3.1 Numerical atmospheric models

COSMO is a non-hydrostatic, fully compressible atmospheric prediction model working on the meso-β and meso-γ scales (Baldauf et al., 2011). The forecasts, inclusive of the complete set of parameters (such as the 3-D wind velocity used here) for eight time steps (from 00:00 to 21:00 UTC), are disseminated daily in two different configurations by the meteorological operative centre–air force meteorological service (COMET): a lower-resolution version (COSMO-ME, 7 km horizontal grid and 45 levels vertical grid, 72 h integration), covering central and southern Europe, and a nudged, higher-resolution version (COSMO-I2 or COSMO-IT, 2.8 km, 65 vertical levels, 2 runs d⁻¹), covering Italy (orange rectangle in Fig. 1a). Owing to the complex topography of the Aosta Valley, and the consequent need to resolve as much as possible the atmospheric circulation at small spatial scales, we used the latter version in the present work (cf. Sect. 4.5 for a discussion of possible effects of the finite model resolution in complex terrain).

FARM version 4.7 (Gariazzo et al., 2007; Silibello et al., 2008; Cesaroni et al., 2013; Calori et al., 2014) is a four-dimensional Eulerian model for simulating the transport, chemical conversion, and deposition of atmospheric pollutants with a 1 km spatial grid, 1 h temporal resolution, and 16 different vertical levels (from the surface to 9290 m). FARM can be easily interfaced to most available diagnostic or prognostic numerical weather prediction (NWP) models, as done with COSMO in the present work. Pollutant emissions from both area and point sources can be simulated by FARM, taking into account transformation of chemical species by gas-phase chemistry and dry and wet removal. Particularly interesting for our study is the aerosol module (AERO3_NEW), coupled with the gas-phase chemical model and treating primary and secondary particle dynamics and their interactions with gas-phase species, thus accounting for nucleation, condensational growth, and coagulation (Binkowski, 1999). Three particle size modes are simulated independently: the Aitken mode (diameter, D<0.1 µm), the accumulation mode (0.1 µm $<D<\mathrm{2.5}$ µm), and the coarse mode (D>2.5 µm). FARM is only run over a small domain (light-blue rectangle in Fig. 1a, b) roughly corresponding to the Aosta Valley. A regional emission inventory (“local sources”, updated to 2015) is supplied to the CTM over the same area to accurately assess the magnitude of the pollution load and its variability in time and space. Data from a national inventory and CTM (QualeAria, here referred to as “boundary conditions”, outer rectangle in Fig. 1a), taken along the border of the inner (light blue) rectangle, are also used to estimate the mass exchange from outside the borders of the FARM domain.

3.2 Trajectory statistical models

The forecasted wind velocity profile from COSMO was here used as an input parameter into the publicly available LAGRANTO Lagrangian analysis tool, version 2.0 (Sprenger and Wernli, 2015), to numerically integrate the 3-D wind fields and to determine the origin of the air masses sampled by the ALC over Aosta–Saint-Christophe. We set up the program to start an ensemble of 8 trajectories in a circle of 1 km around the observing site and at seven different altitudes from the ground to 4000 m a.s.l., for a total of 56 trajectories per run. A backward run time of 48 h was considered sufficient, on average, to cover most of the domain of the meteorological model while minimising the numerical errors.

Back-trajectories calculated with LAGRANTO were then employed in TSMs to provide a general picture of the geographical distribution of the probable aerosol sources. In this broadly used technique, the NWP domain is divided into grid cells (i and j indices), and air parcels arriving at a specific receptor site are analysed. When a cell ij is crossed by a back-trajectory l, the tracer concentration c_l measured at the arrival point of that trajectory (receptor) is considered. Finally, for each cell, a weighted average P_ij is calculated as follows to yield a map of the possible source areas (e.g. Kabashnikov et al., 2011):

where L is the total number of trajectories and τ_ij is the time spent by the trajectory l in the grid cell ij. F(c_l) is a function of the concentration at the receptor, and, depending on the chosen technique, can be the concentration itself (concentration-weighted trajectory (CWT) method, e.g. Hsu et al., 2003), the logarithm of the concentration (concentration field (CF) method, Seibert et al., 1994), or the Heaviside step function H(c_l−c_T), where c_T is a concentration threshold (potential source contribution function (PSCF) method, e.g. Ashbaugh et al., 1985), generally the 75th percentile of the concentration series. In this last case, P_ij can be statistically interpreted as the conditional probability that concentrations above the threshold c_T at the receptor site are related to the passage of the relative back-trajectory through the location ij (e.g. Squizzato and Masiol, 2015). Additional iterative methods exist to reduce trailing effects and to better identify pollution hotspots (e.g. Stohl, 1996). Moreover, the obtained field P_ij can be further weighted as a function of the number of trajectories passing through each cell, thus reducing the impact of cells with limited statistics (Zeng and Hopke, 1989).

Any series of atmospheric measurements can be used as the concentration variable c_l in Eq. (3). In this work, we achieved especially meaningful results employing the scores from the PMF decomposition (Sect. 3.3), i.e. the contributions of the identified sources to the PM₁₀ daily concentration measured at Aosta–Downtown. The same approach, coupling PMF and TSMs, was employed in other recent studies (e.g. Bressi et al., 2014; Waked et al., 2014; Zong et al., 2018). However, since only daily average information is available from the chemical speciation, the PMF output was repeated eight times per day in order to allow correspondence to the eight back-trajectories per day issued by COSMO and LAGRANTO. We present here results obtained with CF, having checked that application of CWT and PSCF methods provides similar outcomes. The COSMO-I2 domain was thus divided into 130×90 grid cells and 48 h back-trajectories were considered. Only points of the back-trajectories at altitudes lower than 2000 m a.s.l. were taken into account, this being the approximate average height of the mixing layer in the Po basin (Diémoz et al., 2019). Similarly, since the arriving air parcels should be able to impact surface PM levels, only trajectories ending close to the surface over Aosta were examined (i.e. altitudes <1500 m a.s.l., the model surface altitude in the Aosta area being about 900 m a.s.l.). To reduce statistical noise, every value P_ij of the resulting map was then multiplied by a weighting factor, w_ij, linearly varying from 0 (for N_ij≤20 end points in a cell) to 1 (for N_ij≥200), i.e. $w_{ij}=\min \mathit{\{}\mathrm{1},\frac{\max \mathit{\{}\mathrm{0},N_{ij}-\mathrm{20}\mathit{\}}}{\mathrm{180}}\mathit{\}}$. To provide an idea of the effect of this weighting procedure, TSM maps without any weighting (i.e. w_ij=1) have been included in the Supplement (Fig. S7).

3.3 Positive matrix factorisation

The positive matrix factorisation (PMF, Paatero and Tapper, 1994; Paatero, 1997) technique, as implemented in the US EPA PMF5.0 software (Norris and Duvall, 2014), was employed to study the 2017 series of aerosol chemical analyses and air-quality measurements in Aosta–Downtown. The method allowed us to identify the possible emission sources of the observed aerosol and notably to discriminate its local and non-local origin. PMF requires a long multivariate series of chemical analyses, matrix X (n×m the n rows being the samples and m columns representing the chemical species), and decomposes it into the product of two positive-definite matrices, i.e. G (n×p, matrix of factor contributions, p being the number of factors chosen for the decomposition) and F (p×m, matrix of factor profiles) plus residuals (matrix E):

A solution is found by minimizing the so-called objective function Q, i.e. the squared sum of E, weighted by the measurement uncertainties. A total of 100 runs were performed in this work for each dataset to find an optimal solution. Here we considered three different datasets X: (a) the overall dataset of anion/cation analyses (data available almost every day in 2017, n=360, “PMF dataset a”), (b) a subset of dataset (a), selecting those measurements also having coincident EC/OC analyses (n=132, “PMF dataset b”), and (c) a subset of (a), selecting those measurements also having coincident metal speciation (n=209, “PMF dataset c”). To make the decomposition of the three series comparable, an “unidentified” contribution, corresponding to the carbonaceous species only included in PMF dataset b, was added to PMF datasets a and c. This was done in the following way: we first estimated the total mass of crustal elements using the measured water-soluble calcium concentration as a proxy, with an empirical conversion factor of 10 (e.g. Waked et al., 2014, note that a more rigorous estimate for the soil component was obtained from the PMF analysis itself and confirmed this factor); we then subtracted the sum of the available chemical components (including the estimated mass from soil components) from the total measured PM₁₀ concentration, thus closing the mass balance. This difference was then added in PMF datasets a and c as the “unidentified” source.

Finally, NO, NO₂, and total PAHs measured at the same site (Aosta–Downtown) were included in the PMF to help identification of local pollution sources. The number of factors for each dataset was chosen based on physical interpretability of the resulting factors (Sect. S4) and on the Q∕Q_exp ratio. The latter is the ratio between the objective function obtained with the selected number of factors (Q, introduced before) and its expected value (Q_exp). Elevated (i.e. >2) Q∕Q_exp ratios could indicate that some samples and/or species are not well modelled and could be better explained by adding another source (Norris and Duvall, 2014). The measured PM₁₀ concentration was selected as the total variable to determine the contribution of each mode to the PM₁₀ concentration.

PMF was additionally performed on volume size distributions measured by the Fidas OPC in Aosta–Saint-Christophe (Sect. 4.3.3), the m columns representing particle sizes instead of chemical species.

4.1 Daily classification schemes based on ALC measurements or meteorology

Information on the vertical dimension obtained by the ALC was a key factor in identifying the events of aerosol pollution transport over the northwestern Alps (e.g. Diémoz et al., 2019). Therefore, a first step of our analysis was to set up a classification scheme of those advection dates based on ALC measurements. To this purpose, we used the longest ALC record available (2015–2017) and coupled it to meteorological measurements/forecasts. Since, to our knowledge, no previous ALC-based classification method is available in the scientific literature, we defined an original, daily resolved classification scheme based on ceilometer measurements to group dates according to specific ALC-observed conditions. A graphical overview with examples for each class identified is given in Fig. 2. Overall, we identified six classes (A–F) of days:

Figure 2Example of ALC images representative of each category (A–F) described in Sect. 4.1. The corresponding dates are 1 December 2015, 19 October 2016, 20 April 2016, 11 April 2015, 1 November 2017, and 27 January 2017. The dashed lines for categories C–F identify the sigmoid interpolation to the SR =3 envelope using the automated algorithm as explained in Sect. 4.2.

Download

Figure 3Frequency distribution of the aerosol layers observed in Aosta–Saint-Christophe based on the ALC classification described in Sect. 4.1. A: no layer, B: short episodes, C: layer detected in the afternoon, D: leaving layer, E: layer dissolving in the morning and a separated structure in the afternoon, and F: persistent layer.

Download

• days “A”: no aerosol layer or only low (i.e. <500 m from the ground) aerosol layers visible for the whole day. These days are intended to represent unpolluted conditions or cases only affected by local sources, since aerosol transport from remote regions generally manifests as more elevated layers, as found by Diémoz et al. (2019);

• days “B”: only brief episodes of layers developing from the surface to altitudes >500 m a.g.l. observed during the 24 h. The origin of the aerosol layers of this intermediate class is uncertain, since they could either result from weak (not completely developed) advections from outside the investigated region or from puffs of locally produced aerosol. For this reason, data in class “B” were used only as a transitional level, but not to draw definitive conclusions;

• days “C”: detection of an elevated, well-developed aerosol layer (usually in the afternoon) and persistent during the night;

• days “D”: detection of the aerosol layer from the previous day dissolving during the morning hours. No layers in the afternoon;

• days “E”: detection of a first aerosol layer from the previous day dissolving (or strongly reducing) in the morning and appearance of a separated structure in the afternoon;

• days “F”: thick aerosol layer persisting all day long.

Note that, although we also developed automatic recognition algorithms for classification purposes, we finally chose a classification based on visual inspection (for a total of 928 daily images over the 3-year period) to ensure the maximum quality of the flagging and avoid further sources of error.

Table 2Classification of weather regimes used in the study, based on wind speed ($\left|\mathit{v}\right|$), wind direction, daily duration of the condition (Δt), and other measured meteorological variables.

^a Both conditions must be met in order to discriminate such diurnal wind systems (inactive at night) from synoptic winds. ^b The directional range for the foehn cases was determined experimentally by comparison with manual classification from a trained weather observer.

Download Print Version | Download XLSX

The ALC classification described above was used to derive the seasonal frequency distribution of each aerosol advection class. Relevant results are shown in Fig. 3. Days not falling into those classes (e.g. complex-shaped layers, presence of low clouds or heavy rain hiding the aerosol layer) or including other kinds of aerosol layers (e.g. Saharan dust) were flagged as “Other” and were not considered for further analyses (this accounting for a maximum 26 % of days, in summer). The results reveal that clearly detected aerosol advections (cases C to F) occur half (50 %) of the days in summer and autumn and with a slightly lower frequency (45 %) in winter and spring. A higher percentage of clear days (A) occurs in winter and spring; persistent layers (F) are mostly found in winter, while cases E are more frequent in summer.

Figure 4Frequency distribution of circulation conditions in Aosta–Saint-Christophe based on the meteorological classification described in Table 2. Easterly winds (diurnal wind systems and channelled synoptic flow from the east) are represented as orange slices, wind calm conditions in yellow, and westerly winds (foehn and channelled synoptic flow from the west) in blue. Precipitation and unclassified days are represented as grey slices and are not used further in the study.

Download

To assess how the ALC classification described above is related to the circulation patterns, we used the surface meteorological observations in Aosta–Saint-Christophe as drivers of a second, independent classification scheme as summarised in Table 2. Seasonal frequency distributions from this meteo-based scheme are displayed in Fig. 4. These show that weather circulation types are remarkably variable during the year and help in interpreting the ALC-based results (Fig. 3). In fact, winter is mostly characterised by wind calm (53 % of the days), which sometimes favours stagnation and persistent aerosol layers (F). Plain-to-mountain diurnal winds gradually increase from winter (only 11 % of the days in that season, which reflects the highest percentage of clear days (A)) to summer (68 % of the days), when the thermally driven regional circulation represents the main mechanism contributing to transport of polluted air masses (E). Foehn winds are fundamental processes, especially in spring (22 % of the days), leading to the removal of pollutants and frequent occurrence of clear days (A) in that season. Finally, channelled synoptic flows from the east (or from the west) are also partly responsible for pushing polluted air masses towards (or away from) the Aosta Valley, although they are markedly less frequent compared to the thermal winds mechanism.

Figure 5(a) Contingency table showing occurrences of the aerosol layer seen by the ALC and wind direction. The blue boxes represent cases when the aerosol layer is not visible (day types A) and pink boxes cases when an arriving or stationary aerosol layer is revealed by the ALC (day types C, E, and F). The numbers inside the boxes refer to the frequency and the number of occurrences for each couple of wind–layer classes. (b) Occurrence of the aerosol layer seen by the ALC and residence time of 48 h back-trajectories in the Po plain.

Download

The association between the ALC- and meteo-based classification schemes was explored using contingency tables (Fig. 5a). To this aim, the ALC classification (A–F) was further simplified and reduced to two main cases: presence of an arriving or stationary aerosol layer (classes C, E, and F) and absence of any thick aerosol layer (class A). Classes B (short and uncertain episodes) and D (layer leaving during the morning) were not used for the next considerations to avoid confounding conditions. Figure 5a shows that the presence of a thick aerosol layer is strongly connected to the wind direction, as expected. In fact, this scheme reveals that the layer appearance can be easily explained using the wind direction as the only information: when air masses come from the east, the probability of detecting an aerosol layer over the Aosta Valley is 93 %, confirming the Po basin as a heavy aerosol source for the neighbouring regions, independently of the day and period of the year. Conversely, this probability reduces to only 2 % of the cases when the wind blows from the west. In the period addressed, we thus detected few exceptions to the perfect correspondence (100 % and 0 %), which can however still be explained. For cases of easterly winds and no aerosol layer found (7 %), possible reasons are

• the diurnal wind, although clearly detected by the surface network, was of too short a duration or too weak to transport the polluted air masses from the Po basin to Aosta–Saint-Christophe (at least ∼40 km must be travelled by the air masses along the main valley). In our record, this was for example the case of 14 and 18 September 2015, and 17 June 2016;

• back-trajectories were indeed channelled along the central valley; however, they did not come from the Po basin, but rather from the other side of the Alps. This was the case, for instance, of 20 August 2015, in which air masses came from the Divedro Valley (northeast of the Aosta Valley) after crossing the Simplon pass (between Switzerland and Italy).

These exceptions also help understand why, in summer, about 70 % of the days feature breeze systems (Fig. 4), while aerosol layers are detected on only about 50 % of the days (Fig. 3). Indeed, some part (7 %, Fig. 5a) of this 20 % discrepancy can be attributed to the above events, the remaining 13 % being likely associated with days with a complex aerosol structure (not classified in Fig. 2) or with days affected by low clouds and/or desert dust events. These days were therefore labelled as “Other” in Fig. 3.

There are few cases (2 %, Fig. 5a) in our record in which, conversely, an aerosol layer was associated with westerly, rather than easterly, winds. These are as follows:

• on 5 February and 15 March 2016, the prevalent wind was from the west and the days were correctly identified as “`Channelled synoptic flow from the west” and “Foehn” respectively. However, as soon as the wind turned and the valley–mountain diurnal circulation started, the aerosol layer arrived;

• on 24 March 2016, a real case of transboundary/transalpine advection occurred and an aerosol-rich air mass was transported from the polluted French valleys on the other side of the Mont Blanc chain to the Aosta Valley. Although heavy pollution episodes can occur also on the French side of the Alps (e.g. Chazette et al., 2005; Brulfert et al., 2006; Bonvalot et al., 2016; Chemel et al., 2016; Largeron and Staquet, 2016; Sabatier et al., 2018), it is very rare to clearly spot an aerosol layer transported from that region to Aosta, since in those cases air masses coming from the west must have crossed the Alps and are almost always mixed with clean air from the uppermost layers; therefore, only a dim aerosol backscatter signal is usually noticed from the ALC.

Overall, the good correspondence between the measured wind regimes and the aerosol layer detection by the ALC (Fig. 5a) suggests that the same association could be employed to forecast the arrival of an advected aerosol layer based on the predicted wind fields. As an example of the simple forecasting capabilities of this phenomenon, Fig. 5b illustrates a contingency table relating the occurrence of an aerosol layer in Aosta–Saint-Christophe to the residence time over the Po Valley of the 48 h back-trajectories arriving at the Aosta–Saint-Christophe observing site. Again, a direct, clear connection between both variables can be seen, with a 100 % correspondence for residence times of air masses over the Po basin >10 h.

Finally, it is worth mentioning that the proven high correlation between the circulation patterns (observed and/or simulated) and the presence of advected polluted layers in the Aosta area may be exploited in the future to reconstruct the impacts of aerosol transport on air quality over the Alpine region back to previous periods not covered by the ALC measurements (a 40-year long meteorological database is available in the region).

4.2 Vertical and temporal characteristics of the advected aerosol layer

The 2015–2017 ALC time series in Aosta–Saint-Christophe is summarised in Fig. 6, showing the 1 h resolved monthly average SR daily evolution. A data screening was preliminarily performed by removing cloud periods with less than 30 min measurements per hour and points <1 week of measurements per month. The plot clearly shows that the maximum aerosol backscatter is generally found in the early morning and in the late afternoon, likely due to coupling between aerosol transport and hygroscopic effects (Diémoz et al., 2019), and not, as usual for plain sites, in correspondence to the midday development of the mixing boundary layer. Also, the altitude of the layer varies with season. Incidentally, months affected by exceptional events can be sharply distinguished: the frequent Saharan dust events in June and August 2017 and the forest fire plumes from Piedmont in October 2017, the latter again affecting the Aosta site in the afternoon because of the thermally driven circulation. A similar climatology, showing the results in terms of average PM concentration profiles retrieved by the ALC, is given in Fig. S1 of the Supplement. In that case, the conversion from backscatter to PM₁₀ was obtained using the methodology described by Dionisi et al. (2018).

Figure 6Monthly averages of the SR daily evolution (hourly measurements, from 00:00 to 24:00 UTC) from the ALC in Aosta–Saint-Christophe (2015–2017). Although ALC measurements extend up to 15 km altitude, we limited the figure to 5000 m a.g.l. to better show the lowermost levels, where aerosol transport from the Po basin occurs (Diémoz et al., 2019). Points with insufficient statistics are not plotted (white areas; cf. main text).

Download

Figure 7(a) Aerosol layer top altitude for the classified days. Marker colour represents the type of the advection, while the marker shape represents the time of the day: morning (a.m.), afternoon (p.m.), or all day (applicable to cases F only). (b) Overall duration of the aerosol layer in a day (morning and afternoon durations were summed up in case E). The boxplots on the right represent synthetic information for each day type (same y scale as the relative charts on the left). Missing measurements in summer 2015 are due to the replacement of the ALC laser module.

Download

To quantitatively explore the spatial and temporal characteristics of the ALC-detected aerosol layer over the long term, we developed an automated procedure, described in detail in the Supplement (Sect. S2), to identify and fit with a sigmoid curve the space–time region of the ALC profiles affected by the aerosol advection (using SR ≥3 as a threshold value). This allowed us to objectively determine, for example, the height and the duration of these advections. The long-term statistics of these two variables is summarised in Fig. 7. The altitude of the polluted aerosol layer (Fig. 7a) displays, as expected, a clear seasonal cycle, with minimum in winter and maximum, up to 4000 m a.g.l., in summer. The overall cycle mimics the variability of the convective boundary layer (CBL) height found in the Po basin by other studies (Barnaba et al., 2010; Decesari et al., 2014; Arvani et al., 2016), with somewhat higher values that reflect the modification of the boundary layer structure in mountainous areas (De Wekker and Kossmann, 2015; Serafin et al., 2018). Notably, stronger, multi-scale thermally driven flows (e.g. the ones developing on the slopes of the valley) further redistribute the aerosol particles in the vertical direction, thus pushing the upper limit of the CBL to the ridge height. Average values of the different classes, represented as boxplots in Fig. 7, are clearly impacted by the season of their maximum frequency over the year (Fig. 3). In fact, minimum altitude of the layer top was found on days F owing to their maximum occurrence in winter, while the top altitude maximum was found on days E being these mostly summertime events. Great variability was also found for the duration of the advection (Fig. 7b), ranging from a few hours in the weakest events to 24 h per day in the extreme cases (full day, by definition, for cases F). The coupling between the duration of the events and their absolute aerosol load determines the impact of the investigated phenomenon on surface air quality, as discussed in the following paragraphs.

4.3 Impact of Po Valley advections on northwestern Alps surface-level PM loads and physico-chemical characteristics

To quantify the impact of the aerosol layers detected by the ALC on the air quality at ground level, we first investigated how the PM concentration daily evolution measured by the surface network changes depending on the ALC-identified classes (Sect. 4.3.1). We then used the positive matrix factorisation of the multivariate dataset of chemical analyses to explore the chemical markers associated with the transport from the Po basin (Sect. 4.3.2). The effectiveness of the identified markers and the coupling between chemistry and meteorology were also confirmed by the use of trajectory statistical models. Furthermore, a link between aerosol chemical and physical properties was investigated in Sect. 4.3.3, where particle chemical composition is coupled to measurements of aerosol particle size. Finally, a preliminary assessment of the effect of the investigated advections on surface pollutants other than particulate matter (e.g. NO_x partitioning) was also performed and is reported in the Supplement (Sect. S5).

4.3.1 Surface PM variability in relation to the ALC profiles' classification scheme

We started investigating the effect of aerosol transport on the Aosta Valley surface air quality by analysing the variations of the daily mean PM concentrations measured by the ARPA surface network as a function of the ALC classes introduced in Sect. 4.1. Figure 8a shows the relevant results, resolved by season. It is quite evident that PM₁₀ concentrations in Aosta–Downtown are in fact highly correlated with the presence/strength of the aerosol layers seen by the ALC. PM₁₀ on days with a persistent aerosol layer (class F) was found to be 2 to 3.5 times higher, on average, than on non-event days (class A, this difference being statistically significant for all seasons, as proven by the p value <0.05 from the Mann and Whitney, 1947, test). Similar behaviour was seen in PM_2.5 concentrations measured at the same station (Fig. S2a) and in the PM_2.5∕PM₁₀ ratio (Fig. S2b), the latter indicating that aerosol particles are smaller, on average, on days when the ALC detects a thick aerosol layer compared to non-event days. Causality between the presence of an elevated layer and variations of the PM surface concentration, however, cannot be rigorously inferred at this point, since, in principle, common environmental conditions affecting both PM at the ground and the aerosol in the layers aloft could generate the observed correlation. Nevertheless, it is interesting to note that this effect is less detectable for other pollutants measured by the network. In fact, concentrations of typical locally produced pollutants such as NO_x and PAHs (Fig. 8b and c) do not change as much as PM among the ALC classes, with only rare exceptions (e.g. class A, which is slightly influenced by foehn winds, and class F, in which temperature inversions/low mixing layer heights may enhance the effect of local surface emissions). These results indicate that the correlation (Fig. 8a) does not trivially originate from common weather conditions or from the uneven distribution of the ALC classes among the seasons.

Figure 8Daily averaged surface concentrations (2015–2017) of PM₁₀ (a), nitrogen oxides (b), and polycyclic aromatic hydrocarbons (c) in Aosta–Downtown as a function of the ALC classes and season. (d) PM₁₀ surface concentration at the Donnas station. The EU-established PM₁₀ daily limit of 50 µg m⁻³ and the NO₂ hourly limit of 200 µg m⁻³ are also drawn as references (horizontal dashed lines).

Download

Furthermore, large correlation was also found between the ALC classification only available in Aosta–Saint-Christophe and the surface PM₁₀ concentration recorded in Donnas (Fig. 8d). Since the two stations are located at 40 km distance, this result suggests that a major role is played by large-scale dynamics rather than by local sources and that the phenomena under consideration have at least a regional-scale extent. As a further test, we correlated the daily PM₁₀ concentrations measured in Donnas with those measured in Aosta. To this purpose, we considered the anomaly (ΔPM₁₀, Eq. 5) with respect to the monthly moving average (〈PM₁₀〉_m), in order to avoid large, fictitious correlations due to the same yearly cycle (maximum PM concentration in winter and minimum in summer), and only focus on the short-term dynamics:

$$\begin{array}{}\text{(5)}& \mathrm{\Delta }{\mathrm{PM}}_{\mathrm{10}}\left(t\right)={\mathrm{PM}}_{\mathrm{10}}\left(t\right)-\langle {\mathrm{PM}}_{\mathrm{10}}{\rangle }_{\mathrm{m}}\left(t\right),\end{array}$$

where t is a specific day. The scatterplot of these anomalies is shown in Fig. 9. It highlights the good correlation between the ΔPM₁₀ at the two sites (ρ=0.73 when considering all classes and ρ=0.76 when only considering advection cases, i.e. classes C to F). Clustering of these results using the ALC classification (circle colours) further shows the different ΔPM₁₀ associated with the different classes and the good correspondence of this effect at both sites. Notably, in the most severe cases (E and F), the anomalies in Donnas are generally larger than in Aosta–Downtown (the best fit line being $y=\mathrm{1.1}x+\mathrm{0.8}$ for cases E and F only), which is compatible with this site being closer to the Po basin (Fig. 1).

The advected aerosol layers were also found to impact the daily evolution of surface PM concentrations. To investigate this aspect, we used hourly and sub-hourly PM₁₀ data from both the Fidas 200s OPC in Aosta–Saint-Christophe and the TEOM in Aosta–Downtown. Figure 10 shows the PM₁₀ daily cycles for cases A (no advection), C, and D, for both Aosta–Saint-Christophe (Fig. 10a) and Aosta–Downtown (Fig. 10b). Despite the instruments' limitations described in Sect. 2, the figure shows that local effects play an important role in the PM₁₀ daily evolution, as visible from the morning and afternoon peaks. In fact, these maxima, common to most pollutants, are the results of the coupled effect of varying local emissions and boundary layer height during the day. The daily cycle of PAH concentrations is added to the plot as a proxy of the diurnal cycle from local sources (dashed line, right y axis). This cycle is similar to the PM₁₀ one for case A. Conversely, the influence of the advections (C and D) is clearly visible in the corresponding PM₁₀ cycles. In fact, Fig. 10 shows the PM₁₀ concentration to markedly increase in the afternoon of days C (by 1.0 and 0.7 µg m⁻³ h⁻¹ in Aosta–Saint-Christophe and Aosta–Downtown respectively) and to markedly decrease on days D (by −1.3 and −0.7 µg m⁻³ h⁻¹). This effect is similar at the two sites, although less marked in Aosta–Downtown, especially at night. This is probably due to TEOM loss of volatile species in the advected aerosol (see also the PMF results on chemical analysis in the following section). Similar results (not shown) were obtained when splitting the data on a seasonal basis, which excludes the observed behaviour originating from the seasonal cycle.

4.3.2 Chemical species within the advected aerosol layers

As a further step to adequately discriminate local and non-local sources, we took advantage of the 1-year long (2017) chemical aerosol characterisation dataset in the urban site of Aosta–Downtown and applied the PMF to the three datasets described in Sect. 3.3 (PMF datasets a, b, c, i.e. anion/cation only, anion/cation with EC/OC, and anion/cation with metals respectively). Results of this analysis show the main factors shaping the composition of PM₁₀ at the investigated site are those given in Fig. 11, all details on this outcome being given in Sect. S4. The question addressed here is, however, how aerosol transport from the Po Valley affects the chemical properties of PM₁₀ sampled in Aosta. In the preliminary analysis of three case studies reported in the companion paper, we found that the advected layers were rich in nitrates and sulfates (accounting, together with ammonium, for 30 %–40 % of the total PM₁₀ mass). This kind of secondary inorganic aerosol was indeed found in high concentrations in the Po basin by previous studies (e.g. Putaud et al., 2002, 2010; Carbone et al., 2010; Larsen et al., 2012; Saarikoski et al., 2012; Gilardoni et al., 2014; Curci et al., 2015). Here we further tested on a longer dataset whether the sum of the nitrate- and sulfate-rich factors (i.e. secondary aerosols) could in fact represent a good chemical marker of aerosol transport from the Po basin (e.g. Kukkonen et al., 2008; Tang et al., 2014). In this respect, it is worth highlighting that these PMF modes are not correlated with typical locally produced pollutants such as NO_x and PAHs (this is revealed by the low percentage contribution of NO_x and PAHs in nitrate-rich and sulfate-rich modes in Figs. S3–S5), which therefore excludes their local origin. We then combined the chemical information (the sum of the secondary components) with the ALC classification. This was done for the year 2017, which is the only overlapping period between anion/cation analyses and ALC profiles (see Table 1). Results are shown in Fig. 12. In particular, Fig. 12a shows the time evolution of the mass concentration carried by the secondary aerosol modes, in which each date is coloured according to the six ALC classes identified. It can be noticed that almost all peak values occur during days of type E or F, i.e. when the advected layer is more persistent and able to strongly influence the local air quality at the ground. The very good correspondence between the sum of the nitrate- and sulfate-rich mode contributions and the ALC classification, as well as the poor correlation between the latter and the role of the other PMF-identified modes (not shown), suggests that the link between the secondary components and the aerosol layer advection is not simply due to particular weather conditions affecting both variables. Rather, the coupling between the chemical and the ALC information indicates that aerosols of secondary origin dominate during the advection episodes from the Po basin.

Figure 9Comparison between daily PM₁₀ anomalies (Eq. 5) registered in Aosta–Downtown and Donnas (40 km apart), relative to a monthly moving average. Circle colour represents the ALC classification (see legend), and the “X” symbol represents the unclassified days. The 1:1 and the best fit line are also represented on the plot. Pearson's correlation index between the two series is ρ=0.73.

Download

Figure 10(a) Daily PM₁₀ cycle sampled by the OPC in Aosta–Saint-Christophe during non-event days (class A) and days with arriving and leaving aerosol layers (classes C and D), plotted together with the daily evolution of PAH as a proxy of the local sources (dashed line, right vertical axis, in ng m⁻³). (b) Same as (a) using the TEOM in Aosta–Downtown.

Download

A summary of the contribution of secondary modes to the total PM₁₀ as a function of the day type, on a statistical basis, is given in Fig. 12b. The Mann–Whitney test, applied to these data, confirms that the contribution of secondary pollutants is significantly lower for class A compared to B (p=0.0005, $\mathrm{3}\times {\mathrm{10}}^{-\mathrm{8}}$), C compared to D ($p=\mathrm{6}\times {\mathrm{10}}^{-\mathrm{8}}$), D compared to E ($p=\mathrm{9}\times {\mathrm{10}}^{-\mathrm{7}}$), and E compared to F ($p=\mathrm{6}\times {\mathrm{10}}^{-\mathrm{7}}$). Figure 12b also allows us to quantify the contribution of non-local sources using as a baseline the results obtained for class A (i.e. no advection, shaded area in Fig. 12a). Note that this baseline is very low on average, about 3 µg m⁻³, as expected from the unfavourable conditions for secondary aerosol production in the area under investigation (e.g. low expected emissions of ammonia, frequent winds) and from the unobserved correlation between secondary aerosol components and their gas-phase precursors. The non-local contribution is calculated as the average difference between the mass from the secondary modes and this baseline, and amounts to about 5 µg m⁻³ on a yearly basis, i.e. 24 % of the total PM₁₀ concentration reconstructed from the PMF. On a seasonal basis, the absolute impact of air mass transport depends on the coupling between emissions (stronger in winter and weaker in summer) and weather regimes (e.g. thermal winds occurring more frequently in summer/autumn with respect to winter/spring, Sect. 4.1). This results in a PM₁₀ contribution of nearly 6 µg m⁻³ in winter and autumn, 4 µg m⁻³ in summer, and 3 µg m⁻³ in spring. In terms of relative contribution, this also depends on the “background” PM₁₀ levels in Aosta. It is therefore highest in summer (32 %), when thermally driven fluxes are more frequent and local emissions lower, and lowest in winter (16 %), when thermal winds are less frequent and local emissions higher. Intermediate relative values are found in spring and summer (27 % and 28 % respectively). In specific episodes, advections from the Po basin may still produce an increase in PM₁₀ concentrations of up to 50 µg m⁻³, i.e. exceeding alone the EU daily threshold. The most important compounds contributing to this increase are nitrate, ammonium, and sulfate, i.e. the species characterising the nitrate- and sulfate-rich PMF modes. However, since the sulfate-rich mode additionally includes organic matter (OM; cf. Figs. S3–S5), the latter species is also relevant in the advection episodes, especially in summer, when the nitrate-rich mode has its lowest contribution. This finding agrees with previous works, identifying the Po basin as a source of organic aerosol (Putaud et al., 2002; Matta et al., 2003; Gilardoni et al., 2011; Perrone et al., 2012; Saarikoski et al., 2012; Sandrini et al., 2014; Bressi et al., 2016; Khan et al., 2016; Costabile et al., 2017).

Figure 11Percentage of aerosol mass concentration carried by each mode for the three PMF datasets considered in the analysis (PMF dataset a: anion/cation only; PMF dataset b: anion/cation together with EC/OC; PMF dataset c: anion/cation together with metals) and their respective confidence intervals from the BS-DISP test. Details are provided in Sect. S4.

Download

Figure 12(a) Temporal chart of the contribution of secondary (nitrate-rich and sulfate-rich) modes to the total PM₁₀. The shaded area represents the estimated local production of secondary aerosol. The difference between each dot and this baseline can thus be read as the non-local contribution to the aerosol concentration in Aosta–Downtown. Since ion chemical speciation started in 2017, only 1 year of overlap with the ALC is available at the moment (see Table 1). (b) Boxplot of the contribution of secondary (nitrate-rich and sulfate-rich) modes to the total PM₁₀ concentration as a function of the day type. PMF dataset “a” was used for both plots.

Download

The Po Valley origin of the secondary components was also further proved by coupling the chemical information to back-trajectories. Figure 13a shows the result of the TSM using the sum of the nitrate- and sulfate-rich modes as a concentration variable at the receptor. It clearly reveals that trajectories crossing the Po basin have a much higher impact on secondary aerosol measured in Aosta compared to air parcels coming from the northern side of the Alps. Interestingly, this is not the case using different source factors from PMF. For example, Fig. 13b reports the same map, but relative to the traffic and heating mode, which is clearly much more homogeneous compared to Fig. 13a and essentially reflects the density distribution of the trajectories. This behaviour highlights the local origin of the PMF source factors other than the two chosen as proxies of the advection (secondary aerosols). As sensitivity tests, similar maps, without any weighting applied, are reported in Fig. S7. No substantial differences can be noticed.

Figure 13(a) Output of the Concentration Field Trajectory Statistical Model, using the sum of the contributions from PMF nitrate- and sulfate-rich modes as a concentration variable at the receptor (Aosta–Downtown). (b) Concentration field based on the traffic and heating mode from PMF. Trajectories are cut at the borders of the COSMO-I2 domain.

4.3.3 PMF of aerosol size distributions and links to chemical properties

Exploring the links between aerosol chemical and physical properties is useful to get insights into the atmospheric mechanisms leading to particle formation and transformation during their transport to the Aosta Valley. Therefore, we investigated correlations between the results of the chemical analyses on samples collected in Aosta–Downtown and particle size distributions (PSDs) measured by the Fidas OPC in Aosta–Saint-Christophe. To ensure comparability between these two datasets, the particle volume distributions from the Fidas OPC (normally extending up to sizes of 18 µm) were weighted by a typical cut-off efficiency of PM₁₀ sampling head (Sect. S6). We then performed a PMF analysis of the hourly averaged PSDs from the Fidas OPC (hereafter referred to as size-PMF). Four principal modes were identified. Their relative contribution to the total particle volume is shown in Fig. 14a. These modes are centred at 0.2, 0.5, 2, and 10 µm diameter respectively and will be thus referred to as 0.2, 0.5, 2, and 10 µm size-PMF modes in the following. A similar figure, showing their dependence on season, wind speed, and direction, is also provided in Fig. S10. The number of factors (p) was chosen based on physical considerations: if p>4, then the last mode (largest sizes) splits into sub-modes, which are not relevant for the present study; conversely, if fewer components (p<4) are retained, they merge into unphysical modes (e.g. very large and very fine particles combined together).

Figure 14(a) Modes identified by the PMF analysis applied on the PSDs measured by the Fidas OPC (size-PMF). (b–d) Comparison between the PMF scores' time series obtained from analysis of chemical speciation (chem-PMF on PMF dataset “a”, showing mass, left-hand vertical axes) and size distributions (size-PMF, showing volume, right-hand vertical axes). The three panels show (b) contributions from chem-PMF nitrate-rich (black) and size-PMF 0.5 µm (green) modes; (c) contributions from the sum of the chem-PMF sulfate-rich and traffic and heating modes (black) and from the 0.2 µm size-PMF mode (blue); (d) sum of contributions from chem-PMF road salting and soil modes (black) and from 2 and 10 µm size-PMF modes (red). Correlation values (ρ) are also reported in each plot.

Download

The scores from the size-PMF were then averaged on a daily basis to be compared to the chemical PMF ones (hereafter, chem-PMF; Sect. S4). Figure 14 compares time series of selected chem-PMF (dataset a) and size-PMF results. It shows that

• a.

  the nitrate-rich mode from chem-PMF and the 0.5 µm size-PMF mode correlate very strongly (ρ=0.81, Fig. 14b);

• b.

  the sum of the sulfate-rich mode and traffic and heating mode correlates very strongly with the 0.2 µm size-PMF mode (ρ=0.82, Fig. 14c). Note that the correlation index decreases (ρ between 0.35 and 0.69) if the sulfate-rich mode and traffic and heating mode are compared separately with the 0.2 µm mode;

• c.

  a moderate, statistically significant correlation was found between the chem-PMF soil plus road salting modes and the size-PMF coarser modes (sum of 2 and 10 µm modes, Pearson's correlation index ρ=0.59, p value $<\mathrm{2.2}\times {\mathrm{10}}^{-\mathrm{16}}$, Fig. 14d). This suggests that both indicate the same source, and likely non-exhaust traffic emissions, such as abrasion from brakes, tire wear, and road (with typical sizes of 2–5 µm), and resuspension from the surface (up to >10 µm), as also found in other studies (e.g. Harrison et al., 2012). This agreement between the two independent datasets is remarkable considering that the particles were sampled at two different sites (Aosta–Saint-Christophe and Aosta–Downtown) with distinct environmental features, and that coarse particles usually travel for short ranges. It is also worth mentioning that the correlation between single modes from chem-PMF (road salting or soil) and size-PMF (2 or 10 µm) separately is weaker (ρ between 0.26 and 0.55) than the one resulting from their sum. Finally, from a preliminary analysis based on back-trajectories and desert dust forecasts (NMMB/BSC-Dust, http://ess.bsc.es/bsc-dust-daily-forecast, last access: 2 August 2019), the 2 µm component seems to be additionally connected to deposition, and possibly resuspension, of mineral Saharan dust in Aosta, an effect already observed in other urban areas in central Italy (Barnaba et al., 2017).

A further interesting feature is the clear size separation of the 0.2 and 0.5 µm accumulation modes, which likely results from different aerosol formation mechanisms in the atmosphere: condensation/coagulation from primary emissions/aging on the one hand (0.2 µm mode, also known as “condensation mode”) and aqueous-phase processes (e.g. in fog during the cold season) on the other hand (forming “droplet mode” aerosol particles of about 0.5 µm diameter, e.g. Seinfeld and Pandis, 2006; Wang et al., 2012; Costabile et al., 2017). Finally, the very good correlation between the fine particles and the nitrate- and sulfate-rich modes at the two stations is a further hint that these kinds of particles are mostly of non-local origin.

Overall, the good agreement between the size- and chem-PMF results further strengthens their independent outputs.

4.4 Impact of Po Valley advections on northwestern Alps columnar aerosol properties

ALC measurements revealed the vertical extent and thickness of the advected polluted layers from the Po Valley. We therefore also investigated whether and how much these layers impact the column-integrated aerosol properties (i.e. those sounded by ground-based Sun photometers, but also by satellites). To this purpose, the ALC day-type classification was applied to the measurements of the Aosta–Saint-Christophe Sun photometer. The average AOD at 500 nm, binned by day type and season, is shown in Fig. 15a. It can be noticed that a general increase in the AOD is found from classes A (about 0.04 on average) to F, for which AOD is more than 4 times higher (0.18). The advections are also found to affect the Ångström exponent (Eq. 2, Fig. 15b), representative of the mean particle size. Results from this column-integrated perspective confirm that the transported particles are on average smaller than the locally produced ones. In fact, the mean Ångström exponents vary from 1.0 for days A in winter and autumn up to 1.6 for days E–F, and show a general increase among the classes for every season. To confirm and fully understand this dependence of the Ångström exponents on the ALC classification, we also investigated the columnar volume PSDs retrieved from the Sun photometer almucantar scans during the Po Valley pollution transport events. This analysis (Fig. S11) confirms what we already found with the surface-level PSDs from the Fidas OPC, i.e. that the advections increase the number of particles in all sizes, notably in the sub-micron fraction. This explains the higher Ångström exponents retrieved by the Sun photometer during the advections.

Figure 15(a) Average aerosol optical depth at 500 nm from the POM Sun photometer for each day type (colour code as in Fig. 12) and season; (b) Ångström exponent calculated in the 400–870 nm band; (c) yearly averaged single scattering albedo at 500 nm. B days in spring were removed from panels (a) and (b) due to insufficient statistics (only 3 d).

Download

A further link between aerosol physical and chemical properties is explored through the variability of the single scattering albedo with day type (Fig. 15c). Yearly averaged data are shown in this case due to the limited number of data points available for this parameter (the almucantar plane must be free of clouds to accurately retrieve the SSA, Sect. 2). Figure 15c reveals that SSA at 500 nm generally increases from class A to class F (0.92 to 0.98), which agrees with the fact that secondary, and thus more scattering, aerosol is transported with the advections. This information is particularly important for follow-up radiative transfer evaluations under the investigated conditions. In this respect, also note that further analysis more focussed on the impact of these advections on particle absorption properties is planned for the future.

4.5 Comparison between observations and CTM simulations

Accurate simulations of air-quality metrics are of utmost importance for environmental protection agencies. Indeed, not only are they essential from a scientific/technical perspective (contributing to better understanding the local and remote pollution dynamics), but they also represent a fundamental duty towards the public, air-quality forecasts being disseminated every day. In this section, we compare long-term (1-year, 2017) daily averages of surface PM₁₀ to the corresponding simulations by FARM in Aosta–Downtown (similar results are found for the other sites in the domain and are not reported here) and next to Ivrea. This exercise was also aimed at evaluating whether and how the information gathered by the ALC can be used to understand deficiencies and thus improve the model performances. The 1-year time series of both the measured and simulated PM₁₀ in the Aosta–Downtown and Ivrea sites are displayed in Fig. 16. Model and measurement show similarities (such as the same seasonal cycle, indicating that local emissions from the regional inventory and general atmospheric processes are correctly reproduced), but also divergences (especially PM₁₀ underestimation by FARM, as already shown and discussed in the companion paper). Table 3 summarises some statistical indicators of the model–measurement comparison, namely mean bias error (MBE), root mean square error (RMSE), normalised mean bias error (NMBE), normalised mean standard deviation (NMSD, i.e. $({\mathit{\sigma }}_{\mathrm{M}}/{\mathit{\sigma }}_O-\mathrm{1})$, where σ_M and σ_O are the standard deviations of the model and observation), and Pearson's correlation coefficient (ρ). The overall performances of FARM are comparable to the results obtained in other studies using CTMs (e.g. Thunis et al., 2012). For the Aosta–Downtown case, we additionally explore the absolute (Fig. 17a) and relative (Fig. 17b) differences between modelled and observed PM₁₀ in relation to the ALC-derived classes. These results reveal that the model underestimation mostly occurs in the cases of strongest advections (F), with extreme model deviations as low as −40 µg m⁻³ (Fig. 17a), i.e. −50 % or lower (Fig. 17b). The observed model–measurement discrepancies might originate from (1) an incomplete representation of the inventory sources (emission component), (2) inaccurate NWP modelling of the meteorological fields, and notably the wind (transport component), or a combination of (1) and (2). Some details on these aspects are provided in the following.

Figure 16Long-term (1-year) comparison between PM₁₀ surface measurements and simulations (FARM) in Aosta–Downtown (a) and Ivrea (b).

Download

Table 3Statistics of comparison between PM₁₀ model forecasts and measurements at the site of Aosta–Downtown. Results with both W=1 and W=4 weighting factors of the PM fraction arriving from outside the boundaries of the regional domain are reported.

Download Print Version | Download XLSX

1. Emissions. Inaccuracies in the (local and national) emission inventories could degrade the comparison between simulations and observations. Based on results shown in Fig. 17a, b, we decided to investigate the sensitivity of our simulations in Aosta–Downtown to the magnitude of the external contributions (boundary conditions). In particular, we used a simplified approach to speed up the calculations: assuming that the contribution from outside the regional domain and the local emissions add up without interacting, we simulated the surface PM₁₀ concentrations turning on/off the boundary conditions (dark- and light-blue lines in Fig. 16) to roughly estimate the only contribution from sources outside the regional domain. Interestingly, the difference between the two FARM runs correlates well with the advection classes observed by the ALC (Fig. 18). This clear correlation between the simulations and the experimentally determined atmospheric conditions is a first good indication that the NWP model used as input to the CTM yields reasonable meteorological inputs to FARM. Then, to further explore the sensitivity to the boundary conditions, we gradually increased, by a weighting factor W, the PM₁₀ concentration from outside the boundaries of the domain trying to match the observed values. In Fig. 17c, d we show the results obtained with W=4. This exercise shows that the overall mean bias error is much reduced compared to the original simulations (W=1, Fig. 17a, b), especially for the winter and autumn seasons, while slight overestimations are now visible for summer and spring. Also, the annually averaged MBE, NMBE, and NMSD improve (Table 3), whilst the other statistical indicators remain stable or even slightly worsen.

   Clearly, our simplified test has major limitations. For example, seasonally and geographically dependent (i.e. relative to the position on the regional domain border) factors W should be used to better correct deficiencies in the national inventory. Also, the high weighting factors needed to match the measurements in winter and autumn suggest not only underestimated emissions, but also missing aerosol production mechanisms (e.g. aqueous-phase particle production as in fog/cloud processing, Gilardoni et al., 2016) during the cold season. In fact, this simplified exercise was only intended to roughly estimate the magnitude of the “outside PM₁₀ tuning factor” necessary to match the observations. Despite this limitation, this numerical experiment still shows quite convincingly that biases in air-quality forecasts can be reduced by revising the emission inventories on the basis of experimental data, among them those from unconventional air-quality systems (e.g. the ALCs used in this study).

2. Transport. Although the COSMO-I2 grid spacing is certainly in line with that of other state-of-the-art operational limited-area models, in our complex terrain it could be insufficient to appropriately resolve local meteorological phenomena triggered by the valley orography (Wagner et al., 2014b), also considering that the actual model resolution is 6–8 times that of the grid cell (Skamarock, 2004). For example, Schmidli et al. (2018) show that at a 2.2 km grid step the COSMO model poorly simulates valley winds, while at a 1.1 km grid step the diurnal cycle of the valley winds is well represented. Similarly, Giovannini et al. (2014) show that a 2 km step can be considered the limit for a good representation of valley winds in narrow Alpine valleys. The smoothed digital elevation model (DEM) used within COSMO and FARM could also play a direct role in the detected underestimation. In fact, as mentioned in Sect. 3.2, the model surface altitude of the Aosta urban area is 900 m a.s.l., whereas the actual altitude is about 580 m a.s.l. (Table 1). The adjacent cells are given an even higher altitude, owing to the fact that the valley floor and the neighbouring mountain slopes are not properly resolved at the current resolution. This results in an apparent lower elevation of the sites located in flatter and wider areas, such as Aosta, while the real topography profile at the bottom of the valley presents a monotonic increase from the Po plain to Aosta. Therefore, it is expected that, just by better reproducing the orography, a higher resolution would better resolve the horizontal advection of aerosol-laden air from much lower altitudes on the plain.

   Most of these issues were extensively addressed in the companion paper (Diémoz et al., 2019). In that study, COSMO-I2 was shown to be capable of reproducing the mountain–plain wind patterns observed at the surface, both on average (cf. Fig. S1 in the companion paper) and in specific case studies (Fig. S13 therein). Nevertheless, it was also found to slightly anticipate in time and overestimate the easterly diurnal winds in the first hours of the afternoon and to overestimate the nighttime drainage winds (katabatic winds, ventilating the urban atmosphere and reducing pollutant loads). These limitations were mostly attributed to the finite resolution of the model.

Figure 17Absolute (a, c) and relative (b, d) differences between simulated and observed PM₁₀ concentrations at the surface in Aosta–Downtown. The mean bias error (MBE) and the normalised mean bias error (NMBE) for each case are reported in the plot titles. (a, b) FARM simulations as currently performed by ARPA. (c, d) The PM₁₀ concentrations from outside the boundaries of the domain were multiplied by a factor W=4.

Download

Figure 18Difference between PM₁₀ surface concentrations in Aosta–Downtown simulated by FARM taking the boundary conditions into account (FARM_tot) and without taking them into account (FARM_local, only local sources), as a function of the advection class observed by the ALC.

Download

To disentangle the role of factors (1) and (2) discussed above, we evaluated the model performances in a flatter area of the domain, where simulations are expected to be less affected by the complex orography of the mountains. We therefore compared PM₁₀ simulations and measurements in Ivrea (Fig. 16b). This test is also useful for operating the CTM at the boundaries of the domain, with special focus on the emissions from the “boundary conditions”. Model underestimation in both the warm and cold seasons is evident. Incidentally, it can be noticed that concentrations simulated without “boundary conditions” are nearly zero, since the considered cell is far from the strongest local sources and most of the aerosol comes from outside the regional domain. As done for Aosta (Fig. 17a, b), Fig. 19a, b present the absolute and relative differences between the model and the observations in Ivrea. The overall NMBE is −0.73, which rather well corresponds to the underestimation factor W=4 (or NMBE $=-\mathrm{0.75}$) found for Aosta–Downtown and other sites in the valley. Since it is expected that in this flat area the NWP model is able to better resolve the circulation compared to the mountain valley, this result suggests poor CTM performances to be mostly related to underestimated emissions at the boundaries rather than to incorrect transport. In addition to this general underestimation, a large scatter between observations and simulations can be noticed in Fig. 16, the linear correlation index between simulations and observation in Ivrea being rather low (ρ=0.54). If such a large scatter, due to the erroneous boundary conditions, is already detectable at the border of the domain, it is likely that at least part of the RMSE reported in Table 3 for Aosta–Downtown can be attributed to inaccuracies in the national inventory. As already mentioned, an additional contribution to the observed RMSE could still be given by errors in modelling transport due to the coarse resolution of the NWP model, although we are not able to quantify the relative role of the two factors. To unravel this issue, high-resolution models (grid step 0.5 km, e.g. Golzio and Pelfini, 2018; Golzio et al., 2019) are being tested on the investigated area and their results will be addressed in future studies.

Figure 19Absolute (a) and relative (b) differences between simulated and observed PM₁₀ concentrations at the surface in Ivrea.

Download

Finally, within the limits of the model performances discussed above, we use FARM to extend our observational findings to a wider domain compared to the few measuring sites. In Fig. 20, we provide some summarising plots of 1-year simulations. In particular, these show the 3-D simulated fields of PM₁₀ concentrations in terms of both surface-level (a, b, and c) and vertical transects along the main valley (d, e, and f), averaged over all non-advection and advection days in 2017. In this latter case, simulations with W=1 (b, e) and W=4 (c, f) are included. Compared to the “background condition” (ALC type A), the advection cases show higher PM₁₀ concentrations and a different geographical distribution, increasing towards the entrance of the valley (Fig. 20b, e). Obviously, the absolute PM₁₀ loads are higher when the non-local contribution is multiplied by W=4 (Fig. 20c, f), which, as discussed above, is the W value providing a better matching with the PM₁₀ concentrations actually measured in Aosta and Ivrea. Vertical profiles of simulated concentrations are also evidently impacted by the advections (e.g. Fig. 20d–f). A more complete set of maps is provided in Fig. S12, in which the contribution of particulate produced inside the regional domain (local sources) and outside the domain (boundary conditions) has been partitioned. An interesting feature in Fig. 20 is that the average maps of local PM₁₀ do not change between advection or non-advection cases, while the (relative) concentration maps of aerosol coming from outside the domain show noticeable differences, thus confirming once again that the polluted air masses detected by the ALC in Aosta–Saint-Christophe are of non-local origin. In conclusion, the excellent correspondence between the model output and the experimental classification based on the ALC, i.e. the remarkable difference of the concentrations and their vertical distribution between days of advection and non-advection, highlights both the rather good performances of the CTM and the ability of the measurement dataset used to reveal the phenomenon.