2.4.1 PMF description

The U.S. EPA PMF 5.0 model was used for VOC source apportionment. The method is described in detail in Paatero and Tapper (1994), Paatero (1997), and Paatero and Hopke (2003).

The general principle of the model is as follows: any matrix X (input chemical dataset matrix) can be decomposed in a factorial product of two matrixes G (source contribution) and F (source profile) and a residual part not explained by the model E. Equation (1) summarizes this principle in its matrix form:

where i is the number of observations, j the number of the measured VOC species, and k the number of factors.

The goal of the PMF is to find the corresponding non-negative matrixes that lead to the minimum value of Q.

where f_kj≥0 and g_kj≥0 and where n is the number of samples, m the number of considered species, and s_ij an uncertainty estimate for the jth species measured in the ith sample.

2.4.2 Preparation of input data

Two input datasets (or matrixes) are required by the PMF: the first one contains the concentrations of the individual VOC and the second one contains the uncertainty associated with each concentration.

The VOC input dataset combines data from both the PTR-MS and the GC–FID. The PTR-MS data have been synchronized with the ones from the GC–FID on its 20 min sampling time every 30 min. The final chemical database used for this study comprises a selection of 23 hydrocarbons and masses divided into seven compound families: alkanes (isobutane, n-butane, n-hexane, n-heptane, isopentane, n-pentane, and 2-methyl-pentane), alkenes (1-pentene, isoprene, and 1,3-butadiene), aromatics (ethylbenzene, benzene, toluene, (m + p)-xylenes, o-xylene, and C9 aromatics), carbonyls (methyl ethyl ketone (MEK), methacroleine + methyl vinyl ketone (MACR + MVK), acetaldehyde, and acetone), alcohol (methanol), nitrile (acetonitrile), and terpenes. Alkanes and alkenes were measured by the GC–FID while benzene, toluene, isoprene, C8 aromatics, carbonyls, alcohol, nitrile, and terpenes were the ones measured by the PTR-MS. For benzene, toluene, and C8 aromatics the PTR-MS data were selected because of the smallest number of missing data.

Since the PMF does not accept missing values, missing data must be replaced. The percentage of missing values ranges from 8 % to 79 % for species measured by the GC–FID. Butanes (iso/n), pentanes (iso/n), and (m + p)-xylenes have the lowest missing value percentage (ranging from 8 % to 12 %) while o-xylene (30 %), 2-methyl-pentane (32 %), ethylbenzene (42 %), 1-pentene (62 %), n-hexane (66 %), n-heptane (72 %), and 1,3-butadiene (79 %) have the highest percentage of missing values. There were only two missing values (0.33 %) for species measured by the PTR-MS.

For those species with a proportion of missing values below 40 %, missing data were replaced by a linear interpolation. For species with a proportion of missing data exceeding 40 %, each missing data point was substituted with the median concentration over the whole measurement period. All the concentrations were above the detection limit.

The uncertainty σ_ij associated with each concentration (Eq. 3) is determined using the method developed within the ACTRIS (Aerosol, Cloud and Trace Gases Research Infrastructure) network (Hoerger et al., 2015) and used in Salameh et al. (2016):

This uncertainty considers the different sources of uncertainty affecting the precision and the accuracy terms. The precision is associated with the detection limit of the instrument and the repeatability of the measurement, while the accuracy includes the uncertainty of the calibration standards and the dilution when needed.

The uncertainty of the PTR-MS ranges between 5 % (toluene) and 59 % (acetaldehyde) of the concentrations while the uncertainty for the GC–FID ranges between 4 % (2-methyl-pentane) and 17 % (o-xylene) of the concentration. Further information about the uncertainties calculation are found in Sect. S1 of the Supplement.

2.4.3 Determination of the optimal solution

The objective of the PMF is to determine the optimal number of factors (p) based on several statistical criteria. Several base runs were performed with a different number of factors from 2 to 12. Statistical criteria were then used to determine the appropriate p value such as Q (residual sum of squares), IM (maximum individual column mean), IS (maximum individual column standard deviation) as defined by Lee et al. (1999), and R² (indicator of the degree of correlation between predicted and observed concentrations). Q, IM, IS, and R² are then plotted against the number of factors (from 2 to 12) (Salameh et al., 2016). The number of chosen factors corresponds to a significant decrease in Q, IM, and IS. In our study, an optimal solution of five factors was retained. In order to ensure the robustness of the solution, a F_peak value of −1 was set by considering the highest mean ratio of the total contribution vs. the model as well as the numbers of independent factors (Salameh et al., 2016).

Table 1Statistical summary of VOC concentrations (ppbv) measured at the urban site of Besiktas from 17 to 30 September. The initials no., GC, P, T, and C stand for the number of samples and instruments measured by the GC–FID, PTR-MS, tubes, and canisters.

Download Print Version | Download XLSX

2.4.4 PMF reference run

A PMF reference run has been performed by removing the period during which there were no GC–FID data (night from 24 to 25 September). In addition, these datasets have been chosen as the PMF reference run because of the higher correlation between observed and reconstructed data by the PMF model (see also Sect. 3.2). A good correlation (R²=0.97) between total reconstructed VOC and measured VOC was obtained. For most compounds the variability is reproduced well with a R² usually higher than 0.70. Poorer correlation was found for alkenes (1-pentene (R²=0.55), 1,3-butadiene (R²=0.22), and isoprene (R²=0.57)) as well as for n-hexane (R²=0.09), MEK (R²=0.41), and acetaldehyde (R²=0.32). The R² of the contribution of the five factors between each other has been calculated; it was found that the value of R² does not exceed 0.28 and is usually less than 0.05, indicating the statistical independence of the five factors. There is therefore no significant correlation between the factors, which means that the factors are independent.

The PMF output uncertainties were estimated by three models: the DISP model (base model displacement error estimation), the BS model (base model bootstrap error estimation), and the DISP + BS model. Further information for the estimation of model prediction uncertainties can be found in Norris et al. (2014) and Paatero et al. (2014). The DISP results of the PMF run show that the five-factor solution is stable and sufficiently robust to be used because no swaps occurred. All the factors were reproduced well through the BS technique at 100 % for factor 1, 96 % for factor 2, 100 % for factor 3, 99 % for the factor 4, and 100 % for factor 5; there were not any unmapped runs. The DISP + BS model shows that the solution is well constrained and stable.

2.4.5 Sensitivity tests to evaluate PMF results

The sensitivity of PMF results to input data has been tested in order to evaluate the representativeness of the PMF reference run results. These tests evaluate the effect of some time period selection described in Sect. 3 (periods 1, 2, and 3, nighttime peaks) and the incorporation of species with a high number of missing data as well as a test where all missing data were directly replaced by the median of species over all the measurement period instead of interpolation.

Table 2Contributions of each factor to each sensitivity test (as a percentage).

^a Displayed in red. ^b Displayed in orange. ^c Displayed in green. ^d Displayed in blue. ^e Displayed in black.

Download Print Version

Table 2 summarizes the contribution of each factor for every sensitivity scenario as well as the values of the correlations between observed and modeled concentrations. The reference run has the best fit with $R_{\mathrm{total}}^{\mathrm{2}}=\mathrm{0.97}$, and more than 82 % of species were reconstructed well by the PMF (r²≥0.5). While the sensitivity tests do not perform as well statistically as the reference run, they all show that the same factors are extracted even though the relative contributions could be slightly modified by ±16 %. Factors will be discussed in Sect. 3.3.

3.2.1 Ambient concentration level and composition

Observed VOCs at the super site of Besiktas are representative of the ones usually encountered at urban background sites. Biogenic compounds are isoprene and terpenes while anthropogenic compounds include both primary (alkanes, alkenes, aromatic compounds) and secondary compounds (oxygenated VOC).

Statistics on the VOC concentrations measured by GC–FID and PTR-MS at the super site of Besiktas are reported in Table 1. VOCs show a high temporal variability with maxima reaching several tens of parts per billion (isopentane, ethylbenzene, methanol, and acetone). Most VOC median concentrations are below 1 ppb except for n-butane (1.48 ppb), toluene (1.25 ppb), and some oxygenated compounds like acetaldehyde (1.39 ppb) and acetone (2.33 ppb). The average composition of VOCs mainly includes OVOCs (0.12–4.92 ppb), which represent 43.9 % of the total VOC (TVOC) observed mixing ratio, followed by alkanes (0.21–2.00 ppb; 26.33 %), aromatic compounds (0.26–2.27 ppb; 20.66 %), alkenes (0.19–0.68 ppb; 4.81 %), terpenes (3.44 %), and acetonitrile (0.84 %). Both OVOCs and alkanes contribute up to 70.2 % of the TVOC concentrations. Methanol (22.40 %, 4.92 ppb on average) is the main oxygenated compound measured in this study, followed by acetone (11 %, 2.63 ppb). N-butane (9.11 %, 2.00 ppb) and isopentane (6.43 %, 1.41 ppb) are the two major alkanes. Toluene is the most abundant aromatic compound (50 %, 2.27 ppb). This was the case for most of the previous VOC studies performed in Istanbul (Demir et al., 2011; Kuntasal et al., 2013; Muezzinoglu et al., 2001; Bozkurt et al., 2018; Elbir et al., 2007).

Levels of alkanes, some alkenes, and aromatics are compared to levels in other European megacities: Paris and London (Borbon et al., 2018) at both urban and traffic sites as well as at an urban site in Beirut (Salameh et al., 2015) during summer (Fig. S4 in the Supplement). Consistency in urban hydrocarbon composition worldwide has already been observed (Borbon et al., 2002; von Schneidemesser et al., 2010; Dominutti et al., 2016). Despite differences in absolute levels, the hydrocarbon composition in Istanbul is quite similar to that in the other cities. Beirut has the highest concentrations of n-butane, isopentane, and 2-methyl-pentane. Higher concentrations in toluene and (m + p)-xylenes were found in Paris, Beirut, and Istanbul. Such a similarity would suggest that the same sources control the hydrocarbon composition, especially traffic in all cities including Istanbul (see also Sect. 3.3 on PMF).

Figure 4VOC composition at various locations in Istanbul from tubes (a) and canister (b) sampling.

Download

Off-line VOC concentrations collected with tubes are reported in Table S5 in the Supplement at various locations including the Besiktas super site. While the number of samples is limited in time, off-line measurements provide a general picture of a wider spectrum of VOCs that are not measured at the super site like light hydrocarbons (C2–C3 alkanes and alkenes), C6–C11 carbonyls, speciated terpenes (camphene, limonene, β-pinene), and C12–C16 intermediate-volatility organic compounds (IVOCs). Therefore, these datasets allow us to address the spatial variability of VOC concentrations and composition within the megacity. The comparison of concentrations between different areas is limited because of the different sampling times. Therefore, the analysis only focuses on VOC relative composition. The relative composition divided into major VOC chemical groups at each site by sorbent tubes and canisters is reported in Fig. 4a and b, respectively. The composition is variable across the megacity for the aliphatic fraction of high- and intermediate-volatility hydrocarbons (C2–C16). As expected, the composition of the 29 September 12:05 UTC+3 (for all times throughout) sample at the Besiktas site is like that derived from the roadside measurements, highlighting the influence of road transport emissions at the super site. Interestingly, the VOC composition of the three samples from sorbent tubes at the Besiktas site are different from the ones at the roadway site with a higher proportion of IVOCs. The VOC composition of the 26 September 10:31 sample by canister is rather similar to that from the seashore sample in Galata (29 September 16:12 sample). In the same way the VOC composition of the samples at the super site derived from tubes is rather like that of the Besiktas seashore. For both of them, the proportion of IVOCs is significant (from 15 % to 40 % in weight). While light VOCs are expected to be of minor importance when considering ship emissions, the higher presence of heavier organics is however expected as observed for alkanes by Xiao et al. (2018) in ship exhaust at berth. The VOC composition comparison would thus suggest not only the impact of road traffic emissions on their composition but also the potential impact of local ship traffic emissions. Finally the composition at Besiktas is not affected by residential emissions, which are enriched in light C2–C3 alkanes (canisters) or aromatics (canisters).

3.2.2 Time series and diurnal variations

The variability of VOC concentrations is driven by several factors: emissions (anthropogenic or biogenic), photochemical reactions (especially with the OH radical during the day and ozone and nitrates at night for alkenes), and the dynamics of the atmosphere (including dilution due to the height of the boundary layer) (Filella and Peñuelas, 2006). The time series of inorganic trace gases (NO_x and CO) and some VOCs representing the diversity of sources and reactivity are reported in Fig. 5. The meteorological periods 1, 2, and 3 described in the previous Sect. 3.1 are also indicated. Because NO_x at the super site was only measured from 25 to 30 September, data from the air quality station in Besiktas were used (see Fig. 1). One should note that the time series of NO_x at the super site and at the Besiktas station are consistent.

Time series of NO_x and CO show high concentrations but a different pattern regardless of the origin of air masses. While a daily cycle of NO_x is depicted, CO does not show any clear pattern. The NO₂∕NO_x ratio fluctuates between 0.34 and 0.93 with an average and median value of 0.53 and 0.55, respectively. These values are very high compared to what is usually found in the literature (Grice et al., 2009; Kousoulidou et al., 2008; Keuken et al., 2012), which is mostly low and below 0.50. However higher values of the NO₂∕NO_x ratio can be found in diesel passenger cars (Grice et al., 2009; Vestreng et al., 2009) and vans (Kousoulidou et al., 2008). This ratio would reflect the impact of the combustion of heavy fuels in a megacity. After road transport, cargo shipping is the second highest contributor to NO_x levels according to the local/regional inventory (Markakis et al., 2012).

Anthropogenic VOC time series (benzene, isopentane, and isobutane) exhibit a high-frequency variability but usually show higher concentrations during the night, especially during period 2. One cause is the very low wind speeds at night, especially during period 2 (Fig. 3), which would reinforce the accumulation of pollutants. Under marine influence (periods 1 and 3), VOC concentrations are the lowest, especially during period 3, which is characterized by rainy days (27 and 28 September), high wind speed, and colder temperatures (Fig. 5). These conditions favor atmospheric dispersion. During transition periods and under continental influence (period 2), VOC concentrations exhibit a strong day-by-day variability with episodic nocturnal peaks in particular on 25 and 26 September. While these peaks are not always concomitant between VOC and are not associated with any increase in NO_x and CO levels, they occur under south and southwestern wind regimes, which are unusual wind regimes according to Fig. 1. This points out the potential influence of industrial and port activities other than fossil fuel combustion. For instance, maximum concentrations of butanes occurred during the period of the marine–continental regime shift with the well-established southwestern wind regime on 22, 23, and 26 September at the end of the day. Maximum concentrations of pentanes occurred during the night of 26 to 27 September like for aromatics (e.g., benzene) (Fig. 5).

Except during transition periods, the background levels of measured trace gases are not affected by the origin of air masses. This strongly suggests that the pollutants measured during TRANSEMED Istanbul were from local and regional sources. Finally, time series would suggest the influence of multiple local and regional sources other than traffic on VOC concentrations, likely industrial and/or port activities, at the super site.

Isoprene and its oxidation products (MACR + MVK) covariate most of the time in periods 1 and 3. They usually show their typical diurnal profiles with higher concentrations during the warmest days and at midday due to biogenic emission processes. Their significant correlation with temperature (R=0.7) implies emission from biogenic sources. Around the Besiktas site, 49.5 % of the vegetation is occupied by hardwood and hardwood mix trees while only 6 % is occupied by softwood and hardwood mix trees. While Quercus (isoprene emitter) only occupies 7.7 % of the total vegetation coverage (Ministry of Forestry, personal communication, 2018), the presence if isoprene at the super site is probably due to the surrounding trees.

Figure 5Time series of NO_x, CO, and a few VOCs. Background colors represent the different periods relative to meteorological conditions (light grey for periods 1 and 3 and dark grey for period 2). Time series of NO_x data from the super site are in pink and the ones from the Besiktas site are in red.

Download

Except during transition periods, the background levels of measured trace gases are not affected by the origin of air masses. The background levels stay constant under continental or marine influence and regardless of the atmospheric lifetime of the species. This strongly suggests that the pollutants measured during TRANSEMED Istanbul were from local and regional sources. Finally, time series would suggest the influence of multiple local and regional sources other than traffic on VOC concentrations, likely industrial and/or port activities, at the super site.

Figure 6Diurnal profiles of some selected gaseous species (ppbv). Blue shaded areas represent the minimum and maximum diurnal concentrations over all the measurement periods, and blue lines show the average diurnal concentrations during periods 1 and 3 and red lines the average concentrations during period 2.

Download

Taking into consideration time series variability, diurnal variations have been split into periods 1 and 3 and period 2 for selected VOCs as well as two combustion-derived trace gases (NO_x and CO). Diurnal profiles of atmospheric concentrations are reported in Fig. 6. Local traffic counts for road transport (personal communication from Istanbul Municipality for fall 2014) and shipping (https://www.marinetraffic.com/en/ais/details/ports/724/Turkey_port:ISTANBUL, 14 January 2019) are also reported in Fig. S6 in the Supplement. Maritime traffic is mostly for passenger shipping (58.02 %) compared with 16 % for cargo shipping. The diurnal profiles of ship and road traffic counts are similar.

Generally, concentrations during period 2 are higher than the ones during periods 1 and 3 and show different diurnal patterns for some compounds. The profile of NO_x is consistent with the one of traffic counts (Fig. S6 of the Supplement). NO_x exhibits higher concentrations during the day and lower concentrations at night for both periods with a morning peak (07:30–08:30) and one early evening peak from 17:30 (Fig. 6a). This is typical of traffic-emitted compounds with morning and evening rush-hour peaks as observed in many other urban areas like Paris, France, in Europe (Baudic et al., 2016) or Beirut, Lebanon, in the eastern Mediterranean (Salameh et al., 2016). As already depicted in time series, the CO diurnal profile is different from that of NO_x. CO concentrations show higher concentrations in the late evening and lower concentrations during the day. During the day, CO is also characterized by a double peak: one in the morning (08:30) and the other one in the middle of the day (Fig. 6b). Both NO_x and CO show quite similar diurnal profiles between the three periods even if morning concentrations tend to be higher.

VOCs show different profiles from that of NO_x. Under marine influence (periods 1 and 3), primary anthropogenic VOCs (i.e. benzene, alkanes, and other aromatics) almost exhibit a constant profile while they show a higher concentration from midnight until 10:00 under continental influence (period 2). For instance, benzene (Fig. 6d) and isopentane (Fig. 6f) nighttime concentrations increase by 4-fold compared to the levels under marine influence. In the middle of the day, the concentration levels are the same as during periods 1 and 3. The profiles of primary anthropogenic VOCs show the complex interaction between local and regional emissions and dynamics. Period 2 shows the influence of VOC emissions other than traffic and combustion processes (no effect on NO_x and CO) at night. While the influence of traffic emissions on CO and VOC cannot be excluded; it seems that their emission level is not high enough to counteract the dispersion effect during the day unlike NO_x. This will be further investigated in the PMF analysis.

Isoprene concentrations increase immediately at sunrise and decrease at sunset during periods 1 and 3 (marine influence), which indicates its well-known biogenic origin (Fig. 6g), which is light and temperature dependent. Isoprene and MACR + MVK concentrations increase at night during period 2 like other alkanes and aromatics, suggesting their potential anthropogenic influence.

Provided some interferences like furans could contribute to isoprene signal by PTR-MS measurements (Yuan et al., 2017), this would suggest an anthropogenic origin for isoprene. While the signals of m∕z 71 are commonly attributed to the sum of MVK and MACR, which are both oxidation products of isoprene under high-NO conditions, more recent GC-PTR-MS studies identified some potential interferences for MVK and MACR measurements, including crotonaldehyde in biomass-burning emissions, C5 alkenes, and C5 or higher alkanes in urban regions (Yuan et al., 2018). Such interferences cannot be ruled out here. During periods 1 and 3, MACR + MVK concentrations follow the same general pattern as isoprene.

With a relatively long atmospheric lifetime (≈68 d), acetone's concentration is quite constant throughout the day within period 1, with a peak in the middle of the day and lower concentrations during the night for period 2 (Fig. 6c). The peak in the middle suggests the presence of a secondary origin. Acetone can have both primary and secondary sources (Goldstein and Schade, 2000; Macdonald and Fall, 1993). Methanol and MEK have the same general pattern as for acetone during both periods without the peak in the middle of the day, suggesting that they might have the same emission source.

3.3.1 Factor identification

The factor profiles have been analyzed using the variability of their contribution together with several external variables (NO_x, SO₂, CO, meteorological parameters, emission profiles). PMF factors are displayed in Fig. 7. The time series and diurnal profiles of their contributions are displayed in Figs. 8 and 9, respectively.

Figure 7Source composition profiles of the PMF factors. The concentrations (ppbv) and the percent of each species apportioned to the factor are displayed as a pale blue bar and a red color box, respectively.

Download

Figure 8Time series of factor contributions (ppbv) extracted from the PMF.

Download

Figure 9Diurnal variation in source contributions (ppbv).

Download

Factor 1: Toluene from solvent use

The speciation profile of factor 1 exhibits high concentrations of toluene with 57 % of its variability explained by this factor (Fig. 7). There is also a small contribution of acetone (18 %), MEK (18 %), and xylenes (17 %). While toluene and xylenes are related to traffic emissions, this factor does not correlate well with any traffic tracer gases (R=0.33 for NO_x,) and combustion trace gases in general ($R=-\mathrm{0.03}$ for CO and 0.13 for SO₂). Moreover toluene also contributes to the road transport factor (factor 5) but to a lesser extent. The time series are highly variable with erratic peaks regardless of the time of the day, especially during period 2 and to a lesser extent during period 1 (Fig. 8). The average diurnal profile looks constant but the relative standard deviation is ±100 % (Fig. 9). Sources related to solvent use are among the expected non-combustive sources for toluene (Baudic et al., 2016; Gaimoz et al., 2011; Brocco et al., 1997; Na and Kim, 2001). In Turkey, toluene was already found in gasoline vehicle, solvent, and industrial emissions (Bozkurt et al., 2018; Yurdakul et al., 2013; Demir et al., 2011).

The toluene ∕ benzene ratio (T∕B) is used as an indicator of non-traffic source influence (Elbir et al., 2007; Lee et al., 2002; Yurdakul et al., 2013). A T∕B ratio ≤2–3 indicates the influence of traffic emissions on measured VOC concentrations (Gelencsér et al., 1997; Heeb et al., 2000; Muezzinoglu et al., 2001; Brocco et al., 1997), whereas T∕B ratios ≥2–3 suggest the influence of sources other than traffic (such as solvent evaporation or industrial sources). The T∕B ratio for this study is between 0.4 (with only four points below 2) and 48.6 (only one point above 29). Only 5.8 % of the ratios were between 2 and 3, 48 % were between 3 and 6, and 45 % were above 6 with 34 % between 6 and 10. This strongly suggests the influence of sources of toluene other than traffic. A high value of the T∕B ratio is mostly found at industrial sites (Pekey and Hande, 2011). The median and mean values of T∕B in this experiment are respectively 5.6 and 6.7, which can also indicate gasoline-related emissions (Batterman et al., 2006). However the absence of other unburned fuel compounds like pentanes excludes this source. These ratios were calculated with toluene and benzene measured by the PTR-MS since the PMF run was performed using those data. By looking at the T∕B ratio measured by the GC–FID, we found approximately the same conclusion: only 1 % of the ratios were between 2 and 3, 47 % were between 3 and 6, and 51 % were above 6 with 38 % between 6 and 10. This factor represents 14.2 % of the total contribution.

Factor 2: Biogenic terpenes

Factor 2 exhibits a high contribution of terpenes with more than 73 % of their variability explained by this factor (Fig. 7). Terpenes are known as tracers of biogenic emissions (Kesselmeier and Staudt, 1999). Isoprene, which is also a biogenic tracer, has only 5 % of its variability explained by this factor. Moreover, the diurnal profile of these two compounds shows opposite patterns as can be seen in Figs. 6 and 9, which indicates that their biogenic emissions are controlled by different environmental parameters: temperature for terpenes and light and temperature for isoprene (Fuentes et al., 2000). Furthermore, the Besiktas site is also surrounded by Pinus, which are terpene emitters and represent the maximum overall vegetation in Istanbul (up to 33 %) (personal communication from Ministry of Forestry). This factor shows high concentrations and contribution during period 1 and the transition periods (Figs. 2 and 7), probably due to a higher temperature while they are almost not significant during period 3. As expected, terpenes do not correlate with any combustion-related gases (R<0.14 for NO_x, CO, and SO₂).

The diurnal profile of this factor is characterized by high concentrations at night and early morning (until 08:30) and low concentrations during daytime (Fig. 9). This type of profile has already been observed at a background site in Cyprus (Debevec et al., 2017), in a forest of Abies borisii-regis in the Agrafa Mountains of northwestern Greece (Harrison et al., 2001), and at Castel Porziano near Rome, Italy (Kalabokas et al., 1997). In a shallower nocturnal boundary layer, low chemical reactions together with persistent emissions lead to the enhancement of their nocturnal mixing ratios. Terpene's lifetime toward OH and ozone is 1.2 to 2.6 h and 5 min (for α-pinene) to 50 min (for camphene), respectively (Fuentes et al., 2000). Dilution processes of the boundary layer in addition to the higher reactivity of terpenes towards the OH radical and ozone could explain the decrease in their concentrations during the day.

This factor is called “biogenic terpenes” and represents 7.8 % of the total contribution.

Factor 3: Natural gas evaporation

Factor 3 is essentially composed of butanes (iso/n) with more than 97 % of their variability explained by this factor (Fig. 7). Isobutane is a typical marker of fossil fuel evaporative sources (Debevec et al., 2017; Na et al., 1998). This factor significantly explains the contributions of 1,3-butadiene (32 %), acetonitrile (23 %), acetaldehyde (24 %), MACR + MVK (21 %), C9 aromatics (22 %), and benzene (20 %). Iso/n butanes correlate poorly with CO (R=0.09) and NO (R=0.33). The alkanes ∕ alkenes ratio of this factor is high (55), which points out the evaporative source of this factor (Salameh et al., 2014). At the same time, the pentanes (n/iso) and the other aromatic compounds are not well represented in factor 3. This suggests that butane evaporation is not related to evaporation from storage, extraction, and distribution of gasoline but rather to natural gas evaporation.

The diurnal profile of factor 3 is characterized by an increase in concentration at night and constant concentration from 10:00 to 18:00 with a slight peak in the middle of the day and in the evening (Fig. 9). This points out a source linked to the use of natural gas as an energy source, especially for cooking at lunchtime with the proximity of many restaurants near the measurement site. This type of profile has already been observed in Paris (Baudic et al., 2016). The time series are characterized by several peaks during period 2, which corresponds to the marine transition by the south-southwest of Istanbul and the Marmara Sea (Fig. 8). The presence of a power plant with a capacity of 1350 MW located in the southwest of Ambarli that uses natural gas as fuel and which can also contribute to the butane loads is noteworthy. This factor does not depend on temperature or on other trace gases (NO_x, CO, and SO₂). The average relative contribution of the natural gas factor is 25.9 %.

Factor 4: Mixed diurnal regional emissions

Factor 4 has a significant contribution of several primary and secondary as well as biogenic and anthropogenic species (Fig. 7). N-hexane is the most dominant species (81.6 % of its variability is explained by this factor). However, knowing that this compound has more than 70 % of the missing values, the analysis related to this compound should be performed carefully. The great contribution of isoprene (66 %) in factor 4 points out its biogenic emissions. Biogenic emissions of isoprene are directly related to temperature as well as solar radiation (Steiner and Goldstein, 2007; Owen et al., 1997; Geron et al., 2000). Note that this factor correlates well with the ambient temperature (R=0.70). Despite the percentage of missing values, the biogenic contribution for this factor can also be explained by the presence of 1,3-butadiene and 1-pentene likewise emitted by plants (Goldstein et al., 1996). Factor 4 is also characterized by the presence of oxygenated compounds such as isoprene oxidation products like MACR + MVK (54 %), acetaldehyde (66 %), acetone (57 %), methanol (59 %), and MEK (59 %). These oxygenated species can have primary sources (both anthropogenic and biogenic) and are also formed secondarily by the oxidation of primary hydrocarbons (Yáñez-Serrano et al., 2016; Millet et al., 2010; Goldstein and Schade, 2000; Singh et al., 2004; Schade et al., 2011).

Figure 10Comparison of speciated profiles issued from canisters (traffic source) and factors 4 and 5 of PMF simulations. The species contributions are expressed in percentage volume.

Download

Aromatic compounds (benzene 40 %, ethylbenzene 29 %, C9 aromatics 24 %, and xylenes 20 % on average) are also well represented in this factor. While they enter in the composition of fossil fuel combustion by constituting the unburned fraction of vehicle exhaust emissions as for C5 alkanes (Buzcu and Fraser, 2006), their proportion in factor 4 does not compare with that of traffic emissions derived from canister measurements along the Barbaros (Bd) (Fig. 10). Solvent use activities from domestic or industrial sectors can also emit aromatics higher than C6. Therefore, the presence of aromatics in factor 4 would be related to solvent use activities. Acetonitrile, highly present in factor 4, is usually used as a biomass burning tracer (Holzinger et al., 1999).

The diurnal profile is characterized by an increase in concentration during the day and a decrease in concentrations at night (Fig. 9). When looking at external variables, this factor correlates well with SO₂ (R=0.5), which is a tracer of industrial emissions and ship emissions (Lee et al., 2011), but with neither NO_x (R=0.25) nor CO ($R=-\mathrm{0.06}$). Indeed the city of Istanbul experiences the highest industrial activities in the country (Markakis et al., 2012) while the Bosporus strait is 500 m away. The diurnal shape is also similar to that of ship traffic counts (Fig. S6 of the Supplement), which follows that of road traffic.

The potential influence of ship emissions has been investigated by looking at the ratio of V ∕ Ni derived from the elemental composition of PM_2.5. This ratio has been commonly used as a tracer of ship emission influence (Viana et al., 2014; Pey et al., 2013; Becagli et al., 2012). While some papers assume that a ratio of 3 usually signifies the impact of ship emissions (Mazzei et al., 2008; Pandolfi et al., 2011), a deeper analysis of the literature suggests that this ratio is highly variable from 0.7 up to 4.5 (Isakson et al., 2001; Agrawal et al., 2008). When plotting particulate V versus Ni concentrations integrated over a 24 h period in September 2014, a ratio of 2.72±0.89 is found (see Fig. S7 in the Supplement). However, the scatterplot of 6 h integrated data reveals a more scattered distribution of points on both sides of the fitting line. First, the derived V-to-Ni value seems to be controlled by the sampling time resolution and a fixed value cannot be used as evidence of ship emission influence. Second, while the Istanbul points lie between the upper and lower limits, some of them are higher than 4.5. Other sources like coal combustion can affect the V and Ni distribution (Oztürk et al., 2019). Finally, by comparing this factor (C5–C7 alkanes, aromatic compounds, 1-pentene, and acetone) to that of ship emissions at berth in Jingtang port (Xiao et al., 2018), no similarity was found. While an influence of ship emissions is not excluded, there is no direct evidence from VOC measurements.

The analysis of the time series shows that the background level of this factor varies as a function of the meteorological period (Fig. 8). During period 2, an increase in minimum concentrations is observed, which may be related to the lower wind speed that favors the stagnation of pollutants. We assume that secondary production affects this factor since the presence of many secondary compounds (MACR + MVK, MEK, acetaldehyde, etc.) is observed. To conclude, this factor can be related to a combination of primary and secondary anthropogenic (combustion, industrial, and ship emissions) as well as biogenic emissions. By considering the different origins of species and their diurnal emission, this factor was labeled “mixed regional emissions factor”. It represents 36.3 % of the total VOC contribution.

By increasing the number of factors, isoprene was isolated into a seven-factor solution. By increasing the number of factors to eight, isoprene was not isolated anymore (Fig. S8). By using only the PTR-MS data (10 min time resolution) for the PMF run, isoprene and its oxidation products have been isolated as well. Synchronizing the PTR-MS data with the GC–FID time step (30 min resolution) degrades the time resolution and smooths the variability of the data and, consequently, the ability of the PMF model to isolate biogenic emissions from other sources.

Factor 5: Road transport

The profile of factor 5 exhibits high contributions of pentanes (iso, n, and 2-methyl) with on average 84 % of their variability explained by this factor (Fig. 7), followed by n-heptane (58 %). A total of 32 % of the variability of 1-pentene is also explained by this factor. Aromatic compounds, such as ethylbenzene (45 %), o-xylenes, (m + p)-xylenes (47 %), C9 aromatics (37 %), benzene (28 %), and toluene (28 %), which are considered typical fossil fuel combustion products (Sigsby et al., 1987), are also predominant species in this factor. Isopentane is one of the most abundant VOCs in the traffic-related sources (Buzcu and Fraser, 2006).

To help in identifying the main sources related to this factor, a comparison between its profile and that obtained from near-source traffic measurements was performed (see Fig. 10). Contrary to factor 4 (mixed regional emissions), both profiles are similar. Factor 5 is much more enriched in aromatics compared to C5 alkanes by almost a factor of 2.

The diurnal profile of factor 5 showed an increase in concentrations from midnight until sunrise and an almost constant concentration during the rest of the day with several small peaks (Fig. 9). Morning peaks (06:30 and 09:30) and a night peak (19:30) are observed. This increase in concentrations corresponds to the morning and evening traffic rush-hour periods. After 18:30 the absolute concentration of this factor stayed high and increased for several reasons: ongoing emissions until 3 h 30 min, lower photochemical reactions, and atmospheric dynamics (the shallower boundary layer leads to more accumulation of pollutants at night). Lower concentrations are observed during late morning until 18:30. The reduction of concentration of this factor during the day could be explained by dilution processes and OH oxidation processes.

The time series shows a period of a peak and a relatively high contribution of this factor in the night of 26 to 27 September (Fig. 8). Factor 5 is the closest to a traffic-related source that covers exhaust emissions and gasoline evaporation emissions. However, the contribution of this factor does not correlate with the traffic tracer NO_x. This factor represents 15.8 % of the total VOC contribution.

As discussed in Yuan et al. (2012), the effect of photochemistry on factor composition had been analyzed by looking at the scatterplots of the contribution of the PMF factors to each VOC as a function of its OH rate constant (k_OH). Nevertheless, no clear evidence from photochemistry was found for the Istanbul PMF factor's contributions.

This study shows that PMF was more easily able to extract some factors (like biogenic terpenes) than others (like diurnal regional factors). These results are consistent with other Turkish cities where sources other than traffic (mostly industrial source) drive the VOC emissions (Yurdakul et al., 2013b; Pekey and Hande, 2011; Civan et al., 2015; Dumanoglu et al., 2014). However, in the EMB, traffic-related emissions are the most dominant source and accounted for 51 % and 74 % in winter and summer, respectively, in Beirut, Lebanon (Salameh et al., 2016). Kaltsonoudis et al. (2016) also found that traffic and biogenic emissions were the dominant source of VOC during summer in Patras and Athens. In Paris, Baudic et al. (2016) found that 25 % of the total VOC contributions were related to traffic, 15 % to biogenic factors and 20 % to solvent use compared with 14.2 % and 23 % to natural gas and background factors that the PMF has not able to dissociate. These differences in contributions with this study could be due to the differences in input data. Thus, PMF results depend strongly on input data. Furthermore, it was shown in McDonald et al. (2018) that source apportionment studies largely underestimated the influence of volatile chemical species (including organic solvents, personal care products, adhesives, etc.) as sources of urban VOCs. This underestimation could be explained by the fact that VOCs are not measured in all their diversity in source apportionment studies in contrast with what was done in McDonald et al. (2018).