Measurement report: Three years of size-resolved eddy-covariance particle number flux measurements in an urban environment

Size-resolved particle number fluxes in the size range 10 nm < particle diameter (Dp) < 200 nm were measured over a 3-year period (April 2017 March 2020) using the eddy covariance technique at an urban site in Berlin, Germany. The observations indicated the site as a net source of particles with a median total particle number flux of FTNC = 0.86 x 10 m s. The turbulent surface-atmosphere exchange of particles was clearly dominated by ultrafine particles (Dp < 100 nm) with a share of 96 % of total particle number flux (FUFP = 0.83 x 10 8 m s). 10 Annual estimates of median FTNC and FUFP slightly decreased by -9.6 % (-8.9 % for FUFP) from the first to the second observation year and a further -5.9 % (-6.1 % for FUFP) from the second to the third year. The annual variation might be due to different reasons such as variation of flux footprints in the individual years, a slight reduction of traffic intensity in the third year or a progressive transition of the vehicle fleet towards a higher share of low-emission standards or electric drive. Sizeresolved measurements illustrated events of bidirectional fluxes, i.e. simultaneous emission and deposition fluxes within the 15 size spectrum, which occurred more often in spring, late summer and autumn than in winter. Multi-year observations of sizeresolved particle fluxes proved to be important for deeper understanding of particle exchange processes with the urban surface and the pronounced influence of traffic at this urban site.


Introduction
Airborne particles, especially ultrafine particles (UFP) with diameters Dp < 100 nm, may enter the circulatory system and are 20 associated to acute and chronic effects on human health (Oberdörster et al., 1995;Nemmar et al., 2002;HEI Review Panel on Ultrafine Particles, 2013;Schraufnagel, 2020). UFP are abundant in cities due to a number of different anthropogenic emission sources such as private and public traffic or commercial and industrial combustion (Tsang et al., 2008;Tie et al. 2009;Bäfver et al. 2011;Weber et al., 2013;Kumar et al., 2014;Gerling et al., 2020). In order to understand the spatio-temporal variation of population exposure to ultrafine particles, information on the strength and location of sources and sinks, and the turbulent 25 exchange between surface and atmosphere are essential (Buzorius et al., 2000;Longley et al., 2003;Mårtensson et al., 2006;Weber et al., 2013).
Urban particle number flux measurements using EC with varying lower cut-off diameters of the particle counting instruments were conducted in a couple of cities across Europe (cf. Table 1). These studies generally report the city to be a net source of particles, i.e. a positive average particle number flux, which by convention indicates an upward directed emission flux to the atmosphere in contrast to a negative (downward directed) deposition flux. A significant contribution to the particle number 35 flux is attributed to vehicle traffic with peak mode diameters of fresh traffic emission in the size range Dp < 30 nm (Wiedensohler et al., 2002;Zhu et al., 2007;Rönkkö et al., 2017). On the diurnal cycle, particle number fluxes are relatively small at night but increase in the early morning hours due to anthropogenic activity and enhanced turbulence, resulting in peak upward fluxes at around noon or in the early afternoon (Dorsey et al., 2002;Mårtensson et al., 2006;Deventer et al., 2013;Conte et al., 2018). Variation in particle number fluxes is shown to be related to different land use within the flux source area 40 with lower emission fluxes from vegetated surfaces or residential areas and higher fluxes from urban land-use with intense traffic (e.g. Järvi et al. 2009, Ripamonti et al. 2013. Although aerosol processes and dynamics such as particle deposition or the influence on human health are strongly related to the particle diameter, information on size-resolved urban particle number fluxes, especially in the ultrafine size range, remains scarce. Size-resolved fluxes, however, help to deepen understanding on particle sources and dynamics in the urban boundary 45 layer. Additionally, the flux data is essential to parameterise dry deposition velocities and to validate urban air quality models (Saylor et al., 2019;Farmer et al., 2021). Size-resolved flux observations may point to situations of bidirectional particle fluxes, i.e. simultaneously occurring upward and downward fluxes in different ranges of the particle size spectrum (Schmidt and Klemm, 2008;Deventer et al., 2015). Usually upward fluxes tend to occur in the smaller size ranges and may point to emission sources such as vehicle traffic whereas downward fluxes mainly occur in larger size ranges, i.e. accumulation mode particles. 50 Particle flux studies usually comprise observation periods of a few weeks or months within different seasons of the year. Only few studies report observations for longer periods such as one-year or multi-annual periods (e.g. Ripamonti et al., 2013;Deventer et al., 2015, cf. Table 1). This is especially true for size-resolved particle number fluxes from urban areas for which multi-annual data has not been available, yet. These longer term studies, however, are important to analyse temporal variation of urban particle number fluxes and their driving mechanisms. 55 Here, we report on three years of size-resolved urban particle number fluxes in the diameter size range 10 nm < Dp < 200 nm.
The measurements were carried out in central Berlin, Germany, in the framework of the urban climate research project [UC]²-3DO (Scherer et al., 2019). We analyse the diurnal, seasonal and annual variation of ultrafine and size-resolved particle number fluxes and study the contribution from surface drivers such as land-use on temporal variation of particle fluxes.

Measurement site 65
Size-resolved particle number fluxes were measured from 01 April 2017 to 31 March 2020 near "Ernst-Reuter-Platz" in central Berlin, Germany, atop the main building of Technische Universität Berlin. The site surroundings comprise residential built-up surfaces, traffic and green areas ( Figure 1). To the north of the flux site, the busy main road "Straße des 17. Juni" with an average daily traffic intensity of about 37,300 vehicles day -1 is located (Umweltatlas Berlin, 2017). To represent traffic intensity in the flux footprint, data from two traffic counting stations at "Hardenbergstraße" (HS, south-west of measurement 70 site) and "Straße des 17. Juni" (S17J, east of the site) were available (cf. Figure 1). Data was provided by the traffic information centre Berlin (VMZ Berlin Betreibergesellschaft mbH) at hourly resolution.

Instrumentation
The particle flux instrumentation consisted of an electric mobility particle sizer (Engine Exhaust Particle Sizer Spectrometer, EEPS 3090, TSI Inc., Minnesota, USA) and a 3D ultrasonic anemometer (USA-1, Metek GmbH, Elmshorn, Germany). Both 75 instruments synchronously sampled at a frequency of 10 Hz. The EEPS measures the particle number size distribution (PNSD) over the size range 5.6 nm < Dp < 560 nm in 32 size channels. Through a stainless steel tube of 0.01 m diameter, the air was sampled at a flow rate of 10 L min -1 resulting in a laminar sampling flow (Reynolds number ≈ 1300; Hinds, 1999). The steel tube was attached to a 10 m meteorological rooftop mast with the sample inlet located next to the sonic anemometer at a height of 57 m above ground level. We sampled dry aerosol using a Nafion dryer MD-700 (Perma Pure LLC, length 0.9 m) to keep 80 relative humidity in sample air < 40 %.
The particle spectrometer EEPS 3090 classifies particles based on their differential electrical mobility. Charged particles entering the electric field are reflected outwards where the charge is delivered to 22 electrodes and converted into particle number concentrations (TSI Inc., 2015). To limit baseline noise in the electrometer readings as can be caused by dirt, the instrument electrometers were zero-checked every second to third day using the instrument firmware. The instrument was 85 primarily developed for measurements of engine exhaust emissions that show higher particle concentrations than usually observed in the urban background. Hence, EEPS readings of specific size channels may fall below the analyser's minimum threshold concentration resulting in non-valid concentration readings. To avoid data gaps within the PNSD, a cubic natural spline interpolation was used as gap-filling strategy to obtain complete PNSDs (cf. section 2.3). To quality-check EEPS measurements for concentrations and fluxes, an on-site inter-comparison to a water-based condensational particle counter (WCPC 3787, TSI Inc., Minnesota, USA) was carried out. 95

Data handling
The post-processing of particle data and fluxes comprised the following procedures: Size-dependent diffusional particle losses within the sampling line and the Nafion dryer were corrected according to Hinds (1999). The Nafion dryer MD-700 was assumed as stainless steel tube of 0.9 m length as no specific correction factors were available for the dryer. 100 Subsequently, gaps in the PNSD (cf. section 2.2) were filled using a natural spline interpolation following Meyer-Kornblum et al. (2019). As gap-filling caused larger uncertainties in the boundary regions of the PNSD, the size spectrum was limited to 21 size channels in the size range 10 nm < Dp < 200 nm. According to Meyer-Kornblum et al. (2019), PNSDs were gap-filled as long as less than 9 gaps in total (minimum of 12 remaining data points within the PNSD) or 5 contiguous gaps (5 neighbouring size channels) occurred in that size range. PNSDs not fulfilling these requirements were discarded. We measured 105 a total of 8.43 x 10 8 PNSDs during the entire study period, of which 85.4 % were gap-filled, 14.0 % were rejected and 0.6 % were without gaps and did not need to be gap-filled. The total particle number concentration (TNC, 10 nm < Dp < 200 nm), ultrafine particles (UFP, 10 nm < Dp < 100 nm) and three modal concentrations, i.e. nucleation mode (NUC, 10 nm < Dp < 30 nm), Aitken mode (AIT, 30 nm < Dp < 100 nm), and accumulation mode (ACC, 100 nm < Dp < 200 nm) were calculated from gap-filled PNSDs. 110 For particle flux calculation, the missing samples allowance for each half-hourly block average was set to 20 %. Wind vectors and particle number concentrations were checked for plausibility concerning a realistic range of absolute values and spikes were removed following Vickers and Mahrt (1997). Additionally, double coordinate rotation for tilt correction and spectral corrections of high-pass (Moncrieff et al., 2004) and low-pass (Moncrieff et al., 1997) filtering effects were applied. For time lag compensation between particle and sonic data, we used covariance maximization with a specified time lag window. 115 Following other urban particle flux studies, we applied linear detrending (e.g. Mårtensson et al., 2006;Vogt et al., 2011a;Deventer et al., 2013;Deventer et al., 2018;Heyden et al., 2018). The particle flux calculation was performed using the software EddyPro® v6.2.2. To calculate particle number fluxes for WCPC data (cf. section 2.2), the same procedure as described above was adopted. By definition, positive flux values indicate upward (emission) fluxes to the atmosphere whereas negative values denote downward (deposition) fluxes to the surface. 120 The response time of the EEPS for fast changes in concentration was reported as 0.5 s (Johnson et al., 2003(Johnson et al., , 2004). This response time is similar to other CPCs or size spectrometers used in particle flux measurements (e.g. Buzorius, 2001;Dorsey et al., 2002;Vogt et al., 2011b;Deventer et al., 2015). Limited sensor response is one reason for underestimation of turbulent particle fluxes. We corrected sensor response according to Horst (1997) resulting in an average flux correction of between +2.4 and +3.0 % in the 21 size channels. Finally, particle number fluxes with quality flags > 6 according to Foken et al. (2004) 125 were discarded from subsequent analysis.
For the flux footprint prediction, we used the two-dimensional parameterisation of Kljun et al. (2015). To analyse the impact of varying amounts of major and minor roads in the annual flux footprints, we calculated a footprint-weighted average daily traffic intensity (ADT). For this, the ADT shape file (data source: Umweltatlas, 2017) was converted into a raster grid (4 km x 4 km with 4 m spatial resolution, comparable to the footprint climatology raster) and weighted with the footprint climatology 130 to calculate ADT.

Data availability
For the entire three-year period, data availability was 59.2 % for total particle number flux (FTNC) (59.3 % for UFP flux, FUFP).
For the individual observation years, data availability of FTNC decreased from 62.9 % (FUFP: 63.0 %) in the first year to 61.1 % 135 (FUFP: 61.2 %) in the second and 53.6 % (FUFP: 53.6 %) in the third year. The data availability varied due to different reasons such as quality checks, PNSDs that not fulfil the requirements for gap-filling, and off-times due to maintenance of the EEPS.
The lower data availability in the third year was due to a 2-month period (22 May -24 July 2019) during which the EEPS was not available at the site due to inter-comparison measurements. Data availability was generally lower during night (higher rejection of EC data due to quality flags as is common in EC measurements due to reduced turbulence and stronger stability) 140 and showed highest diurnal average data availability for FUFP in autumn (69 %), followed by winter (58 %), spring (56 %), and summer (54 %).

Quantification of measurement uncertainty
Atmospheric observations are prone to measurement uncertainty, e.g. sampling errors as well as sensor and gap-filling 145 uncertainty. In the following, we briefly outline the approach to quantify the potential sources of uncertainty.
The gap-filling of size channels below the minimum threshold concentration of the EEPS causes uncertainty due to the natural spline interpolation method (Meyer-Kornblum et al., 2019). Hence, percentage errors resulting from the interpolation process are dependent on the number of gaps within a PNSD and the particle diameter. The highest average percentage errors of 13.2 up to 31.6 % due to gap-filling occur in the size range of 10-20 nm. The corresponding uncertainty estimates for the size ranges 150 of 20-50 nm, 50-100 nm, and 100-200 nm were between 5.6-17.2 %, 2.0-3.7 %, and 2.2-6.7 %, respectively (Meyer-Kornblum et al., 2019).
The random error for eddy covariance flux measurements was estimated according to Finkelstein and Sims (2001). The median random flux error varies between 27 and 40 % for the different size channels (25 % for FUFP and FTNC) and generally increases towards the boundary size regions of the PNSD. This coincides with random flux errors of between 28 and 39 % as reported 155 by Deventer et al. (2018).

Land-use regression analysis
A multiple linear land use regression (LUR) analysis was carried out to estimate the relationship of FTNC and FUFP with landuse in the flux footprint. The land-use types were determined based on a biotope type mapping provided by city authorities (Geoportal Berlin, 2014). We binned data into 16 wind sectors, which consisted of three land-use types, namely "built-up 160 areas", "traffic areas", and "green areas". Green areas (cf. Figure 1) were not classified into more specific subtypes since the surface fractions of forest, bushes/single trees, and garden areas in the flux footprints were small (cf. Figure 5). Additionally, the available data on specific subtypes of green areas was limited and could not differentiate, for instance, between grass areas with or without single trees. The land-use types "water surfaces" and "other areas" were eliminated to avoid collinearity due to high correlation (Pearson's r ≥ 0.6 or r ≤ -0.6) with other land-use types and the small amount of these surface types within 165 the footprint. Since the model is thought to provide an estimate of the variation of particle emission and deposition with landuse, non-significant variables were not eliminated.

Meteorological characterisation of the study period
For an overview of the meteorological situation in the measurement period, data from the German weather service station "Berlin Tempelhof" (7 km SE from the flux site) was used. The average annual air temperatures during the study period were

Quality check of EEPS
The on-site inter-comparison of particle number concentration between EEPS and WCPC (cf. section 2.2) indicated good 185 agreement with an average relative deviation of 10.6 % (slope 0.94, R 2 = 0.84, Figure A2a). It has to be noted that we compare the WCPC data (lower cut-off of 5 nm) with the gap-filled EEPS data (10 nm < Dp < 200 nm). The deviation in particle concentrations is plausible as it is caused by different lower and upper cut-off diameters.
The time series of WCPC and EEPS fluxes were strongly correlated (r = 0.91), although the EEPS underestimated fluxes measured by WCPC (slope 0.66, R² = 0.84, average relative deviation of 26 %, Figure A2b). We argue that the underestimation 190 is due to the lower WCPC cut-off (5 nm) and the dominance of nucleation mode fluxes at the present site (cf. section 3.5).

Spectral analysis
Spectral analysis is used in EC applications to study the frequency response of the measurement setup. TheEither the measured cospectra are eitherwere compared to ideal cospectra (e.g. Kaimal et al., 1972) or to the sonic sensible heat flux cospectrum which often resembles an ideal cospectrum (Aubinet et al., 2012). 195 We calculated normalised daytime cospectra for FUFP for all wind directions under neutral stratification. Most of the cospectra values were positive (blue dots) indicating the frequent occurrence of emission fluxes (Figure 3). The particle flux cospectrum agreed well with sensible heat and the ideal Kaimal spectrum. Furthermore, the cospectrum follows the theoretical -4/3 slope in the inertial subrange indicating reasonable sensor frequency response.

Particle number concentrations
Over the course of the three-year study period, we observed median TNC concentrations of 7,300 cm -3 and 6,450 cm -3 for 205 UFP, respectively (cf. Table 2). The half-hourly average TNC varied between a minimum of 3,099 cm -3 and a maximum of 53,879 cm -3 . Ultrafine particles dominated rooftop concentrations as indicated by an average UFP/TNC rationratio of 0.9. The majority of particles in UFP size range occurred in the NUC mode.

Footprint analysis
The particle fluxes measured at our rooftop site represented a surface area of about 5.3 km² (with respect to the 80 % contour line). The flux peak contribution, however, was from an area situated to the W/SW of the site at a distance of about 50 to 215 350 m (Figure 4). The peak contribution source area slightly shifted northwards in the second year and southwards in the third year. However, the source areas were very similar in the individual years. A detailed analysis of source-area weighted contribution of land-use indicates 60 % of built-up area, 25areas, 26 % traffic areaareas, 11 % green areaareas, and 2 -3 % water surfacesurfaces in the flux footprint ( Figure 5). Green areas were mainly characterised by grass with or without single trees whereas forests, bushes/single trees, and gardens contribute little to the flux footprint. The land-use fractions were similar 220 in the individual study years. However, it has to be taken into account that streets vary in terms of traffic intensity, so that the strength of traffic-related sources might differ between individual years.

Total and ultrafine particle number fluxes 235
The measured particle number fluxes indicated the study area in central Berlin as a net source of particles with a three year median FTNC of 0.86 x 10 8 m -2 s -1 and 0.83 x 10 8 m -2 s -1 for FUFP (cf. Table 3). The majority of 91.4 % of fluxes were upward (emission) fluxes, whereas 8.6 % of data showed downward (deposition) fluxes. Annual estimates of median FTNC and FUFP showed a decrease by -9.6 % (-8.9 % for FUFP) from the first to the second year and a further -5.9 % (-6.1 % for FUFP) from the second to the third year ( Figure 6a, Table 3). We observed a limited number of frequency of higher FUFP (e.g. > 1.2 x 10 8 m -240 FUFP was characterised by distinct seasonal variation, i.e. largest average fluxes occurred in winter whereas fluxes were lower in autumn and summer ( Figure 7, Table 3). Generally, a lower frequency of deposition events and higher emission fluxes 250 prevailed during winter. However, due to consistent positive monthly median values the site footprint was a net particle emission source in every single month. The seasonal pattern is also evident on the mean diurnal cycle (Figure 8a). Strongest emission fluxes occurred in winter with 255 a maximum FUFP of about 2.0 x 10 8 m -2 s -1 , whereas lowest fluxes were observed in autumn and summer. FUFP increased after sunrise and showed two local maxima at around 9:30 LT and in the early afternoon (12:30 to 14:00 LT). The moderate decrease at around noon was related to reduced traffic intensity (cf. Figure 8d). Average FUFP varied between 0.2 x 10 7 m -2 s -1 at night and 2.0 x 10 8 m -2 s -1 during day. The morning increase in FUFP coincided with an increase in atmospheric turbulence and a more unstable boundary layer. In the evening, FUFP decreased as atmospheric stratification changed from unstable to neutral 260 conditions (Figure 8c).
The daily amplitude of FUFP clearly differed between working days and weekends (Figure 8b). Whereas working days showed a strong morning increase of FUFP and a large amplitude over the diurnal course, the weekend was characterised by a reduced morning increase and only half the amplitude in comparison to working days. This difference was in accordance with the variation of traffic intensity between working days and weekends (Figure 8d). 265

Size-resolved particle number fluxes
Size-resolved particle number fluxes were characterised by distinct temporal variation on the diurnal cycle (Figure 9a).
Strongest emission prevailed in the smallest size bin (Dp < 12 nm) with a maximum in the early afternoon. Coarser particles 275 (Dp > 100 nm), however, showed little variation regardless of the day of week. Although deposition fluxes occurred ( Figure   9b), positive average fluxes were prevailing in all size ranges emphasising the dominance of net particle emission from the flux footprint. The frequency of particle deposition events peaked in the ACC mode (Figure 9b). On working days, the deposition frequency of UFP was highest during night, especially in the second half of the night. On weekends, UFP deposition more often occurred in the morning to pre-noon hours. The timing of deposition fluxes was associated with the daily minimum 280 in FUFP, low traffic intensity, and a neutral or stable atmosphere (cf. Figure 8). Generally, NUC mode particles showed higher deposition probabilities than AIT mode particles. This was evident during the entire day and for every day of the week. On the seasonal cycle, highest emission fluxes predominantly occurred in the smallest size bins (Dp < 30 nm) during winter (average FTNC: 1.44 x 10 8 m -2 s -1 and FUFP: 1.38 x 10 8 m -2 s -1 in December, Figure 10a, b). The lowest average particle number fluxes were evident in late summer and autumn. Additionally, the frequency of particle deposition events was higher in summer/autumn with a maximum in August and September (Figure 10c). The variation in the frequency of deposition fluxes 290 over the size spectrum pointed to the occurrence of bidirectional fluxes.
The difference between the maximum and minimum of deposition frequencies in a monthly size spectrum (max. Δ deposition frequency, Figure 10c) might be interpreted as a measure to quantify the occurrence of bidirectional fluxes. Thus, bidirectionality in particle fluxes occurred more frequently in spring, late summer and autumn than in winter. Bidirectionality was often associated with simultaneous emission fluxes in the ultrafine and deposition fluxes in the coarser particle size range. 295 The months of May, June, and July showed less particle deposition events but higher FTNC and FUFP than the other months in spring and summer. This was probably due to the variation of wind direction and, consequently, flux source area. The average three-year wind directions were from NNE in May, WNW in June, and NW in July. In those directions strong particle sources such as the busy road "Straße des 17. Juni" and the traffic circle "Ernst-Reuter-Platz" were located (cf. Figure 1).

Wind sector analysis of modal particle number fluxes
The wind sector-based analysis indicated net emission of particles from every wind sector (Figure 11a). The largest average 305 FTNC were observed for westerly and northerly wind directions (W-N-NE), i.e. the direction in which "Straße des 17. Juni" and the traffic circle "Ernst-Reuter-Platz" are located (cf. Figure 1). In contrast, particle fluxes from easterly and southerly wind directions were smaller. Vegetated greenspaces, built-up areas but also the main road "Hardenbergstraße" were located in this direction. For all wind directions highest particle number fluxes prevailed in NUC mode (57-72 %), followed by particles in AIT (26-38 %) and ACC modes (2-5 %, Figure 11b). 310 When net fluxes arewere distinguished into emission and deposition fluxes, the emission fluxes showindicated similar behaviour such as the average FTNC due to the high frequency of emission events (cf. chapter 3.5, Figure 12a). TheStrongest emission fluxes occurred for wind from W to NE whereas the most frequent emission fluxes occurredwere evident for wind from 180 to 315° (Figure 12c). Strongest deposition fluxes prevailed under north-westerly wind whereas lower deposition was evident for the remaining directions (Figure 12b). However, the most frequent deposition fluxes were observed under winds 315 from southerly and north-westerly directions (Figure 12d). Since particle deposition occurred most often during night (cf. Figure 9b), the wind direction related frequency distribution of deposition fluxes (Figure 12d) was similar to the night time wind rose (not shown here) which was characterised by most frequent winds from southerly and north-westerly directions.
The wind direction related frequency distribution of emission fluxes, however, was similar to the daytime wind rose (not shown here). Generally, ACC particles were more frequently characterised by deposition fluxes than other particle modes. 320

Land-use regression analysis
The three-year data set was analysed with regard to variation of land-use across the flux footprint (cf. Figure 13). We observed the sectors with the highest fraction of traffic areas to match with the highest particle fluxes, e.g. the sector with the highest 330 traffic fraction of 53 % resulted in the largest particle emission flux with 2.05 x 10 8 m -2 s -1 .
To further analyse this relationship, a LUR analysis was carried out for FTNC and FUFP in the 16 wind sectors. At a confidence level of 95 %, only the land-use type "traffic areas" was highly significant with a p-value < 0.001 (Table 4). The LUR analysis based on the three-year data set gives further evidence for the important role of traffic as UFP emission source. Traffic areas emitted FUFP = 3.44 x 10 8 m -2 s -1 , followed by built-up areas with 4.87 x 10 7 m -2 s -1 (not significant, Table   4). Green areas seem to be a sink of particles with FUFP = -1.46 x 10 7 m -2 s -1 although the relationship was statistically not significant. However, the analysis does not take temporal variation of traffic intensity into account. A comparison of hourly 340 traffic data from counting stations with particle number fluxes showed a distinct increase in FUFP with intensifying traffic ( Figure 14). The strongest increase was obvious for FNUC, followed by FAIT. FACC only slightly increased with traffic intensity.

Discussion 350
The particle number fluxes observed at the Berlin site show dominance of particle emission with a three-year median FTNC of 0.86 x 10 8 m -2 s -1 and FUFP of 0.83 x 10 8 m -2 s -1 which makes the site a net emission source of ultrafine particles to the urban boundary layer. Urban particle number flux measurements from different locations are, however, not easy to compare given the different diameter cut-offs of particle counting instruments. Due to the dominating impact of UFP on the number flux, the lower cut-off is of particular importance. However, our results coincide with other studies from cities across Europe using a 355 similar lower cut-off in the particle counting instruments (cf. Table 1). TotalAverage total number fluxes of the studies in Stockholm (Mårtensson et al., 2006), Innsbruck (Deventer et al., 2018;Heyden et al., 2018) orand Lecce (Conte et al., 2018) are in a similar range. the same order of magnitude varying between 0.5 and 3.0 x 10 8 m -2 s -1 .
Although the three observation years point to constant emission of ultrafine particles from the urban surface, annual estimates show a median FUFP decrease of -8.9 % from the first to the second and -6.1 % from the second to the third observation year. 360 The observed decrease might be due to several reasons. Analysing the temporal variation of traffic count data, we found a slight reduction in the third year ( Figure A3A3b). This could be one reason for the lower median FUFP in the third observation year. In addition, the decrease might be associated to the slight spatial shift of the peak contribution location in the flux footprint towards southerly directions. These source areas are characterised by a lower fraction of traffic areas and reduced traffic intensity in comparison to northerly directions. In the second year, the average wind direction and also the peak contribution 365 location in the flux footprint shifted towards north-westerly directions. These were related to stronger particle deposition fluxes (cf. Figure 12) that tend to reduce FUFP. Additionally, aAs traffic areas in the flux footprint have different traffic intensity (i.e. major and minor roads), the annual variation of surface fractions of traffic areas in the flux footprints might not be sufficient to explain particle number flux differences. However, the variation of traffic impact can be estimated using a footprintweighted ADT, which estimates annual variation of traffic influence by quantifying the amount of major and minor roads 370 contributing to the annual flux footprint ( Figure A3a). This indicates a relationship between particle fluxes and footprintweighted traffic contribution, which we assume to be the major reason for the observed variation in annual average particle fluxes. Additionally, the transition of the vehicle fleet towards a higher share of low-emission standards or electric drive might play a role. According to official car registration data, the proportion of electric vehicles in Berlin increased by a factor of three in the time period from 2017 to 2020 (Kraftfahrt-Bundesamt, 2021). Additionally, inIn the third observation year lower 375 traffic intensity due to the COVID-19 lockdown restrictions starting in mid-March 2020 might have influenced our measurements and may explain the local minimum fluxes in March 2020 (cf. Figure 7). However, an in-depth analysis of effects related to reduced traffic intensity during the COVID-19 lockdown is beyond the scope of this study but will be addressed in future work.
The analysis of size-resolved particle number fluxes points to the dominance of emission fluxes over the diurnal cycle. The 380 strongest emission fluxes occur during daytime with a maximum of smallest particles in the early afternoon (e.g. Schmidt and Klemm, 2008;Deventer et al., 2018). The diurnal courses of particle fluxes coincided with traffic intensity demonstrating strong influence of traffic on FNUC. The delayed increase and reduced amplitude of traffic intensity weekends compared to a working day was clearly evident in particle flux data. At this time of low traffic intensity, particle deposition occurred much more frequently. While coarse particles (Dp > 180 nm) were deposited with a frequency > 20 % at any time throughout the 385 day, UFP deposition preferably occurred in the second half of the night or other low-emission periods, e.g. weekend mornings.
On the mean diurnal cycle no such tipping point can be identified at the present site, since our data does not show 390 simultaneously occurring average upward and downward fluxes across the size spectrum.
We found a dominating contribution of FUFP to the total flux (96 %) which corresponds to findings from Innsbruck (99 %; size range: 6 nm < Dp < 637 nm; Deventer et al., 2018) and London (Harrison et al., 2012). In addition, the diurnal course of FUFP showing a nocturnal minimum and an early afternoon maximum that is driven by traffic intensity and atmospheric stability is comparable to other findings (Dorsey et al., 2002;Deventer et al., 2013;Conte et al., 2018). A typical double rush-hour pattern 395 of traffic intensity that is often not occurring in particle flux data (Dorsey et al., 2002;Järvi et al., 2009;Martin et al., 2009;Deventer et al., 2015) is slightly evident in the diurnal cycle of average FUFP in Berlin. Also the differences in the diurnal cycles of FUFP between working and weekend days point to dominating influence of traffic on the measured fluxes at our site. This was confirmed by wind-dependent analysis of flux variation, the regression of fluxes against land-use in the flux source area (LUR analysis) and the temporal variation with traffic data from counting stations. Similar findings of higher fluxes from 400 traffic areas and lower fluxes from areas with higher amounts of green surfaces and buildings were reported by Järvi et al. (2009), Ripamonti et al. (2013), and Mårtensson et al. (2006).
In each wind direction sector, we found the highest particle number fluxes in nucleation mode (FNUC ≥ 57 % of average FTNC, cf. Figure 11b). Road traffic is the main source of particles in the size range below 50 nm (Morawska et al., 2008). Especially rush-hour periods are characterised by a maximum of traffic-related particle emission in the size range between 20 and 30 nm 405 (Wiedensohler et al., 2002). Thus, the large contribution of FNUC confirms the high influence of traffic intensity on fluxes at this site.
The deposition frequency seems to be related to the strength of local sources as indicated by the increase of deposition frequency with decreasing FUFP and FTNC (cf. Figure 10). In addition, theThis may also be the reason for the higher frequency of particle deposition in ACC than UFP, since ACC emission fluxes are significantly lower than AIT and NUC mode fluxes 410 (cf. Figure 9a). The occurrence of bidirectional fluxes in the annual course also seems to depend on the strength of the surrounding sources with a more frequent appearance in seasons with lower emissions. Thus, they occur more often in summer and autumn than in winter (cf. Figure 10c). Deventer et al. (2015) also reported differences in particle deposition between summer and winter and found an increase of deposition fluxes for Dp > 170 nm in spring and summer.
The highest particle number fluxes were observed in winter (cf. Figure 10a) due to a larger number of emission sources (e.g. 415 domestic heating) and reduced atmospheric dilution. However, the LUR analysis illustrates that the relationship between sources from built-up areas and particle number fluxes is statistically not significant and that emission fluxes from built-up areas are one order of magnitude smaller than traffic area fluxes (cf . Table 4). Thus, other local emission sources such as domestic heating in winter are of limited influence for particle number fluxes at this site in contrast to traffic related sources.

Conclusions 420
We report on the first multi-annual data set of urban size-resolved particle number fluxes that were measured over the size range 10 < Dp < 200 nm from April 2017 to March 2020 in an urban area of Berlin, Germany, using the eddy-covariance technique. The Berlin site was a net source of particles with a majority of ultrafine particle emission as indicated by a FUFP/FTNC ratio of 0.96. The magnitude and temporal variation of FTNC and FUFP were rather similar in the individual observation years, however, a reduction of median FUFP of -8.9 % and -6.1 % for the second and third years was evident. This might be related to 425 variation in flux source area, a modification in the vehicle fleet, and an effect of reduced traffic intensity in the third measurement year.
The results clearly point to traffic as the dominant influence on ultrafine particle number fluxes at the Berlin site. Due to the health effects associated with UFP, measures that result in reduced traffic intensity, e.g. by promoting alternative urban mobility concepts, or a transition of the traffic fleet towards a higher share of electric or low-emission vehicles may help to 430 reduce personal exposure towards UFP. Long-term observations of size-resolved particle number fluxes form different urban environments would be an important tool to monitor and asses the success of different mobility strategies and reduction measures.

Data availability 450
The data used in this study are publicly available: "Three years of size-resolved eddy-covariance particle number flux measurements in Berlin, Germany" on https://publikationsserver.tu-braunschweig.de/receive/dbbs_mods_00069675.

Author contribution
SW acted as the supervisor with responsibilities concerning conceptualisation, funding acquisition, project administration as well as reviewing and editing of the paper draft. Together with AS, he designed the experiments. AS took care of data curation, 455 maintaining the flux site, formal data analysis, measurement operation and data collection, methodology, software use and development for data analysis as well as data visualisation. Additionally, AS prepared the original draft of this paper.