Effects of urban density on carbon dioxide exchanges: observations of dense urban, suburban and woodland areas of southern England

Anthropogenic and biogenic controls on the surface e atmosphere exchange of CO 2 are explored for three different environments. Similarities are seen between suburban and woodland sites during summer, when photosynthesis and respiration determine the diurnal pattern of the CO 2 ﬂ ux. In winter, emissions from human activities dominate urban and suburban ﬂ uxes; building emissions increase during cold weather, while traf ﬁ c is a major component of CO 2 emissions all year round. Observed CO 2 ﬂ uxes re ﬂ ect diurnal traf ﬁ c patterns (busy throughout the day (urban); rush-hour peaks (suburban)) and vary between working days and non-working days, except at the woodland site. Suburban vegetation offsets some anthropogenic emissions, but 24-h CO 2 ﬂ uxes are usually positive even during summer. Observations are compared to estimated emissions from simple models and inventories. Annual CO 2 exchanges are signi ﬁ cantly different between sites, demonstrating the impacts of increasing urban density (and decreasing vegetation fraction) on the CO 2 ﬂ ux to the atmosphere.


Introduction
Carbon dioxide concentrations continue to increase globally, reaching 400 ppm on a daily basis at Mauna Loa in 2013 (The Keeling Curve, 2014). Over 70% of global greenhouse gas emissions are from urban areas (IEA, 2012). While a large number of studies have documented the seasonal dynamics of carbon fluxes of vegetated ecosystems (e.g. Schmid et al., 2000;Baldocchi et al., 2001;Aubinet et al., 2012), comparable measurements from urban areas remain relatively limited (see reviews by Velasco and Roth, 2010;Grimmond and Christen, 2012;Christen, 2014;Weissert et al., 2014). The earliest measurements in cities began in the mid-1990s (Grimmond et al., 2002;Nemitz et al., 2002), yet only very recently have multi-year urban fluxes been published (Pawlak et al., 2010;Bergeron and Strachan, 2011;Crawford et al., 2011;J€ arvi et al., 2012;Liu et al., 2012;Peters and McFadden, 2012). Thus understanding of CO 2 exchanges based on direct observations in regions with large urban fluxes is limited. Instead, estimates of emissions are mostly based on fuel consumption inventories, but these tend to have coarse spatial and temporal resolution and do not include biogenic processes such as photosynthetic uptake by urban vegetation Crawford and Christen, 2014). However, to explore the potential impacts of urban planning schemes and policy decisions, or to make predictions about future climates, improved understanding of processes relevant to the urban carbon balance is required. Pataki et al. (2011) highlight the need for more rigorous evaluation of urban greening schemes, which should include both positive and negative impacts on the ecosystem as a whole, realistic cost-benefit analyses and consideration of site-specific and species-dependent behaviour.
Per unit area, annual CO 2 exchanges measured in urban areas greatly exceed those from nearby natural ecosystems: average annual CO 2 release in Helsinki is forty times larger than the uptake by a nearby wetland and eight times larger than the uptake by a boreal forest (J€ arvi et al., 2012). In highly-vegetated Baltimore, however, net CO 2 release is similar in magnitude to the net uptake of nearby forests . Few campaigns have quantified CO 2 exchanges for different urban densities concurrently. Coutts et al. (2007) presented fluxes from two suburban sites in Melbourne, and Bergeron and Strachan (2011) compared fluxes from urban, suburban and agricultural sites in Montreal. Measurements of CO 2 concentration across urban-to-rural gradients in the US include the work of Strong et al. (2011) in Salt Lake Valley and Briber et al. (2013) in Boston. Given the apparent inability of vegetation to assimilate large enough quantities of CO 2 to offset emissions (Pataki et al., 2011;Weissert et al., 2014), quantifying the effect of human behaviour on CO 2 exchange becomes an even more critical area for research. Approaches include long-term observational campaigns which encompass policy changes, for example Song and Wang (2012) assessed the impact of traffic reduction due to the Beijing Olympics in 2008, and combining measurements and models to better inform the attribution of measured CO 2 emissions to various human activities such as building energy use, transport and metabolism (e.g. Christen et al., 2011;Strong et al., 2011).
The objective of this study is to relate observed CO 2 exchanges to physical processes, through consideration of meteorological conditions and surface characteristics. Direct eddy covariance measurements of CO 2 fluxes from three very different land uses (urban, suburban and woodland) over the same period are compared. The sites are located within one of the most densely populated regions of Europe: southern England. This region, which includes London, has been extensively modified by human activities in both rural and urban areas. Atmospheric controls are considered first, by comparing the meteorology observed at each site. After demonstrating the similarity in climatic conditions, links between CO 2 flux and surface characteristics (e.g. land cover, urban density) are explored.

Description of sites
In this paper measurements undertaken at three sites 70e100 km apart and at approximately the same latitude in southern England (Table 1, Fig. 1) are compared. These are a dense urban environment in central London (U); a predominantly residential suburban site in Swindon (S); and a deciduous oak woodland at the Alice Holt Research site (W). Additional details are provided elsewhere (London (Kotthaus and Grimmond, 2012;2014a,b); Swindon (Ward et al., 2013); Alice Holt ).
Across the sites there is a gradient of impervious to pervious land cover, with London having 81% of the plan area covered by roads and buildings, Swindon 49% and Alice Holt effectively 0% (Table 1). The heights of the roughness elements (i.e. buildings and trees) are similar in London and Alice Holt (>20 m) but smaller in Swindon (z6 m). There is very little vegetation at the central London site; trees are mainly London plane (Platanus hispanica) and grass lawns are mainly confined to small public gardens. In Swindon, grass is the predominant surface cover and grows alongside roads as well as in residential gardens, recreational areas and on undeveloped land. Trees comprise a range of species but are mainly deciduous. At Alice Holt the predominant tree species is oak (Quercus robur) with hazel (Corylus avellana) and hawthorn (Crataegus monogyna) making up the understorey . The above and below ground tree biomass is estimated to be 13.4 kg C m À2 (excluding shrubs and ground flora, based on 2009 data) and the mean peak leaf area index is 5.9 m 2 m À2 (1999e2010 data).

Instrumentation and data processing
Net fluxes of CO 2 between the surface and atmosphere were obtained for 30-min intervals using the eddy covariance (EC) technique at each site. The micrometeorological sign convention is used, i.e. negative flux indicates CO 2 uptake by the surface and positive flux indicates CO 2 release. The instrumental setup is summarised in Table 2. Equipment was mounted on towers (a square-section tower at Alice Holt, lattice towers in London and a pneumatic mast in Swindon) to ensure that measurements were made well above the mean height of the roughness elements (z H ) and above the roughness sub-layer (>2 z H for U and S; > 1.3 z H for W, Tables 1 and 2). Sites were carefully selected to ensure the measurements are representative of the local environment. Although the source areas vary with meteorological conditions, footprint models indicate that the majority of the flux usually originates from within a few hundred metres (approximately 200e400 m) of the towers; at night these distances increase (to around 600e700 m) as instability decreases. The variation in land cover around each tower is far smaller than the difference in land cover between the three sites. Full characterisation is provided in the individual site papers.
Raw data from the sonic and gas analyser were processed using LiCOR's EddyPro software (S, W) or ECPACK (van Dijk et al., 2004) (U). The quality control procedures applied were selected based on the requirements of each site, dependent on their different characteristics (see Kotthaus and Grimmond, 2012, Ward et al., 2013. This was judged to be the most appropriate methodology, rather than attempting to apply a single set of tests across all sites which may not be suitable for each environment. All sites were subject to the following standard procedures: adjustment for the lag time between sonic anemometer and gas analyser; correction of sonic temperature for humidity; correction for spectral losses. The planar fit coordinate transformation was applied to the London data; double coordinate rotation was used for Swindon and Alice Holt. Data from all sites were despiked and subjected to physically-reasonable threshold checks and data were removed during times of instrument malfunction. In London the influence of micro-scale building emissions was removed from the local-scale fluxes using an algorithm based on the statistical characteristics of turbulent events (Kotthaus and Grimmond, 2012); this procedure is not required at the less urbanised sites. No friction velocity (u * ) threshold was used to reject CO 2 fluxes at the suburban or urban site because the rough Table 1 Site characteristics (values are those given in the respective publications; surface cover is calculated at U for the average footprint climatology (Kotthaus and Grimmond, 2014b), at S for 500 m around the tower (Ward et al., 2013) and at W for the woodland area ). z H is the average building or tree height; z d zero plane displacement height; z 0 roughness length. nature of these surfaces and anthropogenic energy emissions enhance mixing, resulting in fewer stable periods (Christen and Vogt, 2004) which means that thresholds developed for rural environments are not useful (Bergeron and Strachan, 2011;Crawford et al., 2011;Liu et al., 2012). At Alice Holt, inspection of plots of night-time fluxes against u * suggested a threshold of 0.15 m s À1 . This is similar to the average threshold value derived for the site between 1999 and 2010 by Wilkinson et al. (2012) and is within the range found for woodland sites (e.g. 0.1e0.5 m s À1 , Papale et al. (2006)). CO 2 fluxes were therefore excluded for u * <0.15 m s À1 at W. In this study no corrections were made for CO 2 storage below the EC instruments at any of the sites (accounting for storage would alter the fluxes by <0.5% at W, and the correction for urban areas is estimated to be only a few percent (Crawford and Christen, 2012)). The period analysed comprises January 2011 to April 2013, with all three sites operational from May 2011. Gaps in the data occurred due to a variety of operational reasons (e.g. low power, instrument failure) but the overall data availability is reasonably good. Out of the maximum possible 40848 30-min periods, after quality control CO 2 fluxes were available for analysis 77.0%, 57.1% and 70.3% of the time at London, Swindon and Alice Holt, respectively. Available data were relatively evenly spread between day and night, although less so at Alice Holt where the u * filter excluded a larger proportion of night-time data (76.8% of data are available during the day compared to 63.9% at night).
For calculation of annual totals, CO 2 fluxes for 2011 and 2012 were gap-filled using monthly median diurnal cycles (e.g. J€ arvi et al., 2012). Two sets of median diurnal cycles were calculated for each site, one for working days and the other for non-working days (weekends and holidays), for each of the 28 months. These were used to fill gaps in the datasets based on month, day of week and time of day for each site. As data collection in Swindon did not start until May 2011, the period JanuaryeApril 2011 was filled using the mean of the 2012 and 2013 diurnal cycles for each month (JanuaryeApril) considering working days and non-working days separately. Implications of this procedure are discussed below (Section 4.3).
Other variables measured (Table 2) include photosynthetically active radiation (PAR) at the U and W sites. Comparison of measured PAR and total incoming solar radiation (K Y ) for these sites revealed a high degree of correlation (r 2 ¼ 0.99) with constants of proportionality (PAR/K Y , both in W m À2 ) of 0.38 and 0.42 at U and W, respectively. Gaps in measured PAR were filled using these relations with K Y where possible. For S, PAR was modelled as a proportion (0.40) of incoming shortwave radiation.

Estimating anthropogenic emissions
Carbon emissions for the central London site were estimated with the GreaterQF model (Iamarino et al., 2012) after updating gas (domestic and non-domestic) and other fuel consumption data with 2011e2012 data (DECC, 2013a), and residential and daytime population with 2011 census data (ONS, 2011). Traffic emissions were adjusted for the study period according to the trends in annual totals provided for the relevant central London boroughs (DECC, 2013b). CO 2 emissions from fuel combustion were  calculated by applying conversion factors (NGVA) to heat emissions, as provided by GreaterQF, while CO 2 emissions from human metabolism were estimated following Koerner and Klopatek (2002). Emissions were modelled for the borough of Westminster, which contains the majority of the EC footprint. Estimates of anthropogenic carbon emissions were made for the Swindon site using inventories of domestic energy use, vehicle use and population density (see Appendix A of Ward et al. (2013) for details). Quarterly domestic natural gas consumption statistics (DECC, 2013a) were used to estimate CO 2 emissions due to building energy use (combustion of other fuels was neglected) with subdaily variability modelled based on Hamilton et al. (2009). The total distance travelled by motor vehicles in Swindon (DfT, 2013) was adjusted for the proportion of road in the study area. Equal weight of traffic across all roads (Ichinose et al., 1999), a fixed fuel economy (Sailor and Lu, 2004) and fixed emission factors (Moriwaki and Kanda, 2004) were assumed in the absence of more detailed data specific to Swindon. Temporal variation was modelled using typical daily and monthly profiles (DfT, 2011). CO 2 release from human metabolism was estimated using daytime and nighttime population density (ONS, 2011) following the method of Bergeron and Strachan (2011).

Variability in meteorological conditions
Due to their proximity, the three sites can be expected to experience similar climatic conditions and weather patterns. Southern England has a temperate maritime climate (cool summers, mild winters, cloudy, wet and changeable weather). According to the Met Office (2014a,b), normal (1981e2010) annual rainfall for this region is 782 mm and mean air temperature (T air ) is 10.3 C. In both 2011 and 2012, rainfall was well below normal at the start of the year. Whilst 2011 remained drier than normal, the rest of 2012 was very wet with above average rainfall (AprileDecember). The summers of 2011 and 2012 were generally cloudy, although there were some warm, dry and sunny spells in late May, July and September 2012 (Figs. 2 and 3). Winter 2011/12 was warmer than average but a cold period brought snowfall in February 2012. Winter 2012/13 was cool and temperatures remained low throughout spring, with March 2013 being much colder than normal.
Regional similarities in weather patterns are evident in the data collected at the London, Swindon and Alice Holt sites (Fig. 2). Although instrumental differences have not been accounted for, the mean daily air temperatures at measurement height follow each other closely (r 2 > 0.94). The slightly warmer air temperatures at U compared to S and then W are as expected. Mean daily temperatures are 15.94, 15.29 and 14.85 C in summer and 5.34, 4.99 and 4.82 C in winter at U, S and W, respectively. Fig. 2 also shows that the patterns in relative humidity (RH) are similar between sites, and again show the expected variation with land use (highest humidity at the woodland site and lowest in the city centre).
When T air and RH are combined to give vapour pressure deficit (VPD), periods with high VPD are easily identifiable at all three sites simultaneously (Fig. 3, shading). These warm, dry spells occur when the incoming shortwave radiation is high across the region and skies are near clear (Fig. 3, bars). Diurnal and seasonal behaviour in VPD and PAR are important controls on photosynthesis rates (e.g. Schmid et al., 2000;Flanagan et al., 2002). Comparing across the sites, K Y decreases as the density of urbanisation increases. Although the time series in Fig. 3 are very similar, the colouring reveals that higher transmissivity tends to occur at W. These findings are as expected, as there is evidence of reduced solar radiation receipt in central London (Ryder and Toumi, 2011). Modelled bulk transmissivity was determined as the fraction of observed K Y relative to the top of atmosphere K Y (calculated using the solaR R package (Perpiñ an, 2012) and a solar constant of 1367 W m À2 (Peixoto and Oort, 1992)). The mean bulk transmissivity for the study period was 0.372, 0.382 and 0.415 at U, S and W, respectively. Overall, it can be concluded that the three sites experience generally similar weather conditions and therefore any differences in CO 2 exchanges should be related to land-surface controls.

CO 2 flux comparison between sites
There are clear contrasts in both the magnitude and temporal variability of the observed carbon fluxes (F C ) in London, Swindon and Alice Holt (Fig. 4). At the forested Alice Holt site, the annual cycle is dominated by the balance between photosynthesis and respiration ( Fig. 4c). Net uptake in the daytime by the understorey is evident early in the year , followed by a sudden increase in net uptake which coincides with leaf-out of the oak canopy (Mizunuma et al., 2013). During the night and in winter, the forest acts as a small source of CO 2 . The largest emissions occur during summer nights (5e6 mmol m À2 s À1 ) when ecosystem respiration rates are largest because the vegetation is in leaf and metabolically more active, and temperatures are warm (Figs. 4c and 5c). The behaviour observed at Alice Holt is typical for deciduous forests generally (Falge et al., 2002;Baldocchi et al., 2005) and for other deciduous forest areas in this region (Thomas et al., 2011).
In contrast, central London is a major source of CO 2 all year round, with larger emissions during winter than summer (Fig. 4a).
Emissions are 10e30 mmol m À2 s À1 larger during the day than the night ( Fig. 5a) and the shape of the diurnal pattern in the CO 2 flux does not differ considerably with season, being slightly asymmetrical and tending to remain higher in the evening approaching midnight compared to the early hours of the morning (Fig. 6). This is in accordance with expected typical patterns of human behaviour and similar to those documented in Tokyo (Moriwaki and Kanda, 2004) or Marseille (Grimmond et al., 2004), for example. In densely built-up areas both the abundance of sources and sustained daytime activity (e.g. from traffic, human metabolism and  commercial building use) cause emissions to stay high throughout the day (e.g. Helfter et al., 2011;Iamarino et al., 2012;Lietzke and Vogt, 2013). Average daytime values peak at about 35 mmol m À2 s À1 in summer and 80 mmol m À2 s À1 in winter. Note the availability of CO 2 data at U is very low for January 2013 (4.8%, all occurring on weekdays, Fig. 4a). The monthly values for January 2013 in Figs. 5a and 6 are therefore very likely overestimates of the true CO 2 fluxes.
Suburban Swindon, having a mixture of vegetation cover plus built areas, has fluxes that respond to both biogenic and anthropogenic controls. As for London, Swindon's emissions are largest in winter (Fig. 4b), when there is increased fuel combustion for building heating and vegetation is largely dormant. During the growing season photosynthetic uptake is greatest in the middle of the day, associated with maximal PAR. The pattern of daytime CO 2 fluxes is similar to W (Fig. 6). However, during winter the diurnal pattern of F C is markedly different: two clear peaks are observed which coincide with road traffic and building heating demand during the morning and evening rush hours (Ward et al., 2013). The role of human activities resulting in differences in F C between U and S is evident. In U the emissions remain high throughout the daytime, reflecting the intensity of activities. In low density residential  areas, such as Swindon, rush-hour peaks are more typically observed, for example in Melbourne (Coutts et al., 2007), Montreal (Bergeron and Strachan, 2011), Helsinki (J€ arvi et al., 2012) and Mexico City (Velasco et al., 2013). Typical wintertime fluxes are approximately five times smaller at S (9 g C m À2 day À1 ) than U (50 g C m À2 day À1 ), whilst summertime fluxes are roughly ten times smaller at S (2 g C m À2 day À1 ) than U (22 g C m À2 day À1 ).
Uptake of CO 2 in towns and cities is often masked by anthropogenic emissions, partly as the proportion of land covered by vegetation tends to be considerably lower (e.g. 44% at S) than in the surrounding countryside or before development. Despite the significant amount of vegetation at S, the net daily (24-h) flux is usually positive. Nevertheless, the potential of urban vegetation to reduce the overall release of CO 2 into the atmosphere is evident in the Swindon data (Fig. 6). From spring until autumn, F C is negative during the middle of the day and 24-h emissions are kept low (z2 g C m À2 d À1 ). Photosynthesis continues into the autumn, but increased emissions from heating at S mean that uptake is no longer seen in the middle of the day (compare fluxes at S and W during October and early November, Fig. 4).
As there are no anthropogenic emissions at the woodland site, the photosynthetic uptake is a strong signal in the observed CO 2 flux (znet ecosystem exchange (NEE)). The ecosystem respiration (R eco ) at this site is around 5e10 mmol m À2 s À1 during summer daytimes , so gross primary production (GPP) may typically be 40 mmol m À2 s À1 around midday during the peak of the growing season. At W, the diurnal range in the CO 2 flux is small in winter (F C z 0e3 mmol m À2 s À1 ) but there is still a small yet discernible reduction in CO 2 release during the middle of the day (Fig. 4c). These estimates of R eco are consistent with long-term averages at this site and other similar sites . Chamber measurements made at W in 2007e2010 (Heinemeyer et al., 2012) indicated soil CO 2 effluxes are equal to about half of R eco .
At all times the CO 2 release from the U site exceeds that from S and W. The CO 2 fluxes are most comparable during the early hours of the morning in summer (e.g. June, July and August 2012) at around 3e10 mmol m À2 s À1 ; these represent amongst the smallest emissions from U and largest emissions from W. The spread of observed values is greatest at U, small and stays the same all year round at S, while W data are much more variable in summer.
Despite the additional sources of CO 2 at S compared to at W, the observed fluxes were lower at S during daytimes in spring 2012 (March and April), and similar around midday during May 2012 and April 2013 (Figs. 4bec and 6). This may be due to the slightly warmer temperatures within urban areas which cause leaf development to occur earlier, but also reflects the large soil decomposition and tree respiration rates in woodland environments. Phenological differences of extended growing season in urban environments have been observed in a number of areas (e.g. Imhoff et al., 2004;Zhang et al., 2004). Unfortunately no S data are available before May 2011, so it is not possible to extend this comparison over more years. There is also some species dependency: photosynthetic activity occurs in the shrub layer at W from late February  data) but the tree canopy does not develop until late AprileMay; most of the vegetation around the residential S site is grass that can begin photosynthesising early in the year (Hiller et al., 2011). Although both sites have a small proportion of evergreen vegetation (other than grass), it is unlikely to be a significant contribution at either site.
Inter-annual differences in the CO 2 fluxes are mainly attributable to variations in the weather. Largely warm and sunny conditions in April and early May 2011 (Fig. 3) promoted development of vegetation and led to a rapid increase in the rate of photosynthesis.
In 2012, a dry, sunny spell in March was followed by a very wet and dull April, which resulted in rapid leaf-out of deciduous vegetation at the end of May (Fig. 3). In early 2013, positive F C observed across all sites is attributed to the cold start to the year (Fig. 2) suppressing vegetation growth and generating a greater demand for building heating and thus increased CO 2 emissions.

Relation to biophysical processes and anthropogenic activities
To assess the roles of biophysical processes and anthropogenic activities at the three very different sites, relations with explanatory variables are investigated here. Whilst Alice Holt essentially has no emissions from human activity, the tiny amount of vegetation around the London site is insufficient to be detectable compared to the strong anthropogenic signal. Swindon lies between the two with fluxes that relate to a mixture of controls. Analysis over relatively long time series (28 months) also permits exploration of seasonal trends and inter-annual variability.

Photosynthetically active radiation
During the growing season photosynthetic uptake by vegetation exhibits a clear dependence on the amount of PAR received by leaves. As PAR increases, observed F C at both S and W decreases. When vegetation is fully leafed-out, data closely follow the expected functional dependence (Fig. 7). Seasonally, the F C -PAR relation varies: as vegetation approaches senescence the relation between PAR and rate of CO 2 uptake decreases as the leaf area changes at both W and S (not shown). Naturally, the woodland ecosystem has the strongest dependence on PAR as the footprint is completely vegetated and during summer daytimes photosynthesis constitutes the main component of F C . As the largest PAR values are approached, the rate of CO 2 uptake decreases and the curves level off as the light saturation threshold for photosynthesis is reached. At times with high PAR (i.e. sunny weather, typically warm with a high VPD, Fig. 3), when there is limited water available, plants may reduce transpiration rates to conserve water. Levelling off of the light response curve, or even a reduction in CO 2 uptake at high PAR, may also indicate stomatal closure. Separating morning and afternoon data reveals some diurnal hysteresis, with slightly higher uptake at moderate to large PAR in the morning compared to the afternoon at S and W. At W the uptake in 2011 (drier than average, dashed line in Fig. 7) is lower than in 2012 (a very wet summer, dotted line) across mid-to-high PAR values. There is also some asymmetry in the diurnal cycles of F C in late summer 2011 (Fig. 6), which may be due to limited water availability. However, note that there are large data gaps in September 2011 for W (Fig. 4).
At U there is no clear relation with PAR. The spread of the data is much greater and mean values are higher than the medians as a result of some large F C values and a slight positive skew (more so in summer 2012, Fig. 5a). The observed increase in F C at low PAR values is probably related to the coincident increase (decrease) in anthropogenic activities following sunrise (before sunset). The lack of evidence for photosynthetic uptake in central London indicates that the role of vegetation is negligible at this densely urbanised site.

Temperature
In natural environments, the nocturnal CO 2 flux is equated to ecosystem respiration. At W the expected exponential increase in ecosystem respiration with air temperature (e.g. Lloyd and Taylor, 1994;Schmid et al., 2000) is observed across the seasons (Fig. 8c). The variability increases with temperature and the CO 2 emissions level off above about 15 C. This levelling-off may be caused by the grouping of data at different periods in the seasonal cycle, variability in soil moisture conditions or the changing relation between soil temperature and air temperature. Respiration rates are higher when soil conditions are wet at warmer temperatures, presumably due to higher decomposition rates, as indicated by the subset April 2012eApril 2013 (a very wet period, Section 3).
During summer nights at S, when CO 2 emissions from traffic and building heating are minimal, F C also shows an increase with T air (subset, Fig. 8b), suggesting a contribution from ecosystem respiration. Again, emissions are larger during summer 2012 (dotted line) than during summer 2011 (dashed line), particularly for warm temperatures. Anthropogenic emissions dominate at cooler temperatures in response to increased heating demand. The coolest period occurred in February 2012 and included snowfall (Section 3). The rate of increase of F C with decreasing T air is greater at U than at S due to the difference in urban density. At U, R eco (excluding human respiration) is a very small contributor to the total carbon emissions. Thus, an exponential dependence of R eco on T air is not seen, even in summer. Instead F C decreases with increasing T air and the relation levels off for the warmest temperatures (Fig. 8a).
At urbanised sites, monthly total carbon emissions are often negatively correlated with air temperature (e.g. Liu et al., 2012). Fig. 9 combines the effects of seasonal variation in vegetation activity and demand for building heating with temperature (assuming that traffic load is independent of temperature). At U a decrease in emissions with increasing temperature is seen for all months, as well as for winter and summer separately. The gradient (À1.95 mmol m À2 s À1 C À1 ) is larger in magnitude than the value given for Beijing (À0.34 mmol m À2 s À1 C À1 (Liu et al., 2012)), reflecting the greater seasonal variability and larger contribution of building activities to the total emissions at the London site compared to the Beijing site. As U is much more densely built, the  total emissions are also greater than in Beijing (15% vegetation fraction, P V ). Building-scale F C is estimated to reach around 60 mmol m À2 s À1 in the morning hours (November 2009eJune 2011) on average (Kotthaus and Grimmond, 2012). Emissions at the warmest temperatures are associated with traffic (assuming traffic does not vary seasonally) and metabolism, plus gas combustion for non-heating purposes. The London Olympics occurred during the warmest month, August 2012 (circled in Fig. 9), when central London traffic was less and more people used public transport (TfL, 2012). Thus a more typical estimate associated with the slightly cooler summer months may be around 20 mmol m À2 s À1 .
An overall trend of reduced emissions with warmer temperatures is seen for S, with a shallower gradient than for U. However, when winter and summer are considered separately at S, an increase in F C with temperature is observed, possibly associated with increased respiration rates (Fig. 8b). Similar relations are seen for W, but the woodland site goes from being a source of CO 2 to a sink of CO 2 with increasing temperature (i.e. with season), whereas Swindon is always a net source of CO 2 on a monthly basis.

Day of week
The major impact of anthropogenic activities on the carbon fluxes is apparent when data are separated into working days and non-working days, i.e. weekends and holidays (Fig. 10). Over a long enough study period there is no reason for weather conditions or biogenic processes to differ substantially between working and non-working days. Thus, as expected, CO 2 fluxes from the woodland site do not show any significant systematic difference between these two subsets.
For the suburban site, however, F C tends to be larger on working days compared to non-working days (Fig. 10). This is particularly evident in winter when F C is dominated by anthropogenic controls. The two 'rush-hour' peaks seen in the morning and evening are more pronounced on working days, while the diurnal pattern on non-working days is flatter with a reduced morning peak and a tendency for increased emissions during the middle of the day. These differences closely resemble patterns of human behaviour and energy consumption (e.g. Sailor, 2011). In winter, when vegetation is dormant, trends in observed F C are modelled reasonably well by the estimated anthropogenic emissions, although the modelled emissions overestimate the observations. Emissions from traffic tend to show little seasonal variation, while emissions from building heating vary from month to month (due to temperature). Reduced building emissions are partly responsible for lower F C observed in summer, but at this time of year photosynthesis dominates the shape of the diurnal cycle. Emissions from human metabolism are a small contribution (~1 mmol m À2 s À1 ) to total F C .
For the London site, emissions are mostly determined by energy use in buildings, the majority of which is non-domestic gas combustion according to inventory data (DECC, 2013a). Emissions are considerably higher on working days than on non-working days all year round (by about 15e30 mmol m À2 s À1 at midday), with the largest fluxes usually occurring in the middle of the day. This is typical for the area, which is strongly service-oriented and has a large daytime population made up almost entirely of nonresidents. Model results for the borough of Westminster (area 22.2 km 2 ) are presented here. Most of the EC source area is located within this borough, which has a resident population density of less than 10,000 persons km À2 compared to a workday population density of 31,000 persons km À2 (ONS, 2011). However, the total daytime population density (including tourists) is even higher, estimated at 46,000 persons km À2 (GLA, 2013). For such densely populated areas CO 2 emissions from human metabolism are a significant contribution to the total emissions, particularly during the middle of working days. The model assumes that the daytime population applies for times between 1000 and 1600, which is probably an oversimplification. In reality, people arrive earlier and leave later for work, while others may travel into the area in the evening. This difference may explain why the observed diurnal cycles are more peaked than the modelled diurnal cycles. The GreaterQF model does not account for seasonal variation in population density, although the daytime population is generally lower in summer coinciding with the holiday season for workers and students. As seen in Swindon, the traffic load at U is fairly steady all year round. The lowest emissions are associated with non-working days during summer. Vehicular emissions are estimated to contribute an average of 6.7/5.2 mmol m À2 s À1 on working/nonworking days to F C ; building emissions are more significant at 23.7/22.6 mmol m À2 s À1 .
Overall, the GreaterQF model captures the shape of the diurnal cycle in F C and differences between working days and non-working days reasonably well, in particular the fluxes on non-working days remaining high late into the evening. The model estimates for the borough of Westminster are slightly lower than the observations at U in winter and higher in summer. At smaller scales, GreaterQF emission estimates became very spatially variable, dominated by variation in non-domestic gas consumption. Emission estimates for some areas were more than twice the size of the observations at U. Such spatial contrast raises the question of whether energy use inventories are an appropriate method to estimate carbon fluxes at these scales. The location of individual buildings, roads or neighbourhoods with respect to the boundaries used for reporting energy use, traffic and population can have a major impact on the corresponding emissions estimated for that area. Validation would require far more detailed information than is available at present, both spatially and temporally. Considering the land cover at the London site in comparison with Westminster borough, we would expect traffic emissions to contribute to the observations to a greater extent than suggested by the model estimates in Fig. 10.

Annual totals and inter-annual variability
Cumulative carbon fluxes indicating net carbon uptake or release (Fig. 11) were obtained by gap-filling with monthly median diurnal cycles (Section 2.2). Annual totals for the two complete years studied (2011/2012) are 12.08/13.36, 1.62/1.87 and À0.46/ À0.37 kg C m À2 yr À1 for U, S and W, respectively. (Note, as S began operating from May 2011 earlier data are all gap-filled.) Mean annual totals are: 12.72, 1.75 and À0.42 kg C m À2 yr À1 for U, S and W, respectively (based on 2011 and 2012 data). The amount of carbon released at U and S was greater (by 10e15%) in 2012 compared to 2011 and the carbon uptake at W was smaller (by about 20%). Higher emissions in 2012 at U and S likely result from larger heating demand: cold weather including snowfall in February, a cool and wet April and a cooler autumn than normal. The timing of these differences between 2011 and 2012 can be seen in Fig. 11 (most clearly for U but the difference in autumn is also visible at S during October and November). F C is noticeably smaller for August 2012 (16 g C m À2 day À1 ) than August 2011 (24 g C m À2 day À1 ) at U (see also Figs. 5a and 6), which may be partly due to reduced traffic during the London Olympics. It is not possible to quantify the impact of the Olympics on the CO 2 emissions with this dataset, as inter-annual variability can be driven by many factors, for instance August 2012 was also slightly warmer than 2011. But these results demonstrate the potential for human behaviour to affect emissions of CO 2 , either indirectly via responses to synoptic weather conditions or due to entirely anthropogenic factors. Despite the later start to the growing season in 2012, the magnitude of daytime uptake at W was greater in 2012 than 2011, however, emissions from respiration were also larger in 2012 than 2011 (Fig. 4c). The annual carbon uptake at W during 2011 and 2012 is comparable to, but slightly smaller than, the 1999e2010 average of À0.49 kg C m À2 yr À1  and well within the range observed in previous years: À0.30 kg C m À2 yr À1 (2010, outbreak of defoliating caterpillars) to À0.63 kg C m À2 yr À1 (2007, long growing season).
J€ arvi et al. (2012) compare various gap-filling techniques for an urban site in Helsinki, and find that the gap-filling method has only a small effect (<5%) on the estimates of annual CO 2 exchange. The continuous period JanuaryeApril 2011 before measurements began at S needs to be considered differently (Section 2.2), because gapfilling using data from other years can introduce bias due to inter-annual variability of the fluxes. Temperatures in April 2012 were lower than April 2011 and February, March and April 2013 were cooler than the same months in 2011. Gap-filling the JanuaryeApril 2011 data with the average of the diurnal cycles for Fig. 11. Cumulative carbon flux over for the two complete years (2011, 2012) within the study period (a) for all sites together and (bed) for each site individually. Grey points indicate data that have been gap-filled; the legend includes the proportion of data available for 2011e2012. these months in 2012 and 2013 has therefore likely resulted in an overestimation of CO 2 release in 2011, as it is reasonable to expect the heating demand was lower in 2011 compared to in April 2012 and FebruaryeApril 2013. The modelled emissions support this hypothesis: emissions from building heating are estimated to be 37% smaller in April 2011 than in April 2012 and 23% smaller in FebruaryeApril 2011 than in FebruaryeApril 2013. Additionally, the very cool and cloudy weather in April 2012 and cool temperatures in FebruaryeApril 2013 are thought to have delayed photosynthetic uptake compared to spring 2011 (e.g. Fig. 5c), i.e. using the average of 2012 and 2013 data for spring 2011 likely underestimates uptake for this period. Again, this potential bias due to the gap-filling procedure would lead to an overestimation of CO 2 fluxes for JaneApr 2011. However, the estimated annual CO 2 release in 2011 is still smaller than for 2012. These year-to-year variations highlight the need for long-term measurements to evaluate emissions inventories.
On an annual basis, S emits four times more carbon per unit area than is absorbed at W, despite its relatively large proportion of vegetation cover. Per unit area, carbon emissions from U are thirty times larger than the uptake at the woodland site and seven times larger than emissions from the suburban site. These results can be considered in terms of the impact of land use change. In vegetated areas, deforestation followed by urbanisation can change a region from a net CO 2 sink to a net CO 2 source, which may emit ten or more times more CO 2 than was previously being taken up. The opposite effect may be observed in arid areas, however, where urbanisation can increase the amount of vegetation relative to the surroundings (Oke et al., 1989;Seto and Shepherd, 2009;Ramamurthy and Pardyjak, 2011).
Values from the National Atmospheric Emissions Inventory (NAEI, 2011) compare reasonably well with the results presented here, although, contrary to EC observations of the net CO 2 exchange, they do not account for biogenic processes such as ecosystem (including human) respiration and photosynthesis. Inventory data for the four 1 km Â 1 km grid squares around the locations of each EC tower were considered. NAEI total emissions for the London site amounted to 13.00 kg C m À2 yr À1 , almost all of which is due to non-industrial combustion (63%) and road use (34%). This value is similar to the observed F C (12.72 kg C m À2 yr À1 ) and similar to the emission estimates from the GreaterQF model for Westminster (13.5 kg C m À2 yr À1 , or 11.3 kg C m À2 yr À1 if human metabolism is not included). As discussed in Section 4.2.3, the traffic emissions from GreaterQF are thought to be an underestimate for the study site. The slightly lower emission estimates from GreaterQF may also be explained by land cover differences with the study site, as about 20% of Westminster borough is open green-space (but the borough also includes some of the busiest streets in London).
In Swindon, non-industrial combustion and road use comprise nearly all (97%) of the total NAEI value of 1.92 kg C m À2 yr À1 (value calculated omitting one of the four grid squares with a large contribution from industrial processes which is outside the footprint of the EC system). The difference between the NAEI value and observed average annual total is 0.17 kg C m À1 yr À1 , which is mainly attributed to the influence of vegetation on F C (uptake of À0.12 kg C m À2 yr À1 ). The contribution of vegetation to F C was approximated based on typical annual uptake from European grassland sites of 0.24 kg Cm À2 yr À1 (Soussana et al., 2007) and uptake by trees of À0.42 kg C m À2 yr À1 (from the Alice Holt observations), scaled by the proportion of grass (36%) and tree (9%) cover in the study area.
The same comparison at Alice Holt offers little insight, as the NAEI estimates are based on processes which are largely unimportant at the woodland site. NAEI estimates vary around the EC tower, from 0.0027 kg C m À2 yr À1 for a grid square that is almost entirely vegetated with no roads or houses (note that this is most representative of the footprint of the EC system) to >0.3 kg C m À2 yr À1 when a main road is found within the grid square. The footprint of the EC system is mainly concentrated over the woodland , so the emissions from these roads are not thought to impact F C . However, it is remarkable that a few roads within a largely undeveloped and vegetated landscape result in modelled emissions of a similar order of magnitude to the woodland uptake.
A wide range of annual carbon fluxes from other urban areas around the world have been observed (Table 3). Both globally and across the three study sites in southern England, decreasing vegetation cover and increasing population density are associated with enhanced emissions (Fig. 12). Reduced vegetation cover coincides with an increased area of buildings and roads, and therefore greater associated emissions, as well as reduced capacity to take up CO 2 . As the vegetation fraction of urban areas decreases, the building density increases for both the plan area and the height of buildings, and the population density increases. Hence F C exponentially increases as vegetation cover decreases (Fig. 12a). The relation with population density appears to be more linear (Fig. 12b) and more scattered, as was also shown by Nordbo et al. (2012). There is considerable uncertainty associated with the population estimates. As already discussed, there is huge variation in the number of people in central London throughout the day, both spatially and temporally, and accurate quantification is challenging. Therefore this descriptor is often not reported consistently in the literature. Workday population densities according to the 2011 census are Table 3 Annual observed carbon fluxes (F C ), vegetation cover (P V ) and population density (r pop ) from urban sites in the literature (with measurements for at least one year). Population densities given are for a daytime, b the cities of Florence (Matese et al., 2009) and Ł od z (Fortuniak et al., 2013), and c the municipality of Beijing (Liu et al., 2012). given for the London studies in Table 3 and Fig. 12b; resident population densities would be much smaller and total daytime population densities including tourists would be much higher (Section 4.2.3). The population estimate given in Liu et al. (2012) is for the whole Beijing municipality and is very low for the compact, built-up nature of their study site; 10,000e20,000 persons km À2 seems more appropriate for the measurement footprint. Similarly, 3470 persons km À2 (Matese et al., 2009) is a very low estimate for the particular area around the central Florence site which has a 'very high population density', as discussed in Gioli et al. (2012). Such variation in population density therefore limits its usefulness as an estimator for F C . The distribution of points in Fig. 12b highlights the need for more observations in very densely urbanised areas. The two studies with the largest fluxes and highest population densities took place in London. The central London annual carbon release measured in this study is the largest to date, corresponding to the site with the highest population density and low vegetation cover. The second largest emissions are also from London, and are associated with a slightly higher vegetation fraction and slightly lower population density (Helfter et al., 2011). These measurements were made at a height of 190 m and so had a much larger footprint than the present study. Highly-vegetated low-density suburban Baltimore has the smallest F C of the urban sites of 0.361 kg C m À2 yr À1 .
Given the diversity of sites in Fig. 12, the vegetation fraction is a useful proxy for approximating carbon emissions from urban areas. Besides the plan area of vegetation cover, species and biomass may also be relevant (Velasco et al., 2013), for example evergreens have greater potential for carbon assimilation compared with deciduous trees (Falge et al., 2002). Variation in net ecosystem exchange among vegetated ecosystems is about 1 kg C m À2 yr À1 , ranging from small CO 2 release to CO 2 uptake approaching À0.9 kg C m À2 yr À1 (Falge et al., 2002;Soussana et al., 2007;Thomas et al., 2011;Wilkinson et al., 2012). This variation, however, is only about 10% of the variation in studied urban sites (>12 kg C m À2 yr À1 ). Some studies have suggested that urbanisation reduces natural variability and has a homogenising effect on ecology as a result of similar land use, land management and species selection across the globe (McKinney, 2006;Groffman et al., 2014;Polsky et al., 2014).
Some of the spread within Fig. 12 can be attributed to climatic differences, for example emissions in urban Montreal are relatively high given the vegetation fraction of the site due to fossil fuel combustion for building heating (Bergeron and Strachan, 2011), whilst the year-round warm climate in Singapore means emissions from buildings are limited to combustion for cooking (Velasco et al., 2013). The net exchange for the Swindon, Helsinki and Singapore sites are almost identical (Table 3). Swindon and Helsinki have broadly similar climates, the same amount of vegetation cover (P V ¼ 44%) and similar population densities, whilst the effects of Singapore's tropical climate with no heating requirement balances the smaller proportion of vegetation (P V ¼ 15%) and higher population density. Contrary to findings elsewhere, summertime carbon release in semi-arid Salt Lake Valley was smaller for a highlyvegetated suburban site (P V ¼ 49%) compared to the surrounding natural grasslands, partially attributed to the dormancy of the grasslands during dry conditions compared to irrigated suburban vegetation and abundance of non-native trees in the suburban area (Ramamurthy and Pardyjak, 2011).
Compared with the differences between sites, variation between measurement years is small. In Beijing, F C was found to be closely related to traffic load, ranging from 4.61 to 5.40 kg C m À2 yr À1 (2006e2009). Traffic restrictions during the 2008 Olympic Games reduced the number of vehicles by 60% and coincided with the lowest annual emissions (Liu et al., 2012). These traffic restrictions were estimated to reduce daily carbon fluxes by about 30% (Song and Wang, 2012). In Helsinki, observed F C ranges from 1.58 to 1.88 kg C m À2 yr À1 (2006e2010), with inter-annual variability attributed to synoptic weather conditions (both the exceptionally warm and sunny conditions in summer 2006 and the influence of wind direction on flux footprint) (J€ arvi et al., 2012).
Differences in F C between sites can occur due to footprint composition, particularly when strong localised sources such as industry or roads are located nearby. Annual flux estimates over a suburban lawn (Hiller et al., 2011) were adjusted by 0.13 kg C m À2 yr À1 to account for the impact of a nearby road (carrying approx. 10,000 vehicles day À1 ), hence even a small area of road (small change in P V ) can lead to appreciable differences in annual CO 2 exchange. The higher-than-expected net exchange for Vancouver given the vegetation cover and population density (Fig. 12) is attributed to the proximity of major roads to the measurement tower . Urban sites generally consist of a complex spatial distribution of sources and sinks of CO 2 . Whilst the woodland site is fairly homogeneous, in Swindon larger CO 2 release is observed from the area south-west of the tower where there is a busy crossroads and greater CO 2 uptake is seen to the north-east when the source area contains more gardens and large trees. Differences in CO 2 exchange between 90 sectors due to local heterogeneity may be about 0.6 kg C m À2 yr À1 at S, which is of a similar magnitude as would be expected from Fig. 12 for a change in P V of z10%. The prevailing wind direction at S is south-westerly, hence F C is expected to be biased to higher values compared with  Table 3 for references).
the study area as a whole (Ward et al., 2013). In London, F C tends to be higher when the source area includes a busy junction to the north and smaller when the source area includes the River Thames to the south (see Kotthaus and Grimmond, 2014b for analysis of the spatial variability in energy fluxes at the central London site). Measurements of CO 2 concentration within streets indicate higher variability close to the river, where the air is generally cleaner but CO 2 'hotspots' occur in areas of high traffic load, however, there is little spatial variation in concentrations for roads that are not adjacent to the river. Differences in F C related to land cover variation at each site individually are far smaller than the differences between sites.

Conclusions
In this paper, we examine direct observations of the carbon dioxide exchange from three very different sites in close proximity monitored over the same multi-seasonal time period: a woodland site at Alice Holt, a suburban site in Swindon and a dense urban site in central London. These sites are subject to the same meteorological conditions, and thus related drivers have little effect on the observed differences (as they are effectively constant across the sites). Differences in the exchange of CO 2 between surface and atmosphere can therefore be attributed to processes related to land use. Comparison of three sites of different land cover is useful because trends begin to form when the sites are ranked by urban density. The results shown here support previous findings on the relation between CO 2 fluxes and building fraction (or the close to inverse relation with vegetation fraction), as shown in Grimmond and Christen (2012), Nordbo et al. (2012) or Weissert et al. (2014) for example. The huge anthropogenic emissions associated with cities means that variation in annual CO 2 exchange among urbanised study sites is about ten times that observed among vegetated ecosystems.
Signatures of different anthropogenic and biogenic controls were investigated at various timescales. At sites with a significant proportion of (deciduous) vegetation, fluxes are very different between the summer growing season and winter. Photosynthetic uptake dominates the diurnal and seasonal cycle at the forested Alice Holt site, but also plays a key role in the carbon cycle of suburban Swindon (44% vegetated), where daily summertime net carbon emissions are small. That photosynthesis is an important component of the total measured flux is illustrated by tight F C -PAR curves exhibited by the suburban and woodland data. Net CO 2 uptake is not seen in central London. Here, the times of smallest CO 2 release are summer night times, coinciding with minimal human activity. Nocturnal emissions are largest during winter for sites where building heating is important, whilst at the woodland site nocturnal emissions peak in summer when warm temperatures and photosynthetic uptake during the day promote respiration.
Total observed carbon exchange for the London site is similar to NAEI values, whilst at Swindon and Alice Holt vegetative and biogenic processes (not included in the NAEI data) play a significant role in the carbon balance. For the urban and suburban site, anthropogenic emissions were modelled at finer time resolution using the GreaterQF model and statistics of energy and vehicle use respectively. Comparison with the observed fluxes suggests the anthropogenic emissions are overestimated by this approach for Swindon, while estimates for London are highly dependent on the spatial resolution of the inventory data. However the models effectively replicate the impact of patterns in human behaviour on carbon exchange. Validation of the inventory approach to estimating emissions at smaller scales in heterogeneous environments would require consumption data at higher resolution than currently available and an experimental design whereby the spatial extent of the inventory and observational data could be well matched.
Urban and suburban patterns of CO 2 fluxes at the sub-daily, weekly and seasonal cycles can be directly attributed to patterns in human behaviour. Traffic and building energy use at both sites closely resemble the single-and double-peaked diurnal cycles observed in F C in London and Swindon, respectively, and largely explain reduced emissions on weekends compared to weekdays. The major difference between CO 2 release on working days and non-working days illustrates the potential impact that reducing anthropogenic activities (or cutting emissions by improving efficiency) could have. Reductions in traffic intensity, improved energy efficiency in buildings, and preferential use of electricity over combustion of fossil fuels could make a significant impact in limiting carbon release from towns and cities.
The main result of this study is the order of magnitude differences in the net CO 2 exchange at the three very different sites, based on direct contemporaneous observations. It has been shown that the diurnal and seasonal behaviour varies significantly with urban density. At annual timescales, these results provide an independent evaluation of inventory-based assessments. These findings demonstrate, and begin to quantify, the major role urban areas have in global CO 2 emissions.