Environ. Sci. Technol. 2005, 39, 4953-4960
Sources of Fine Particles in a Rural Midwestern U.S. Area E U G E N E K I M , † P H I L I P K . H O P K E , * ,‡ DONNA M. KENSKI,§ AND MICHAEL KOERBER§ Departments of Civil and Environmental Engineering and Chemical Engineering, Clarkson University, Box 5708, Potsdam, New York 13699, and Lake Michigan Air Directors Consortium, Des Plaines, Illinois 60018
Ambient PM2.5 (particulate matter e 2.5 µm in aerodynamic diameter) samples collected at a rural monitoring site in Bondville, IL on every third day using Interagency Monitoring of Protected Visual Environments (IMPROVE) sampler were analyzed through the application of the positive matrix factorization (PMF). The particulate carbon fractions were obtained from the thermal optical reflectance method that divides particulate carbon into four organic carbon, pyrolyzed organic carbon (OP), and three elemental carbon fractions. A total of 257 samples collected between March 2001 and May 2003 analyzed for 35 species were used and eight sources were identified: summer-high secondary sulfate aerosol (40%), secondary nitrate aerosol (32%), gasoline vehicle (9%), OP-high secondary sulfate aerosol (7%), selenium-high secondary sulfate aerosol (4%), airborne soil (4%), aged sea salt (2%), and diesel emissions (2%). The compositional profiles for gasoline vehicle and diesel emissions are similar to those estimated in other U.S. areas. Backward trajectories indicate that the highly elevated airborne soil impacts were likely caused by Asian and Saharan dust storms. Potential source contribution function analyses show the potential source areas and pathways of secondary sulfate aerosols, especially the regional influences of the biogenic as well as anthropogenic secondary aerosol.
Introduction Advanced source apportionment studies for ambient PM2.5 (particulate matter e 2.5 µm in aerodynamic diameter) are needed to assist in state implementation plan development, regional haze rule planning, and source-specific community epidemiology. Positive matrix factorization (PMF) (1) has been used successfully to assess contributions from ambient PM2.5 sources in many studies (2-4). However, PMF could not always resolve the carbonaceous particle sources in the previous analyses, especially traffic related combustion sources. These sources were extracted as a mixed carbonrich source combined with others because they had similar chemical profiles and similar temporal emission patterns. In recent studies, PMF was applied to ambient PM2.5 compositional data sets of daily integrated samples including * Corresponding author fax: (315)268-4410; e-mail: hopkepk@ clarkson.edu. † Department of Civil and Environmental Engineering, Clarkson University. ‡ Department of Chemical Engineering, Clarkson University. § Lake Michigan Air Directors Consortium. 10.1021/es0490774 CCC: $30.25 Published on Web 05/24/2005
2005 American Chemical Society
eight individual carbon fractions collected at four monitoring sites across the U.S.: Atlanta, GA (5), Washington, DC (6), Brigantine, NJ (7), and Seattle, WA (8). PM2.5 carbon was analyzed using the Interagency Monitoring of Protected Visual Environments/Thermal Optical Reflectance (IMPROVE/ TOR) method that divides carbon into four organic carbon (OC), pyrolyzed organic carbon (OP), and three elemental carbon (EC) fractions (9, 10). In these studies, gasoline vehicle emissions could be distinguished from diesel emissions based on the differences in the abundances of the carbon fractions between the two sources. Also, the compositional profiles for these two source types show similarities among the four sites. Temperature resolved carbon fractions also enhanced separations of carbon-rich secondary sulfate aerosols indicating that the temperature resolved carbon fractions can be utilized to enhance source apportionment of ambient PM2.5. The objectives of this study are to examine the use of carbon fractions to identify PM2.5 sources in Midwestern U.S. area and estimate their contributions to the PM2.5 concentrations. In the present study, PMF was applied to an ambient PM2.5 compositional data set of 24-h integrated samples including eight individual carbon fractions collected at the IMPROVE monitoring site in Bondville, IL. The PMF derived PM2.5 sources and their seasonal trends are discussed. The resolved fractional carbon profiles are compared with measured profiles from chassis dynamometer tests. Also, the similarities among PMF resolved fractional carbon profiles as well as weekday/weekend variations in five monitoring sites are discussed. The potential source contribution functions were calculated to help identify the likely source locations and pathways of the PMF identified secondary sulfate aerosols.
Experimental Methods Sample Collection and Chemical Analysis. The PM2.5 samples were collected on every third day at the rural IMPROVE (11) monitoring site located in Bondville, IL (latitude: 40.051, longitude: -88.372). As shown in Figure 1, this monitoring site is located 8 km southwest of Champaign, IL, 130 km east of Springfield, IL, and 240 km south of Chicago. Highways are situated to the north and east of the monitoring site. The prevailing winds at the monitoring site were from the south. PM2.5 samples were collected on Teflon, Nylon, and quartz filters. The detailed filter analyses are shown in Malm et al. (11) and Cahill et al. (12). The IMPROVE/TOR method (9) was used to analyze the quartz filter for eight PM2.5 carbon fractions. This method volatilizes OC by four temperature steps in a helium environment: OC1 at 120 °C, OC2 at 250 °C, OC3 at 450 °C, and OC4 at 550 °C. After the OC4 response returns to baseline or a constant value, OP is oxidized at 550 °C in a mixture of 2% oxygen and 98% helium environment until the reflectance returns to its original value. Then three EC fractions are measured in oxidizing environment: EC1 at 550 °C, EC2 at 700 °C, and EC3 at 850 °C. Hourly measured wind direction and speed were obtained from Springfield Capital Airport, IL. Samples for which PM2.5 concentrations were not available were excluded from this analysis (4.4%). Samples in which the PM2.5 concentration error flag was not ‘NM’ (normal) were also excluded (1.5%). A total of 94.5% of the original data was used for this study. SO42- was not included, and only S was used in this study because they showed good correlations (slope ) 3.5 ( 0.05, r2 ) 0.95). The reported EC1 concentration in IMPROVE/TOR in method includes the OP concentration. In this study, the OP was subtracted from EC1 and utilized as an independent variable. Thus, EC1 in VOL. 39, NO. 13, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
4953
FIGURE 1. Location of the IMPROVE monitoring site in Bondville, IL. this study did not include OP. A total of 257 samples collected between March 2001 and May 2003 and 36 species including PM2.5 were used in this study. Table 1 shows a summary of PM2.5 species used in this study. Multivariate Receptor Modeling. The receptor modeling problem can be expressed in terms of the contribution from independent sources to all chemical species in a sample (13, 14). PMF provides a matrix of source profiles and a matrix of time series of source contributions without prior knowledge of PM2.5 sources. The detailed receptor modeling and PMF equations are specified in refs 1 and 4. The application of PMF depends on the estimated uncertainties for each of the measured data. The uncertainty estimation provides a useful tool to decrease the weight of missing and below detection limit data in the solution. To assign measured data and the associated uncertainties as input data to the PMF, the procedure of Polissar et al. (15) was used. In addition to the standard uncertainty estimation, the uncertainty must take into account the measurement uncertainty as well as the temporal variability in the source profiles over the monitoring period. In several cases, to take the temporal variability into account, larger uncertainties were used to decrease the weight of some specific variables in the model fit (16). In this study, it was found necessary to increase the estimated uncertainties of S, NO3-, and Se by a factor of 4 to take the high seasonal variability into account. The estimated uncertainties of OC1 were increased by a factor of 2 to down-weight the influence of the known positive artifact from the adsorption of gaseous OC (17). Similarly, the estimated uncertainties of EC1 were increased by a factor of 3 to account for the additional uncertainty from the subtraction of OP. The initial data analysis produced a factor with high Al and Ca and a factor with high Si which is highly unlikely (18). To obtain sensible profiles, it was found necessary to increase the estimated uncertainties of Al and Si by a factor of 4 and Ca by a factor of 2. The measured PM2.5 mass concentration was included as an independent variable in the PMF modeling to directly obtain the mass apportionment without the usual multiple regression (6). The estimated uncertainties of the PM2.5 4954
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 39, NO. 13, 2005
concentrations were set at four times their values so that the large uncertainties decreased their weight in the model fit. When the measured PM2.5 concentration is used as a variable, the PMF apportions a PM2.5 contribution for each source according to its temporal variation. There are an infinite number of possible combinations of source contribution and profile matrices to the multivariate receptor modeling problem due to the free rotation of matrices (19). PMF uses non-negativity constraints on the factors to decrease rotational ambiguity. Also, the parameter FPEAK and the matrix FKEY are used to control the rotations (20, 21). By setting nonzero values of FPEAK, the routine is forced to add one source contribution vector to another and subtract the corresponding source profile factors from each other and thereby yield more physically realistic solutions. PMF was run with different FPEAK values to determine the range within which the scaled residuals remains relatively constant (21). The optimal solution should lie in this FPEAK range. In this way, subjective bias was reduced to a large extent. External information can be imposed on the solution to control the rotation. If specific species in the source profiles are known to be zero, then it is possible to pull down those values toward lower concentration through appropriate settings of FKEY resulting in the most interpretable source profiles. Each element of the FKEY matrix controls the pullingdown of the corresponding element in the source profile matrix by setting a nonzero integer values in FKEY matrix (20). To determine the number of sources, it was necessary to test different numbers of sources and find the optimal fit with the most physically reasonable results. Thus, the final PMF solutions were determined by experiments with different numbers of sources, different FPEAK values, and different FKEY matrices with the final choice based on the evaluation of the resulting source profiles as well as the quality of the species fits. The global optimum of the PMF solutions was tested by using multiple random starts for the initial values used in the iterative fitting process (1, 21). Conditional Probability Function Analysis. The conditional probability function (4, 22) analyzes local source impacts, especially gasoline vehicle and diesel emissions,
TABLE 1. Summary of PM2.5 and 35 Species Mass Concentrations Used for PMF Analysis concentration (ng/m3) species
geometric arithmetic meana mean
PM2.5 OC1 OC2 OC3 OC4 OP EC1 EC2 EC3 S NO2NO3Al As Br Ca Cl ClCr Cu Fe H K Mg Mn Na Ni Pb Rb Se Si Sr Ti V Zn Zr
9514 11086 2148 36555 73 103 2.8 502 226 272 50 939 320 446 20 2120 363 445 92 2431 156 193 3.0 626 268 329 20 1150 84 96 8.8 213 9.1 11 2.9 41 886 1127 170 5194 16 21 0.10 158 1609 2471 138 12553 61 110 12 1694 0.46 0.55 0.15 3.3 2.1 2.5 0.53 10 36 46 6.2 359 4.1 5.7 1.1 8.5 84 94 1.9 356 0.30 0.75 0.03 9.5 0.61 0.73 0.09 3.9 32 44 5.5 897 474 552 109 2121 37 44 9.2 342 33 46 13 146 1.3 1.9 0.16 23 179 274 13 2691 0.22 0.26 0.05 1.3 2.8 3.3 0.51 18 0.20 0.24 0.03 1.4 1.3 1.5 0.24 6.2 114 156 26 2948 0.33 0.43 0.04 6.2 3.0 5.4 0.36 102 0.54 0.92 0.06 5.3 7.4 9.2 1.0 168 0.23 0.28 0.05 0.85
min.
max.
no. of BDLb values (%) 0 95 11 32 0 29 10 8 147 0 204 0 193 105 0 0 254 208 116 11 0 0 0 229 41 178 145 0 77 0 8 37 16 104 0 198
no. of missing values (%)
0 (37.0) 0 (4.3) 0 (12.5) 0 0 (11.3) 0 (3.9) 0 (3.1) 0 (57.2) 0 0 (79.4) 1 (0.4) 1 (0.4) (75.1) 0 (40.9) 0 0 0 (98.8) 0 (80.9) 1 (0.4) (45.1) 0 (4.3) 0 0 0 0 (89.1) 0 (16.0) 0 (69.3) 0 (56.4) 0 0 (30.0) 0 0 (3.1) 0 (14.4) 0 (6.2) 0 (40.5) 0 0 (77.0) 0
a Data below the limit of detection were replaced by half of the reported detection limit values for the geometric mean calculations. b Below detection limit.
from varying wind directions using the source contribution estimates from PMF coupled with the time-resolved wind directions. The CPF estimates the probability that a given source contribution from a given wind direction will exceed a predetermined threshold criterion. CPF is defined as
m∆θ CPF ) n∆θ
(1)
where m∆θ is the number of occurrence from wind sector ∆θ that exceeded the threshold criterion, and n∆θ is the total number of data from the same wind sector. The same daily contribution was assigned to each hour of a given day to match to the hourly wind data. In this study, 24 sectors were used (∆θ ) 15 degrees). Calm winds (
