Quantitative Source Apportionment of Size-segregated Particulate Matter at Urbanized Local Site in Korea

The suspended particulate matter (PM) has been collected by a cascade impactor for 2 years (September 2005– September 2007) in Kyung Hee University-Global Campus located on the border of Yongin and Suwon Cites in Korea. PM was separated into 9 fractions with following aerodynamic size ranges: ST1 (> 9 μm), ST2 (5.8–9.0 μm), ST3 (4.7–5.8 μm), ST4 (3.3–4.7 μm), ST5 (2.1–3.3 μm), ST6 (1.1–2.1 μm), ST7 (0.7–1.1 μm), ST8 (0.4–0.7 μm), ST9 (< 0.4 μm). The 20 chemical species (Al, Mn, Si, Fe, Cu, Pb, Cr, Ni, V, Cd, Ba, Zn, Na, NH4, K, Mg, Ca, Cl, NO3, SO4) were analyzed by ion coupled plasma atomic emission spectrometry and ion chromatography after proper pretreatments of each sample filter. Based on the chemical information, positive matrix factorization (PMF) was used to identify size-segregated PM sources except for ST9. A total of 11 sources were identified and their contributions were intensively estimated. Further conditional probability function (CPF) was used to examine the potential location of identified sources after PMF modeling, A result of 2-year average source contribution showed that aged sea salt, road dust, long-range transport, and soil sources were most dominant in ST1 (PM > 9.0 μm); mixed automobiles and coal combustion sources in ST5 (2.1 μm < PM < 3.3 μm); oil combustion, secondary aerosol, and incineration in ST8 (0.4 μm < PM < 0.7 μm); and biomass burning source in ST7 (0.7 μm < PM < 1.1 μm), respectively. The relative contribution of most abundant sources was 32.4% of longrange transport source in the coarse particle mode and 34.5% of secondary aerosol in the fine particle mode, respectively. It seems that the size-resolved analysis by PMF provides useful information on controlling local/regional emission sources and on acquiring scientific knowledge for size-resolved aerosol compositions emitted from specific sources.


INTRODUCTION
Rapid growth of industrial activities and urban population is a main issue causing serious air pollution problems in Korea.High level of particulate matter (PM) including fine particles affects to reduce visual range, to increase the rate of mortality rate and the number of hospitalization, and especially to increase the perception of air pollution.The PM holds a variety of physical characteristics such as size, shape, density, hygroscopicity, and so on, and chemical characteristics such as contents of heavy metal, watersoluble ingredients, and so on (Hidy, 1972).Among many parameters, the PM size must be closely related to respiratory deposition and visibility.Thus it is considered to be a key parameter in terms of air pollution control and human health/wealth (Appel et al., 1985;John et al., 1990).
There are numerous aerosol sources including natural and anthropogenic sources in local and regional areas.Various statistical methods like receptor models were playing an important role in identifying those sources.Main purposes for those models are to calculate quantitatively the source contribution at the receptor site and finally to suggest reasonable air pollution control strategies and management policies.However, despite a first study on receptor modeling was introduced about two decades ago (Kim et al., 1990) and many studies have been performed in Korea (Kim andLee 1993, Hwang et al., 2001;Hwang et al., 2002;Hwang and Kim 2003;Han et al., 2006a, b;Shin et al., 2006;Kang et al., 2008), the local/central government still consider automobile source as a unique and predominant contributor to be controlled and handled.It is because of ignoring scientific basis on the emissions from various sources.Up to the present time, they have estimated only source-oriented PM emissions by summing up all the fuel combustion activities multiplied by simple emission factors and thus failed to notice the emissions from the other man-made sources, natural sources, and secondary sources.Due to the inflation of the automobile source contribution for many years in Korea, the government has established blind policies by spending huge budgets to inefficiently control the single source.
Receptor modeling starts from sampling PM, analyzing physicochemical species in the PM, and ends with applying various models to apportioning PM mass.Two receptor models, CMB (chemical mass balance model) and PMF (positive matrix factorization model), have been widely used in many countries.To apply the former model, one needs a priori information on existing emission sources for the study area and further the corresponding source profile obtained by either direct sampling or by literatures.In the urbanized local site like our study area, since there are no source profiles properly designed for receptor modeling, we have applied PMF instead of CMB to reasonably identify local/regional sources.
The purpose of this study is to estimate source contribution to size-resolved PM mass in an urban mixed site in Korea.PM was collected by a cascade impactor and 20 chemical species were analyzed by ICP-AES and IC.In our present study, PMF and CPF (conditional probability function) were intensively used to classify PM sources by checking their direction.Similar previous studies using TTFA (target transformation factor analysis) and PMF had been performed to obtain contributions to TSP, PM 10 , PM 2.5 at this sampling site (Kim et al., 1993;Hwang, et al., 2002;Hwang, et al., 2003).However, those apportionment studies were performed by using insufficient numbers of chemical species or applied without size-segregation scheme.

Study Area and Sampling Site
The airborne PM sampling was conducted on the roof of Engineering Building, Kyung Hee University-Global Campus located on the border of two big cities, Suwon City and Yongin City.The two cities are located on the central part of Gyeonggi Province, in the center of the Korean Peninsular.Populations of Yongin and Suwon were 0.86 million and 1.09 million on February 2010, respectively (Suwon City, 2010;Yongin City, 2010).The sampling site (37°14 N, 127°04 E, 20 m above street level) was carefully determined to collect PM samples large enough to minimize the external effect of surrounding barriers or physical environment.A map of Korea showing the location of Suwon and Yongin cities with sampling site is given in Fig. 1.
The two cities are urbanized and mixed with industrial and agricultural areas.There were 181 industrial point sources at the end of 2007 including the Samsung Semiconductor and Samsung Electronics Company, located a few km south from the sampling site (Suwon, 2008).There have been paddy fields inside campus and around the Lake Shingal (52.3 km 2 ) located on the east.Besides them, there have been distinct local emission sources such as line sources (i.e. 2 major express highways and many local paved roads) and area sources (i.e.fuel combustion for cooking and heating from residential area and fugitive emissions from broad construction areas and from illegal open-burning activities).It was also reported that there were various types of sources affecting the local air quality around the study area (Hwang, 2003).
PM of each stage was collected on 80 mm membrane filter (Gelman Science Co. Model GN-6, USA) and PM of back-up stage was collected on glass-fiber filter (Adventec Co. Model GB 100R, Japan).The filters right after sampling were put in polyethylene plastic bags and preserved in a refrigerator.All the filters were weighed before and after sampling with an analytical balance (A&D Co., Model HM-202, reading precision of 10 g) after conditioning at constant temperature and humidity.

PM Extraction and Analytical Methods
A microwave pre-treatment method under Clean Water Act issued by USEPA and a HNO 3 -HCI pre-treatment method with Questron's Model Q-15 MicroPrep were used to analyze inorganic elements in PM.To do this, each filter was cut to several pieces with fixed size by a stainless steel cutter, and one of pieces was soaked by HNO 3 (61%) and HCI (35%) solution and heated for 5 minutes.After extraction the solution was filtered by a filter paper (No. 5B, 110 mm, Advantec MFS Inc.), and diluted to 50 mL with deionized water.Each filtrate was analyzed to determine 12 inorganic elements (Al, Mn, V, Cr, Fe, Ni, Cu, Zn, Cd, Ba, Pb, and Si) by inductively coupled plasma atomic emission spectroscopy (ICP-AES, DRE ICP, Leeman Labs Inc., USA).Further a piece of the filter was extracted with ultra pure water and ionic components were ultrasonically extracted.After passing through microporous membrane filters (pore size, 0.45 m; diameter, 25 mm), each filtrate was used to analyze 5 cations (Na + , K + , NH 4 + , Mg 2+ , and Ca 2+ ) and 3 anions (NO 3 -, SO 4 2-, and Cl -) by ion chromatography (Dionex, Model DX-400), which consists of a separation column (Dionex Ionpac AS12A for anion and CS12 for cation) and a guard column (Dionex Ionpac AG 11 for anion and AG12A for cation).
Prior to use ICP-AES and IC, several standard solutions with similar level of sample concentration were injected to examine quality assurance.An accuracy check was performed by calculating a relative error (RE) resulting from standard solutions and a precision check was performed by obtaining relative standard deviation and a coefficient of variation (CV) based on 3 repetitive measurements.Table 2 shows RE and CV for each chemical species.REs for Si by ICP-AES and for NH 4 + by IC were 17.6% and 9.6% in terms of measurement accuracy, respectively.Similarly, CVs for Si by ICP-AES and for NO 3 -by IC were 11.1% and 12.5% in terms of measurement precision, respectively.When analyzing Si, ICP-AES provided relatively low accuracy and precision.Furthermore Zn and Al appeared uniquely at ST9 on which glass-fiber filter was using.It must be because the detection limits of ICP-AES using glass-fiber filter were extremely high for Ba, Zn, and Al.Thus in our PMF modeling, we decided to use the chemical data only obtained at ST1 to ST8 with excluding ST9.

DATA ANALYSIS
The number of sources (or number of factors) statistically determined is highly dependent on number of aerosol samples, number of chemical variables, temporal and spatial variability, various local and regional environments, and study scopes.The number of sources determined by CMB model (used when source profiles are provided or when the number of dominant sources are known) is generally larger number than those by TTFA or PMF model (used when a priori information on source inventory is not provided).However, PMF model is being more practically applied than CMB model since source apportionment results are extremely sensitive to source profiles used or created (Kim et al., 2006;Lee et al., 2007;Lee et al., 2008).

PMF
Factor analysis is used to look for key correlation among measured variables and mainly applied to identify emission sources by interpreting groups of chemical variables with strong correlation.However, when traditional factor analysis depending on covariance matrix is applied, researchers experience physical difficulties due to insufficient information, factor loading of negative value, uncertainty on factor rotation, etc (Hwang, 2001).Such limitations might provide possibilities to bring subjective results when determining a proper number of emission sources and thus they stimulated to develop more improved PMF methodology than traditional factor analysis (Paatero and Tapper, 1994).
PMF analysis always provides positive factor loadings and it depends not only on information of correlation matrix but on the algorithm of least square minimization, namely, on information of error estimation for each data.As all receptor models are keeping the rule of mass balance and mass conservation, so PMF is also expressed as the following mass balance Eq. ( 1).
where x ij is the ambient concentration of jth chemical species of ith PM sample, g ik is the source contribution of kth source in ith sample, f kj is the source profiles, and e ij is residual error.After the model applies a least squares fitting, Q value expressed in Eq. ( 2) is used to minimize differences between measured and estimated concentration in PMF modeling.The value is useful to determine the proper number of factors.
where ij is the uncertainty of ambient concentration of jth chemical species of ith PM sample.Uncertainty can be flexibly estimated by raw input concentration data based on sampling and analytical environment.This might be an advantage of PMF model compared to the existing factor analysis.

Input Data
A total of 32 sets of PM data were obtained from a 9-stage cascade impactor that had been consecutively operated for approximately 2 years.Backup stage (ST9) was excluded for PMF analysis.Two-year average concentration of each chemical species for each size range is shown in Table 3.
Major inorganic species for all size ranges were Fe, Al, Si, Zn, Ni, and Pb.Concentrations of Fe, Al, Si, Mn were much higher in the coarse fraction, while those of Ni, Cu, Cd, V, Pb, Cr, Zn were much higher in the fine fraction.The former must be mainly influenced by crustal sources and the latter must be mainly emitted by man-made sources.Fig. 2 shows Table 3.Average concentrations of PM mass ( g/m 3 ), inorganic elements (ng/m 3 ), and ionic components (ng/m 3 ) for each size range during the study period of September 2005 to September 2007.the variation of PM-2.1 and PM-9.0 concentration during the sampling period.The temporal patterns of PM-2.1 and PM-9.0 were quite similar each other.The levels were generally high in winter and low in summer.Some of high peaks appeared during the Asian dust storm episodes, which were in the plot with shadow areas.

Data Treatment and PMF Modeiling
A total of 8 input data sets, consisting of 20 chemical variables and 32 samples for 8 size ranges, were prepared for PMF.PMF modeling was separately performed for each size range.In our study, Q-mode analysis was applied to obtain a matrix of correlations among samples by arranging the samples in rows and chemical variables in columns on each data sheet.
When there are incomplete data sets containing data below detection limit or missing data, there are several methods to substitute by approximated values.In this study, values below detection limit were substituted by the half value of detection limit and missing values are substituted by the geometric average of relevant chemical species described in the previous reports (Polissar et al., 1998;Lee et al., 2002).As a matter of fact, PMF model requires a pair of data for each variable consisting of a measured value plus a corresponding uncertainty.It was reported that the uncertainty, u ij , could be calculated like the following Eq.(3) (Polissar et al., 1998).
where u ij is the uncertainty of jth chemical species of ith sample, x ij is the ambient concentration of jth chemical species of ith sample, h is the fractional error, and DL is the detection limit.The h can be presumed by linear relationship between the two values of measured concentration and analytical uncertainty (Kim et al., 2005).Thus, uncertainty for each variable was approximated by the above Eq.( 3).
In any factor analysis, one of important steps is to choose proper number of factors.The optimum number of factors can be determined by trial and error (Song et al., 2001).A method to find out the number is using Q value described in Eq. ( 2).This step minimizes the difference between true measured value and theoretical value.Then the theoretical Q values should be equal to the values subtracting all elements in a factor matrix from data matrix and further standardized residuals in residual matrix R should be within -2.0 and +2.0 with the probability of 0.8 (Hopke 2000;Ramadan et al., 2000;Polissar et al., 2001).In addition, the maximum individual column mean (called IM in PMF model) and standard deviation (called IS in PMF model) can be calculated from the matrix R to determine proper factors.Generally IM and IS remarkably decrease as Q decreases (Lee et al., 1999).In our study, IM and IS were examined for each relevant number of factors.Details were described in the previous paper (Oh et al., 2009).
As results given in Table 4, 4 to 6 factors were determined for each of 8 size-segregated data sets.After choosing the number of factors, modeling was conducted by repeatedly increasing 0.1 of Fpeak between -1.0 and 1.0.The rotational matrix (called Rotmat) explaining uncertainty of factor rotation is used as a criterion by determining degree of freedom.The largest element in Rotmat indicated maximum uncertainty of rotation at the corresponding Fpeak value (Lee et al., 1999).Generally optimum degree of rotational freedom can be determined with preserving a certain limit of Q values (Song et al., 2001;Kim et al., 2003;Han et al., 2006).Optimum modeling conditions for Fpeak and Q value used in our study are given in Table 4.

Conditional Probability Function
Conditional probability function (CPF) was used to examine the relationship of source contributions with wind direction described in Kim and Hopke (2004a).To estimate the potential location of identified sources after PMF modeling, the CPF described in Ashbaugh et al. (1985) and Kim et al. (2004b) was extensively applied.The technique estimates the probability that a given source contribution from a given wind direction exceeds a given threshold concentration.CPF can be mathematically expressed as the following Eq.( 4).where m is the number of times air parcels from the wind sector exceeded a given criterion and n is the total number of data points from the wind sector .Identified sources are then potentially located in the directions which have high CPF values (Kim and Hopke, 2004c, d).
As reported in the previous study (Guerra et al., 2006), wind direction surely contributes to temporal variation in air pollutants emitted from specific sources.In our study, we calculated probabilities that PM concentrations coming from 16 sectors of wind direction divided by 22.5° per sector exceeded average values.After exploring several levels of threshold criteria, we decided to use the upper 20th percentile for the fractional contribution from each source.Calm wind conditions with wind speed less than 1 m/sec were excluded in our analysis.

Size Segregated and Profile
As shown in Table 4, the number of factors at each size range were determined as follows; four factors at ST1, ST2, five factors at ST3, ST4, ST5, ST6, ST8, and six factors at ST7, respectively.According to various information such as mass fraction of chemical species, ratio among specific species, pattern of time series, and reference source profiles, each statistical factor by PMF can be restored as an emission source that is physically interpretable in the real world.
A total of 39 source profiles were initially prepared since each of 8 stages provided 4 to 6 sources.However, only 11 sources were finally determined in this study because many sources at one stage were repeatedly duplicated at other stages as shown in Table 5.For example, the oil combustion source designated as one of 11 sources appeared 6 times at ST3 to ST8 (0.4 m < PM< 5.8 m) and both long-ranged transport and soil sources appeared 5 times at ST1 to ST5 (PM > 2.1 m).The sources such as sea salt, road dust, soil, long-ranged transport, and NH 4 NO 3 related sources dominantly appeared from ST1 to ST5 where coarse particles larger than 2.1 m were collected.On the other hands, the sources such as coal combustion, incineration, and biomass burning sources appeared from ST5 to ST8 where fine particles smaller than 2.1 m were collected.In case of mixed automobile source, it appeared at ST3 to ST7 with the size range of 0.7 to 5.8 m.
The repetition of sources consecutively appearing at next size ranges provided useful information when classifying undefined sources into potential sources.Thus an obscure source at a certain stage if any could be easily determined by examining temporal patterns of contribution at neighbor stages.As a matter of fact, sources can be misclassified when construction of source inventory (or source library) is poor, when numbers of measured fingerprint variables are not enough to trace sources, or when many complex sources exist.Despite only 4 to 6 sources were obtained at each stage, the results of contribution produce affluent information to control a specific emission source in local area, to assess their impacts on health and wealth, and to study size-segregated aerosol compositions from specific sources.
Size-resolved source profiles and their corresponding temporal contributions are shown in Figs. 3 to 13.The time series distributions were plotted by about 2 weeks contributions to PM mass concentrations in each size-range.Conditional probabilities of potential source directions for 11 sources are plotted in Fig. 14.

Size Segregated Source Profile
As the first factor, aged sea salt source was identified at ST1 and ST2 (PM > 5.8 m).Fig. 3(a) shows the source profiles separately determined at ST1 and ST2 and Fig. 3(b) shows the time series of contribution at the corresponding stage.The temporal patterns on the figure show almost same shapes with different strengths.The sea salt is well known coarse particles naturally generated from the Yellow Sea (or called Hwanghae), which is lying between the Korean peninsula on the east and China on the west and north.It is about 700 km wide on an average from east to west.Since the sea derived its name from the color of the silt-laden water discharged from the major Chinese rivers (http://www.britannica.com/EBchecked/topic/652686/Yellow-Sea), some of crustal species are deservedly expected in this source.Marker species of natural sea salt source were reported as Na + , Cl -, SO 4 2-, Mg 2+ , K + , and Ca 2+ by (U.S. EPA, 1999).It was reported that Na + and Cl -had a fraction of more than 30 percent and Mg 2+ , K + , and Ca 2+ had a fraction of more than 1 percent in a sea salt source (Lewis and Schwartz, 2004).
The source in this study was characterized by Na + , Cl -, Mg 2+ , Ca 2+ and the concentrations of Mg 2+ and Ca 2+ were especially high as expected.It was also reported that some of NO 3 -existed in aged sea salt as a form of NaNO 3 in the coarse fraction (Jonson et al., 2000;Anlauf et al., 2006).According to a temporal pattern of contribution in Fig. 4(b), it was considered as the construction debris source nearby sampling site since the pattern of source strength agreed well with the remodeling periods of the Engineering Building.The CPF plot for the source in Fig. 14(b) demonstrated direction of the remodeling source located east from the monitoring site.The source contribution to total mass concentration was very small less than a single percent.
The third factor was identified as long-range transport source mostly laden the Asian dust (or called Yellow Sand or Hwangsa).The source appeared at ST1 thru ST5 (> 2.1 m) in the coarse fraction as shown in Table 6.In Fig. 5(a), the source was dominated by NO 3 -, SO 4 2-, NH 4 + , Fe, Si, Ca 2+ , Cl - in order.It seemed that various sources such as soil, secondary aerosol, and industrial sources were well mixed during long-range transport from the source origin (Mori et al., 2003;Takahashi et al., 2010).The 5 plots in Fig. 5(b) show temporal patterns of its contributions from ST1 to ST5.The strength of the contribution gradually decreases as particle size decreases.As the size decreases, many watersoluble ions appear and their concentrations increase based on the source profiles.There are distinguished peaks on April and May of 2006 in the temporal contribution plots.As a matter of factor, the Asian dust storms were reported on April and May 2006.Thus the source contribution was considerable during the springtime.According to the sizeresolved profiles in Fig. 5   direction of the regional source located southwest from the monitoring site, even though CPF analysis has a limit when identifying long-range transport sources.
The fourth soil and road dust source was characterized by Al, Fe, Si, Cl -, NO 3 -, SO 4 2-, Ca 2+ as shown in Fig. 6(a) The source appeared at ST1 to ST5 (> 2.1 m) in the coarse fraction.It was reported that Si had a fraction of more than 10 percent in a typical soil source for coarse particles (Chow, 1995).However, the fractions of both Al and Fe were slightly higher than that of Si on the average in this study.It seems that high analytical uncertainty of Si by ICP-AES may cause those differences.In Fig. 6(b), the patterns of temporal distribution at ST1, ST2, ST4 were somewhat different from those at ST3 and ST5.This source contribution to mass concentrations at ST3 and ST5 was much smaller than that at ST1, ST2, ST4 as presented in Table 6.The CPF plot in Fig. 14(d) pointed two distinct directions such as southwest and northeast.Huge soil digging works for building apartment complexes had been continued about 500 m southwest as well as about 2 km northeast from the sampling site.Two different source locations might be causing two different temporal distribution in Fig. 6(b).In addition, it seemed probable that some materials such Cl -and Ca 2+ as snow-melting agent sprayed in winter were resuspended into the air.Thus this source can be interpreted as the mixed source of typical soil and resuspended road dust in local area.
NH 4 NO 3 related source was identified as the fifth factor.This source appeared at ST3 and ST4 (3.3 m < PM < 5.8 m) in the coarse fraction.As shown in Fig. 7(a), NO 3 -, NH 4 + , Si, Fe, Na + , SO 4 2-were dominant in order.NO 3 -is generally produced in polluted urban areas from various combustion activities and it is also a well known compound found in the fine fraction.However NO 3 -in this source was dominant in the coarse fraction.In our previous study (Oh et al., 2009) and from Table 4, distinct bimodal distribution was observed for both NO 3 -and NH 4 + , with a small peak at ST3 in coarse fraction and a big peak ST7 in fine fraction.Likewise two peaks were also observed for Si, Fe, and Na + , but with a big peak at ST3 in coarse fraction and a small peak at ST7 in fine fraction.In the other study of frequency distribution in various particle sizes in southern California (John et al., 1990), three clear peaks were observed for NO 3 -, NH 4 + , and SO 4 2-with two peaks in fine fraction and one in coarse fraction.The presence of nitrate in larger particles characterized by crustal material can be due to the reaction of gas-phase nitric acid with mineral dust (Feng et al, 2007;Hwang et al, 2006).It seemed that this coarse NH 4 NO 3 related source was associated with emissions from agricultural activities near sampling site.There had been cultivating activities in campus paddy field located southeast and additional soil works converting huge paddy field into apartment complexes located about 500 m southwest from the site.These activities agreed well with source direction shown on a CPF plot in Fig. 14(e).
The species of SO 4 2-, NO 3 -, Na + , Ca 2+ , Fe, Si, Ni, V indicated various types of oil combustion as the sixth factor.This source was observed over many size ranges from ST3 to ST8 (4.7 m < PM < 0.4 m).In general, Ni and V are well-known marker elements of oil combustion in the fine fraction (Hopke, 1985;Chow, 1995;Song et al., 2001;Lee et al., 2002;Morawska and Zhang, 2002).Such markers are mainly emitted from diesel oil or residual oil combustion and the other markers like Na, Ca, NO 3 -, SO 4 2are also released by burning processes using various liquid oils (Hopke, 1985;Schroeder and Dockery, 1987;Chow, 1995).Also Fe is an affluent maker element in oil emissions.The 6 plots in Fig. 8 strength of the contribution is gradually increasing as particle size decreases.Even though there are many oil combustion sources near the study area, the overall contribution of the source has been decreased because many industries have quickly replaced energy supply systems using liquid fuels into electricity or gaseous fuels according to a government regulation of preventing high sulfur B-C oil in urban area.The CPF plot in Fig. 14(f) showed direction of this local source located north to northeast, where many scattered small-scale industries using various oil fuels are located.
As the seventh factor, mixed automobile source was determined by SO 4 2-, NO 3 -, Ca 2+ , Na + , Fe, Zn, Si at ST3 to ST7 (0.7 m < PM < 5.8 m).Unfortunately Pb as a marker element was not involved in this source.It was observed that Ca 2+ together with SO 4 2-and NO 3 -had high mass fraction on profiles shown in Fig. 9(a).It was reported that calcium was generally emitted by wear of brake linings, tires (associated with zinc emission), clutch plates and by asphalt road surface mainly in the coarse mode (Monte and Rossi, 2000;Kupiainen et al., 2005).However, it was also considerably emitted from various lubricating oils and diesel powered vehicles in the fine mode (Hopke, 1985).According to emissions from mixed motor vehicles measured in two tunnels in Milwaukee (Lough et al., 2005), PM 10 metal emissions were characterized by crustal elements Si, Fe, Ca, Na, etc. Similar results were obtained in a freeway tunnel study in Taiwan (Chiang and Huang, 2009), they reported that SO NO 3 -, Ca 2+ , Na, Fe (presumed to be emitted from mixed automobiles) had high concentrations in/outside the tunnel in fine and coarse fractions as well.In addition, air pollution control regulations by the Korean government had been allowed only a small amount of lead in gasoline since July 1987, unleaded gasoline was completely substituted for leaded gasoline after early 1993.Even though the ambient air quality standard for lead was strengthened from 1.5 g/m 3 /3-month to 0.5 g/m 3 /year in 2001, the annual average lead concentration showed to be satisfactory with a steadily descending trend in the range of 0.304 to 0.008 g/m 3 during the monitoring period of 1989 to 2007 in this study area (Lee et al., 1995;Kim et al., 1997).Thus, lead and bromine are no longer useful tracers for gasoline emission source in Korea like many other countries.According to Fig. 14(g), the source must be situated northwest from the sampling site.In fact, there have been a densely populated residential area shortly and downtown Suwon City distantly both located northwest from the sampling site.Further campus main gate has been located about 200 m northwest from the sampling site and a 8 lane broad traffic road has been running in front of the gate.
The eighth factor was designated as coal combustion source since NO 3 -, SO 4 2-, NH 4 + , Fe, Si, and other trace heavy metals were dominantly observed at ST5 to ST8 (0.4 m < PM < 3.3 m).In general fossil fuel combustion emission dominates submicron particles (Seinfeld and Pandis, 1998), but the coal combustion source in this study was partly observed in the coarse particles.It seemed to be aged and grown during transferring to the receptor.The crustal elements together with various anthropogenic species including SO 4 2-, NO 3 -, NH 4 + were known as markers of coal burning (Hopke, 1985;Chow, 1995;Wilson et al., 2002;Hwang, 2003).After the economic crisis of 1998 in Korea, coal consumption in Korea has been rapidly increased in spite of a regulation for solid fuel restriction in use issued by the Ministry of Environment in 1985.As a matter of fact, a consumption ratio of coal energy to the total energy showed a minimum of 19.3% in 1997 and increased continuously by 22.3% in 200022.3% in , 23.8% in 200322.3% in , 24.3% in 200622.3% in , and 27.4% in 200822.3% in , respectively (KEEI, 2009)).According to the plots in Fig. 10(b), the source strength was strong during the winter time.The CPF plot for the source in Fig. 14(h) demonstrated direction of this local source located north to northwest from the site similar to oil combustion source.It is noted that Seoul is located almost 30 km north and there had been many scattered small-scale industries using various fossil fuels between Seoul and Suwon.
Secondary aerosol source as the ninth factor was characterized by SO 4 2-, NO 3 -, NH 4 + at ST6 to ST8 (0.4 m < PM < 2.1 m).This source was considered to be combined with local and regional sources because it takes time to form particulate SO 4 2-, NO 3 -, NH 4 + from gaseous SO 2 , NO 2 , NH 3 .Once releasing to the ambient, the source exists in the dominant form of NH 4 NO 3 and (NH 4 ) x SO 4 (x = 0 to 2) by homogeneous or heterogeneous photochemical processes (Watson and Chow, 1994;Khoder, 2002;Wilson et al., 2002).Eatough et al. (2007)   dominant with average wind speed of 2.5 and 1.9 m/sec, respectively.Incineration source as the tenth factor was characterized by NO 3 -, SO 4 2-, Cl -, NH 4 + , Fe, Zn, Si at ST6 to ST8 (0.4 m < PM < 2.1 m) according to species abundances in aerosol mass described in the previous papers (Chow, 1995;U.S.EPA, 1999).The associated emissions were mainly considered from a huge municipal waste incinerator (treating 600 ton/day of waste) located about 2 km northwest from the sampling site.It seemed that it was also influenced by various burning facilities nearby small and mid-size industries.As the last eleventh factor, biomass burning source was identified by SO 4 2-, NO 3 -, NH 4 + , K + , Ca 2+ , Na + , Cl -at ST7 and ST8 (0.4 m < PM < 1.1 m).Similar chemical abundances were reported from agricultural biomass (sugar cane trash) burning in Brazil (Rocha et al., 2005) and wintertime biomass burning identified by PMF in Canada (Jeong et al., 2008).K + is also one of well-known markers for wood burning source (Hopke, 1985;Song et al., 2001).From Fig. 14(k), it is noted that various illegal burning activities have been conducted on rural areas located southwest and east from the site.
Consequently, a total of 11 sources were finally extracted even though only 4 to 6 sources were obtained at each stage.As mentioned above, ST9 (backup stage: PM < 0.4 m) was not involved for PMF.Thus it was difficult to estimate contribution to total aerosol mass including ST9.Nevertheless the average source contributions for two years assorted in specific size ranges could be calculated in Table 6 and average seasonal contributions in Table 7. Seasonal differences were seen in its mass contribution.Long-range transport source was dominated during the springtime by the Asian dust storms.On the whole, aged sea salt, road dust, long-range transport, and soil sources are most dominant in ST1 (PM > 9.0 m); mixed automobiles and coal combustion sources in ST5 (2.1 m < PM < 3.3 m); oil combustion, secondary aerosol, and incineration in ST8 (0.4 m < PM < 0.7 m); and biomass burning source in ST7 (0.7 m < PM <1.1 m), respectively.Major contributors at this site were long-range transport source from the southwest, secondary aerosol source from the southwest and south-southeast, and mixed automobile source from the northwest by contribution sequence.

CONCLUSIONS
A total of 32 sets of PM samples were collected for two years by a 9-stage cascade impactor at an urbanized local site in Korea.The 20 chemical species in PM were analyzed by ICP-AES and IC.Based on the chemical data (except data from backup stage), PMF was used to identify PM sources and CPF was used to examine the potential location of the identified sources.Major contributors at the sampling site were long-range transport source from the southwest, secondary aerosol source from the southwest and south-southeast, and mixed automobile source from the northwest.On the other hand the building debris source as a minor was identified near the site.The relative contribution of most abundant sources was 32.4% of longrange transport source in the coarse particle mode and 34.5% of secondary aerosol in the fine particle mode, respectively.
The source apportionment study based on size segregation was useful to control specific emission sources in local area, to assess their impacts on health and wealth, and further to examine size-resolved aerosol compositions from them.Above all benefits obtained from PMF study, the sizeresolved analysis provided most useful information when classifying statistically undefined sources into potentially defined sources by examining temporal dependence and distribution of contributions at nearby particle size ranges.
For further studies to understand more detailed local/regional emission sources, it is necessary to expand more chemical variables such as temperature-resolved carbon compounds and the other inorganic elements, to develop more marker species for specific sources, and to improve quality assurance and quality control.
Fig. 1.Map showing two big neighbor cities in Korea with sampling site.

Fig. 2 .
Fig. 2. Temporal variation of PM-9.0 and PM-2.1 concentrations observed from September 2005 to September 2007 at the sampling site.
Fig. 14(a) shows a CPF plot for the aged sea salt source demonstrated well its location of the Yellow Sea.The second factor was characterized by SO 4 2-, NO 3 -, Ca 2+ , Mg 2+ , Na + , Cl -, Si, and it appeared at ST1 and ST2 (PM > 5.8 m) in the coarse fraction as shown in Fig 4(a).

Fig. 3 .
Fig. 3. (a) Aged sea salt source profiles obtained at ST1 and ST2, respectively.(b) Time series of corresponding contributions by PMF model.

-
Fig. 4. (a) Construction debris source profiles obtained at ST1 and ST2, respectively.(b) Time series of corresponding contributions by PMF model.

Fig. 5 .
Fig. 5. (a) Long range transport source profiles obtained at ST1 to ST5, respectively.(b) Time series of corresponding contributions by PMF model.

Fig. 7 .
Fig. 6.(a) Soil and road dust source profiles obtained at ST1 to ST5, respectively.(b) Time series of corresponding contributions by PMF model.

Fig. 10 .
Fig. 10.(a) Coal combustion source profiles obtained at ST5 to ST8, respectively.(b) Time series of corresponding contributions by PMF model.

Fig. 14 .
Fig. 14.The CPF plots of the contribution at sampling site.

Table 2 .
Analytical uncertainties of ICP-AES and IC in this study.

Table 4 .
Optimum PMF modeling conditions for each stage.

Table 5 .
Sources identified in each size range.

Table 7 .
A comparison of seasonal source contributions to the aerosol mass with particle size larger than 0 0. .4 4 m.