that the green roof at NO 2 removal, owing to higher NO 2 concentrations within the canyon.

Top-of-canyon fluxes
The recent UK National Ecosystem Assessment assumes that urbanization of vegetated land has reduced the potential sinks for pollutants 36 . To investigate this hypothesis a single-box CiTTyCAT run was carried out for a 100%-vegetated planar land surface and compared the calculated NO 2 deposition fluxes with those for our CiTTy-Street runs ( Figure S7). A street canyon with 100% green walls was a more efficient NO 2 sink than a planar vegetated surface at all wind speeds. Therefore it may be concluded that 100% greening is not required in order for urban areas to provide a pollutant sink that is as efficient as those green spaces they replace. For the central London wind speed scenario used in this work, the level of wall greening required was ~50%.