Figure 1a shows a map of eastern China with provincial borders and land cover, in which Shanghai city (provincial level) sits in the middle portion of China's eastern coast and its metropolitan area (indicated as the densely populated area in Fig. 1b) concentrated on the south edge of the mouth of the Yangtze River. The municipality borders the provinces of Jiangsu and Zhejiang to the north, south and west, and is bounded to the east by the East China Sea (Fig. 1a). Shanghai has a humid subtropical climate and experiences four distinct seasons. Winters are chilly and damp, with northwesterly winds from Siberia sometimes causing nighttime temperatures to drop below freezing. In summer, the airflow carries moist air from the Pacific Ocean to mainland China, and it also brings the main precipitation. The city is also susceptible to typhoons in summer and the beginning of autumn. Air pollution in Shanghai is low compared to other cities in northern China, such as Beijing, but still substantial by world standards, especially in winter (Han et al., 2015; Chang et al., 2017).

Field measurements were performed at the rooftop (∼18 m a.g.l.) of the Pudong Environmental Monitoring Center (PEMC; 31.2331∘ N, 121.5447∘ E; ∼7 m a.s.l.) in the Pudong New Area of southwestern Shanghai, a region with dense population (Fig. 1b). Pudong New Area is described as the “showpiece” of modern China due to its height-obsessed skyline and export-oriented economy. For the PEMC, there were no metal-related sources (except for road traffic) or high-rise buildings nearby to obstruct observations, so the air mass could flow smoothly. More broadly, as indicated in Fig. 1c, the PEMC is surrounded by a multitude of emission sources such as coal-fired power plants (CFPPs) in all directions and iron and steel smelting in the northwest. Furthermore, a high level of ship exhaust emissions in 2010 such as V (Fig. 1d) and Ni (Fig. 1e) in the YRD and the East China Sea within 400 km of China's coastline was recently quantified based on an automatic identification system model (Fan et al., 2016). Therefore, the PEMC can be regarded as an ideal urban receptor site of diverse emission sources. More information regarding the sampling site has been given elsewhere (Chang et al., 2016, 2017).

From 1 March 2016 to 28 February 2017, hourly ambient mass concentrations of 18 elements (Si, Fe, K, Ca, Zn, Mn, Pb, Ba, V, Cu, Cd, As, Ni, Cr, Ag, Se, Hg and Au) in PM2.5 were determined by a Xact multi-metals monitor (Model Xact^™ 625, Cooper Environmental Services LLT, OR, USA) (Phillips-Smith et al., 2017; Jeong et al., 2016; Fang et al., 2015; Yanca et al., 2006). Specifically, the Xact sampled the air on a reel-to-reel Teflon filter tape through a PM2.5 cyclone inlet (Model VSCC-A, BGI Inc., MA, USA) at a flow rate of 16.7 L min-1. The resulting PM2.5 deposit on the tape was automatically advanced into the analysis area for nondestructive energy-dispersive X-ray fluorescence analysis to determine the mass of selected elemental species as the next sampling was being initiated on a fresh tape spot. Sampling and analysis were performed continuously and simultaneously, except during advancement of the tape (∼20 s) and during daily automated quality assurance checks. For every event of sample analysis, the Xact included a measurement of pure Pd as an internal standard to automatically adjust the detector energy gain. The XRF response was calibrated using thin film standards for each elements of interest. These standards were provided by the manufacturer of the Xact, produced by depositing vapor phase elements on a blank Nuclepore filter (Micromatter Co., Arlington, WA, USA). The Nuclepore filter of known area was weighed before and after the vapor deposition process to determine the concentration (µg cm-2) of each element. In this study, excellent agreement between the measured and standard masses for each element was observed, indicating a deviation of <5 %. The 1 h time resolution minimum detection limits (in ng m-3) were: Si (17.80), K (1.17), Ca (0.30), V (0.12), Cr (0.12), Mn (0.14), Fe (0.17), Ni (0.10), Cu (0.27), Zn (0.23), As (0.11), Se (0.14), Ag (1.90), Cd (2.50), Au (0.23), Ba (0.39), Hg (0.12) and Pb (0.13).

As a reference method to validate the Xact on-line measurements, daily PM2.5 samples were also collected at the PEMC site using a four-channel aerosol sampler (Tianhong, Wuhan, China) on 47 mm cellulose acetate and glass fiber filters at a flow rate of 16.7 L min-1. The sampler was operated once a week with a 24 h sampling time (starting from 10:00 am). In total 48 filter samples (26 cellulose acetate filter samples and 22 glass fiber filter samples) were collected, in which eight paired samples were simultaneously collected by cellulose acetate and glass fiber filters. In the laboratory, the elemental analysis procedures strictly followed the latest national standard method “Ambient air and stationary source emission–Determination of metals in ambient particulate matter–Inductively coupled plasma/mass spectrometer (ICP-MS)” (HJ 657-2013) issued by the Chinese Ministry of Environmental Protection. A total of 24 elements (Al, Fe, Mn, Mg, Mo, Ti, Sc, Na, Ba, Sr, Sb, Ca, Co, Ni, Cu, Ge, Pb, P, K, Zn, Cd, V, S and As) were measured using the Inductively coupled plasma-mass spectrometer (ICP-MS; Agilent, CA, USA). The comparisons of different measurement methods (Xact vs. ICP-MS) and different collection substrates (glass filter vs. cellulose filter) are reported in Table S1 in the Supplement. In general, the data for elements like K, Cr, Mn, Fe, Ni, Cu and Au proved to be of high quality, while elements like As, Cd and Ba have relatively poor data quality.

Meteorological data, including ambient temperature (T), relative humidity (RH), wind direction (WD) and wind speed (WS), were provided by the Shanghai Meteorological Bureau at Century Park station (located approximately 2 km away from the PEMC). The hourly mass concentrations of PM2.5 at the PEMC were measured by a particulate monitor (Thermo, FH62C-14). The routine procedures, including the daily zero or standard calibration, span and range check, station environmental control and staff certification, followed the Technical Guideline of Automatic Stations of Ambient Air Quality in Shanghai based on the national specification HJ/T193-2005. This was modified from the technical guidance established by the US EPA. Quality Assurance and Quality Control (QA/QC) for the Xact measurements was implemented throughout the campaign. The internal Pd, Cr, Pb and Cd upscale values were recorded after the instrument's daily programmed test, and the PM10 and PM2.5 cyclones were cleaned weekly.

The Positive matrix factorization or PMF is an effective source apportionment method to identify and quantify possible emission sources of measurements using the bilinear factor model (Paatero and Tapper, 1994) xij=∑k=1pgikfkj+eij, where xij is the jth species concentration measured in the ith sample, gik is the contribution of the kth source to the ith sample (factor time series) and fkj is the concentration of the jth species in the kth source (factor profiles). The part of the data remaining unexplained by the model is represented by the residual matrix eij. The entries of gik and fkj (required to be non-negative) are fit using a least-squares algorithm that iteratively minimizes the objective function Q: Q=∑i=1n∑j=1meijσij2. where σij are the measurement uncertainties.

In this work, the US EPA PMF version 5.0 was applied to attribute PM2.5 trace elements to specific factors/sources. 1 year hourly resolved measurements (n=8784) of 18 elements in the PM2.5 fraction were obtained and included for PMF analysis. The measurements (n=1265) with missing data were excluded. An estimated fractional uncertainty of 10 % was used to derive the uncertainty data set (Kim et al., 2005; Kim and Hopke, 2007), which did not impact the interpretability of the PMF results. The missing values of individual elements were replaced by their geometric mean of the remaining observations, and their accompanying uncertainties were set to 4 times the geometric mean. The measurements below detection limit (BDL) were set to half the detection limit, with uncertainties set at five-sixths the detection limit (Polissar et al., 1998). The EPA PMF 5.0 has three uncertainty estimation methods, including bootstrapping (BS), displacement (DISP) and bootstrapping enhanced with DISP (BS-DISP) (Norris et al., 2014; Brown et al., 2015; Paatero et al., 2014; Wang et al., 2017). BS-DISP analysis is time consuming due to the huge data set (7519×18), and only BS and DISP analysis were conducted individually. Details of the uncertainty analysis are described in the supporting information (Sect. S1). In this study, PMF solutions using 3–10 factors were considered, and the final factor number is determined based on the interpretability as well the uncertainty analysis with the BS and DISP methods.

The determination of the geographical origins of trace elements in Shanghai requires the use of diagnostic tools such as the conditional probability function (CPF) and bivariate polar plot (BPP), which are very useful in terms of quickly gaining an idea of source impacts from various wind directions and have already been successfully applied to various atmospheric pollutants and pollution sources (Chang et al., 2017; Carslaw and Ropkins, 2012). In this study, the CPF and BPP were performed on the one-year data set for the major trace elements with a similar source. The two methods have been implemented in the R “openair” package and are freely available at http://www.openair-project.org/ (last access: 1 August 2018) (Carslaw and Ropkins, 2012).

The CPF is defined as CPF =mθ/nθ, where mθ is the number of samples in the wind sector θ with mass concentrations greater than a predetermined threshold criterion, and nθ is the total number of samples in the same wind sector. CPF analysis is capable of showing which wind directions are dominated by high concentrations and with which probability. In this study, the 90th percentile of a given element species was set as threshold, and 24 wind sectors were used (Δθ=15∘). Calm wind (<1 m s-1) periods were excluded from this analysis due to the isotropic behavior of the wind vane under calm winds.

The BPP demonstrates how the concentration of a targeted species varies synergistically with wind direction and wind speed in polar coordinates, which is thus essentially a non-parametric wind regression model to alternatively display pollution roses but include some additional enhancements. These enhancements include plots being shown as a continuous surface and surfaces being calculated through modeling using smoothing techniques. These plots are not entirely new as others have considered the joint wind speed–direction dependence of concentrations (see for example Liu et al., 2015). However, plotting the data in polar coordinates and for the purposes of source identification is new. The BPP has been described in more detail in Carslaw et al. (2006) and the construction of BPP has been presented in our previous work (Chang et al., 2017).

Factor 1 was characterized by a large mass fraction of Ca, Fe, Ba and Si, which explained 90 %, 50 %, 77 % and 63 % of the concentration, respectively. This mixed factor is similar to that reported by Amato et al. (2009, 2013), Bukowiecki et al. (2010), Harrison et al. (2012) and Visser et al. (2015b). In the urban atmosphere, Fe can be released from engine oil or catalyst equipped gasoline vehicles (Chen et al., 2007). Besides, Fe is linked to non-exhaust emissions such as brake wear because it is the support material for brake pads and the agents present in brake linings typically consist of Ba, Mn and Cu (Lough et al., 2005; Hjortenkrans et al., 2007; Dall'Osto et al., 2016). Therefore, Fe and Ba can be regarded as chemical tracers for a traffic-related source (exhaust and non-exhaust) (Thorpe and Harrison, 2008; Lin et al., 2015). Ca and Si are known as two of the most abundant elements in the upper continental crust, and their atmospheric origin is typically attributed to wind-blown dust. Located on the eastern coast of China, Shanghai rarely receives long-range transport of crustal matter from aeolian dust and the Gobi Desert in northwestern China (Huang et al., 2013). Sampling in the urban area of Shanghai, airborne Ca and Si should be dominated by anthropogenic activities like road fugitive dust or urban construction works. In Fig. S10, significant correlations are observed for Ca, Si, Fe and Ba, suggesting that the measured Ca and Si during our study period were more likely derived from road fugitive dust. Therefore, factor 1 can be assigned to a traffic-related source and it was the largest source in Shanghai, accounting for 46 % (680 ng m-3) of the total measured elemental mass in PM2.5.

Hourly measurements over one-year periods provide a unique opportunity to examine the diurnal profile of factor 1. As reported in Fig. 4, the concentration of the trace elements contributed by factor 1 shows a marked bimodal diurnal cycle, with average values at rush hours that are over 2 times higher than at nighttime. Such variation pattern agrees well with the diurnal variation of the traffic flow in Shanghai (Chang et al., 2016), further confirming that factor 1 can be interpreted as traffic-related emissions.

In Fig. 3, V (100 %) and Ni (74 %) come almost exclusively from factor 2, while factor 2 contributes to less than 10 % of any other elemental species. V is typically emitted from oil and petrochemical refining and combustion, and natural gas extraction and processing (Duce and Hoffman, 1976; Hope, 1994; Shafer et al., 2012). From CPF and BPP analysis (Fig. S11), higher concentrations of both V and Ni were observed when winds originated from the east, northeast and southeast directions. The most dominant directions were east and southeast, suggesting the influence from the coastal port cluster or petroleum refinery industry located east/southeast of Shanghai (Fig. 1). Gathering evidence revealed that the ratio of V / Ni can serve as a robust indicator of shipping emissions (Tao et al., 2013; Celo et al., 2015; Liu et al., 2017; Viana et al., 2009). A recent study in Shanghai port suggested that the ratio of V / Ni in aerosols emitted from heavy oil combustion of ocean-going ship engines was 3.4 on average (Zhao et al., 2013). Here measured in the urban area, the average ratio of V / Ni in our study was 3.2 with slight seasonal changes (Fig. 5), indicating V- and Ni-containing aerosols from shipping emissions subject to minor atmospheric transformation. In short, factor 2 likely corresponds to shipping emissions (instead of petrochemical refining), which is consistent with the results of many previous source apportionment works (e.g., Liu et al., 2017; Zhao et al., 2013; Cesari et al., 2014; Healy et al., 2010).

Although shipping emissions only contribute to 5.9 % of the trace elements in the Shanghai urban center, their share can be expected to greatly increase in the harbor district (Zhao et al., 2013). The good news is that since 1 January 2016, the sea areas of Shanghai and the neighboring ports were designed as shipping emission control area, requiring use of lower sulfur fuels in place of heavy fuel oil in the main engines of the ships (Zhen et al., 2018). Therefore, it is critically important to assess the impacts of fuel changes on the air quality in Shanghai in the future through continuous measurements of trace elements.

The predominant elements found in factor 3 were Au (100 %), Cd (65 %) and Ag (63 %) with 37 % of Hg. These four heavy metals are important associated elements in Cu, Pb and Zn ores. In fact, Cu, Pb and Zn smelting represent the three most common forms of nonferrous metal smelting in China (Tian et al., 2015). Due to high temperatures during the roasting, sintering and smelting process for the extraction of Cu, Pb and Zn from ores, metals like Au, Cd, Ag and Hg in nonferrous metal ores will inevitably be vaporized and released into the flue gas (Pacyna and Pacyna, 2001; Wu et al., 2012). Therefore, factor 3 was interpreted as nonferrous metal smelting emissions and the contribution of this factor was 15 % (220 ng m-3) to the total measured elemental mass in PM2.5.

To further pinpoint the specific subsector of nonferrous metal smelting, here we calculate percentile concentration levels of Au and Hg, and plot them by wind direction in Fig. 6 (and Ag, Cd in Fig. S12). It clearly shows that Au and Hg largely share the same source region which is different from that for Ag and Cd, indicating that Au and Hg were emitted from a similar subsector of nonferrous metal smelting. In Shanghai, Zn smelting is the most important contributor of Hg emissions from the nonferrous metal smelting sector. Therefore, the element Au resolved in factor 3 during our study period can be expected to be originated from Zn smelting.

The most abundant elements found in factor 4 were As, Se, Pb, Hg (explaining 56 % to 95 % of the concentration) with some contributions of Cu (30 %), Zn (29 %) and an unexpected large amount of K (53 %). As, Se, Pb, Hg and Cu are typical marker elements for coal combustion (Morawska and Zhang, 2002). In China, 73 % of As, 62 % of Se, 56 % of Pb and 47 % of Hg were found to be emitted from coal combustion (Tian et al., 2015). Coal consumption in southern China (including Shanghai) is mainly driven by industrial boilers and power plant, while in northern China, coal-based heating is also a major sector of coal consumption (Tian et al., 2015). Seasonally, the average mass concentration of coal combustion-related PM2.5 trace elements during winter (410 ng m-3) was much higher than that during spring (270 ng m-3), summer (150 ng m-3) and fall (210 ng m-3) (Fig. 7). This seasonal pattern was not observed for other sources (not shown). Shanghai has a humid subtropical climate and experiences four distinct seasons. Winters are chilly and damp, with northwesterly winds from northern China transporting air pollutants (including trace elements) caused by coal-based heating to the Shanghai atmosphere (Huang et al., 2013; Chang et al., 2017). As the largest city-scale coal consumer in China, coal combustion contributed to 280 ng m-3 or 19 % of the PM2.5 trace elements during our study period.

Traditionally, K in particles is considered to originate from biomass burning along with some contribution of fugitive dust (Zhang et al., 2010; Hueglin et al., 2005; Fang et al., 2015). Here we show that over half of the element K in urban Shanghai was derived from coal combustion. The reason for this discrepancy may be that in most previous studies, K in particles was pretreated using deionized water to extract (X. Wang et al., 2013). In fact, K has a high mineral affinity (elements associated with aluminosilicates, carbonates and other minerals in coal ash), and in some extreme cases, only about 1 % of K in fly ash from coal combustion can be extracted by water (Querol et al., 1996). For example, particles collected from coal combustion by X. Wang et al. (2013) were extracted with deionized water, then atomized and measured by an ATOFMS. The ATOFMS mass spectrum contained a relatively low K peak. The observation by X. Wang et al. (2013) was not consistent with that of Suess (2002), who observed larger K peaks in ATOFMS spectra for coal combustion particles in an in situ measurement (i.e., freshly emitted particles were directly introduced into the ATOFMS and measured).

Factor 5 was distinguished by high levels of Cr, Mn and Zn representing 100 %, 56 % and 52 % of the concentration, respectively. These elements are typically emitted from ferrous metal smelting. For example, the steel production industry represents the dominant contributor to Zn emissions, accounting for about 60 % in China (Tian et al., 2015). Driven by rapid modernization of its infrastructure and manufacturing industries, China produced more than 49 % of world steel production in 2017 (around 830 million tons), and 6 of 10 of the largest steel producers are in China (data retrieved from https://www.worldsteel.org, last access: 1 August 2018). Headquartered in Shanghai (20 km northwest of the sampling site), the Baosteel is the fifth-largest steel producer in the world measured by crude steel output, with an annual output of around 35 million tons. Meanwhile, there are several factories of ferrous metal processing located in western Shanghai (Fig. 1). As the element Cr is reported to be transported over substantial distances by the air (Perry et al., 1999), the presence of ferrous metal smelting activities in the west–northwest of the sampling site is inferred to be associated with this factor based on the results of CPF and BBP in Fig. 8. Overall, ferrous metal smelting contributed 220 ng m-3 or 15 % of the PM2.5 trace elements in Shanghai.

The average mass concentration of each elemental species before, during and after every precipitation event is presented in Fig. S14. If precipitation effectively scavenges and removes aerosol, then the mass concentrations of trace elements during a precipitation event should be lower than that before and after this precipitation event. However, there is no uniform variation pattern in Fig. S14, indicating that precipitation may not be the predominant factor to influence the ambient elemental mass in some cases. For example, most elemental species had a relatively higher mass concentration during the 12th precipitation event (which lasted from 09:00 25 December to 22:00 26 December; Fig. S14). This can be explained by the much less anthropogenic activities during the periods prior to (03:00 to 08:00) and after (23:00 to 03:00 the next day) the 12th precipitation event.

For each elemental species, the variation of mass concentration before, during and after every precipitation event was aggregated and reported in Fig. 9a. Before the precipitation events, the mass concentrations of all species except Cr and Au were higher than those during the precipitation events (notably V, Zn, Fe), suggesting that water spray could generally help to reduce the PM2.5 trace elements load in the atmosphere. After the precipitation events, there were six species (notably V and Au) with their mass concentrations lower than during the precipitation events, indicating a potential long-lasting aftereffect of precipitation scavenging. Among all elemental species, the mass concentrations of Zn and Se fluctuate as a most ideal V-shape, which properly reflects the cycle of precipitation. However, as shown in Fig. 9b, a linear relationship cannot be observed between the decrease rate of the Zn concentration and the rainfall amount of each precipitation event. Although we failed to pinpoint the exact value in this study, our results imply that there is a threshold of precipitation amount to lower the ambient PM2.5 trace elements mass.

The variation of the overall mass concentration of trace elements contributed by each source and their relative contributions before, during and after the total of 12 precipitation events is shown in Fig. 10a to e and f, respectively. The mass concentration of traffic-related trace elements experienced the sharpest decrease during the transition of no-rainy hours to rainy hours (159 %), and a moderate rebound after precipitation (35 %). Fang et al. (2015) found that mobile source emissions generated through mechanical processes (re-entrained road dust, tire and break wear) and processing by secondary sulfate were major contributors to water-soluble metals. In our study, traffic-related sources mainly include road dust and brake wear, which can not only be easily removed through precipitation but also can hardly be blown up from wet road surfaces after raining. In comparison, the mass contribution of the coal combustion source was also wet removed rapidly first (139 %) due to its tracer elements like As, Se, Pb and Hg having a larger water-soluble fraction. However, after precipitation, the contribution of the coal combustion source dramatically increased over 2 times (Fig. 10d and f). This can be explained that different from the traffic-related source, the coal combustion-related trace elements are generally emitted through elevated chimneys in the sectors of industrial broilers and power plants. The mass concentrations of trace elements contributed by nonferrous and ferrous metal smelting during the three periods remained quite flat (Fig. 10c and d), suggesting that precipitation has little effect on ambient trace elements emitted from metal smelting activities. Nevertheless, given that traffic-related and coal combustion are the dominant contributors to ambient PM2.5 trace elements, our results validate that water spray could be an effective approach to help curb the severe atmospheric metal pollution in many Chinese cities.

In Fig. 10b and f, the contribution of shipping emissions to ambient trace elements (mainly V and Ni) during the three periods reduced continuously. Mostly transported from the Eastern China sea, V and Ni almost exclusively originated from the east of the sampling site. In other words, the contribution of shipping emissions to the urban atmosphere is supposed to be very sensitive to wind speed and wind direction in Shanghai. The wind roses for the three periods are presented in Fig. 11. It shows that before precipitation events, the average wind speed (±1σ) was the lowest (2.3±1.3 m s-1), and easterly winds prevail in most times. These factors are favorable to the transportation of shipping emissions from the Eastern China sea which then accumulated in the Shanghai urban atmosphere. In contrast to the period before precipitation events, the average wind speed after precipitation events was the highest (3.1±2.1 m s-1) with northwesterly and northerly winds from mainland China which can dilute shipping-related trace elements to the lowest levels (Fig. 10b). In brief, the mass concentration of shipping-related trace elements in the Shanghai urban atmosphere is more likely to be influenced by winds instead of precipitation.