Source Apportionment of Submicron Particle Size Distribution and PM2.5 Composition during an Asian Dust Storm Period in Two Urban Atmospheres

Asian dust storms (ADS), coming from deserts of China and Mongolia, have serious environmental impact on particulate matter (PM) and other pollutants in Taiwan. This study selected two urban sites, Taipei and Kaohsiung, to evaluate the influence of ADS on air quality. During the ADS periods, the hourly PM10 mass concentrations were 800 μg m in Taipei and the 400 μg m in Kaohsiung, which was three to five times higher than PM episodes during the non-ADS periods. By using the principal component analysis (PCA) manner, the potential sources, the dust storm contained, can be successfully identified during ADS periods. The other potential sources can be identified as vehicular emission and secondary organic aerosols from local area. There have been many studies conducted on the impact of ADS on airborne coarse particle concentration, but very few on fine particle concentration. This study focused on, using PCA for analysis and discussion, the impact of ADS on submicron particle size distribution. The results showed that there was no close relationship between the ADS and Aitken mode (30–100 nm or D30–100nm) or accumulation mode (from 0.1–1 μm). However, it was found that strong correlation existed between the ADS and nucleation mode (10–30 nm or D10–30nm). In addition, it was found that nucleation mode appeared first, followed by an air plume of dust particles twelve (12) hours later. The nucleation component from the PCA could be used as predictors of arrival time for an ADS. Taking into account the effects of meteorological conditions and employing technique of backward trajectories, PCA can be utilized as a powerful tool to better identify the source of dust storms and provide accurate results.


INTRODUCTION
Asian dust storms (ADS), which come from the deserts of China and Mongolia (Natsagdorj et al., 2003;Aoki et al., 2005), increase the airborne coarse (size distribution > 2.5 µm) particle concentration and bring anthropogenic pollutants into China, Japan, Korea, Taiwan and sometimes northern Pacific Ocean areas.(Guo et al., 2004;Ma et al., 2005;Liu et al., 2006;Han et al., 2008;Hwang et al., 2008).Dust storms also contain natural mineral particles, sulfates and nitrates (Talbot et al., 1997;Wang et al., 2007;Huang et al., 2013;Malaguti et al., 2015), black carbon (Ramana et al., 2010), hazardous heavy metals (Kim et al., 2003), and carbonaceous aerosols (Ramana et al., 2010;Stone et al., 2011) from biomass burning (Lin et al., 2012;Tsai et al., 2012).High particulate matter (PM) concentrations in Taiwan usually result from ADS, which occur several times annually, especially in the winter and spring (Zhou et al., 1996;Ma et al., 2001;Chen et al., 2004;Lee et al., 2006).During transport this dust can mix with primary particulate emissions from industries, transportation, power generation and secondary materials formed by the gas-to-particle conversion mechanism (Dentener et al., 1996;Song and Carmichael, 2001;Zhang and Iwasaka, 2001).The particle number concentration, especially in the sub-micrometer size range, also increases during the dust storm period, but decreases during times of precipitation (Park, 2006).There have been many studies conducted on the impact of ADS pertaining to increasing coarse particle mass concentrations, but fewer on the appearance of fine particles and ultrafine particles in the early stage of an ADS.Uematsu et al. (2002) studied a dust event in Japan and found that fine aerosol particles (0.3-0.5 µm and 0.5-1.0µm) appeared first, followed by an air plume of dust particles 12 hours later.
Dust effects on human health have recently become a major concern (Cook et al., 2011).Numerous studies have documented evidence of an association between concentrations of airborne particles and increased daily mortality (Dockery and Pope, 1994;Holgate et al., 1999).The sources of ultrafine particles (UFPs; particles size distribution < 100 nm) in the urban ambient atmosphere include vehicle emissions and coal-fired power plants (Dippel et al., 1999;Zhu et al., 2002;Morawska et al., 2008) and secondary particle nucleation (Brock et al., 2002;Holmes, 2007;Kulmala and Kerminen, 2008).Kasumba et al. (2009) found that traffic, industrial emissions and nucleation were the major sources of UFPs in urban areas.UFPs present the greatest health concern because they have high particle concentration numbers and a large surface area, with the potential to form adsorbed or condensed toxic air pollutants.When compared with larger particles, UFPs can be inhaled most deeply into the lungs (Miller et al., 1979).The most common way to present particle size distribution data for the urban aerosol is in terms of the three way, nucleation, Aitken, and accumulation modes.Each mode has different sources, size range, formation mechanisms and chemical compositions.Typically, UFPs account for about 90% of the total number concentration in urban areas.
Over the last decades, multivariate analysis has been widely used to identify the sources of ambient particles.The application of principal components analysis (PCA) manner to data sets produces statistically independent linear combinations of original variables, and also explains most of the total variation with reduced dimensions.This method has been successfully applied to identify potential sources of air pollutants and analyse relationships between distinct air pollutants in measured data (Horel, 1981;Yu et al., 2000;Shi et al., 2009;Abhishek et al., 2010;Wang et al., 2010;Kothai et al., 2011;Amodio et al., 2013;Huang et al., 2013;Liang et al., 2013).Some research studies have been conducted with the aim of predicting the ozone generation based on the meteorological conditions and the pollutants-precursor appearing at a particular point in time (Abdul-Wahab et al., 2002, 2005).The HYSPLIT model, established by the NOAA (National Oceanic and Atmospheric Administration) has been widely used in many scientific research applications and emergency scenarios that require modelling the transport and dispersion of hazardous air pollutants (Draxler and Rolph, 2003;Rasheed et al., 2015).This study uses HYSPLIT model to determine source locations and transport path of the ADS.By combining the HYSPLIT model and PCA manner, this study tries to identify potential sources from local areas or external areas, examine the relationship between the particle size distributions and investigate spatiotemporal features of air pollutants during ADS and non-ADS episodes.

SAMPLING LOCATIONS AND MEASUREMENTS
Data was collected from Taipei and Kaohsiung Supersites.

Taipei Supersite
As shown in Fig. 1 the area of our study, Taipei Supersite, is located in New Taipei City -a 2,052 km 2 basin with a population of 3.96 million.The city has 1,150 motorized vehicles per square km, with 930,000 cars and more than 2.36 million motorcycles.The Taipei Supersite is located between two major traffic hubs in the Hsinchuang area: (1) the intersection (north) of the Chungshan Expressway and Highway 64 and (2) the intersection (south) of the Highway One and Chungzheng Road.Both locations were determined to be the potential major potential sources of traffic pollution.The Linkou petrochemical park lies 9 km northwest of the station, which is believed to be the potential source of industrial pollution.According to the Taiwan Emission Database 8.1 (Taiwan EPA, 2010), the 2010 PM 2.5 emissions were 8,560 metric tons year -1 , of which 21%, 28% and 7% came from vehicles, home cooking and factories, respectively.Table 1

Kaohsiung Supersite
As shown in Fig. 1 the area of our study, Kaohsiung Supersite, is located in Kaohsiung City -a 2947 km 2 area with a population of 2.77 million.The city has 775 motorized vehicles per square km, with 825,000 cars and more than 2.28 million motorcycles.Mobile and stationary sources are important in the Kaohsiung city and Pingtung County area (the so-called Kao-Ping air basin).According to the Taiwan Emission Database 8.1 (Taiwan EPA, 2010), the 2010 PM 2.5 emissions were 5,708 metric tons/year, of which 13% come from vehicles and 39% from factories.Comparing to Taipei urban area the potential pollutant are more complicated.Vehicle emissions are concentrated in Kaohsiung center area.Stationary sources are located primarily in three industrial parks: Linhai, Renda and Linyuan (Fig. 1).Linhai, the largest industrial park, is located south of Kaohsiung city.Linhai has a coal-fired power plant (2,700 MW), steel mills, and petroleum refineries.The Kaohsiung Supersite, located at Fooyin University (22.603°N, 120.388°E), is approximately at the center of the industrial activities in the Kao-Ping air basin.The station is 7 km from Kaohsiung city center to the west, 8 km from the Linhai industrial park to the southwest, 15 km from the Renda industrial park to the north and 13 km from the Linyuan industrial park to the south.Liang et al., Aerosol and Air Quality Research, 15: 2609-2624, 20152612 (9) meteorological variables (i.e., wind speed, wind direction, temperature and relative humidity).

Data Analysis
This study defined the period of April 24-26, 2009 as the ADS episode and the periods of April 2-4 and 22-24, 2009 as the non-ADS episodes.The non-ADS episodes were defined as when PM 10 concentrations did not exceed the annual average.The periods of February 10-12 and April 6-8, 2009 were defined as the PM episodes which were highest PM 10 concentration during non-ADS periods in this year.The dust storm episodes observed in 2009 were analyzed using the HYSPLIT model, producing two trajectories that were calculated for 60 hours duration.This paper focuses mainly on the ultrafine particle composition and size distribution in the nucleation mode, Aitken mode and accumulation mode in the ADS and non-ADS periods, whereby a non-rotational PCA manner was applied to the data set.The original numbers of variables for the sampling data were 16, including air pollutant concentration items, particle number concentrations and size distributions ranging from 14 to 500 nm.Hourly values at individual stations were normalized as Eq. ( 1) where Z ik denotes the normalized value of the k th observation at i th pollutant, C ik represents the k th values at the i th pollutant, µ i is the mean value at the i th pollutant, and S i denotes the standard deviation at the i th pollutant.
The non-rotational principal component model is Eq. ( 2) where n denotes the number of air pollutant, L ij is the factor load for the i th pollutant on the j th non-rotational component, and P jk represents the component score for the kth value for the j th non-rotational component.
Four unrotated factors at the supersite were isolated based on the following criteria.First, the number of factors was selected such that the cumulative percentage variance explained by all chosen factors more than 90%.Secondly, the factors with Eigenvalues greater than one may obtain the final factor matrix.High factor loads (> 0.70) for ultrafine particle composition and size distribution in a principal component can help identify potential sources.

RESULTS AND DISCUSSION
Table 3 summarizes the non-rotation results for the four principal components together with the amount of variance explained by each component for the ADS periods and non-ADS periods.In practice, only factor loads with absolute values greater than 0.5 are selected for the principal component interpretation (Jolliffe, 1986).In this study, factor loads greater than 0.7 show a strong positive linear relationship.The positive correlations between 0.4 and 0.7 indicate a moderately positive linear relationship.The positive correlations between 0 and 0.4 indicate a weak positive linear relationship.In this study, atmospheric sub-micrometer particles were comprised of three modes, nucleation mode, Aitken mode and accumulation mode.After analyzing the particle size distribution the results show that the three modes are similar to those found in most urban areas (Whitby et al., 1978;Morawska et al., 1999;Wehner et al., 2002).This will be elaborated in the later section as well as in Fig. 4. The representative size distribution that had the strongest positive relationship was collected.

Asian Dust Storm Periods
Figs. 2 and 3 show the 60 hr backward trajectories at the Taipei Supersite and Kaohsiung Supersite from the time when the maximum PM 10 concentrations took place.When dust storms arrived, the PM 10 concentration exceeded 800 µg m -3 at 11:00 on April 25, 2010 at the Taipei Supersite and exceeded 400 µg m -3 at 02:00 on April 26, 2010 at the Kaohsiung Supersite.To further trace the features of air pollutants, the non-rotational PCA manner was employed to identify the principal components.
At Taipei Supersite, the three PCs were vehicle emissions (VE, 69% of total variance), nucleation mode (D 10-30 nm, 15% of total variance) and coarse particle from Asian dust storms (ADS, 6% of total variance) during the ADS periods.But the major quantitative source contributions are coarse particle from ADS.Note that the three principal components accounted for 90% of the total variance.The first component (PC1), with 69% of the total variance, was vehicular emissions with a high factor load (> 0.7) in PC1 for PAH, NH 3 , NO x , OC, EC, and NO 3 -.In Fig. 4(a), the particle size distribution in the Aitken and accumulation modes is strongly related (factor load > 0.95) to PC1.The Aitken and accumulation modes contain primary particles from transportation, combustion sources and secondary aerosol materials (sulfate, nitrate, ammonium, secondary organics) formed by chemical reactions resulting in gas-toparticle conversion.Fig. 6 displays the time series for the three PC scores.According to Figs. 6(g)-6(i), PAHs, OC, EC concentrations and the number concentration for D 168 nm peaked very close to one another on April 24, as did the FAC1 score.It can be certified that the main PC1 is vehicular emissions.
The second component (PC2), with 15% of the total variance, shows a higher load for the nucleation mode (D 10-30 nm) in Fig. 4(a).ADS brought some of the ultrafine particles along with the coarse particles but this time it was not obvious.
During ADS periods, the nucleation mode is sometimes transported downwind with dust particles.Mixing of anthropogenic pollution and dusty air parcels over downwind cities has been observed in Taiwan (Tsai et al., 2014).In addition, the specific weather conditions also play an important role in the nucleation mode.Some studies in urban environments claimed that nucleation mode is formed through nucleation in the atmosphere after rapid cooling and dilution of emissions (Charron and Harrison, 2003;  Kittelson et al., 2006;Morawska et al., 2008).The number concentration for the nucleation mode is much less than that in non-ADS periods because of rain and/or not being detected.As found in previous research, the aerosol diameter distribution increased and the number concentration decreased in atmospheres with high relative humidity in Central Taiwan (Tsai and Cheng, 1999).The third component (PC3), with 6% of the total variance, shows a higher load for PM 10 , PM 2.5 , SO 4 2-.Comparing the time series plot in Figs.6(d)-6(f), the FAC3 accurately reflects the PM 10 , PM 2.5 , SO 4 2-during the ADS periods.The third potential sources could be identified from ADS.This is coincided with the findings by some researchers indicating that high mineral element and pollutant concentrations such as sulfate were produced by industries in China (Talbot et al., 1997;Park et al., 2013).
During the ADS periods, four potential source categories are identified at the Kaohsiung Supersite and they are mainly from vehicle emissions (VE, 47% of total variance), nucleation mode from Asian dust storms (D 10-30 nm, 22% of total variance), secondary organic aerosol (SOA, 13% of total variance) and coarse particle from Asian dust storms (ADS, 6% of total variance).Note that the four principal components accounted for 88% of the total variation.The first component (PC1), with 47% of the total variance, was vehicle emissions with a high factor load (> 0.7) in PC1 for PM 2.5 , SO 4 2-, NO 3 -, OC, EC, NO x , and CO.The particle size distribution in the Aitken and accumulation modes is strongly related (factor load > 0.86) to PC1 in Fig. 5(a).In addition, the time series in Figs.7(e)-  A Liang et al., Aerosol and Air Quality Research, 15: 2609-2624, 20152615 Asian 7(g), the FAC1, OC and EC concentrations present similar peaks.The OC and EC were the source makers for vehicle exhausts (Yuan et al., 2006).The diesel dominant factor contributed the most to EC and approximately one-third to OC vehicle .The gasoline dominant factor contributed the least to EC but the most to OC vehicle (Huang et al., 2014).Thus, PC3 can be identified from vehicle emissions.
The second component (PC2), with 22% of the total variance, shows a higher load for the nucleation mode in Fig. 5(a).According to the time series in Figs.7(c), 7(e) and 7(g), ADS brought some of the ultrafine particles along with the coarse particles and mixed them with local motor vehicle emissions.The number concentration of the nucleation mode arrived Southern Taiwan 12 hours ahead of ADS, as shown in Fig. 7.This finding can help to predict the arrival time for a forthcoming ADS.In conformity with Tsai et al. (2014), it was found that the descent of dusty air from the free troposphere lagged the arrival of polluted air by 7 hours, and was partially mixed with polluted aerosol when the transport decelerated over Taiwan.
The third component (PC3), with 13% of the total variance, shows a higher load for O 3 (factor loan = 0.68).Local photochemical production of O 3 due to large emissions of precursors plays an important role in the O 3 distribution.According to Chou et al. (2006), the O 3 concentration over Taiwan has increased consistently over the last 13 years.In Kaohsiung City, the potential sources were more complicated than Taipei area because of the heavy industries including petroleum refinery emissions.The potential sources could be the secondary aerosols from vehicle, factory, homecooking, power plant and petroleum refinery emissions.
The fourth component (PC4), with 6% of the total variance, obtains a higher load for PM 10 , PM 2.5 , and a moderate load for SO 4 2-.Figs.7(c)-7(e) show the PM 10 , PM 2.5 , SO 4 2concentrations and FAC4 scores peaked almost in the same time frame on April 26.Therefore, this factor (PC4) can be considered as the potential sources of ADS.

PM Episodes
At Taipei Supersite during PM episodes, 92% of the variance is represented by three PCs, which account for 64%, 24%, and 4% of the total variance.The first component.(c), 8(f), and 8(g) also show that the time series for the PAH, OC, EC concentrations peak almost in the same time frame.Previous researchers have found that the aerosol number distribution may be characterized by three modes at diameters 20 nm, 100 nm and 2 µm.These particle size ranges are strongly related to vehicle exhaust emissions (Schauer et al., 1996) and have been observed on and near the roadway (Whitby et al., 1975;Joumard and Perrin, 1988;Harrison et al., 1999;Kittelson et al., 2000).Hence, PC1 could be identified from vehicle emissions.
The second component (PC2), with 24% of the total variance, shows higher loads for nucleation mode (D 10-30 nm) in Fig. 4(b).Most fresh particles emitted by motor vehicle exhausts are ultrafine particles (< 100 nm), and tend to exhibit a size distribution with a nucleation mode (< 20 nm) (Palmgren et al., 2003;Pey et al., 2009), indicating that primarily gasoline vehicle emissions could be the potential source for the second component.
The third component (PC3), with 4% of the total variance, obtains higher load for PM 10 , SO 4 2-, NO 3 -.As shown in Fig. 8(b), the PM 10 and PM 2.5 concentration increase, when the wind speed decreases, which indicate that PC3 is coming from road dust.
At Kaohsiung Supersite during PM episode periods, 84% of the variance is explained by three factors, which account for 42%, 35%, and 7% of the total variance.The first component (PC1) , with 42% of the total variance, shows higher loads for PM 10 , PM 2.5 , SO 4 2-, OC, EC, NH 3 , NOx, CO (factor load > 0.7).Fig. 5(b) presents higher loads for the Aitken, and accumulation modes.Figs.9(c)-9(f) shows the concentration of PM 10 , PM 2.5 , OC, EC and SO 4 2-peaks together.This indicates that some potential sources are vehicle emission and road dust.
The second component (PC2), with 35% of the total variance, shows higher loads for nucleation mode.This indicates that primarily gasoline vehicle emissions could be the potential source for the second component.The third component (PC3), with 7% of the total variance, shows a   higher load for O 3 (factor load = 0.72).The component can be indicated from road dust, secondary organic aerosol.

Non-ADS Periods
At Taipei Supersite during non-ADS periods, 86% of the variance is represented by two factors, which account for 46%, and 40% of the total variance.The first component (PC1), with 46% of the total variance, shows higher load for NO 3 -, PAHs, NH 3 , NOx, BC (factor load > 0.7).In addition, the first component presents higher loads for nucleation, Aitken, and accumulation modes.Therefore, the PC1 could be identified as vehicle emissions.The second component (PC2), with 40% of the total variance, shows higher loads for PM 10 , PM 2.5 , NO 3 -.This indicates that road dust is a potential source.
At Kaohsiung Supersite during non-ADS periods, 83% of the variance is expressed by three factors, which account for 47%, 23%, and 13% of the total variance.The first component (PC1), with 47% of the total variance, shows higher loads for PM 10 , PM 2.5 , NO 3 -, OE, EC, NO, NO x , CO (factor load > 0.7).Also, PC1 presents higher loads for the Aitken and accumulation modes.Therefore, the PC1 could be identified as vehicle emissions and road dust.The second component (PC2), with 23% of the total variance, shows a higher load for the nucleation mode.This indicates that fresh emissions from primarily gasoline vehicle are potential sources.The third component (PC3), with 13% of the total variance, shows a higher load for O 3 (factor load = 0.74).The potential sources could be the secondary organic aerosols from local pollutants such as vehicle, factories, power plant and petroleum refinery emissions.

CONCLUSIONS
The results revealed that the mass concentration in ADS was roughly three to five times higher than that of the highest concentration measured in Taiwan during the non-ADS seasons.Combining PM 2.5 chemical composition and size distribution from 14 to 500 nm data sets have produced a better identification of potential pollutant sources during ADS periods and non-ADS periods.The size distribution of nucleation mode, Aitken mode, and accumulation mode can be clearly separated by PCA manner.According to these three modes, it can provide more information about the potential sources.
During the ADS periods, HYSPLIT model shows the pollution was transported from Mongolia through China's industrial coastline before coming down to Taiwan.The potential source in Taipei and Kaohsiung areas are local vehicular emissions, nucleation mode from ADS and urban emission, coarse and fine particulates of dust storms.The result shows that the local vehicular emissions are significant pollution in urban area, but the major source contributions are coarse particle of the ADS.In additions, it was found that a strong relationship existed between the Aitken mode, accumulation mode and vehicular emissions.Few studies focus on the submicron particles during ADS periods, especially nucleation mode.In this study, it was found that nucleation appeared first, followed by an air plume of dust particles 12 hours later.Combining with meteorological parameters, emitted sources, and concentrations of air pollutants, PCA could be an effective approach to identify possible sources.During PM periods, the potential sources in Taipei are vehicular emissions, primarily gasoline vehicle, and road dust.In Kaohsiung, the potential sources are vehicular emissions, primarily gasoline vehicle and secondary aerosol and road dust.During non-ADS periods, the main potential sources are vehicular emissions, and road dust in Taipei.In Kaohsiung, the potential sources are vehicular emissions and road dust, primarily gasoline vehicle, secondary organic aerosol.The results show that the potential pollutants in the urban area are vehicular emissions and road dust.The above results show that the potential pollutants in the urban area are vehicular emissions and road dust during PM and non-ADS periods.
As previously mentioned, the PM 10 concentration was in the neighborhood of 800 µg m -3 in Taipei and 400 µg m -3 in Kaohsiung during the same ADS.The lower particulate loadings in Kaohsiung could be attributed to the facts of 1) some particulates being washed out by rain in Taipei area during the ADS periods and 2) some particulates dropping out during the process of traveling from Taipei to Kaohsiung.By integrating the PCA manner, pollutants concentration, and particle size distribution is expected to provide more complete and significant physical interpretation for the impact of ADS.It is the cherished expectation that more studies and case analysis be done in the future to provide more information and accurate predicting model for the incoming ADS.
The sites were established by the Taiwan Environmental Protection Administration in 2000 and began operations in November of 2001 in Taipei and April of 2005 in Kaohsiung.Fig. 1 displays the sampling site locations.

Fig. 1 .
Fig. 1.Location of supersite in New Taipei city and Kaohsiung city.

Fig. 4 .
Fig. 4. Correlations of three principal component (PCs) with size distributions of particles at Taipei Supersite.

Fig. 5 .
Fig. 5. Correlations of three principal component (PCs) with size distributions of particles at Kaohsiung Supersite

Table 1 .
Summary of data used at Taipei Supersite.

Table 2 .
Summary of data used at Kaohsiung Supersite.

Table 3 .
Factor loads between the principal components and measured concentrations during ADS and non-ADS periods.