During a study period of 2 years, non-methane hydrocarbons (NMHCs) and OVOCs (carbonyl compounds) were measured, routinely employing complementary offline methods. Four-hour-integrated (09:00–13:00 or 12:00–16:00 UTC) ambient air samples were collected bi-weekly (every Monday and Thursday) into steel canisters and on sorbent cartridges. The inlets were roughly 1.5 m above the roof of a container mainly housing trace gas analysers. Table 1 describes VOC measurements set up throughout the observation period and Fig. S1 specifies their collection periods.

As generally realized in the EMEP network, 24 C2–C9 NMHCs were collected into Silcosteel canisters of a volume of 6 L, conforming to the TO-14 technique, which is considered adequate for the measurement of many non-polar VOCs (US-EPA, 1997); 152 air samples were realized with a homemade device (PRECOV) for sampling air at a steady flow rate regulated to 24 mL min-1 by canisters previously placed under vacuum. NMHC analysis was performed by a gas chromatograph coupled with a flame ionization detector (GC-FID) within 3 weeks following the sampling. Separation was performed by a system of dual-capillary columns supplied with a switching device: the first column was a CP Sil5CB (50 m × 0.25 mm × 1 µm), suitable for the elution of VOCs from six to nine carbon atoms, and the other one was a plot Al2O3 / Na2SO4 (50 m × 0.32 mm × 5 µm), in order to effectively elute VOCs from two to five carbon atoms. Four main steps constituted the quality assurance/quality control programme: (i) the implementation of standard operating procedures, (ii) canister cleaning and certification (blank levels < 0.02 ppb), (iii) regular intercomparison exercises and (iv) sampling tests carried out in field conditions and concomitant with in situ measurements (Sauvage et al., 2009).

About 150 air samples were gathered using sorbent cartridges (63 air samples on multi-sorbent cartridges and 89 additional ones on 2,4-dinitrophenylhydrazine – DNPH – cartridges), by means of an automatic clean room sampling system (ACROSS, TERA Environment, Crolles, France). C1–C16 VOCs were collected via a 0.635 cm diameter 3 m long PFA line. They were then trapped into one of the two cartridge types: a multi-sorbent one consisting of carbopack C (200 mg) and carbopack B (200 mg; marketed under the name of carbotrap 202 by Perkin-Elmer, Wellesley, Massachusetts, USA) or a Sep-Pak DNPH-Silica one (proposed by Waters Corporation, Milford, Massachusetts, USA). These offline techniques are further characterized in Detournay et al. (2011), and their satisfying use in situ has already been discussed by Detournay et al. (2013) and Ait-Helal et al. (2014). Succinctly here, the sampling of 44 C5–C16 NMHCs, comprising alkanes, alkenes, aromatic compounds and six monoterpenes, as well as six C6–C11 n-aldehydes, was conducted at a flow rate fixed at 200 mL min-1 and using the multi-sorbent cartridges. These latter ones were preliminary prepared by means of a RTA oven (French abbreviation for “régénérateur d'adsorbant thermique” – manufactured by TERA Environment, Crolles, France) in order to condition them during 24 h with purified air heated to 250 ∘C and at a flow rate regulated at 10 mL min-1. Concomitantly, 15 additional C1–C8 carbonyl compounds were collected at a flow rate fixed at 1.5 L min-1 using the DNPH cartridges. During the field campaign, several ozone scrubbers were successively inserted into the sampling lines in order to limit any eventual ozonolysis of the measured VOCs: a MnO2 ozone scrubber was retained for the multi-sorbent cartridges, while KI ozone scrubber was placed upstream of the DNPH cartridges. Moreover, stainless-steel particle filters of 2 µm diameter porosity (Swagelok) were installed in order to prevent particle sampling. Then, VOC samples were transferred to the laboratory to be analysed within 6 weeks of their collection using a GC-FID (for the multi-adsorbent cartridges) or by a high-performance liquid chromatograph connected to an ultraviolet detector (HPLC-UV; for the DNPH cartridges).

The reproducibility of each analytical instrument was frequently checked by analysing a standard and examining results by plotting them on a control chart realized for each compound. The VOC detection limit was determined as 3 times the standard deviation of the blank variation. Detection limits in this study were all below 0.05 µg m-3 for the steel canisters and the DNPH cartridges and 0.01 µg m-3 for the multi-sorbent cartridges. The uncertainties for each species were evaluated respecting the ACTRIS-2 guidelines for the uncertainty evaluation (Reimann et al., 2018), considering precision, detection limit and systematic errors in the measurements. Relative uncertainties assessed in this study ranged from 7 % to 43 % for the steel canisters, from 7 % to 73 % for the multi-sorbent cartridges and from 6 % to 41 % for the DNPH cartridges. Finally, the VOC dataset was validated following the ACTRIS protocol (Reimann et al., 2018).

Among the 71 different VOCs monitored at Ersa during the observation period, 35 VOCs were finally selected in this study following the methodology described in Sect. S1 of the Supplement.

Other trace gases (CO and O3) and meteorological parameters were ancillary monitored at the Ersa site during the observation period. CO was measured from 22 November 2012 to 16 December 2013 by a cavity ring-down spectroscopy analyser (G2401; Picarro, Santa Clara, California, USA) at a time resolution of 5 min. O3 was measured from 31 May 2012 to 26 December 2013 using an ultraviolet absorption analyser (TEI 49i manufactured by Thermo Environmental Instruments Inc., Waltham, Massachusetts, USA) at a time resolution of 5 min. Meteorological parameters (temperature, pressure, relative humidity, wind speed, wind direction and total – direct and diffuse – solar radiation) were measured every minute from 8 June to 14 August 2012 and every 5 min from 15 August 2012 to 11 July 2014, with a weather station (CR1000 manufactured by Campbell Scientific Europe, Antony, France) placed at approximately 1.5 m above an adjacent container roof to the one which housed trace gas instruments. Note that ancillary trace gas and meteorological parameter results presented in this study are 4 h averages concurrent with VOC sampling periods (see Fig. S1).

In order to identify and classify air-mass origins, we analysed back-trajectories calculated by the online version of the HYSPLIT Lagrangian model (the Hybrid Single Particle Lagrangian Integrated Trajectory Model developed by the National Oceanic and Atmospheric Administration – NOAA – Air Resources Laboratory; Draxler and Hess, 1998; Stein et al., 2015) using Ersa as the receptor site (arrival altitude: 600 m a.s.l.). For each 4 h-atmospheric data point of the field campaign used for the factorial analysis, five back-trajectories of 48 h were computed using GDAS one-degree resolution meteorological data, in order to follow the same methodology as Michoud et al. (2017). The first back-trajectory of a set corresponds to the hour when the air sampling was initiated (i.e. 09:00 or 12:00 UTC – see Table 1) and the four other ones were calculated every following hour. The time step between each point along the back-trajectories was fixed at 1 h.

Then, back-trajectories were visually classified. Having several back-trajectories per sample allowed us to check whether air masses transported to the station over 4 h were of the same origin. Samples associated with air masses showing contrasted trajectories (e.g. due to a transitory state between two different origins) were classified as being of mixed origins (9 % of the air masses) and discarded from this study. Remaining air masses were then manually classified into five trajectory clusters (marine, Corsica–Sardinia, Europe, France and Spain – Fig. 2 and Table 2) depending on their pathway when they reached the Ersa station, their residence time over each potential source region and the length of their trajectories. Additionally, air masses of each cluster were subdivided depending on their distance travelled during their 48 h course in order to highlight potential more distant sources from local ones. This subdivision is also given in Table 2 in order to pinpoint differences in transport times.

PMF source contributions were coupled with back-trajectories in order to investigate potential emission regions contributing to long-distance pollution transport to the Ersa site. To achieve this, the concentration field (CF) statistical method established by Seibert et al. (1994) was chosen in the present study. The CF principle has already been presented in Debevec et al. (2017) and is therefore mentioned in Sect. S3 of the Supplement.

Seventy-two-hour back-trajectories together with meteorological parameters of interest (i.e. precipitation) were retrieved from the GDAS meteorological fields with a PC-based version of the HYSPLIT Lagrangian model (version 4 revised in April 2018), following the same methodology as that used for the 48 h back-trajectories previously presented. The arrival time of trajectories at the Ersa station corresponds to the hour when half of the sampling was carried out (i.e. 11:00 or 14:00 UTC – see Table 1). Note that longer back-trajectories were considered for CF analyses than those for air-mass origin classification (Sect. 2.4.1), in order to be in the same conditions as Michoud et al. (2017) and hence to have comparative results between the two Ersa VOC studies.

CF analyses applied to VOC source contributions were carried out by means of the ZeFir tool (version 3.70; Petit et al., 2017). Back-trajectories were shortened (i.e. the ZeFir tool considered shorter back-trajectories than 72 h) when precipitation higher than 0.1 mm was encountered along the trajectory (Bressi et al., 2014). As done by Michoud et al. (2017), back-trajectories have also been shortened when air-mass altitudes exceeded 1500 m a.s.l. in order to discard biases related to the significant dilution impacting air masses reaching the free troposphere. A better statistical significance of the CF results is commonly considered for grid cells with a high number of crossing trajectory points. As a result, some studies applied an empirical weighing function so as to limit the possible influence of elevated concentrations, which may be observed during occasional episodes (e.g. Bressi et al., 2014; Waked et al., 2014, 2018), on cells with a low number of trajectory points. We initially tried to apply this weighing function in this study. Exploratory tests revealed that CF results with the empirical weighing function only highlighted local contributions, given the number of air masses of this study. The farther a cell was from the Ersa station, the lower was its corresponding nij value (number of points of the total number of back-trajectories contained in the ijth grid cell, Sect. S3 of the Supplement) and the more the weighing function tended toward downweighting the results related to this cell. Therefore, CF results discussed in this study were realized without weighing, and these limitations should be taken into account when examining CF analyses, which are hence considered to be indicative information.

Finally, the spatial coverage of grid cells was set from (9∘ W, 32∘ N) to (27∘ E, 54∘ N), with a grid resolution of 0.3∘ × 0.3∘. Allocated contributions were smoothed following a factor (corresponding to the strength of a Gaussian filter) set to 5 to take into account the uncertainties in the back-trajectory path (Charron et al., 2000).

Concentration variations of three selected BVOCs, isoprene, α-pinene and camphene, were analysed at different timescales (monthly/interannual variations; Fig. 4). These BVOCs exhibited high concentrations from June to August consistently with temperature and solar radiation variations (see Sect. 3.1). Indeed, throughout the summer 2013 observation period, Michoud et al. (2017) and Kalogridis (2014) observed that emissions of isoprene and the sum of monoterpenes were mainly governed by temperature and solar radiation, supported both by the diurnal variations in their concentrations (Geron et al., 2000a, b; Guenther et al., 2000) and their correlations with environmental parameters. Additionally, significant concentrations of α-pinene were noticed from September to November (Fig. 4), while isoprene concentrations were close to the detection limit and temperature and solar radiation were decreasing. However, solar radiation decreased much faster than temperature during these months (Fig. 3), which could suggest that additional emissions (Laothawornkitkul et al., 2009), dependent only on temperature contrary to those prevailing in summer, have influenced α-pinene concentrations during these months. Ozone concentrations, lower in autumn (i.e. the months of October to December) than in summer (O3 concentration variations are depicted in Fig. S2 of the Supplement), also pointed out a weaker degradation of α-pinene in autumn.

Biogenic compounds showed significant interannual variations over the study period, linked to temperature and solar radiation variations. Higher June concentrations of isoprene and α-pinene were noticed in 2012 (average concentrations of 1.0 ± 1.1 and 2.6 ± 1.4 µg m-3 for isoprene and α-pinene, respectively) and 2014 (0.7 ± 0.5 and 0.2 µg m-3) than in 2013 (0.2 ± 0.2 and < 0.1 µg m-3). Higher June concentrations of camphene (and α-terpinene; not shown) were also noticed in 2014 than in 2013 (Fig. 4). These concentration levels may be related to the fact that temperature and solar radiation were more favourable to enhancing June biogenic emissions in 2012 and 2014 than in 2013 (Sect 3.1). Due to air relative humidity values observed in June (Sect. 3.1), we cannot rule out that an increase in BVOC concentrations may be linked to a transient modification of BVOC emissions induced by drought stress (Ferracci et al., 2020; Loreto and Schnitzler, 2010; Niinemets et al., 2004). Moreover, isoprene and α-pinene concentrations in July and August were higher in 2013 (average concentrations of 0.5 ± 0.3 and 1.1 ± 0.4 µg m-3 for isoprene and α-pinene, respectively) than in 2012 (0.3 ± 0.2 and 0.6 ± 0.3 µg m-3). High concentrations of camphene and α-terpinene were also noticed in August 2013 (0.2 ± 0.1 and 0.3 ± 0.3 µg m-3, respectively; Fig. 4). Solar radiations were lower in July and August 2012, temperatures were slightly lower in July 2012 and mean wind speed was slightly higher in July 2012 (Fig. 3). These meteorological conditions may have affected biogenic emissions and have favoured their dispersion and their dilution by marine air masses, owing to the position of the Ersa station (Sect. 2.1).

Note that the interpretation of interannual variations in BVOC measurements is based on a limited number of sampling days during the study period and different collection times (Table 1 and Sect. 2.2.1). It should therefore be considered cautiously, given variable day-to-day and strong diurnal BVOC concentration variations which were observed during the summer 2013 observation period (Kalogridis, 2014).

Variations of a selection of NMHCs are depicted in Fig. 5. These compounds illustrate contrasted reactivity, based on their atmospheric lifetimes (estimated according to their photochemical reaction rates with OH radicals defined in Atkinson, 1990, and Atkinson and Arey, 2003). Despite lifetimes in the atmosphere ranging from a few hours to some days, all selected NMHCs were characterized by similar seasonal variations, with an increasing winter trend followed by a decrease in spring/summer (Fig. 5 and Table 4), with the exception of n-hexane, propene and C8 aromatics (which were the most reactive species of the NMHCs selected in this study and had the lowest concentrations; Tables 3 and 4). NMHC concentrations were higher in winter than in summer, up to 5 times higher in the case of acetylene (Table 4). Note that ethane concentration levels were relatively important during summer (mean concentration of 1.0 ± 0.2 µg m-3), while other NMHCs showed concentrations below 0.4 µg m-3.

NMHCs exhibited different concentration levels during the two studied winter periods (Fig. 5). Mean winter NMHC concentrations were higher in 2013 than in 2014 by up to 0.3 µg m-3 in the cases of propane, acetylene and benzene (relative differences of 15 %, 42 % and 42 %, respectively). These compounds and ethane had the longest lifetimes among those selected in this study. However, ethane concentrations recorded at Ersa did not show any interannual variation over the study period (Fig. 5).

Variations of selected OVOCs, illustrating contrasted reactivity (following the same methodology as that applied to NMHCs; Sect. 3.4.2), are presented in Fig. 6. Formaldehyde, acetaldehyde, glyoxal, methylglyoxal and C6–C11 aldehydes have relatively short lifetimes into the atmosphere, and hence they can result from relatively close sources. On the other hand, acetone and methyl ethyl ketone (MEK) have the longest atmospheric lifetime among the OVOCs selected in this study, and hence they can also result from distant sources and/or be formed within emission-enriched air masses advected to the Ersa station.

Firstly, formaldehyde, methylglyoxal and n-hexanal have shown similar seasonal variations (Fig. 6), with high summer and spring concentrations (Table 4), suggesting an important contribution of primary/secondary biogenic sources to their concentrations. Fu et al. (2008) found that the largest global sources for methylglyoxal were isoprene and to a lesser extent acetone. Besides photochemical production, n-hexanal and formaldehyde can be notably emitted by many plant species (Guenther et al., 2000; Kesselmeier and Staudt, 1999; Wildt et al., 2003). Interannual variations in formaldehyde, methylglyoxal and n-hexanal summer concentrations confirmed their links with biogenic sources. For instance, the highest concentrations of methylglyoxal were observed in June 2012 (average concentration of 0.7 µg m-3), similarly to isoprene (Sect. 3.4.1). Monthly concentrations of n-hexanal peaked at 0.7 µg m-3 in August 2013, in agreement with monoterpenes, especially camphene and α-terpinene (Sect. 3.4.1). Formaldehyde showed high concentrations in both June 2012 and August 2013 (2.9 and 3.6 µg m-3, respectively).

Acetaldehyde and acetone have shown an increase in their concentrations more marked in summer than in winter (Fig. 6), suggesting they were probably mainly of both secondary (anthropogenic/biogenic) and primary biogenic origins. Acetaldehyde is known to be mainly produced through the chemical transformation of anthropogenic and biogenic VOCs (Rottenberger et al., 2004; Schade and Goldstein, 2001; Seco et al., 2007; Wolfe et al., 2016), particularly in clean and remote areas. Acetaldehyde can also be released by plants (Jardine et al., 2008; Rottenberger et al., 2008; Winters et al., 2009). Acetone emissions are thought to be globally of biogenic rather than anthropogenic origin (Goldstein and Schade, 2000; Schade and Goldstein, 2006). Acetone can also result from the oxidation of various VOCs (Goldstein and Schade, 2000; Jacob et al., 2002; Singh et al., 2004), and roughly half of its concentrations measured at diverse urban or rural sites have been assigned to regional background pollution by several studies (e.g. Debevec et al., 2017; de Gouw et al., 2005; Legreid et al., 2007; with regional contributions at a scale of hundreds of kilometres). Additionally, acetaldehyde and acetone concentration variations in winter (e.g. mean February concentration in 2013 was 0.5 and 2.4 µg m-3 higher than in 2014, respectively) also pinpointed primary/secondary anthropogenic origins (Sect. 3.4.2).

Glyoxal and MEK showed an increase in their concentrations in both summer and winter (Fig. 6 and Table 4), suggesting they were probably produced by several biogenic and anthropogenic sources. Glyoxal increases were in similar proportions (Fig. 6 and Table 4), while the MEK increase in winter was more marked than in summer, which may indicate that primary/secondary anthropogenic sources primarily contributed to MEK concentrations. Fu et al. (2008) found that the largest global sources for glyoxal were isoprene and to a lesser extent acetylene and direct emissions. MEK can be emitted from both vegetation and numerous anthropogenic sources and can result from the oxidation of various VOCs (Yáñez-Serrano et al., 2016, and references therein). Glyoxal and MEK exhibited different concentration levels during the two studied winter periods since their mean February concentration in 2013 was 65 %–75 % higher than in 2014, confirming their links with anthropogenic sources. These two OVOCs also showed different interannual variations in late spring and summer (Fig. 6). The highest concentrations of glyoxal were measured in June 2012 (average concentration of 0.2 µg m-3), similarly to isoprene and α-pinene (Sect. 3.4.1), when MEK concentrations were among their lowest (0.3 µg m-3).

Finally, n-nonanal did not show clear seasonal variations. Its March and April concentrations were higher in 2014 than in 2013. An increase in MEK concentrations was also noticed in March and April 2014 in lower proportion. These findings suggest contributions from different sources in winter/early spring, compared to the rest of the year. Moreover, a slight increase in n-nonanal concentrations was noted in August 2013, which is consistent with a biogenic source contribution (Matsunaga et al., 2003; Wildt et al., 2003).

Factor 1 contributions to measured concentrations of the NMHCs selected in the factorial analysis were 0.2 ± 0.4 µg m-3 on average (corresponding to 4 % of the sum of these NMHCs – Table 5) and peaked at 3.1 µg m-3 on 20 June 2012. In late spring/summer, it was one of the main factors observed (16 % on average and up to 53 %; Fig. 8). The chemical profile of factor 1 depicts an elevated contribution of isoprene, known as a chemical marker for biogenic emissions, with its variability fully related to this factor. The isoprene relative contribution to factor 1 is 70 %. The estimated tropospheric lifetime of isoprene was quite short (winter: 6 h and summer: ∼ 1 h), indicating that this compound was emitted mostly by local vegetation. Consequently, factor 1 is labelled “local biogenic source”. Note that the factor 1 contribution to selected NMHC concentrations may be slightly influenced by the two different sampling times used during the 2-year observation period (Table 5) as well as the number of VOCs and data points considered in the PMF analysis (see Sects. 3.4.1 and S2). However, Michoud et al. (2017) provided additional information on this local primary biogenic source in summer, like the contributions of additional primary BVOCs (the sum of monoterpenes) and some OVOCs (carboxylic acids, methanol and acetone), and presented its clear diurnal variations.

Factor 2 contributed 19 % to the sum of the selected NMHCs (Table 5; average contribution of 1.0 ± 0.5 µg m-3). This factor mainly consists of toluene (73 % of its variability attributed to this factor; Fig. 7), C8 aromatic compounds (93 %), ethylene (48 %) and propene (83 %), typically emitted by combustion processes and with short-to-medium lifetimes (winter: 1-4 d; summer: 5–20 h). The cumulative relative contribution of these VOCs to factor 2 is 66 %. Besides road traffic, toluene is also a good marker for solvents generated by industrial sources (Buzcu and Fraser, 2006; Leuchner et al., 2015; Zhang et al., 2014), suggesting that this profile could also be imputed to industrial sources. Additionally, a significant proportion of C5–C6 alkanes, i.e. i-pentane (32 % explained), n-pentane (37 %) and n-hexane (51 %), typically emitted by gasoline evaporation and with medium lifetimes (winter: 4–6 d; summer: 21–31 h), also contributes 19 % to this factor. Factor 2 did not show a good correlation with CO (Pearson correlation coefficient only of 0.2). As a conclusion, factor 2 is hence attributed to the grouping of several short-lived anthropogenic sources, partly related to gasoline combustion and/or evaporation and solvent use. Note that the PMF model did not reconstruct well 5 of the species comprising this factor (propene, n-pentane, n-hexane, toluene, and C8 aromatics; Sect. S2 in the Supplement), especially their concentration variations over short periods. As a result, factor 2 contributions over short periods may be underestimated. However, given the objectives of this study, the examination of factor 2 contribution variations will be limited to seasonal and interannual ones (Sect. 4). Factor 2 contribution variations over a short period were previously investigated in Michoud et al. (2017).

Regarding factor 2 contributions coupled with air-mass origins (Fig. 9), more elevated contributions were noticed at Ersa under the influence of continental air masses coming from France and Europe. The distinction of short trajectories from long ones (see Sect. 3.2) highlighted that factor 2 was potentially influenced by relatively close sources when Ersa received air masses from continental France, whereas other continental European sources were probably more distant (Fig. 9). Furthermore, CF analysis applied to factor 2 contributions (see Fig. S3 of the Supplement) confirmed that this factor was influenced by various potential emission areas, either located in Italy (the Po Valley and central Italy), France (south-eastern region) or possibly in central Europe (western Hungary, Croatia and Slovenia). Given the CF results, ship emission contribution cannot be discarded as well, as already suggested in Michoud et al. (2017). Indeed, ship emissions are predominantly composed of light alkenes, aromatic compounds and heavy alkanes (> C6 compounds; Eyring et al., 2005).

Factor 3 contributions to the selected NMHC concentrations were 1.2 ± 1.0 µg m-3 on average (22 % of their sum; Table 5). The profile of this anthropogenic factor displays an important contribution from alkanes, principally i-/n-butanes (with lifetimes of 10–11 d in winter and ∼ 2 d in summer), with more than 69 % of their variabilities explained by factor 3 (Fig. 7), along with i-/n-pentanes, n-hexane and propane (42 %–59 %; lifetimes of 4–21 d in winter and 21 h–4 d in summer). The cumulative relative contribution of these alkanes to factor 3 is up to 88 %. C3–C6 alkanes are identified in the gasoline composition and evaporation sources (storage, extraction and distribution of gasoline or liquid petroleum gas; Sauvage et al., 2009, and references therein). Additionally, propane can be viewed as a relevant profile signature of natural gas transport, storage and use (Leuchner et al., 2015). Consequently, this factor can be viewed as “evaporative sources”.

Regarding factor contributions as a function of air-mass origins (Fig. 9), more elevated contributions were noticed at the Ersa station when aged air masses originated from France and Europe and were probably transported toward Ersa by relatively distant sources. According to the CF analysis applied to factor 3 contributions (Fig. S3 of the Supplement), the Po Valley (especially Emilia-Romagna, an Italian region centre for food and automobile production), central Italy, the south-east of France and the Sardinian region seemed to be identified as the main potential emission areas for factor 3, as well as possibly more distant emission areas in central Europe such as western Hungary (i.e. the western Transdanubian region specializing in the automotive and machinery industries).

The average contribution of factor 4 to the selected NMHC concentrations is evaluated at 0.9 ± 0.7 µg m-3 (corresponding to 16 % of their sum; Table 5). Its profile displays an important contribution from acetylene (100 % explained; Fig. 7), benzene (49 %) and propane (37 %), with lifetimes of 19–26 d in winter and 4–5 d in summer. The average cumulative relative contribution of these VOCs to factor 4 is 80 %. Aromatic compounds and acetylene are generally associated with combustion sources, such as vehicle exhaust (e.g. Badol et al., 2008; Pang et al., 2014). However, factor 4 is characterized by a loading of benzene far superior to toluene's (49 % and 3 %, respectively), suggesting it is more related to a residential heating source than a traffic one (Elbir et al., 2007; Leuchner et al., 2015; Sauvage et al., 2009). The factor 4 profile, mainly composed of long-lived species together with a low contribution of shorter-lived species, may indicate that it is related to partly aged air masses which were advected towards the Ersa station. These suggestions are consistent with the fact that this factor correlated particularly well with CO (Pearson correlation coefficient of 0.8). As a result, this factor can be viewed as “long-lived combustion sources”, including residential heating.

Factor 4 showed higher contributions when the Ersa station received European air masses (see Fig. 9), particularly those with a long trajectory, which is consistent with factor 4 profile examination. The CF analysis depicted in Fig. S3 of the Supplement only pointed out western Hungary and to a lesser extent the Po Valley as the main potential emission areas for factor 4.

Contributing 39 % to the selected NMHC concentrations (average contribution of 2.2 ± 1.0 µg m-3), factor 5 corresponds to the dominant source detected at Ersa during the study period. Its profile is principally composed of ethane (with its variability fully explained by factor 5; Fig. 7) and propane (18 % explained; lifetimes of 21–93 d in winter and 4–19 d in summer). Their relative contribution to factor 5 is up to 96 % and they typically result from the transport, storage and use of natural gas. Additional anthropogenic NMHCs with shorter lifetimes are attributed to this factor, including ethylene and propene (12 %–16 % explained; lifetimes of 21 h–3 d in winter and 4–13 h in summer), despite a low contribution to factor 5 mass (∼ 3 %). Given the high abundance of long-lived species in its profile, factor 5 can be viewed as a regional contribution of diverse remote sources of the Mediterranean region, thus indicating the continental regional background (Hellén et al., 2003; Leuchner et al., 2015; Sauvage et al., 2009; regional contributions at a scale of hundreds of kilometres). As a conclusion, we associated factor 5 with the “regional background”.

Factor 5 showed slightly higher contributions when the Ersa station was under the influence of European air masses, particularly those with a long trajectory (potentially connected to distant emission areas; Fig. 9). As expected, the CF analysis applied to factor 5 contributions did not clearly pinpoint a specific potential emission area (Fig. S3 of the Supplement), apart from maybe western Hungary and to a lesser extent the Po Valley.

In this study, we decided to restrain the number of VOCs in the factorial analysis in order to have a better temporal representativeness of factor contributions (see Sect. S2). This choice has undoubtedly had some consequences for the PMF solution results of this study (in terms of factor number, chemical profiles and variations). Consequently, to support our results, we compared our five-factor PMF solution modelled with the 2-year NMHC dataset (14 variables, 152 4 h resolution observations), with the six-factor PMF solution modelled with the summer 2013 VOC dataset (42 variables, 329 1.5 h resolution observations; Michoud et al., 2017). The comparative study (see Sect. S5 in the Supplement) pointed out a good representativeness of the primary sources identified in this study. Overall, sources identified in both factorial analyses as influencing NMHC concentrations at Ersa had similar chemical compositions, regarding NMHCs in common. This finding supports the relevance of the 17 NMHCs selected in the factorial analysis of this study to identify major sources which have influenced Ersa NMHC concentrations over 2 years and suggests that they include key source tracers.

In order to support these statements, we computed a summer 2013 PMF solution only with the NMHCs selected in the factorial analysis of this study, and we compared it with the summer 2013 PMF solution modelled with a higher VOC number by Michoud et al. (2017). Results (see Sect. S6 in the Supplement) revealed a global consistency of the two summer 2013 PMF solutions in terms of factor chemical composition, regarding NMHCs in common between the two factorial analyses, and contribution variations, with the exception of those of short-lived anthropogenic sources (which may also depend on additional VOCs; Sect. S6). PMF factorial analyses computed with a limited VOC number therefore show limitations in the investigation of short-lived anthropogenic sources influencing Ersa VOC concentrations given the remote location of the Ersa station.

The two studies on Ersa VOC measurements also helped to discuss here the best experimental strategy to characterize variations in VOC concentrations measured at a remote background site surrounded by vegetation such as Ersa (Sect. 2.1). Firstly, the consideration of monoterpenes in the 2-year PMF analysis would significantly increase the contribution of primary biogenic sources (see Sect. 4.1). A better temporal resolution and coverage of the VOC measurements would have been beneficial in this study to further assess the contribution of local biogenic sources, considering BVOC day-to-day variations (as assessed with the summer 2013 dataset; Sect. 3.4.1). Additional measurements of speciated monoterpenes, sesquiterpenes and BVOC oxidation products would even help in identifying different local sources associated with various emitting Mediterranean plants (Owen et al., 2002). This statement is supported by the fact that different interannual variations in BVOC concentrations were noticed between isoprene and selected speciated monoterpenes in this study (Sect. 3.4.1). Contrarily, concentrations of non-speciated monoterpenes and isoprene measured in summer 2013 were mainly attributed to the same primary biogenic source in Michoud et al. (2017). However, concentration variations of the sum of monoterpenes may have mirrored those related to abundant biogenic species such as α-pinene and thus masked those characterizing more reactive species such as α-terpinene (Kalogridis, 2014; Debevec et al., 2018).

The limited number of VOCs included in our PMF study had the advantage of a better deconvolution by the PMF model of primary sources and a better reconstruction of measured VOCs with low concentrations (Sects. S2 and S6), without the need to minimize the weight of species with elevated background levels (unlike ethane, methanol and acetone in Michoud et al., 2017). These low concentrations thus represented a higher proportion of the total VOC concentrations that the model was supposed to explain. The longer timescale of VOC measurements (i.e. the 2-year period) even helped to deconvolve long-lived combustion sources from regional background (Sect. S5). However, PMF comparison results also raised the importance of a finer time resolution and a higher temporal coverage of VOC measurements conducted over a long period at a remote background site such as Ersa to support source apportionment results, in terms of decorrelation of anthropogenic sources. Indeed, anthropogenic sources identified in this study with bi-weekly observations have shown some significant correlations between them (Sect. S5) as a result of similar seasonal variations in concentrations of NMHCs composing them (Sects. 3.4.2 and 4). The consideration of diurnal variations in the 2-year factorial analysis would help in limiting this potential statistical bias.

The incorporation of 19 OVOCs in the source apportionment of Michoud et al. (2017) had little impact on the deconvolution of the main sources influencing NMHC concentrations observed at the Ersa station (Sects. S5 and S6). However, it can modify their relative contributions to the total VOC concentration, emphasizing those of local biogenic/anthropogenic sources and decreasing those of regional anthropogenic sources (Michoud et al., 2017). VOCs were largely dominated at Ersa by OVOCs (Sect. 3.3), especially in summer (Michoud et al., 2017). OVOC abundance observed at Ersa was further increased by the measurement of alcohols and carboxylic acids, carried out only in summer 2013 with automatic analysers (Sect. S4). The incorporation of OVOCs in the 2-year PMF analysis would have helped to better identify co-variations between these species (Sect. 3.4.3) and their proportion resulting from primary sources, but the limited number of samples realized with DNPH cartridges did not favour it (Fig. S1 and Sect. S2). Formaldehyde has shown in this study high concentration levels (Table 3) and clear seasonal variations (Sect. 3.4.3) and would have been of high interest in the summer 2013 PMF analysis.