Ambient aerosol samples considered in the present work come from different research and monitoring programs, conducted over the last 5 years in France (Fig. 1). Each program includes at least 1 year of field sampling, providing a total number of 5343 daily filter samples available for the sake of the present study. These sites offer diverse conditions in terms of typologies (rural, traffic, urban sites, Alpine valley environments, etc.), local climate, and vegetation types, and were selected in order to cover the complex and variable national environmental conditions. These sites are assumed to represent typical environmental conditions in France, and our observations and general tendency could therefore be extrapolated to neighboring western European countries presenting quite homogeneous environmental conditions.

The site of OPE-ANDRA is a specific monitoring observatory in a rural environment, without any village or industry within several kilometers (description available from http://www.andra.fr, last access: 12 February 2019). All other sites correspond to stations of regional air quality monitoring networks (AASQA). The availability of filter samples was variable from one site to the other one, depending on the sampling frequency (typically every third or sixth day). Filter collection was conducted within the PM10 or the PM2.5 aerosol size fractions, depending on the investigated site (Table S1 in the Supplement). Moreover, co-located and simultaneous PM10 and PM2.5 samplings were conducted at OPE-ANDRA and Revin, allowing us to investigate the distribution of primary sugar compounds between the fine and the coarse aerosol size fractions at these two sites.

Ambient aerosols were collected onto quartz fiber filters (Tissuquartz PALL QAT-UP 2500, 150 mm diameter) and preheated at 500 ∘C for 4 h minimum before use. After collection, all filter samples were wrapped in aluminum foils, sealed in zipper plastic bags, and stored at < 4 ∘C until further chemical analysis. Field blank filters were also collected, at least once a month, using the same handling procedure as for PM samples. More detailed information on the sampling periods, air sampler, number of filters, and nature of PM samples are provided in Table S1 and Fig. S1 in the Supplement.

PM samples were analyzed for various chemical species using subsampled fractions of the collection filters. In the frame of the present study, the carbonaceous matter (OC and EC) was analyzed using a thermo-optical method on a Sunset Lab analyzer (Birch and Cary, 1996) as described by Aymoz et al. (2007) using the EUSAAR2 temperature program (Cavalli et al., 2010), except for the five sites of northern France where the NIOSH870 protocol was employed (Birch and Cary, 1996). OM contents were then estimated by multiplying the organic carbon mass concentrations by a fixed factor, with OM=1.8×OC. This OM-to-OC ratio value of 1.8 was chosen based on previous studies performed in France (Favez et al., 2010; Petit et al., 2015, and reference therein) and around the world. (e.g., Aiken et al., 2008; Li et al., 2018; Ruthenburg et al., 2014; Vlachou et al., 2018), with a typical range of 1.2–2.4 for these values.

For the analysis of anhydrosugars, sugar alcohols, and primary saccharides, filter punches (typically of about 10 cm2) were first extracted into ultrapure water, then filtered using a 0.22 µm Acrodisc filter. Depending on the site, analyses were conducted either by the IGE (Institut des Géosciences de l'Environnement) or by the LSCE (Laboratoire des Sciences du Climat et de l'Environnement) (Table S1). At IGE, extraction was performed during 20 min in a vortex shaker and analyses were achieved using high-performance liquid chromatography (HPLC) with pulsed amperometric detection. A first set of equipment was used until March 2016, consisting of a Dionex DX500 equipped with three columns (Metrosep Carb 1-Guard + A Supp 15–150 + Carb 1–150); the analytical run being isocratic with 70 mM sodium hydroxide eluent, followed by a gradient cleaning step with a 120 mM NaOH eluent. This analytical technique enables us to detect anhydrous saccharides (levoglucosan, mannosan, galactosan), polyols (arabitol, sorbitol, mannitol), and glucose (Waked et al., 2014). A second set of equipment was used after March 2016, with a Thermo Fisher ICS 5000+ HPLC equipped with 4 mm diameter Metrosep Carb 2×150 mm column and 50 mm pre-column. The analytical run is isocratic with 15 % of an eluent of sodium hydroxide (200 mM) and sodium acetate (4 mM) and 85 % water at 1 mL min-1. This method allows for additional separation and quantification of erythritol, xylitol, and trehalose. At LSCE, extraction was performed during 45 min by sonication and analyses were achieved using an ion chromatography (IC) instrument (DX600, Dionex) with pulsed amperometric detection (ICS3000, Dionex). A CarboPAC MA1 column was used (4×250 mm, Dionex), the analytical run being isocratic with 480 mM sodium hydroxide eluent. This analytical technique enables us to detect anhydrous saccharides (levoglucosan, mannosan, galactosan), polyols (arabitol, mannitol), and glucose.

Field blank filters were handled as real samples for quality assurance. The present data were corrected with field blanks. The reproducibility of the analysis of primary sugar species (polyols, glucose), estimated from the analysis of extracts of 10 punches from the same filters, is generally in the range of 10 %–15 %.

Additional chemical analyses were conducted for most of the sites, allowing us to quantify up to 130 different chemical species (Calas et al., 2018). Then 30–35 chemical species were selected in order to achieve PMF analyses as discussed hereafter.

Species concentration measurements were first analyzed for normality using Shapiro–Wilk's method with the statistical program R studio interface (version 3.4.1). Since data were generally not distributed normally, we used nonparametric statistical methods. The strength of the relationship between species concentrations was investigated using the nonparametric Spearman rank correlation method. Multiple mean comparison analyses were performed with the Kruskal–Wallis test method. Statistical significance was set at p<0.05.

Positive matrix factorization for the source apportionment of the PM was previously performed at several sites of this study, as part of the SOURCES (Favez et al., 2016) and DECOMBIO (Chevrier, 2017) projects. We used the US EPA PMF 5.0 software (US EPA, 2015), following the general recommendation guidelines of the European Joint Research Centre (JRC; Belis et al., 2014). Briefly, the SOURCES program aimed at performing source apportionment at 15 sites using a harmonized methodology, i.e., using the same chemical species, uncertainties, constraints, and criteria for factor identification. The PMF conducted within the SOURCES project uses about 30 different species (Table S6), including carbonaceous fraction (OC, EC), ions (Cl-, NO3-, SO42-, NH4+K+, Mg2+, Ca2+), organic markers (polyols, i.e., sum of arabitol, mannitol, and sorbitol; levoglucosan; mannosan), and metals (Al, As, Ba, Cd, Co, Cs, Cu, Fe, La, Mn, Mo, Ni, Pb, Rb, Sb, Se, Sn, Sr, Ti, V, Zn). The PMF conducted within the DECOMBIO project, for the sites of Marnaz, Chamonix, and Passy, used aethalometer (AE 33) measurements instead of EC measurements (Chevrier, 2016). This complementary measure gives the total black carbon (BC), thus enabling the deconvolution of BC concentrations into its two main constituents: wood-burning BC (BCwb) and fossil-fuel BC (BCff) (Sandradewi et al., 2008). For graphical simplicity, BCwb and BCff were summed up and labeled as EC in the following figures. PMF modeling was performed separately for each site. Statistical significance was validated with bootstrapping higher than 80 % for each factor. Detailed methodology and results about these studies are given in their respective papers (Chevrier, 2017; Favez et al., 2016). It should be noted that glucose was not included in the final solution for any of these PMF, since it generally produced statistical instability of the solutions (this point is further discussed in Sect. 3.2).

The PMF analysis took advantage of the ME-2 algorithm to add constraints to different chemical profiles (see Tables S3 and S4 for details). Mainly soft constraints were applied in order to add some prior knowledge about the emission sources and “clean” the different profiles without forcing the model toward an explicit solution. In particular, the polyol concentrations were “pulled up maximally”, while levoglucosan and mannosan were set to zero, and EC was “pulled down maximally” in the PBOA factor. This was achieved to avoid mixing with the biomass burning factor as well as possible influences of unrealistic high contributions of EC to PBOA. Other constraints were added parsimoniously to other factors, targeting specific proxies of sources (Table S4).

As for the general results of this large PMF study, we identified some well-known sources for almost all the sites (biomass burning, road traffic, secondary inorganics, dust and sea salt). Two other less-common factors were identified for all sites: secondary biogenic aerosols (probably from marine origin), traced mainly by the presence of methanesulfonic acid (MSA), and PBOA, traced by the presence of more than 90 % of the polyols total mass in the factor. Table S5 and Fig. S4 present a more detailed description of the chemical tracers in each factor, together with their yearly average contribution for each site. Hereafter, only the PBOA chemical profile will be extensively investigated. The uncertainties in this PBOA factor are discussed below and its stability is presented in Fig. S5. Bootstrap analysis based on 100 resampling runs evidenced the very high stability of this PBOA factor since the PBOA initial constrained factor was mapped to the PBOA bootstrap factor more than 99 % of the time.

Figure 2 presents an overview of the relative mass concentration distributions of individual chemical species quantified at two sites with very different characteristics, an urban site in Grenoble and the rural site of OPE-ANDRA. Data are presented for the warmer season (e.g., during summer and fall), when concentrations were at their maximum (see Sect. 3.4). Glucose is the most abundant species measured (average 37.6±26.4 ng m-3), accounting on average for 25 % of primary sugar compounds (SC) total mass at both sites. Mannitol (37.3±24.6 ng m-3) and arabitol (32.0±22.2 ng m-3) are the second and third most abundant species, accounting for 25 % and 23 % of SC mass, respectively. Trehalose is relatively abundant in samples from these two sites (20.1±16.2 ng m-3), accounting for 14 % of SC mass, but in general its concentration is frequently below the limit of quantification for samples from other sites in France. The other identified polyols (i.e., erythritol, inositol, glycerol, sorbitol, and xylitol) present lower concentration levels (4.9±2.1 ng m-3), corresponding altogether to 13 % of SC total mass.

Such ambient mass concentration distribution patterns are similar (but with variable intensities) to those previously reported for aerosol samples collected at various locations around the world. For example, Verma et al. (2018) found that glucose, and arabitol together with mannitol, contributed to 16.7 % and 48.1 %, respectively, of total primary sugar compounds in aerosols from Chichijima island. Similarly, Yttri et al. (2007b) showed that glucose and the pair arabitol and mannitol were the main contributors of total primary monosaccharides and sugar alcohols in aerosols collected from four various background sites in Norway. In addition, Carvalho et al. (2003) reported that arabitol, mannitol, and glucose are the most dominant primary sugar compounds in aerosols from rural background and boreal forest sites in Germany and Finland, respectively.

Although various primary sugar alcohols and saccharides have been detected and quantified for most of the investigated sites, the following study focuses only on the three major and ubiquitous species, namely arabitol, mannitol, and glucose. Figure 3 presents their average relative contributions at all sites, for the warmer period, displaying very similar features at a first glance. However, discrepancies could be observed from site to site, as discussed in the following sections.

Figure 4 summarizes linear correlations obtained between arabitol and mannitol concentrations at each site during the warmer period. Medium to very high coefficients of determination could be observed (0.58≤R2≤0.93; 30≤n≤143 or 45≤n≤341 for PM2.5 and PM10 series, respectively), with slopes in a rather narrow range (between 0.59 and 1.10), and quite low intercepts (always below 9 ng m-3). Such covariations indicate that both species are most probably co-emitted, by one or several type(s) of sources, at each site during the summer–autumn period. These observations are in agreement with previous studies also showing strong covariations between arabitol and mannitol (Kang et al., 2018; Verma et al., 2018; Zhu et al., 2015). Therefore, it seemed reasonable to consider both species together so that their concentrations are summed up and labeled as “polyols” in the following sections.

Conversely, linear correlations between glucose and polyol concentrations are generally weaker (0.10<R2≤0.78), with slopes varying over a much larger range (between 0.12 and 0.94), and variable intercepts (between -5.6 and 16.4 ng m-3). This suggests that glucose concentrations might follow a different pattern compared to that of polyols, either due to different emission sources, or different chemical stability in the atmosphere. It is therefore reasonable to keep glucose as a separate chemical species in the following discussion.

It should be emphasized that the variability in the slope of the regressions between the chemical concentrations is most probably related to the emissions and atmospheric processing. Particularly in the case of mannitol and arabitol, they may be influenced by biogenic or biotope characteristics. Nevertheless, no evident relationship between the slope values and the typology or the geographical location of the sites could be observed (Fig. 4).

Figure 5 shows the average PM10 and PM2.5 concentrations of polyols and glucose at OPE-ANDRA and Revin during the summer and autumn seasons. The polyol mass concentrations ranged from 7.5±10.9 to 27.8±33.3 ng m-3 in PM2.5, and from 48.9±38.2 to 73.5±61.8 ng m-3 in PM10, in Revin and OPE-ANDRA sites, respectively. PM10-to-PM2.5 ratios were then on average of about 3 to 5. Similar size distribution patterns, with variable intensity, were observed for glucose (Fig. 5). These results indicate that polyols and glucose are mainly associated with the coarse PM fraction. This observation is in good agreement with several previous investigations where polyols (especially arabitol and mannitol), together with glucose, were prevalent in the coarse fraction (Fu et al., 2012; Fuzzi et al., 2007; Pio et al., 2008; Yttri et al., 2007b). However, Carvalho et al. (2003) reported different size distributions for polyols and glucose, with variable fine or coarse mode maxima depending upon sampling location. For instance, maximum atmospheric concentrations of mannitol were associated to fine and coarse aerosols from boreal forest (Finland) and rural background sites (Germany), respectively. The authors hypothesized that these observations are due to different assemblages of dominant fungal biota (with variable aerodynamic characteristics) at different sites. Some other previous studies showed aerodynamic diameters typically ranging from 2 to 10 µm, even though a few airborne bacterial/fungal spores could exceed that size (Bauer et al., 2008a; Elbert et al., 2007; Huffman et al., 2012; Zhang et al., 2015).

Hence, although the precise mechanisms of atmospheric emission of particulate polyols and glucose are not fully resolved, our observations are in good agreement with ambient mass concentrations of polyols and glucose being likely associated with biological particles, as already suggested elsewhere (Fu et al., 2012; Verma et al., 2018; Zhang et al., 2015). These species could enter into the atmosphere through either natural or anthropogenic resuspension of surface soils and associated bacterial/fungal spores (containing polyols and primary sugar compounds), or via a direct input resulting from microbial activities (e.g., sporulation). Another hypothesis would be the abrasion of leaves and the subsequent release of microbial organisms and plant debris (Fu et al., 2012; Medeiros et al., 2006; Simoneit et al., 2004b).

The sum of polyols (arabitol + mannitol) represents only a small fraction of the total OM. However, as proxies of PBOA, they are most probably emitted with other chemical species. Emission from biological particles is a complex topic since it may include a wide variety of compounds, both organic and inorganic (Elbert et al., 2007; Zhang et al., 2015). Moreover, it is not clear if polyols are mainly emitted directly in the atmosphere or are linked to other materials, e.g., with soil dust during resuspension processes. To investigate the relationship between polyols and other molecular tracers of emission sources, it is possible to perform simple correlation analysis with individual chemical species. This approach has the disadvantage of being a one-to-one relation and thus highly sensitive to the dynamics of all PM emission sources, not only the one we are interested in. Alternatively, another way is to use a PMF approach, also based on correlations but including much more information on the temporal variations of the different sources influencing the PM chemistry at a given receptor site.

As mentioned in Sect. 2.3, the PMF results used in this study include sites of different typologies (rural, traffic, urban sites in Alpine valley environments, etc.) for 16 different locations spread over France and part of the current dataset. At each site, the PMF studies allowed us to identify a PBOA factor, characterized by the presence of more than 90 % of the total polyol content (sum of arabitol, mannitol, and sorbitol), as presented in Table S5 and Fig. S6. Moreover, the sensitivity of this factor to random noise in the data was investigated thanks to randomly resampling the input matrix of observation. In PMF analysis, this is done via the bootstrap method (Paatero et al., 2014) in the constrained run. The PBOA factor was always mapped to itself for 13 of the sites and nearly always (97 %) for the last three ones. It means that the PBOA factor does have a very high statistical stability since it never swaps with another factor (see Fig. S5). Hence, the chemical composition of this factor may be informative to investigate the PBOA source components (Table S6), and to evaluate the importance of PBOA emissions in terms of OM mass apportionment.