Introduction

There is a significant association between increased concentrations of atmospheric particulate matter (PM) and morbidity and mortality of the exposed population (WHO 2016). The toxicity of PM depends on their mass concentration, size, composition, solubility, oxidative properties, etc. (Xing et al. 2016). Due to their size, fine PM (PM2.5, aerodynamic diameter of 2.5 μm or smaller) can penetrate and accumulate deep inside the respiratory tract (alveolar region) and therefore promote cardiovascular and respiratory diseases as well as cancer (Du et al. 2016). PM2.5 are largely comprised of combustion particles from traffic and the coal, wood and fuel oil burning, but they may also contain particles from finely pulverized road dust and soils (Laden et al. 2000). It has been reported that health effects of these particles vary with sources of emissions, which define their chemical composition, and that mobile and stationary combustion sources pose the greatest threat in that regard (Laden et al. 2000; Bell et al. 2014; Kioumourtzoglou et al. 2015). Large cities represent the most vulnerable areas due to the high density of both the exposed population and the emission sources, while a long atmospheric lifetime allows PM2.5 to impact areas up to several thousand kilometres away from the emission source as well (Anenberg et al. 2014). Therefore, the allocation of the major emission sources and definition of their chemical profiles is an important task not only on the local but also on the regional and international scale.

This study integrates multiple statistical receptor-based models to improve the accuracy of geographical allocation of PM2.5 sources. Data on elemental and ionic composition of PM2.5 daily samples were analysed to assess major emission source types and their contributions and to explore the influence of air mass movements on the regional and long-range scale. Positive matrix factorization (PMF), conditional bivariate probability function (CBPF), trajectory cluster analyses (TCA), concentration weighted trajectory (CWT) and the existing pollutant emission inventories have been combined in an approach that aims to recognize areas from which different sources affect PM2.5 concentrations and composition. The study area, Belgrade metropolitan area, is located in the continental part of Balkan Peninsula and it may be an illustrative example of the under-investigated urban settlements in South-Eastern Europe (Belis et al. 2017). This region is a part of Europe’s air pollution ‘hot spots’ where developing countries like Serbia, Bosnia and Herzegovina and Northern Macedonia face serious air quality issues. Although in a process of the EU accession, those countries have not completed the transposition of EU legislation, including those related to the air quality management. The most of regulations in Serbia are in compliance with EU Directives (Belis et al. 2017) while certain difficulties in the implementation process still exist.

Reports on allocation of PM2.5 sources are also scarce in this part of Europe. The increase in number of air quality studies may be observed for Greece in recent years (Grigoratos et al. 2014; Tolis et al. 2015; Manousakas et al. 2015, 2017; Aleksandropoulou and Lazaridis 2017; Diapouli et al. 2017; Argyropoulos et al. 2017; Saraga et al. 2019; Taghvaee et al. 2019), while for the rest of the countries these studies are fairly rare (Karagulian et al. 2015; Perrone et al. 2018). Studies on the atmospheric particulate matter in Belgrade have started in the early 2000s, based on the short seasonal sampling campaigns and composition analysis by various techniques and procedures. Later, multi-component analyses and receptor modelling had been applied for source apportionment revealing several common factors for examined sites: traffic, oil combustion, coal/wood combustion and industry (Bartonova and Jovasevic-Stojanovic 2012). PM2.5 measurements in urban areas have been set by national air quality monitoring program, but due to frequent technical issues, there is a lack of data. Several research studies have been published about fine PM fraction in Belgrade (Tasic et al. 2008; Rajšić et al. 2008; Joksic et al. 2009; Đuričić-Milanković et al. 2018a) but this one is the first to encompass a year-round sampling and gravimetric measurements of PM2.5 on a suburban location with 3/7 frequency. Besides, non-destructive PIXE nuclear analytical technique was applied for simultaneous determination of about 20 trace elements as well as cation and anion species by ion chromatography. The suburban sampling site had been selected as suitable for observation of both wider regional and long-range transport (Đuričić-Milanković et al. 2018b). This was done with an objective to make the study a basis for further action towards mitigation of the most harmful sources of pollution and their health impacts.

Materials and methods

Study area

Belgrade is the capital of Serbia, situated in the continental part of Balkan Peninsula. Its population is about 1.6 million inhabitants in an urban area of about 360 km2. In this study, the sampling site (Zeleno Brdo) was selected to be in a suburban residential area of Belgrade, approximately 5 km to the E-SE direction from the city centre. At the same location was settled one of the air quality and meteorological stations classified ‘suburban background’ in the national monitoring network, operated by the Serbian Environmental Protection Agency. The Zeleno Brdo sampling site and its surroundings, as well as its placement in regard to the city centre, are presented in Figures1S and 2S, Online Resource.

Belgrade is a most industrially developed part of the country and the most of energy sector is located in the vicinity of the city. A steel manufacturing complex Smederevo is placed about 40 km SE from the Belgrade residential area. Two oil refineries are located about 7 km N and 12 km NE from the sampling site as well, the second one being a part of the Pancevo chemical industrial complex. The heating system of Belgrade consists of 14 generation plants and 45 boiler units with a total capacity of 2917 MW, which mainly use natural gas or heavy oil as a fuel (AQP 2016). Individual heating units rely mainly on the biomass and lignite coal. Five thermal power plants (TPPs) using low-caloric lignite are located about 30 and 50 km SW, 40 km SSW, 50 km SEE and 80 km SE from the sampling site. Nameplate capacities of these TPPs are 1745 MW, 1240 MW, 271 MW, 921 MW and 125 MW, respectively, while their age varies between 30 and 63 years. Together, they produce about 70% of the countries’ electric energy and represent the main sources of SO2 in the region. Regarding city transport, the total number of registered vehicles in Belgrade is about 600,000. Vehicle fleet has a high share of old cars, where about 40% personal vehicles are older than 16 years and another 38% are between 15 and 10 years old (Domazet and Stosic 2017).

Belgrade metropolitan area has a moderately continental climate with four seasons. During the sampling period, average temperatures in November and December 2014 and January and February 2015 were higher than the respective means in the 1980–2010 time interval (Figure 3S, online source). The mean wind speed was 2.4 m s−1, with 0.8% calm hours, and the prevalent wind directions were SEE and SSE (Figure 3Sa, online source). Precipitation in the area was unusually high in May, July and September 2014 and in March 2015 (Figure 4S, online source) when monthly amounts of precipitation were 280.4 mm, 250.6 mm and 126.0 mm, respectively. The heaviest rain fell in May 2014 when daily amounts of rainfall were the highest in the last 120 years (RHMZ 2014).

Sampling and analytical procedures

Sampling of PM2.5 aerosol fraction was performed from May 2014 until May 2015 at suburban monitoring station Zeleno Brdo (Figure 1S and 2S, online resource) in Belgrade, Serbia, using MVS6 Sven Leckel reference sampler, with 2.3 m3 h−1 airflow rate. For sampling were used PTFE membrane filters, exposed for 24 h (starting 00.01 h each day) on average three times per week (frequency 3/7). Mass concentrations of PM2.5 deposits were determined by the standard gravimetric procedure according to SRPS EN 12341:2015. Multi-elemental analysis of PM2.5 aerosol was performed by a non-destructive nuclear technique, particle-induced X-ray emission (PIXE), in the Laboratory of Ion Beam Applications, Institute of Nuclear Research, Hungarian Academy of Sciences (setup described in Borbély-Kiss et al. 1985). The procedure involved the 1.8-MeV energy H+ ion beam directed to the certain area of aerosol deposits and detection of emitted characteristic X rays. PTFE filters were used as carriers in order to treat samples as ‘thin’ targets and optimize the ratio between H+ path through the target and its ability for multiple interactions (cross-sections) in the sample. Elemental concentrations of Al, Si, P, S, Cl, K, Ca, Ti, V, Cr, Mn, Fe, Ni, Cu, Zn, Br, As, Ba and Pb were determined in nanograms per cubic metre based on the airflow data. For quality assurance, NIST SRM2783 thin aerosol reference material was analysed using the same setup. Concentrations of major ions (NH4+, NO3−, SO42−) were determined by the ion chromatography method following SOP MDL 064 (CalEPA 2002). Details about each of the analytical procedures are given in the Online Resource.

For data analysis, concentrations lower than LOD were replaced by LOD/2. Mann-Whitney U tests (p < 0.05) were employed for comparisons of average seasonal mass concentrations, i.e. of heating (HS) and non-heating season (NHS) means.

Source apportionment by PMF

Source categories (in further text sources) and their contributions to PM2.5 were determined by the PMF analysis technique that has been described in detail elsewhere (Paatero and Tapper 1994; Belis et al. 2019a). The elemental and ionic composition of PM2.5 daily samples were used as input data. Sample concentrations below LOD were replaced by LOD/2 and the matching uncertainties by 5/6 × LOD/2 (Polissar et al. 1998). Missing data were replaced by the geometric mean of the sampling period and the uncertainties by four times the corresponding measured concentration (Norris et al. 2015).

The US EPA PMF (version 5.0) software that implements the ME-2 algorithm developed by Paatero (1999) was used for modelling. It includes bootstrap (BS) and displacement (DISP) error estimation methods that describe the uncertainty of PMF analyses due to both random errors and rotational ambiguity (Norris et al. 2015). To support the assignment of source categories, quantitative comparison of factors’ chemical profiles with the reference source profiles was performed by DeltaSA tool (available at http:// source-apportionment.jrc.ec.europa.eu/). DeltaSA tool tests similarity by calculating two coefficients: Pearson distance (PD) and standardized identity distance (SID). If estimated PD is less than 0.4 and/or SID is less than 1, the solution is considered acceptable (Simon et al. 2010; Pernigotti and Belis 2018).

Statistical analysis

Examining the history of incoming air masses may point to the potential source areas that have a short- or long-range impact on the composition of PM2.5 at the receptor site.

The short-range ground-wind circulation influence on sources’ contributions was investigated using the CBPF method. The CBPF combines ordinary conditional probability function with the wind speed and allocates probabilities of high contribution (above 70th percentile as specified by authors) to the cells defined by intervals of both wind speed and wind direction (Carslaw and Beevers 2012; Carslaw and Ropkins 2012). Plots computed this way provide more information on the nature and the remoteness of the sources because different source categories may have different wind speed dependences (Uria-tellaetxe and Carslaw 2014).

Exploration of transboundary source areas was done through a simultaneous analysis of non-hierarchical k-means TCA and CWT maps based on 96-h air mass back trajectories. The arrival height was set to 100 m above the ground level. Details about these trajectory statistical methods (TSA) and the reasons for using each of them are given in the supplementary material. As a part of TCA, i.e. to check the influence of each cluster on the PM2.5 composition at the receptor site, average contributions of the US EPA PMF resolved sources with bootstrap evaluated 95% confidence intervals were calculated for each cluster. As a form of sensitivity analysis, trajectories were coupled with planetary boundary layer height (PBLH) at the receptor site and those arriving when PBLH was less than 100 m were excluded. The TSA methods were repeated with this new database in order to see if there are differences with and without application of such criteria (Lee et al. 2019).

Results and discussion

PM2.5 mass and elemental composition

Results of the gravimetric, elemental and ion analysis, obtained for the observed suburban monitoring site, are presented in Table 1 and compared with available data from other cities. The average mass concentration of PM2.5 was above the annual air quality guideline (AQG) value set by WHO (10 μg m−3) but below the annual AQG proposed by European Commission (25 μg m−3) (WHO 2006). WHO AQG for daily levels (25 μg m−3) was exceeded during 33% of the sampling days, almost exclusively during the HS. In comparison with previous studies of particulate matter in Belgrade, the PM2.5 mass concentrations were lower (Rajšić et al. 2008; Joksic et al. 2009) and in the same concentration range as the eleven cities presented in Table 1. Average levels of most of the trace elements in PM2.5 were significantly lower than those in Beijing, while the S, K and Pb concentrations were among the highest within the cities in the Balkans and Europe (Yu et al. 2013; Cuccia et al. 2013; Wang et al. 2013; Amato et al. 2015; Samek et al. 2017; Perrone et al. 2018).

Table 1 Mass concentrations of PM2.5 (μg m−3) and its inorganic components, elements (ng m−3) and ions (μg m−3), in the suburban area of Belgrade in period May 2014–May 2015 compared to other cities
Full size table

Long-term decreases on some anthropogenic components—S, Ni, V, Ba, Pb, Zn and Cr—in Belgrade urban area could be attributed to both regional and local changes in pollutant emissions. Significant progress was made on the local scale since 2007, while at the same time large reductions were seen in the EU countries as well (Guerreiro et al. 2018). Since 2010, the structure of the vehicle fleet in the Belgrade public city transport has changed and, on the busiest traffic sections, NOx emissions have been reduced by about 35% and PM10 by about 65% on average. Levels of Pb and S in liquid fuels were limited except for sulphur content in heavy oil used in heating plants, that is still significantly high (3%). Many boiler rooms around the city have been closed, buildings connected to the district heating network, while, at the same time, some of the heating plants have been renovated (AQP 2016). This has significantly reduced the consumption of liquid and solid fuels with high percentage of sulphur. Note though that the position of the sampler in the previous study was characterized as urban background with dominant winds coming from the city centre.

Significant seasonal differences may be noticed in the elemental composition of PM2.5 (Figure 5S, online source) showing higher concentration values during the HS. This can be attributed to higher emissions from heating plants and individual units in the city and region. Emissions from power plants and traffic were also higher throughout the HS (EPS 2017).

Seasonal changes in the PBLH are also important drivers of pollutants’ concentration variations. Low PBLH means reduced vertical diffusion and vice versa. It often occurs together with the low wind speed and high relative humidity, which lead to high PM2.5 concentrations. In figure 4S, it is evident that PBLH was the lowest during the winter season (Dec, Jan and Feb) when most of the pollution episodes occurred (PM2.5 > 25 μg m−3). Except for the elements of crustal origin (Fe, Mn, Ca, Si and Ti), Pearson correlations between the elements and ions and mean daily PBLH were all lower than − 0.40. It was reported that disadvantaged dispersion conditions go together with enhanced secondary PM2.5 formation chemistry (Lv et al. 2020). Pollutants’ vertical dilution also could have elevated their ground-level concentrations.

Major ions and neutralization ratio

Three major ions accounted for 29.6% of PM2.5 mass concentration: SO42− for 16.2%, NO3− for 8.1% and NH4+ for 5.4%. The share of three major ions’ concentrations in PM2.5 mass obtained in this study was in the same range as the share reported for urban and suburban background sites in other European cities (20.6–36.4%, Table 1). Concentrations of SO42− were among the highest in the European region, but still considerably lower than those in the Beijing or in the Belgrade area about 7 years ago (Table 1).

Concentrations of SO42−, NO3− and NH4+ showed seasonal variations (Figure 6S, online source) which may be attributed to the higher release of NOx and SOx during the HS. In the HS, temperature inversions support the accumulation of gases under meteorological conditions, like low temperature and high humidity, that favour the ion formation (Rogula-Kozłowska et al. 2013). The seasonal difference of NO3− concentrations is also related to thermal instability of semi-volatile NH4NO3 (Seinfeld and Pandis 2016) which, during the summer months, may also result in losses of NO3− captured on the filter (Schwarz et al. 2016).

Neutralization ratio (NR) was calculated to describe aerosol acidity. NR was calculated as [NH4+]/([SO42−] + [NO3−])—concentrations are in ionic equivalents—and it illustrates the degree of neutralization of SO42− and NO3− by ammonia (Bencs et al. 2008; Engelhart et al. 2011). Average NR during the sampling period was 0.8 ± 0.1 (mean ± 95% CI) suggesting that PM2.5 samples were mostly acidic. NR values during the HS and NHS were 0.6 ± 0.1 and 1.1 ± 0.2, presumably due to the high concentrations of both SO42− and NO3− during the HS. Aerosol acidity cannot be excluded even during the NHS because NRs were lower than 1 for most of the days with a few exceptional samples. These samples showed unusually low levels of SO42− that corresponded to the extreme precipitation events and consequential floods in 2014 as discussed in the “PMF source apportionment” section (EC RNA team 2014).

PMF source apportionment

The EPA PMF 5.0 model was applied on databases with daily mass concentrations of PM2.5, analysed elements and ions, and their respective uncertainties. The maximum number of selected physically meaningful factors was 5. Error diagnostics, namely low Q/Qexp values, a high number of correctly mapped BS factors and no swaps in DISP, indicated that the solution was stable (Table S3, online source). Factors representing primary PM2.5 sources passed the DeltaSA similarity test for at least one of the two distances (Pernigotti and Belis 2018). The uncertainties of factors’ contributions to the PM2.5 were calculated as a standard deviation over 100 model runs (Norris et al. 2015). Estimated uncertainties were lower than 10% for all factors except the TRAFF where it was 11%.

Five dominant source groups defined by PMF analysis were as follows: biomass burning (BB), traffic (TRAFF), regional combustion/secondary sulphates (RC/SS), local combustion/secondary nitrates (LC/NO3) and soil (SOIL), as shown in Fig. 1. Their average contributions to PM2.5 mass concentrations (in %) are presented in Table 2.

Fig. 1
figure 1

Chemical profiles and contributions (ng m−3) of PMF EPA resolved sources

Full size image
Table 2 Average annual PMF source contributions to PM2.5 (in %)
Full size table

High loads of K and Cl and a significant presence of S and Zn characterized biomass burning. Potassium is a tracer of BB process and it is found in the form of KCl in the fresh smoke (Manousakas et al. 2017). Burning of plastic waste together with wood can form Cl-rich particles as well (Kostenidou et al. 2013). In the aged smoke, K can be found as K2SO4, which could explain a medium load of S (Viana et al. 2013), while presence of Zn could be attributed to the burning of painted and treated wood, as well as different kinds of combustible waste (Pacyna and Ottar 1989; Tolaymat et al. 2000). BB had higher contributions during the HS (Fig. 2) as expected since about 40% of the households in the studied area use wood for heating (SO/RS 2013). Mean contribution of BB was 3 μg m−3.

Fig. 2
figure 2

Seasonal variations of PMF source contributions (μg/m3)

Full size image

The second factor was characterized by high loads of Fe and Pb and the medium presence of Mn, Cu and Zn. Fe and Zn (from additives in lubricating oil) are abundant in gasoline engine emissions (Morawska and Zhang 2002). Ntziachristos et al. (2007) reported that urban traffic contributed to enriched levels of ultrafine particles in Cu, Fe and Zn. Pb is reported to originate from gasoline exhaust (even in trace amounts) and wheel weights. Elements such as Fe, Mn and Zn may be present in brake linings and tires as well (Thorpe and Harrison 2008). Zinc, Cu and Pb are also associated with coal combustion, incineration and high-temperature metal processing (Minguillon et al. 2007; Dai et al. 2015). Therefore, impact of the steel plant emissions and various small metal-processing workshops distributed around the city could not be entirely excluded. Since the sampling was done on the outskirts of a large urban area, this factor was attributed to traffic emissions (TRAFF). The contributions were higher during the HS probably because of lower inversion layer which limits the dilution of vehicular emissions during the winter season (Manousakas et al. 2017), while the elevation of the sampling site probably made resuspension processes less influential. This assumption is supported by hourly variations of PM2.5 concentrations recorded at the “Zeleno Brdo” automatic monitoring station (Figure 7S, Online resource). The impact of TRAFF was 0.81 μg m−3. It should be noted that traffic emissions were most probably incorporated in the nitrate-containing factor as well, but without the specific tracers such as OC/EC unambiguous identification and quantification of this source is very difficult.

A high abundance of S and NH4+ determined secondary sulphate from regional combustion sources (RC/SS). Secondary sulphates are associated with aged air masses due to the slow conversion of SO2 to SO42− (Seinfeld and Pandis 2016). However, high loads of Ti and the medium loads of Cu and Pb indicated the contribution of coal combustion emissions as well (Rogula-kozłowska et al. 2016; Kursun Unver and Terzi 2018). Emissions of inorganic precursors (SOx and NOx) and particulate matter from TPPs and industrial zones in the region affect PM2.5 concentrations in the Belgrade area year-round (Popovic and Djordjevic 2008; Bartonova and Jovasevic-Stojanovic 2012; Đorđević et al. 2012; Nikezić et al. 2017). Higher contributions in the NHS were probably caused by fast photochemistry processes (Fig. 2) as it was previously reported in Europe (Belis et al. 2013). It should be mentioned that this is not in contradiction with higher concentrations of S, SO4+ and NH4+ during the HS (Section 3.2) since 30% of S is included in the BB factor and 16% of NH4 in the LC/NO3 factor.

Decreasing trend starting from May and ending in October 2014 (Fig. 1), with particularly low contributions from the end of July, was most likely caused by unusually high precipitation and consequential floods in May, July and September 2014. Rain represents one of the main sinks of SO42− in the atmosphere (Seinfeld and Pandis 2016). The local lignite mines were flooded and power plants, the main sources of SO2 in the region, were running with a limited capacity until December 2014 (RNA team 2014). Concentrations’ rise at the beginning of October was most probably connected to the start of the HS and the increased use of coal and liquid fuels in the individual households (SO/RS 2013).

The fourth factor, characterized by Ni, V, NO3− and NH4+, was attributed to the mix of local combustion sources and ammonium nitrate (LC/NO3). Ni and V are typical tracers of fuel oil burning (Barwise 1990). Heavy fuel oil is used in the ignition phase and as an auxiliary fuel in district heating plants, boilers and industrial furnaces in the area (MME/RS 2016). Nitrates in urban areas come mainly from NOx from traffic, but a substantial amount can come from domestic heating as well. In the case of Belgrade, heating plants can be one more important source. This factor showed both seasonal and episodic behaviour (Fig. 1) with high contributions at the beginning of HS and during the coldest intervals when energy demand was higher than usual. Pearson correlation with heating degree days was 0.7.

A larger availability of NOx in HS combined with a rapid reaction between gaseous-phase NH3 and HNO3 under the low-temperature and high-humidity conditions favours the formation of NH4NO3. Therefore, similar temperature dependence of heavy fuel oil usage and NH4NO3 formation could have prevented the US EPA PMF model from separating them as factors. Fugitive emissions from a nearby industrial complex that includes a refinery, and fertilizer and nitrate production facilities, were also potential contributors (Ragothaman and Anderson 2017).

Soil (SOIL) was characterized by crustal elements—high loads of Si and Ca, and medium loads of Ti, Fe and Mn. Crustal elements originate from various places such as soil, road dust and fieldwork dust resuspension, as well as industrial and building sources (Belis et al. 2013; Amato et al. 2014). The contribution of Saharan dust also may have been considerable (Basart et al. 2012). Less pronounced seasonal difference than expected may have been the consequence of mentioned precipitation events that kept the ground moist during the 3 months of the NHS.

It is visible in Fig. 1 that four out of five factors were extremely low (down to zero) from the end of July until the beginning of October 2014. In addition to the generally low PM2.5 levels in this location during the summer (Figure7S, Online resource), extreme precipitation events in May 2014 followed by very to extremely rainy periods in July and September 2014, and consequential floods that heavily affected the energy, mining and manufacturing industries in the area, could possibly explain this behaviour. Agricultural activities and manufacturing industry were also minimized in these areas. Details about peculiarities of the sampling campaign are given in the Online Resource.

Table 2 shows that the largest contributors to PM2.5 were LC/NO3 and RC/SS, while the relative BB contribution was much lower. Compared to the studies in the table, the proportion of nitrates is similar to that in Milan and Krakow, while it is significantly higher than values in southern European cities. This could be partly explained with the frequent temperature inversions during the winter season. The share of RC/SS is similar to the shares recorded in Athens, Florence and Barcelona, while it is higher than the median for Europe. Relatively high share of RC/SS indicated a high influence from coal combustion or other S-bearing fuels (Amato et al. 2015). The share of traffic is lower than in most of the presented studies, most likely due to the reasons already discussed above, although the shares of non-exhaust emissions in southern European cities were similar. Unaccounted mass in this study was somewhat higher due to the lack of OC and EC measurements, especially since the emission sources in Belgrade area are generally combustion heavy. This assumption is supported by higher UM contributions during the HS when the combustion of different fuels takes place. Note that the absolute contributions (in μg m−3) might differ significantly among presented studies because of different PM2.5 concentrations.

Influence of local atmospheric circulation

The variations of PM2.5 mass concentrations and the related sources’ contributions may be understood better if they are associated with different air circulation patterns. In that regard, average contributions, and the respective bootstrap evaluated 95% CIs of the mean, were calculated for each 100 m PBL height interval. The influence of ground-wind circulation was examined by CBPF analysis (conc. > 70th perc.) and the plots were drawn to tell apart potential emission sources in the vicinity of the sampling site from those under the influence of long-range transport (Perrone et al. 2018).

The highest concentrations of PM2.5, BB and LC/NO3 were mainly connected to low PBLH, which favours accumulation processes (Figure 8S, online resource) (Seinfeld and Pandis 2016). This was less pronounced with TRAFF, and almost not at all with SOIL and RC/SS. The highest contributions of these three sources were associated with high PBLH suggesting the intrusion of polluted air masses from remote areas.

Figure 3 shows that the highest probabilities for BB occurred with SE and SW winds at low to medium speeds and with north winds at medium to high speeds. Narrowness of the ‘high-probability areas’ suggested that the sources were mainly local. Settlements to the SE from the sampling site are mainly of semi-rural and rural type that use wood for heating in between 60 and 70% of households (SO/RS 2013).

Fig. 3
figure 3

CBPF plots of PM2.5 mass concentrations and the related sources’ contributions above the 70th percentile

Full size image

For TRAFF, high probabilities came with SE winds at high speeds, and somewhat lower probabilities were associated with lower wind speeds from the same direction. CBPF map suggested some influence from the SW direction as well. High contributions associated with high wind speeds pointed to the steel industry complex and a TPP located further away in the SE direction, but also could have originated in traffic resuspension (Belis et al. 2013). Removing data points with only one recorded hour from CBPF analysis highlighted the SW impact—possible E75/urban traffic impact—and the proximity of the SE source at low wind speeds (Figure 9S, Online resource). Details about excluding and weighting of data points in CBPF modelling can be found in Carslaw and Ropkins (2012). Analysis of CPF probabilities of PM2.5 (above 70th percentile) in the HS (automatic monitoring data distributed over hours of the day) further indicated that the impact of traffic emissions might dominate this factor (Figure 10S, Online resource).

Probabilities for RC/SS grew with increasing speeds of the SE winds where two lignite powered TPPs are located. Besides, about 16% of households in the Belgrade metropolitan area use coal for heating (SO/RS 2013).

The highest probabilities for LC/NO3 occurred with the west winds at medium to high speeds pointing to the densely populated city area and the biggest district heating plant. Several central heating units that run on fuel oil only are distributed around that territory as well. Nitrates formed by the conversion of NOx (emitted from traffic and various heating units) due to accumulation processes also could have originated from this direction (Masiol et al. 2015).

The highest probabilities for SOIL source appeared with the SE and north winds at high speeds that could possibly reflect the influence of transported air masses carrying resuspended mineral dust from deserts, semi-arid areas, and agricultural and bare lands (Belis et al. 2019b).

The influence of the three sources (BB, LC/NO3 and SSULPH) with the highest contribution was visible in CBPF plots considering PM2.5 mass concentrations. Overall, CBPF plots suggested that the emission sources were not strictly local.

Long-range air transport

Trajectory cluster analysis was used to examine history of air masses arriving at the receptor site and CWT concentration maps to highlight the relative significance of potential source areas. To check the correspondence between the back-trajectory analyses and emissions, TCA and CWT maps were interpreted by visual comparison with the pollutant emission maps. The influence of orographic interferences in the analysed trajectories—due to the low arrival height—was examined by repeating analyses with 96 h backward trajectories starting at 500 m, and although some differences were noticed, they did not look substantial for the following discussion. Results of these analyses along with short discussion are presented in Figures 18S, 19S and 20S, online resource.

Trajectory cluster analysis defined five clusters. In Figure 11S and Figure 12S, online resource, gridded frequency maps of unique back-trajectory crossings are shown, and the average sources’ contributions associated with each cluster. CWT maps are presented in Fig. 4.

Fig. 4
figure 4

CWT plots (based on 96-h back trajectories) for PM2.5 and EPA PMF resolved sources’ contributions (μg m−3)

Full size image

The first cluster, shown in Figure 11S, included 21.3% of trajectories most frequently passing over Poland, Slovakia and Hungary. Cluster 2 (18.2%) comprised trajectories coming from Western Europe, covering Central Europe, Italy and Western Balkans, while cluster 3 (32.3%) was the most frequent, and it involved air masses circulating mainly over Serbia and the Balkan Peninsula. Cluster 4 (20.3%) originated in Eastern Europe with trajectory crossings being the most frequent over south Romania, Moldavia and Ukraine, while the fifth cluster (7.8%) carried the fast air masses coming from the Atlantic Ocean and North-Western Europe.

High contributions of BB were associated with clusters 2, 4 and 5, i.e. with the air masses covering Central Europe and Italy, Eastern Europe and North-Western Europe. These clusters and the regions highlighted by CWT were mostly related to cold airflows that lead to worsening weather conditions in Belgrade (Figure 13S) and therefore most probably associated with high local emissions. The use of wood increased in Europe in the last decade (Guerreiro et al. 2018), but a noticeable impact of long-range transport is unlikely due to a high percentage of households in both Serbia and Belgrade district use wood for heating. Note though that open fires in February and March 2015 in the north of Ukraine potentially affected BB contributions at the studied site (Perrone et al. 2018).

The highest TRAFF contributions were found within the cluster 4. Although this source is expected to be mainly local, trajectory frequencies and CWT maps reflected the spatial distribution of metal-processing industries in Eastern Europe (Figure 14S, online resource). It is thus possible that a long-range transport affected this source, carrying air masses probably enriched by emissions from the regional metal-processing industry (about 50 km to the SE from the sampling site).

RC/SS levels were the highest in cluster 3 and close to the sampling period mean in clusters 1, 2 and 4. Cluster 3 included slow air masses circulating over Serbia and the Balkan Peninsula and was also related to the highest share of calm hours and the highest temperatures (Figure 15S, online source). Levels of SSULPH in this cluster highlighted the impact of lignite burning emissions as several countries in the region (Serbia, Bosnia and Herzegovina, Northern Macedonia, Greece, Romania, Bulgaria and Montenegro) use it as the main fuel for electricity generation (Koukouzas et al. 2004; Jovančić et al. 2011; Manousakas et al. 2015; Belis et al. 2017). Two of the Serbian TPPs (Nikola Tesla and Kostolac) represent some of the main SO2 polluters in the region (Belis et al. 2017). CWT maps pointed to Northern Macedonia, Bulgaria and Northern Greece and to a lower extent to Romania and the areas north of the Black Sea. More than several TPPs that run on locally produced lignite are located in these regions. Total capacities of TPPs in Northern Greece and Western Bulgaria are higher than 2000 MW each. High lignite consumption and low efficiency in these TPPs are associated with high SO2 emissions (Vazquez-Hernandez et al. 2018) thus suggesting the influence of regional transport on secondary sulphate concentrations.

The highest LC/NO3 contributions were associated with cluster 5, while in clusters 1 and 2 they were around the sampling period mean. Cluster 5 was connected to the highest wind speeds and the coldest days during the HS and had almost no contribution in the NHS (Figures 15S and 16S, online source). The other two covered mainly Central Europe and Italy, while Poland and Italy were highlighted by the CWT map. The usage of fuel oil for transformation and heating purposes in Poland and Italy is higher than that in most of Europe (EC 2019), but there are several reasons why a substantial impact of long-range transport is unlikely in this case. High contributions were, as in cluster 5, mainly associated with cold air masses (Figure 13S, online resource) and, during the coldest days, district heating plants in Belgrade mostly run on heavy fuel oil while some local boilers use it all the time. Besides, if there were significant long-range contributions from oil burning sources, these source regions should have been common with the SS source and vice versa (Pacyna et al. 1984). Low temperatures accelerate nitrate particle formation as well, so emissions from heating units within the city and surroundings and the enhanced winter traffic were probably the most significant contributors.

Soil factor was mainly associated with clusters 2 and 3. Both clusters coincided with high wind speeds (> 4 m s−1) from the south-east direction, and both included trajectories coming from north and north-western Africa. CWT values were high to the south-east of the studied area and over the Mediterranean and the Adriatic Sea. It appears that a part of the SOIL factor could have originated in the region possibly from resuspended and/or transported agricultural soil (Belis et al. 2019). High CWT values over the Mediterranean region were considered only indicative of the transport path and not as a source area as suggested by Perrone et al. (2018). The DREAM and NOAA HYSPLIT model predictions pointed to several Saharan dust episodes (Figure 17S, online resource), which are relatively frequent in this region (Basart et al. 2012). For the 17th of March 2015, NOAA HYSPLIT trajectories indicated contribution from the region east of the Caspian Sea, possibly the Turkmenistan Desert.

Conclusion

The comprehensive analysis of mass and composition of daily PM2.5 deposits, collected in May 2014–May 2015 with 3/7 frequency, in a suburban area of Belgrade, Serbia, has been a subject of the study. Extraordinary precipitation events marked the sampling campaign influencing the results and contributing significantly to adequate explanations.

Results have pointed to a relatively high exceedance frequency of WHO AQG value; i.e. daily PM2.5 mass concentrations were higher than 25 μg m−3 in about 33% of the samples almost exclusively during the heating season. The highest portion of PM2.5 was attributed to SO42− ion while the aerosol acidity had been demonstrated by neutralization ratio calculation. Mass concentrations of S, K and Pb were among the highest in the European region, indicating extensive use of fossil fuels and wood for electricity and heat generation in this area. Nevertheless, a long-term decrease (since 2007) on some anthropogenic components—S, Ni, V, Ba, Pb, Zn and Cr—was visible, probably due to both local and regional abatement measures that took place in the observed timespan.

Based on data obtained by PIXE and ion chromatography techniques, several receptor-based statistical tools have been combined to allocate possible sources of emission on the regional and long-range scale. Source apportionment, performed by PMF EPA receptor model, distinguished five dominant types of sources: biomass burning (14.5%), traffic (3.9%), regional combustion/secondary sulphates (28.8%), local combustion/secondary nitrates (29.7%) and soil (5.4%). High share of unaccounted mass (17.8%) was attributed to the lack of OC and EC measurements since emission sources in the area are combustion heavy, while the impact of traffic emissions was minimal probably due to the position of a sampling place. Local and long-range transport analyses combined suggested that the BB, and the LC/NO3 originated from the wider urban area. This is in accordance with a high number of households and heating units using wood and fuel oil for heating. Emissions of NOx from natural gas combustion in district heating plants and enhanced traffic during the heating season may be considered precursors for the local nitrate formation. The highest contributions of secondary sulphates were attributed to days when slow air masses were circulating above Serbia and the Balkan Peninsula. This was not surprising since developing countries in the region use locally excavated lignite for electricity and heat production. Temporary shutdown of TPPs and lignite mines in Serbia coincided with the decrease of RC/SS contributions suggesting that a significant share originated in the oxidation of SO2 emitted in the ‘local’ TPPs as well.

Presented results suggested that the future efforts on the local-urban level should be directed mainly towards mitigation of emissions from various heating units and traffic, while for the reduction of secondary sulphate levels, emissions from thermal power plants in Serbia and the region should be addressed.