The monitoring station was located in Foshan, Guangdong Province, China (Fig. 1), on top of the Foshan Environmental Monitoring Center (Foshan EMC; 23.0025∘ N, 113.1035∘ E; approximately 35 m above ground level). Foshan is one of the most important manufacturing hubs in China, characterized by the ceramics industry and household appliance industry, and produces ∼ 15 % of China's home appliances and ∼ 30 % of the world's ceramics (Guo et al., 2011). Adjacent to the monitoring site is a continuous sampling system measuring NOx (Advanced Pollution Instrumentation, API, model 200E), CO (API, model 300E), SO2 (API, model 100E), O3 (API, model 300E) and PM2.5 and PM10 (Thermo Scientific, model FH62C14), operated by the Foshan Environmental Monitoring Center. The station is surrounded by residential buildings and business offices on flat terrain. Meteorological parameters, including wind speed (WS), wind direction (WD), temperature (T), relative humidity (RH) and rainfall were obtained from Foshan Meteorological Bureau. The meteorological station is situated in the Foshan No. 1 Middle School, about 5 km north of Foshan EMC. The time resolution for the meteorological data is 10 min.

Hourly time-integrated samples of size-segregated coarse (PM2.5–10) and fine (PM2.5) PM samples were collected using a modified Streaker sampler (PIXE International Corporation, USA ). The Streaker sampler has previously been described in detail (Annegarn et al., 1988), and used in a number of studies (Annegarn et al., 1992, 1996; Filippi et al., 1999; D'Alessandro et al., 2004). The Streaker sampler system used in this study is described in Zhou et al. (2016). Briefly, the Streaker sampler consists of a pre-impactor that removes particles larger than PM10 from the incoming air flow, a thin Kapton foil that collects coarse particles through impaction and a Nucleopore filter (0.4 µm pore size) that collects fine particles. An electronic control system regulates the Streaker sampler pneumatics mass flow (1 L min-1) and stepper motor rotation of the filter between each user-defined sampling period. The hourly samples were collected with discrete spacing between each deposit to ensure that each deposit only consisted of particulate matter collected during the intended hour. A total of 60 samples or 60 h can be collected on each filter.

The high-resolution particulate matter sampling system was installed on the roof of Foshan Environmental Monitoring Center. The sampler was set to collect hourly samples beginning at 12:00 UTC+8 on 24 October 2014 and ending at 10:00 on 14 December 2014. A total of 47 samples or 2 days were collected per filter (coarse and fine) with the sampling ending at 11:00 (47 h later), allowing an hour to change over filters before the sampling was restarted again at 12:00. In all, 1127 samples each of PM2.5 and PM2.5–10 were collected, representing 1127 h or 7 weeks of sampling.

Ion beam analysis (IBA) was used to measure the concentrations of elements with atomic numbers above neon in the PM samples. The IBA was performed using a 3 MeV accelerator proton beam with standards (SrF2, NaCl, Cr, Ni, SiO, KCl, Al) run before and after each analytical cycle. Spectral X-ray peak deconvolution was performed using Gupix software (Maxwell et al., 1995). The number of pulses (counts) in each peak for a given element is used by the Gupix software to calculate the concentration of that element. The background and neighboring elements determine the statistical error and the limit of detection. Note that Gupix provides a specific statistical error and limit of detection (LOD) for each element in each PM sample and these have been used to provide the uncertainty matrix using in the PMF analysis. IBA measurements were carried out at the New Zealand National Isotope Centre operated by the Institute of Geological and Nuclear Sciences (GNS) in Gracefield, Lower Hutt, New Zealand (Trompetter et al., 2005). Further details on the IBA techniques used, analytical uncertainties and limits of detection have been reported previously (Ancelet et al., 2012). Black carbon was measured using a M43D digital smoke stain reflectometer (Ancelet et al., 2011).

Receptor modeling and apportionment of PM mass by PMF was performed using the EPAPMF version 5.0.14 program in accordance with the user's guide (https://www.epa.gov/air-research/epa-positive-matrix-factorization-50-fundamentals-and-user-guide). With PMF, sources are constrained to have non-negative species concentrations; no sample can have a negative source contribution and error estimates for each observed point are used as point-by-point weights. This is a distinct advantage of PMF, since it can accommodate missing or below detection limit data that is a common feature of environmental monitoring (Song et al., 2001). Another advantage of PMF is that PM mass concentrations can be included in the model as another variable and the results are directly interpretable as the covariant PM mass contributions associated with each factor (source). Prior to the PMF analyses, data and uncertainty matrices were prepared in the same manner as previous studies (Polissar et al., 1998; Song et al., 2001). Data screening and the source apportionment were performed in accordance with the protocols and recommendations set out by Paatero et al. (2014) and Brown et al. (2015). Due to the effect that random analytical noise can have on the receptor modeling process, variables with low signal-to-noise ratios were examined by alternate inclusion and exclusion in a modeling run and only those variables that could be explained in association with source emissions were included in the final results (Paatero and Hopke, 2003).

In order to locate the local sources and estimate the wind direction impacts on each source identified using PMF, the CPF method was applied. The CPF calculates the probability that a source is located with a particular wind direction sector (Pekney et al., 2006). The probability that a source originates from a given wind direction is estimated by comparing the wind direction distribution for the upper 25 % (or 75 percentile) of source contributions relative to the total wind direction distribution in this study: CPFΔθ=mΔθnΔθ, where mΔθ is the number of occurrences from wind sector Δθ for the upper 25 % of concentrations and nΔθ is the total number of occurrences from the same wind sector.

Using the high-temporal-resolution data, bivariate polar plots were generated with the R statistical and Openair software packages (Carslaw and Ropkins, 2012). Using bivariate polar plots, source concentrations can be shown as a function of both wind speed and direction, providing invaluable information about the direction of potential sources and the influence of wind speed on concentrations. In this study we produced bivariate polar plots using the conditional probability function analysis (CPF) to identify the directions from which high source concentrations are likely to originate. A full description of CPF analysis can be found in Carslaw and Ropkins (2012).

Among the detected elements, As, Pb, Mn, Si and Cr were identified as non-carcinogenic risks, while inorganic As, Pb and Ni were considered carcinogenic (USEPA, 2017). The reference levels and cancer risk slopes for Cr vary with its valences (e.g., Cr(VI) and Cr(III)) and exposure pathways (e.g., oral intake, inhalation). However, neither reference concentration (Rfc) nor cancer risk slope for inhalation of Cr(III) are available in the Integrated Risk Information System (IRIS), so we assumed that the total Cr in our study was Cr(VI), the Rfc and cancer risk slope of which are provided in IRIS (USEPA, 2017).

The adjusted air concentrations (Cair-adj) for the toxic elements contained in PM2.5 was calculated from Eq. (1), modified based on the equations in the risk assessment guidance (USEPA, 2009). Cair-adj=∑124Cair×ETi×(1day/24h)×ED×EFAT×1000, where Cair is the hourly ambient trace elements concentration detected in PM2.5 (µg m-3); ETi is the exposure time (1 h day-1); ED is the exposure duration (6 years for children and 30 years for adults); EF is exposure frequency (350 days year-1 in this study); AT is the averaging time (30 years × 365 days year-1 for non-carcinogens; 75 years × 365 days year-1 for carcinogens; M. Huang et al., 2014, 2016).

The obtained adjusted air concentrations of the toxic trace elements were subsequently divided by their respiratory reference concentrations (RfCs) to yield the respective hazard quotient (HQ; USEPA, 2017). Hazard index (HI) was produced by summing up all the hazard quotients (HQs) based on the hypothesis that the adverse effect is proportional to the sum of hazard quotients (HQs). Hazard index of less than 1 indicated no significant adverse non-carcinogenic effect.

For the carcinogenic risks assessment, the obtained adjusted air concentrations of the toxic elements of cancer risks from Eq. (1) were multiplied by their slope factors to calculate the carcinogenic risks (CRs; USEPA, 2017). The carcinogenic risk refers to a person's chance of developing cancer from the exposure to any carcinogenic agent. CR equal to 1 × 10-4 is considered as the least stringent level, above which indicating significant cancer risk.

The hourly variations of PM2.5 and PM10 mass concentrations during the observation are displayed in Fig. S1 in the Supplement. The PM2.5 and PM10 concentrations varied from ∼ 8 to ∼ 270 and ∼ 16 to 363 µg m-3, with an average value of 62.5 ± 33.5 and 91.6 ± 47.2 µg m-3, respectively. Around 25 % PM2.5 mass concentrations exceeded the second grade of new issued National Ambient Air Quality Standard (NAAQS, GB 3095-2012) of China, with daily average value of 75 µg m-3. In addition, the mean ratio of PM2.5 / PM10 was estimated to be about 0.67, which is consistent with the ratio reported by the previous studies in the Pearl River Delta region, North China Plain and Yangtze River Delta region (Wang et al., 2006; Sun et al., 2004; Hu et al., 2014). This result indicated that fine particles, PM2.5, accounted for a majority of PM10.

In order to compare the relationship between the range of analytes measured in PM and the total PM mass, a mass closure approach was used. Ideally, when elemental analysis and organic compound analysis have been undertaken on the same sample one can reconstruct the mass using the following general equation for ambient samples as a first approximation (Cahill et al., 1989; Cohen, 1999; Malm et al., 1994): Reconstructed mass=[Soil]+[OM]+[BC]+[Smoke]+[Sulfate]+[Seasalt], where [Soil] = 2.20[Al] + 2.49[Si] + 1.63[Ca] + 2.42[Fe] + 1.94[Ti], [OM] = Σ[Concentrations of organic compounds], [BC] = Concentration of black carbon (soot), [Smoke] = [K] - 0.6[Fe], [Seasalt] = 2.54[Na], [Sulfate] = 4.125[S].

The reconstructed mass (RCM) is based on the fact that the six composite variables or “pseudo” sources given in Eq. (2) are generally the major contributors to fine and coarse particle mass and are based on geochemical principles and constraints. As a measure of [OM] was not available in this study, it was assumed that it composed part of the “remaining mass” (the difference between RCM and measured PM (BAM)) that includes water and nitrates as major components (Cahill et al., 1989). As Na was below the LOD across most of the samples, the reciprocal calculation of [Seasalt] = 1.65[Cl] was substituted, though care was taken with the interpretation of this pseudo source due to the potential for Cl loss by volatilization from aerosol (or from filters) in the presence of acidic aerosol species, particularly in the fine fraction (Lee et al., 1999; Chen et al., 2016). Most fine sulfate particles are the result of oxidation of SO2 gas to sulfate particles in the atmosphere (Malm et al., 1994). For RCM, it is assumed that sulfate is present in fully neutralized form as ammonium sulfate. [Sulfate] therefore represents the ammonium sulfate contribution to aerosol mass with the multiplicative factor of 4.125[S] to account for ammonium ion and oxygen mass (i.e., (NH4)2SO4= ((14 + 4) × 2 + 32 + (16 × 4) / 32); Cahill et al., 1989; Cohen, 1999; Malm et al., 1994).

The RCM and mass closure calculations using the pseudo-source and pseudo-element approach are a useful way to examine initial relationships in the data and how the measured mass of species in samples compares to total PM mass, an important consideration for receptor modeling studies. Figure 2 indicates that the RCM of measured elements accounts for approximately 30 % of total mass for both PM2.5 and PM10 and that there was significant correlation (r2 = 0.64 and 0.69 for PM2.5 and PM10, respectively), which indicates that the unmeasured PM components (OM, nitrate, bound H2O) are likely to be strongly associated with those species that were measured. Analysis of the data showed that most of the remaining mass (RM) for PM10 was in fact PM2.5, as would be expected for such aerosol species as OM (including secondary organic aerosol) and nitrate. Therefore, an extra variable was calculated, RMPM2.5 (where RMPM2.5 = PM2.5 - RCMPM2.5), to be included in the PMF analysis as a placeholder for the unmeasured components, an approach that has been successfully applied in other studies (Belis et al., 2013). The advantage of such an approach is that all PM mass is accounted for in the PMF model. However, the limitation is that it still does not define exactly what aerosol species the RMPM2.5 variable includes, although some inferences can be made depending on the source association.

Using the combined fine and coarse PM elemental data sets along with the gaseous data from the Foshan EMC, PMF was used to identify six factors or sources contributing to PM concentrations during the monitoring period. PMF modeling diagnostics and other parameters are detailed in the Supplement. Multiple PMF model runs were performed choosing fewer and more factors to examine the effect on modeling diagnostics and interpretability of the source profiles coupled with the advantage of high-resolution data to examine the diurnal concentration variations. The final six-factor solution adopted included an Fpeak rotation (-3, %dQ(Robust) = 2.91) that provided good separation of the minor marine aerosol source as evidenced by the Fpeak bootstrapping results. Note that the gaseous components (SO2, NO2 and CO) have been included to aid with source identification and to examine those PM emission sources or secondary aerosol sources they are closely associated (covariant) with. This method has been also adopted in previous studies, and proved to enhance the resolving power of source apportionment studies (Zhou et al., 2005; Liu et al., 2006). The sources identified were marine aerosol, biomass combustion, construction dust, motor vehicles and road dust, secondary inorganic aerosol and industrial coal combustion. These sources were found to explain 89 and 91 % of the PM2.5 and PM10 mass, respectively (Fig. S3). The factor profiles are presented in Fig. 3, where elements in the coarse fraction are denoted as X-C (e.g., Al-C). The first factor represents biomass combustion because of the presence of BC, K and fine Cl as primary species along with Zn, which is consistent with previous studies and represents the burning of wood or plant material (Maenhaut, 2018; Ancelet et al., 2012). Motor vehicles and road dust were identified as the source of the second factor based on the presence of NO2, BC and crustal matter components Al, Si, Ca, Fe and Zn from the coarse fraction as the significant elements in the profile. This profile represents both exhaust (tailpipe) emissions and non-exhaust (road dust and brake and tire wear) emissions, hence the combination of coarse and fine elemental species and the higher contribution of the source to PM10 concentrations relative to PM2.5. Ambient source profiles derived for motor vehicles generally include particulate matter from all engine types as emissions tend to be co-mingled by turbulent air movement at street level due to road traffic and are therefore temporally and spatially covariant (Amato et al., 2009; Pant and Harrison, 2013). The third source contains most of the black carbon, a substantial amount of CO, SO2, NO2, S and fine fraction heavy metals (Cr, Mn, Cu, Zn, As, Pb) and has been attributed to coal combustion which is likely to include coal-fired power station emissions that are probably mixed with industrial process emissions (Song et al., 2007; Tian et al., 2014). The fourth factor contains the majority of both coarse and fine S, which dominates the elemental mass and is considered to represent the secondary inorganic aerosol component from gas-to-particle transformation in the atmosphere. Vanadium and nickel were also associated with this source suggesting contributing source emissions from the combustion of sulfur-containing fuels such as heavy fuel oil or oil refinery emissions (Querol et al., 2008; Maenhaut, 2018). Secondary aerosols are primarily fine particles from the gas-to-particle conversion process, but some of the particle size range does extend into the coarse fraction (Anlauf et al., 2006), particularly where heterogeneous atmospheric chemistry takes place on the surface of particles or in aerosol droplets during the reaction of sulfur gaseous species to form secondary sulfate particle species (George and Abbatt, 2010). The fifth factor, construction dust, with a high Ca loading in the profile along with crustal matter components (Al, Si, Fe) in both the fine and coarse fractions, has been attributed to activities that generate cementitious (hence the high Ca content) and crustal matter dusts in the area such as construction/demolition of buildings and other structures (i.e., cement mixing, concrete pouring, concrete cutting or drilling and soil excavation), which significantly distinguishes the source from that which might be associated with motor vehicles/road dust emissions (Owega et al., 2004; Chueinta et al., 2000; Maenhaut, 2018). The sixth source was characterized as marine aerosol due to the predominance of Cl-C. Interestingly, as indicated in Fig. 3 most of the calculated variable, RMPM2.5 (derived in Sect. 3.1), was associated with the motor vehicle and biomass combustion sources (45 and 17 %), respectively (Fig. S4), most likely indicating the association of OC with these sources (Querol et al., 2013; Pósfai et al., 2004), and the remainder of RMPM2.5 (36 %) was associated with the secondary aerosol source and is probably indicative of secondary nitrate and secondary organic aerosol concentrations (Huang et al., 2011; Freney et al., 2014).

Average source contributions to particulate mass concentrations are presented in Fig. 4, which shows that industrial coal combustion (26 %, 16.0 µg m-3), secondary inorganic aerosol (24 %, 14.8 µg m-3) and motor vehicle emissions (20 %, 12.5 µg m-3) dominated PM2.5 concentrations during the monitoring period. Mass contributions to PM10 from coal combustion and secondary inorganic aerosol were essentially unchanged from the fine fraction (16.8 and 16.9 µg m-3, respectively), while the contributions from motor vehicles (30.2 µg m-3) and construction dust (15.1 µg m-3) increased due to the significant coarse particle component associated with these latter sources.

When considering the gases, SO2 was strongly associated with the industrial coal combustion source. Meanwhile, the majority of NO2 was split between motor vehicles and the industrial coal combustion source. Interestingly CO was primarily associated with the industrial coal combustion and secondary inorganic aerosol sources. CO, NOx and SO2 emissions are associated with the light-duty motor vehicles (i.e., petrol vehicles), which emit few particles directly. The petrol vehicles as well as coal combustion-emitted NOx and SO2 transform to nitrate and sulfate particles through chemical reactions. The association of CO with secondary inorganic aerosol is explained by co-emission of CO with the gaseous combustion product precursors (e.g., SO2, NOx) of secondary inorganic aerosol, and they are therefore present (covariant) in the same air mass. Diesel-powered vehicles produce very little CO initially (Rhys-Tyler et al., 2011). It is assumed that diesel vehicle tailpipe emissions are primarily responsible for motor-vehicle-related PM2.5 concentrations in the Foshan urban airshed, consistent with international research (Kirchstetter et al., 2008; Kim et al., 2009; Wang et al., 2012; Targino et al., 2016). Therefore, CO is not associated with motor-vehicle-emitted particulate matters (primary aerosol). Also included in the source contribution graphs (Fig. 4) is the remaining mass (i.e., measured PM mass minus predicted PM mass). The time series plots presented in Figs. 5 and 6 for PM2.5 and PM10, respectively, show that the biomass burning and marine aerosol sources were episodic and that the concentrations of the motor vehicles and construction dust sources increased in PM10 due to the coarse particle content.

Peak PM2.5 and PM10 episodes were associated with motor vehicle and coal combustion, particularly when PM from biomass combustion was also a significant contributor. This suggests that biomass combustion sources were additive over the other sources and forced the PM2.5 concentrations to extreme highs. Analysis of satellite imagery (MODIS fire spot) indicates that biomass combustion activities were outside of the region to the north and northeast of Foshan and were advected over the city during the peak PM pollution events (Fig. S2). High concentrations of potassium were also detected during the biomass burning episodes (Zhou et al., 2016). It is most likely that the biomass combustion source is associated with agricultural burn-off around Foshan. The time averaged daily patterns in source concentrations were examined and as shown in Fig. 7a, the industrial coal combustion source concentrations were generally higher at night, perhaps due to a stable nocturnal boundary layer/mixing height effect. The construction dust concentrations (Fig. 7b) were significantly higher during the day reflecting the pattern of daytime activities generating the dusts. The motor vehicles and road dust sources (Fig. 7c) showed a bimodal pattern with peaks in the morning and evening rush hours, which was ascribed to the morning and evening commuter traffic and lower boundary layer. We could see the distinguishing diurnal variations between motor vehicle/road dust and construction dust sources.

The secondary inorganic aerosol source (Fig. 7d) demonstrated slightly higher concentrations at night than at daytime. Precursor gaseous emissions are oxidized via various gas or aqueous pathways and are then converted to sulfate aerosol (Seinfeld and Pandis, 2006; Sun et al., 2014). However, the impacts of changes in the boundary layer tend to build up the air pollutants during the nighttime. This is also consistent with the fact that there is a considerable amount of precursor gas emissions from upwind regions and that processing has already occurred there, or while the parcel has undergone dynamical transport. This implies that there is a higher chance of being mixed throughout the boundary layer and lower troposphere as well and therefore it is not as sensitive with respect to a ground-based measurement as more local emissions. The marine aerosol and biomass combustion sources are not shown in Fig. 7 due to their episodic nature.

Figure 8 presents bivariate CPF polar plots for the source contributions to PM. Fig. 8a shows that the highest 25 % of industrial coal combustion source contributions originated from the west to northwest quadrants under light winds (1–3 m s-1).

A highly resolved spatial anthropogenic emission inventory for PRD was conducted by South China University of Technology for the base year 2012 (Yin et al., 2017). The data shows that the highest density of PM2.5 combustion sources (including power plants and industrial combustion) and industrial processes emissions were concentrated within 30 km to the west and northwest of the Foshan EMC as presented in Fig. S5. Three other large coal fired power stations were also located approximately 20 km to the west–southwest of the monitoring site.

Peak marine aerosol (Fig. 8b) arrived at the site from the southeast at higher wind speeds (5–7 m s-1) in the direction of the South China Sea. Secondary inorganic aerosol (Fig. 8c) came from the north-northeast direction at moderate wind speeds (3–5 m s-1) with a smaller component from the south-southeast. Construction dusts (Fig. 8d) were highest during low to moderate wind speeds out of the northeast sector, arriving from the direction of the Foshan city center, where the majority of construction activities were occurring. The upper 25 % of biomass combustion concentrations (Fig. 8e) were from the northwest sector. It is likely that the biomass combustion emission sources were intermittent and regional, related to agricultural activities or possibly wildfires. Peak motor vehicle source concentrations (Fig. 8f) occurred during south-southwest winds. The PM sampler was located 35 m above Fen Jiang Nan Road, which is an arterial route with six lanes running north–south past the monitoring site. The CPF result suggests that winds from the south-southwest were most effective in transporting roadway emissions to the sampler.