TECHNICAL PAPER
ISSN 1047-3289 J. Air & Waste Manage. Assoc. 53:731–739
Copyright 2003 Air & Waste Management Association
Source Identification of Atlanta Aerosol by Positive Matrix
Factorization
Eugene Kim and Philip K. Hopke
Department of Chemical Engineering, Clarkson University, Potsdam, New York
Eric S. Edgerton
Atmospheric Research and Analysis, Inc., Durham, North Carolina
ABSTRACT
Data characterizing daily integrated particulate matter
(PM) samples collected at the Jefferson Street monitoring
site in Atlanta, GA, were analyzed through the application
of a bilinear positive matrix factorization (PMF) model. A
total of 662 samples and 26 variables were used for fine
particle (particles ⱕ2.5 ␮m in aerodynamic diameter)
samples (PM2.5), and 685 samples and 15 variables were
used for coarse particle (particles between 2.5 and 10 ␮m
in aerodynamic diameter) samples (PM10 –2.5). Measured
PM mass concentrations and compositional data were
used as independent variables. To obtain the quantitative
contributions for each source, the factors were normalized
using PMF-apportioned mass concentrations. For fine particle data, eight sources were identified: SO42⫺-rich secondary aerosol (56%), motor vehicle (22%), wood smoke
(11%), NO3⫺-rich secondary aerosol (7%), mixed source
of cement kiln and organic carbon (OC) (2%), airborne
soil (1%), metal recycling facility (0.5%), and mixed
source of bus station and metal processing (0.3%). The
SO42⫺-rich and NO3⫺-rich secondary aerosols were associated with NH4⫹. The SO42⫺-rich secondary aerosols also
included OC. For the coarse particle data, five sources
contributed to the observed mass: airborne soil (60%),
NO3⫺-rich secondary aerosol (16%), SO42⫺-rich secondary
aerosol (12%), cement kiln (11%), and metal recycling facility (1%). Conditional probability functions were computed
using surface wind data and identified mass contributions
IMPLICATIONS
PMF has been applied to PM2.5 and PM10 –2.5 composition
data derived from samples collected in Atlanta. The results for
the two sets of samples illustrate the use of PMF as a data
analysis tool for identifying and apportioning PM sources in
these size ranges. With the likelihood of a PM10 –2.5 standard
being promulgated along with the PM2.5 standard after the
current round of National Ambient Air Quality Standards review, such methods will be needed as part of the planning
process for implementing the air quality standards for PM.
Volume 53 June 2003
from each source. The results of this analysis agreed well
with the locations of known local point sources.
INTRODUCTION
An association between particulate matter (PM) concentrations and adverse health effects has been found in
many studies.1–3 Since the U.S. Environmental Protection
Agency promulgated new National Ambient Air Quality
Standards for airborne PM, a number of additional air
quality and epidemiology studies have been undertaken.4 – 6
The Southeastern Aerosol Research and Characterization
(SEARCH) study was begun in early 1998. In this study,
instrumentation for daily sampling and continuous PM
mass and composition, gases, and meteorology measurements were deployed at eight monitoring stations over a
broad geographical region in the southeastern United
States. The objectives of the SEARCH study are to obtain
an understanding of PM composition and its seasonal and
regional variability in southeastern U.S. states, develop
PM climatology, and estimate source contributions.
As a part of the SEARCH study, advanced source apportionment methods for airborne PM, such as positive
matrix factorization (PMF), are being applied to the data.
PMF7 has been shown to be a powerful alternative to
traditional receptor modeling of airborne PM.8 –10 PMF
has been successfully used to assess particle source contributions in the Arctic,11 Hong Kong,12 Phoenix, AZ,13 Thailand,14 Vermont,15 and three northeastern U.S. cities.16
The objectives of this study are to identify PM sources
and estimate their contributions to the particle mass concentrations. In the present paper, PMF was applied to a
particle compositional data set of daily samples collected
during a two-year period at a single monitoring site in
Atlanta, GA. The resolved particle sources of ambient fine
(particles ⱕ2.5 ␮m in aerodynamic diameter) particles
(PM2.5) and coarse (particles between 2.5 ␮m and 10 ␮m
in aerodynamic diameter) particles (PM10 –2.5) and their
seasonal trends are discussed. To help identify the likely
locations of the PMF-identified sources, a conditional
Journal of the Air & Waste Management Association 731
Kim, Hopke, and Edgerton
probability function (CPF) was calculated. The results of
this study could be useful in the understanding of local
particle sources and could assist in air quality management.
SAMPLE COLLECTION AND CHEMICAL ANALYSIS
The particle compositional data utilized in this study consisted of measurements taken between August 1998 and
August 2000 at a monitoring site (Jefferson Street) located 4
km northwest of downtown Atlanta. The Jefferson Street site
is located in an industrial and commercial area, shown in
Figure 1. Coarse particle samples were collected using a
conventional dichotomous sampler (Andersen Instruments,
Inc.). Daily integrated fine particle samples were collected
using the particulate composition monitor (PCM) (Atmospheric Research and Analysis, Inc.), which has three independent sampling lines, shown in Figure 2. Each sampling
line has a 10-␮m cyclone followed by a Well Impactor
Ninety-Six, which has a 2.5-␮m cutoff size in particle aerodynamic diameter.17,18 The PCM permits simultaneous sampling on a three-stage filter pack (Teflon, nylon, and cellulose filter), nylon filter, and quartz filter. The PCM includes
carbonate denuders and citric acid denuders upstream of
both the three-stage filter and the nylon filter. The quartz
filter includes an upstream carbon denuder to remove gaseous organic materials. Chemical analysis was performed on
the Teflon filters of the three-stage filter pack samples via
energy-dispersive X-ray fluorescence (XRF)19 by Chester
LabNet, Inc. Nylon and cellulose filters of the three-stage
filter pack samples were analyzed via ion chromatography
(IC) for NO3⫺ and NH4⫹. The nylon filters of the independent sampling line were analyzed via IC for SO42⫺, NO3⫺,
and NH4⫹. The quartz filter was cut in half and analyzed via
thermal optical reflectance/Interagency Monitoring of Protected Visual Environments protocol for organic carbon
(OC) and elemental carbon (EC) by Desert Research Insti-
Figure 1. Location of the Jefferson Street monitoring site in Atlanta, GA.
732 Journal of the Air & Waste Management Association
Figure 2. Schematic of the PCM.
tute. For the fine particle data, 662 daily samples collected
between August 1998 and August 2000 and 26 species were
used. XRF S and SO42⫺ showed excellent correlations
(slope ⫽ 3.1, r2 ⫽ 0.99), so it is reasonable to exclude XRF S
from the analysis.
Analysis of the compositional data revealed a mass closure problem for the fine particle samples. In Figure 3, the
fine mass concentrations measured using the three-stage
filter are compared with the summations of fine particle
species excluding crustal elements (Al, Si, K, Ca, Fe) that
would normally be associated with O2. Approximately 27%
of the measured mass concentrations were smaller than the
summations of species concentrations. This mass closure
problem is thought to be caused by the loss of semivolatile OC.20,21 The loss of NO3⫺ mass on the Teflon filter
was accounted for by the subsequent nylon filter in the
Figure 3. Time-series plot of measured mass concentration minus
summation of measured species for fine particles.
Volume 53 June 2003
Kim, Hopke, and Edgerton
three-stage filter pack. This additional mass of NO3⫺ was
added to the measured particle mass from the Teflon filter. This
mass closure issue (sum of species ⬎ particle mass) presents
a problem for the multilinear regression analysis that has
generally been used in PMF analysis. Therefore, the alternative approach described in the next section was employed.
For the coarse particle data, 685 samples collected
between July 1998 and December 2000 and 15 species
were available. OC and EC were not measured. In coarse
particle data, approximately 53% of the XRF S values were
smaller than SO42⫺/3. This result suggests that XRF S
concentrations were underestimated where it is surmised
that the problem is X-ray penetration losses. For this
reason, the XRF S was excluded from the analysis of coarse
particle data. Summaries of the fine and coarse PM species
used in this study are shown in Tables 1 and 2, respectively.
冘
p
x ij ⫽
gikfkj ⫹ eij
(1)
k⫽1
where xij is the jth species concentration measured in the
ith sample, gik is the particulate mass concentration from
the kth source contributing to the ith sample, fkj is the jth
species mass fraction from the kth source, eij is residual
associated with the jth species concentration measured in
the ith sample, and p is the total number of independent
sources. The corresponding matrix equation is
X ⫽ GF ⫹ E
(2)
where X is an n ⫻ m data matrix with n measurements and
m number of elements; E is an n ⫻ m matrix of residuals;
G is an n ⫻ p source contribution matrix with p sources;
and F is a p ⫻ m source profile matrix. As pointed out by
Henry,25 there are an infinite number of possible solutions to the factor analysis problem (rotations of G matrix
DATA ANALYSIS
The general receptor modeling problem can be stated in
terms of the contribution from p independent sources to
all chemical species in a given sample as follows:22–24
Table 1. Summary of fine particle mass and 25 species concentrations used for PMF modeling.
Concentration (ng/m3)
Species
Geometric Mean
Arithmetic Mean
15871
4488
2485
877
82
1643
4024
1.0
17
2.9
2.0
1.4
3.7
2.0
0.93
3.9
3.6
12
0.54
0.51
0.62
11
191
65
61
102
18029
5579
2920
1124
108
1982
4495
1.5
19
4.2
3.8
2.0
6.6
3.5
1.4
4.4
5.0
17
0.61
0.57
0.46
29
253
81
78
132
PM2.5
SO42⫺
NH4⫹
NO3⫺
Cl
EC
OC
As
Ba
Br
Cu
Mn
Pb
Sb
Se
Sn
Ti
Zn
Cr
Ni
V
Al
Si
K
Ca
Fe
a
a
Number of BDLb
Min
1930
526
298
127
21
227
878
0.47
13
0.26
0.59
0.37
1.1
1.9
0.32
3.2
2.0
0.42
0.47
0.47
0.27
5.7
23
7.7
4.6
13
Max
Values (%)
Number of Missing
Values (%)
49264
20851
10314
6014
613
10234
22089
11
57
204
42
13
78
105
10
17
55
211
3.5
2.3
3.2
1700
3966
996
702
1502
0
0
0
0
115 (17.4)
0
0
337 (50.9)
0
29 (4.4)
214 (32.3)
166 (25.1)
195 (29.5)
321 (48.5)
284 (42.9)
551 (83.2)
363 (54.8)
3 (0.5)
603 (91.1)
635 (95.9)
611 (92.3)
406 (61.3)
0
0
1 (0.2)
0
0
5 (0.8)
6 (0.9)
44 (6.6)
44 (6.6)
22 (3.3)
21 (3.2)
35 (5.3)
34 (5.1)
34 (5.1)
34 (5.1)
34 (5.1)
34 (5.1)
34 (5.1)
34 (5.1)
34 (5.1)
34 (5.1)
34 (5.1)
13 (2.0)
12 (1.8)
4 (0.6)
34 (5.1)
34 (5.1)
34 (5.1)
34 (5.1)
34 (5.1)
Data below the limit of detection were replaced by half of the reported detection limit values for the geometric mean calculations; bBDL ⫽ below detection limit.
Volume 53 June 2003
Journal of the Air & Waste Management Association 733
Kim, Hopke, and Edgerton
Table 2. Summary of coarse particle mass and 14 species concentrations used for PMF modeling.
Concentration (ng/m3)
Number of BDLb
Species
Geometric Mean
Arithmetic Mean
Min
Max
Values (%)
Number of Missing
Values (%)
PM10–2.5
NO3⫺
SO42⫺
NH4⫹
Cr
Cu
Fe
Mn
Ni
V
Al
Si
K
Ca
Ti
8611
349
285
33
0.73
1.31
4.72
1.20
0.74
0.84
581
1380
139
375
46
10116
457
393
68
0.75
1.2
8.0
1.5
0.75
0.81
767
1816
162
480
54
511
8.6
15
3.5
0.73
1.1
2.3
0.59
0.73
0.79
12
49
7.9
12
2.1
34701
2403
2796
795
1.1
5.6
189
6.9
2.5
2.1
3148
7653
711
3458
240
0
4 (0.6)
24 (3.5)
139 (20.3)
635 (92.7)
606 (88.5)
366 (53.4)
327 (47.7)
636 (92.8)
627 (91.5)
0
0
0
0
3 (0.4)
0
11 (1.6)
11 (1.6)
12 (1.8)
50 (7.3)
52 (7.6)
52 (7.6)
54 (7.9)
47 (6.9)
56 (8.2)
48 (7.0)
56 (8.2)
56 (8.2)
48 (7.0)
56 (8.2)
a
a
Data below the limit of detection were replaced by half of the reported detection limit values for the geometric mean calculations; bBDL ⫽ below detection limit.
and F matrix). To decrease rotational freedom, PMF uses
nonnegativity constraints on the factors. The parameter
FPEAK is used to control the rotations.26 PMF provides a
solution that minimizes an object function, Q(E), based
on uncertainties for each observation.7,27 This function is
defined as
冘冘
n
Q共E兲 ⫽
m
i⫽1 j⫽1
冤
xij ⫺
冘
p
k⫽1
uij
冥
2
gikfkj
(3)
where uij is an uncertainty estimate in the jth element
measured in the ith sample.
The application of PMF depends on the estimated
uncertainties for each of the data values. The uncertainty
estimation provides a useful tool to decrease the weight of
missing and below-detection-limit data in the solution, as
well as accounting for the variability in the source profiles. The procedure of Polissar et al.28 was used to assign
measured data and the associated uncertainties as the
input data to the PMF. The concentration values were
used for the measured data, and the summation of the
analytical uncertainty and 1/3 of the detection limit value
was used as the overall uncertainty assigned to each measured value. Values below the detection limit were replaced by half of the detection limit values, and their
overall uncertainties were set at 5/6 of the detection limit
values. Missing values were replaced by the geometric
mean of the measured values, and their accompanying
uncertainties were set at 4 times this geometric mean value.
734 Journal of the Air & Waste Management Association
The uncertainty must take into account both the
measurement uncertainty and the variability in the
source profiles. It also has to help to take into account the
differences in scale between major species as compared
with the lower-concentration species. In northeastern
U.S. aerosol studies,16 PMF separated the S into a high
photochemistry source and a low photochemistry source
with seasonal differences of the Se/S concentrations. In
this study, without adequate Se data, it was found that it
was necessary to increase the estimated uncertainties of
SO42⫺ and NH4⫹ by a factor of 4 to take the high photochemistry variability into account. Raising the estimated
uncertainty of SO42⫺ and NH4⫹ decreases the weight of
SO42⫺ and NH4⫹ in the solution.
Because of the mass closure problem noted previously, the measured fine particle mass concentrations
were included as an independent variable in the PMF
modeling to directly obtain the mass apportionment
without the usual multilinear regression. The estimated
uncertainties of the mass concentrations were set at 4
times their values so that the large uncertainties decreased
their weight in the model fit.
The results of PMF modeling were then normalized
by the apportioned particle mass concentrations so that
the quantitative source contributions for each source were
obtained. Specifically
冘
冉冊
p
x ij ⫽
k⫽1
共ckgik兲
fkj
ck
(4)
Volume 53 June 2003
Kim, Hopke, and Edgerton
where ck is directly apportioned mass concentration by
PMF for the kth source.
To analyze point-source impacts from various wind
directions, the CPF29 was calculated using source contribution estimates from PMF, coupled with wind direction
values measured on site. To minimize the effect of atmospheric dilution, daily fractional mass contribution from
each source relative to the total of all sources was used
rather than the absolute source contributions. The same
daily fractional contribution was assigned to each hour of
a given day to match the hourly wind data. Specifically,
the CPF is defined as
CPF ⫽
m ⌬␪
n ⌬␪
(5)
where m⌬␪ is the number of occurrence from wind sector
⌬␪ that exceeded the threshold criterion, and n⌬␪ is the
total number of data from the same wind sector. In this
study, ⌬␪ was set at 11.3°. Calm winds (⬍1 m/sec) were
excluded from this analysis. The threshold was set at the
upper 10th percentile of the fractional contribution from
each source. The sources are likely to be located in the
directions that have high conditional probability values.
RESULTS AND DISCUSSION
To determine the number of sources, it is reasonable to
experiment with different numbers of sources and find
the one with the most physically meaningful results. Also,
because rotational ambiguity exists in factor analysis
modeling, PMF was run several times with different
FPEAK values to determine the range within which the
objective function Q(E) value in eq 3 remains relatively
constant.26 The optimal solution should lie in this FPEAK
range. In this way, subjective bias was avoided to some
extent. The final PMF solutions were determined by trial
and error with different numbers of sources as well as
different FPEAK values and were based on the evaluation
of the resulting source profiles.
In this study, the robust mode was used to reduce the
influence of extreme values on the PMF solution. A data
point was classified as an extreme value if the model residual
exceeded 4 times the error estimate. The estimated uncertainties of those extreme values were then increased so that
the weights of the extreme values in the solution were decreased. The global optima of the PMF solutions were tested
using five different seeds of pseudorandom values.
Fine Particle Data
Eight sources were identified for the fine particles. A value of
FPEAK ⫽ 0.2 provided the most physically reasonable source
profiles. A comparison of the daily reconstructed fine mass
Volume 53 June 2003
contributions from all sources with measured fine mass concentrations is shown in Figure 4. When the uncertainties
associated with this data set are considered, the squared
correlation coefficient of 0.83 indicates that the resolved
sources effectively account for most of the variation in the
particle mass concentration. Figure 5 presents the identified
source profiles and Figure 6 shows time-series plots of estimated daily contributions to fine particle mass concentrations from each source. Figure 7 shows conditional probabilities of source locations for local point sources.
The SO42⫺-rich secondary aerosol has a high concentration of SO42⫺ and shows strong seasonal variation with
high concentrations in summer when the photochemical
activity is highest. This source includes NH4⫹, which explains 30% of its mass. It also includes OC, which typically becomes associated with the secondary SO42⫺ aerosol. This OC association is consistent with previous
Phoenix13 and northeastern U.S.16 aerosol studies.
The motor vehicle source is represented by high OC
and EC.30 –32 It also includes some soil dust constituents
(Si, Fe). This indicates the motor vehicle source is not just
vehicle tailpipe source but also street emissions, including
resuspended road dust. The ratio of OC/EC is 1.49 for this
source, whereas it is more typically 2.05 in fresh gasoline
exhaust and 0.72 in diesel emissions for PM10.33 Therefore, this motor vehicle source is a mixture of diesel emissions and particles produced by gasoline-powered vehicles. Watson et al.31,32 reported 0.05– 8.6 wt % of SO42⫺
concentration in motor vehicle exhaust from 1992 and
1995 roadside measurements. In this study, the motor
vehicle source has a negligible amount of SO42⫺ that
might arise from SO2 emissions where there would not be
time for substantial conversion to SO42⫺.
Figure 4. Measured vs. predicted fine particle mass concentration.
Journal of the Air & Waste Management Association 735
Kim, Hopke, and Edgerton
Figure 5. Source profiles resolved from fine particle samples.
coal-fired cement kiln source profiles.34 The high OC concentration of this source indicates that the cement kiln and
an OC-rich source are collocated and that their daily emission
patterns are similar. Also, a rail yard located approximately 2
km northwest of the site may have mixed with this source.
The airborne soil factor contains the characteristic
elements Si, Al, Fe, K, and Ti.31,32 It has a few random high
peaks. Unpaved roads, construction sites, and windblown soil dust could produce particles of these crustal
elements. Those were not separated in this study because
their chemical profiles were similar and their daily variations were not enough to separate them.
For the seventh source, a metal recycling facility located
approximately 0.7 km east of the site is suggested because of
the factor characterized by its high mass fraction of Zn, Fe,
and Cl.34,35 The Zn/Fe ratio of this source is 1.1. This source
has short-term peaks in winter. The facility is used mainly
for storage, grinding, shredding, and loading onto railcars.
In addition, several metal processing facilities are located
approximately 6 km southeast of the site. In Figure 7, there
are indications of contributions from the east and southeast.
Mixed source II includes high values of EC, Fe, K, Cu,
and Pb. The CPF plot (Figure 7) suggests that this factor
comes from sources in the southeasterly direction, which
include a bus station30,36 located only approximately
200 m southeast of the site. This diesel emission seems to
The wood smoke source is characterized by OC, EC, and
K.32 This source has a seasonal trend with high values in
winter and short-term peaks in spring and summer. The winter
peaks indicate residential wood burning, and the spring and
summer events are likely to be caused by forest fires. There
are also prescribed burnings to the south of Atlanta in winter.
The NO3⫺-rich secondary aerosol has a high concentration of NO3⫺. It shows seasonal variation with maxima in
winter. These peaks in winter indicate that low temperature
and high relative humidity help the formation of NO3⫺
aerosols in Atlanta. This result is consistent with a previous
study of aerosol data from three northeastern U.S. sites.16
This source includes NH4⫹, which becomes associated with
the secondary NO3⫺ aerosol. The NO3⫺ concentration is
0.98 ␮g/␮g for this source. In a given source, the summation
of species concentrations is often slightly more than 100%
because PMF modeling uses a statistical approach and there
are uncertainties associated with any of these values.
Mixed source I has high OC and Ca concentrations.
This factor is likely to include contributions from a cement
kiln located approximately 7 km northwest of the monitoring site and an unknown OC-rich source. It is consistent
with the conditional probability plot for this source in Figure 6. The cement kiln emission contains OC because this
stationary source combusts fossil fuels. However, the ratio of
OC/Ca is 2.43 for this source, whereas it is typically 0.54 in
Figure 6. Source contributions for fine particle samples.
736 Journal of the Air & Waste Management Association
Volume 53 June 2003
Kim, Hopke, and Edgerton
Table 3. Average source contributions (␮g/m3) to fine particle mass concentration.
Average Source Contribution
(standard error)
SO42⫺-rich secondary aerosol
Motor vehicle
Wood smoke
NO3⫺-rich secondary aerosol
Mixed source I
Airborne soil
Metal recycling
Mixed source II
Total
Average
Summer
Average
Winter
Average
8.85 (0.21)
3.53 (0.11)
1.72 (0.05)
1.15 (0.03)
0.31 (0.01)
0.18 (0.01)
0.08 (0.003)
0.05 (0.003)
12.1 (0.30)
3.28 (0.12)
1.35 (0.05)
0.80 (0.03)
0.36 (0.02)
0.26 (0.02)
0.06 (0.003)
0.05 (0.003)
6.17 (0.12)
3.74 (0.17)
2.03 (0.08)
1.45 (0.05)
0.27 (0.01)
0.12 (0.006)
0.10 (0.006)
0.06 (0.004)
fine particle mass concentration. The SO42⫺-rich secondary aerosol, mixed source I, and airborne soil contributed
more to the fine particle mass in summer. The motor
vehicle, wood smoke, NO3⫺-rich secondary aerosol, and
metal recycling sources contributed more in winter. The
slightly winter-high trend of the motor vehicle source is
likely caused by the temperature inversions that concentrate pollutants in the surface mixing layer. Mixed source
II did not show a strong seasonal trend.
Figure 7. Hourly CPF plots for the highest 10% of the mass contibution
from fine particle point sources.
be mixed with metal processing sources34,35 located in the
same direction as the bus station from the monitoring site.
The average contributions of each source to the fine
particle mass concentration are summarized in Table 3.
The seasonal average contributions are also presented
(summer: April–September; winter: October–March). The
SO42⫺-rich secondary aerosol has the highest source contribution to the fine particle mass concentrations (56%).
It is consistent with a study of three northeastern U.S.
sites, which identified its contributions of 47, 55, and
51% to the fine particle mass concentration.16 The second
contributor is motor vehicles, accounting for 22% of the
Volume 53 June 2003
Coarse Particle Data
For the coarse particles, the PMF identified five sources
with an FPEAK of 0. Table 4 summarizes the five sources
and their average mass contributions. As presented in
Figure 8, a comparison of the coarse mass contributions
from all sources with measured coarse mass concentrations shows good agreement (r2 ⫽ 0.92). Figures 9 and 10
show the identified source profiles and time-series plots of
daily contributions to coarse mass concentrations from
each source, respectively. Figure 11 shows conditional
probability plots for point sources.
The airborne soil source is represented by Si, Al, Ca, K,
and Ti.31,32 Resuspended road dust is typical in urban
areas. There could be transfer of cement-derived Ca to soil,
Table 4. Average source contributions (␮g/m3) to coarse particle mass concentration.
Average Source Contribution
(standard error)
Airborne soil
NO3⫺-rich secondary aerosol
SO42⫺-rich secondary aerosol
Cement kiln
Metal recycling
Total
Average
Summer
Average
Winter
Average
5.94 (0.16)
1.61 (0.05)
1.22 (0.05)
1.06 (0.04)
0.13 (0.008)
5.93 (0.21)
1.53 (0.07)
1.39 (0.07)
0.94 (0.05)
0.15 (0.01)
5.95 (0.25)
1.68 (0.08)
1.07 (0.08)
1.16 (0.06)
0.12 (0.01)
Journal of the Air & Waste Management Association 737
Kim, Hopke, and Edgerton
Figure 10. Source contributions for coarse particle samples.
Figure 8. Measured vs. predicted coarse particle mass concentration.
particularly in dry weather. Nitrate-rich and SO42⫺-rich
secondary aerosols are characterized by NO3⫺ and SO42⫺,
respectively. The cement kiln source has high Ca levels.
As shown in Figure 10, the conditional probability plot of
cement kiln points to northwest of the monitoring site,
which is consistent with the fine particle plot (see Figure 7).
The fifth source has high Si, Al, and Fe levels. This source
may be the metal recycling facility. The conditional probability plot of this source shows relatively high contributions from the northwest, northeast, and southeast.
The main source of coarse particles is airborne soil
(60%), although it is likely that much of the soil is resuspended by the action of motor vehicles. Most urban areas
report soil as the major contributor to coarse particles.13
Figure 9. Source profiles resolved from coarse particle samples.
738 Journal of the Air & Waste Management Association
The five sources of coarse particles do not have strong
seasonal trends, as shown in Table 4. Nitrate-rich and
SO42⫺-rich aerosols have seasonal trends that are weak but
similar to those of the comparable fine particle sources.
Figure 11. Hourly CPF plots for the highest 10% of the mass contribution from coarse particle point sources.
Volume 53 June 2003
Kim, Hopke, and Edgerton
The cement kiln contributed more in winter. In contrast,
the metal recycling source contributed more in summer.
CONCLUSIONS
Daily integrated PM compositional data measured at a monitoring site in Atlanta were analyzed through PMF. Including particle mass concentrations as independent variables,
the PMF effectively resolved eight and five sources for fine
particles and coarse particles, respectively. SO42⫺-rich secondary aerosol contributed to the fine particles the most,
which is consistent with the results of northeastern U.S.
aerosol studies.16 This source impact is higher in summer.
Airborne soil contributed to the coarse particles the most.
The impacts from the point sources are more clearly seen
using PMF results combined with plots of CPFs. Those plots
of both fine and coarse particles clearly show the direction
of the cement kiln. In addition, this approach revealed the
direction of the metal recycling source and the bus station.
ACKNOWLEDGMENTS
This study was supported by the Southern Company,
Atlanta, GA.
REFERENCES
1. Dockery, D.W.; Pope, C.A.; Xu, X.P.; Spengler, J.D.; Ware, J.H.; Fay,
M.E.; Ferris, B.G. An Association between Air-pollution and Mortality
in 6 United States Cities; New Engl. J. Med. 1993, 329, 1753-1759.
2. Samet, J.M.; Zeger, S.L.; Dominici, F.; Curriero, F.; Coursac, I.; Dockery,
D.W.; Schwartz, J.; Zanobetti, A. The National Morbidity, Mortality,
and Air Pollution Study. Part II: Morbidity and Mortality from Air Pollution in the United States; Res. Rep. Health Effect Inst. 2000, 94 (Pt 2), 5-79.
3. Pope, C.A.; Burnett, R.T.; Thun, M.J.; Calle, E.E.; Krewski, D.; Ito, K.;
Thurston, G.D. Lung Cancer, Cardiopulmonary Mortality, and LongTerm Exposure to Fine Particulate Air Pollution; J. Am. Med. Assoc.
2002, 287 (9), 1132-1141.
4. Clarke, R.W.; Coull, B.A.; Reinisch, U.; Catalano, P.J.; Killingsworth,
C.R.; Koutrakis, P.; Kavouras, I.; Lawrence, J.; Lovett, E.G.; Wolfson,
J.M.; et al. Inhaled Concentrated Ambient Particles Are Associated
with Hematologic and Bronchoalveolar Lavage Changes in Canines;
Environ. Health Perspect. 2000, 108 (12), 1179-1187.
5. Laden, F.; Neas, L.M.; Dockery, D.W.; Schwartz, J. Association of Fine
Particulate Matter from Different Sources with Daily Mortality in Six
U.S. Cities; Environ. Health Perspect. 2000, 108 (10), 941-947.
6. Sarnat, J.A.; Koutrakis, P.; Suh, H. Assessing the Relationship between
Personal Particulate and Gaseous Exposures of Senior Citizens Living
in Baltimore, MD; J. Air & Waste Manage. Assoc. 2000, 50, 1184-1198.
7. Paatero, P. Least Square Formulation of Robust Nonnegative Factor
Analysis; Chemomet. Intell. Lab. Systems 1997, 37, 23-35.
8. Huang, S.; Rahn, K.A.; Arimoto, R. Testing and Optimization Two
Factor-Analysis Techniques on Aerosol at Narragansett, Rhode Island;
Atmos. Environ. 1999, 33, 2169-2185.
9. Willis, R.D. Workshop on UNMIX and PMF as Applied to PM2.5; EPA
600-A-00 – 048, U.S. Environmental Protection Agency: Research Triangle Park, NC, 2000.
10. Qin, Y.; Oduyemi, K.; Chan, L.Y. Comparative Testing of PMF and CFA
Models; Chemomet. Intell. Lab. Systems 2002, 61, 75-87.
11. Xie, Y.L.; Hopke, P.K.; Paatero, P.; Barrie, L.A.; Li, S.M. Identification of
Source Nature and Seasonal Variations of Arctic Aerosol by Positive
Matrix Factorization; J. Atmos. Sci. 1999, 56, 249-260.
12. Lee, E.; Chun, C.K.; Paatero, P. Application of Positive Matrix Factorization in Source Apportionment of Particulate Pollutants; Atmos.
Environ. 1999, 33, 3201-3212.
13. Ramadan, Z.; Song, X.H.; Hopke, P.K. Identification of Sources of
Phoenix Aerosol by Positive Matrix Factorization; J. Air & Waste Manage. Assoc. 2000, 50, 1308-1320.
14. Chueinta, W.; Hopke, P.K.; Paatero, P. Investigation of Sources of Atmospheric Aerosol at Urban and Suburban Residential Area in Thailand
by Positive Matrix Factorization; Atmos. Environ. 2000, 34, 3319-3329.
15. Polissar, A.V.; Hopke, P.K.; Poirot, R.L. Atmospheric Aerosol over Vermont: Chemical Composition and Sources; Environ. Sci. Technol.
2001, 35, 4604 – 4621.
Volume 53 June 2003
16. Song, X.H.; Polissar, A.V.; Hopke, P.K. Source of Fine Particle Composition in the Northeastern U.S.; Atmos. Environ. 2001, 35, 5277-5286.
17. Peters, T.M.; Vanderpool, R.W.; Wiener, R.W. Design and Calibration
of the EPA PM2.5 Well Impactor Ninety-Six (WINS); Aerosol Sci. Technol. 2001, 34, 389-397.
18. Vanderpool, R.W.; Peters, T.M.; Natarajan, S.; Gemmill, D.B.; Weiner,
R.W. Evaluation of the Loading Characteristics of the EPA WINS PM2.5
Separator; Aerosol Sci. Technol. 2001, 34, 444-456.
19. Dzubay, T.G.; Stevens, R.K.; Gordon, G.E.; Olmez, I.; Sheffield, A.E.;
Courtney, W.J. Composite Receptor Method Applied to Philadelphia
Aerosol; Environ. Sci. Technol. 1988, 22 (1), 46-52.
20. Apple, B.R.; Tokiwa, Y.; Kothny, E.L. Sampling of Carbonaceous Particles in the Atmosphere; Atmos. Environ. 1983, 17, 1787-1796.
21. Van Vaeck, L.; Van Cauwenberghe, K.; Janssens, J. The Gas-Particle
Distribution of Organic Aerosol Constituents: Measurement of the
Volatilization Artifact in Hi-Vol Cascade Impactor Sampling; Atmos.
Environ. 1984, 18, 417-430.
22. Miller, M.S.; Friedlander, S.K.; Hidy, G.M.A. Chemical Element Balance for the Pasadena Aerosol; J. Colloid Interface Sci. 1972, 39, 165-176.
23. Hopke, P.K. Receptor Modeling in Environmental Chemistry; Wiley &
Sons: New York, 1985.
24. Hopke, P.K. Receptor Modeling for Air Quality Management; Elsevier:
Amsterdam, The Netherlands, 1991.
25. Henry, R.C. Current Factor Analysis Models Are Ill-Posed; Atmos. Environ. 1987, 21, 1815-1820.
26. Paatero, P.; Hopke, P.K.; Song, X.H.; Ramadan, Z. Understanding and
Controlling Rotations in Factor Analytic Models; Chemomet. Intell.
Lab. Systems 2002, 60, 253-264.
27. Paatero, P. User’s Guide for Positive Matrix Factorization Programs
PMF2 and PMF3, Part 1: Tutorial. ftp://rock.helsinki.fi/pub/misc/pmf
(accessed Apr 2003).
28. Polissar, A.V.; Hopke, P.K.; Paatero, P.; Malm, W.C.; Sisler, J.F. Atmospheric Aerosol over Alaska 1. Elemental Composition and Sources; J.
Geophys. Res. 1998, 103 (D15), 19045-19057.
29. Ashbaugh, L.L.; Malm, W.C.; Sadeh, W.Z. A Residence Time Probability Analysis of Sulfur Concentrations at Ground Canyon National
Park; Atmos. Environ. 1985, 19 (8), 1263-1270.
30. Watson, J.G.; Chow, J.C.; Lowenthal, D.H.; Pritchett, L.C.; Frazier, C.A.
Differences in the Carbon Composition of Source Profiles for Diesel and
Gasoline Powered Vehicles; Atmos. Environ. 1994, 28 (15), 2493-2505.
31. Watson, J.G.; Chow, J.C. Source Characterization of Major Emission
Sources in the Imperial and Mexicali Valleys along the US/Mexico
Border; Sci. Total Environ. 2001, 276, 33-47.
32. Watson, J.G.; Chow, J.C.; Houck, J.E. PM2.5 Chemical Source Profiles
for Vehicle Exhaust, Vegetative Burning, Geological Material, and
Coal Burning in Northwestern Colorado during 1995; Chemosphere
2001, 43, 1141-1151.
33. Cadle, S.H.; Mulawa, P.A.; Hunsanger, E.C.; Nelson, K.E.; Ragazzi, R.A.;
Barrett, R.; Gallagher, G.L.; Lawson, D.R.; Knapp, K.T.; Snow, R. Composition of Light-Duty Motor Vehicle Exhaust Particulate Matter in
the Denver, Colorado Area; Environ. Sci. Technol. 1999, 33, 2328-2339.
34. EPA SPECIATE version 3.2. U.S. Environmental Protection Agency:
Research Triangle Park, NC, 2002.
35. Small, M.; Germani, M.S.; Small, A.M.; Zoller, W.H.; Moyers, J.L.
Airborne Plume Study of Emissions from the Processing of Copper
Ores in Southeastern Arizona; Environ. Sci. Technol. 1981, 15, 293-299.
36. Lowenthal, D.H.; Zielinska, B.; Chow, J.C.; Watson, J.G. Characterization of Heavy-Duty Diesel Vehicle Emissions; Atmos. Environ. 1994, 28
(4), 731-743.
About the Authors
Eugene Kim has been a research associate at Clarkson
University since December 2001, when he completed a
postdoctoral fellowship in the Northwest Particulate Matter
and Health Center. He received a Ph.D. in environmental
engineering from the University of Washington. Philip K.
Hopke is the Bayard D. Clarkson Distinguished Professor at
Clarkson University. He has studied airborne PM for 30
years. Eric S. Edgerton is president and chief scientist of
Atmospheric Research and Analysis, Inc., in Durham, NC.
Address correspondence to: Dr. Philip K. Hopke, Department
of Chemical Engineering, Clarkson University, P.O. Box 5705,
Potsdam, NY 13699; e-mail: hopkepk@clarkson.edu.
Journal of the Air & Waste Management Association 739