Source identifications of airborne fine particles using positive matrix factorization and U.S. Environmental Protection Agency positive matrix factorization.

ABSTRACT

The widely used source apportionment model, positive matrix factorization (PMF2), has been applied to various air pollution data. Recently, U.S. Environmental Protection Agency (EPA) developed EPA positive matrix factorization (PMF), a version of PMF that will be freely distributed by EPA. The objectives of this study were to conduct source apportionment studies for particulate matter less than 2.5 [micro]m in aerodynamic diameter ([PM.sub.2.5]) speciation data using PMF2 and EPA PMF (version 1.1) and to compare identified sources between the two models. In the present study, ambient [PM.sub.2.5] compositional datasets of 24-hr integrated samples collected at EPA Speciation Trends Network monitoring sites in Chicago, IL, and Portland, OR, were analyzed. Both PMF2 and EPA PMF extracted eight sources for the Chicago data and 10 sources for the Portland data. The model-resolved source profiles were similar between two models for both datasets. However, in several sources, the average contributions did not agree well and the time series contributions were not highly correlated. The differences between PMF2 and EPA PMF solutions were caused by the different least-square algorithm and the different nonnegativity constraints. Most of the average source contributions resolved by both models were within 5-95% uncertainty provided by EPA PMF, indicating that the sources resolved by both models were reproducible.

INTRODUCTION

There is growing interest in source apportionment studies for the airborne particulate matter (PM) with an increased focus on the control of the sources of airborne PM, because the statistical association between PM and adverse health effects was shown in many studies. (1-5)

Among many source apportionment methods, positive matrix factorization (PMF2) (6) has been shown to be a powerful alternative to traditional multivariate receptor modeling of airborne PM. (7-10) PMF2 requires one to purchase a license. PMF2 has been used to analyze ambient PM measurements in the Arctic, (11) in Hong Kong, (12) in Thailand, (13) in Toronto (Ontario, Canada), (14) in Vietnam, (15) and in the United States. (16-21)

A more flexible tool to fit multilinear models, the multilinear engine (ME), (22) was developed to solve any problem that can be expressed as a sum of products. It has been used to analyze the standard bilinear factor analysis model (23,24) and multiway models. (25-27) Recently, as one of the efforts to provide free source apportionment tools for the development and implementation of air quality standard, U.S. Environmental Protection Agency (EPA) developed EPA positive matrix factorization (PMF; version 1.1), (28) adopting a bilinear model solved by ME with a graphical user interface platform.

The objective of this study was to examine the source apportionment results using PMF2 and EPA PMF. In this study, the major sources of PM less than 2.5 [micro]m in aerodynamic diameter ([PM.sub.2.5]) were identified, and their contributions were estimated for two selected EPA Speciation Trends Network (STN) sites located in Chicago, IL, and Portland, OR. The identified source compositions and source contributions were compared for each site.

EXPERIMENTAL WORK

Data Collection

STN [PM.sub.2.5] samples were collected on a one-in-three-day schedule with a Mass Aerosol Speciation Sampler (URG) and Spiral Aerosol Speciation Samplers (Met One Instruments) at the monitoring sites located in Chicago and Portland, respectively. The Chicago monitoring site (Aerometric Information Retrieval System [AIRS] site code: 170310076; latitude: 41.754; longitude: -87.714) is located in urban residential area approximately 10 km southwest of downtown Chicago. Highways are situated around the monitoring site. The Portland monitoring site (AIRS site code: 410510246; latitude: 45.561; longitude: -122.668) is located at the urban residential area approximately 5 km southwest of the Portland International Airport. Highway 5 is located 750 m west of the site.

[PM.sub.2.5] samples were collected on three different filters: the Teflon filter was used for mass concentrations and for the elemental analysis via energy dispersive X-ray fluorescence (XRF) spectrometers. The Nylon filter was analyzed via ion chromatography (IC) for sulfate (S[O.sub.4.sup.2-]), nitrate (N[O.sub.3.sup.-]), ammonium (N[H.sub.4.sup.+]), sodium ([Na.sup.+]), and potassium ([K.sup.+]). The quartz filter was analyzed for organic carbon (OC) and elemental carbon (EC) via the National Institute for Occupational Safety and Health/Thermal Optical Transmittance protocol. (29)

Because the reported particulate OC concentrations were not blank corrected, (30) and carbon denuders were not used in the sampling line with the quartz filter, the integrated OC blank concentrations, including trip and field blank, as well as OC-positive artifact, were estimated using the intercept of the regression of OC concentrations against [PM.sub.2.5]. (21,31,32) The estimated OC blank values 1.38 [micro]g/[m.sup.3] at Chicago and 0.87 [micro]g/[m.sup.3] at Portland were subtracted from the reported STN OC concentrations before further analyses.

Receptor Modeling

PMF2 is a multivariate receptor model providing source profiles and their contributions based on a weighted least-square method that uses uncertainties for each measurement as the data point weights. (6) ME provides an approach to the fitting process that is more general and flexible to solve a variety of receptor modeling problems. (22) ME uses a structural equation input along with a set of constraints and can solve widely different multilinear and quasimultilinear problems. In EPA PMF, ME solves the standard bilinear model. Detailed explanations and equations are presented in previous publications. (26,33)

There are infinite numbers of possible solutions to the factor analysis problem because of the free rotation of matrices. (34) Both PMF2 and EPA PMF use nonnegativity constraints on the factors, which decrease this rotational freedom. (6,22) Also, PMF2 estimates uncertainties associated with source contributions and profiles from alternating regression fits where each row of source contribution is determined while source profile keeps constant, and each column of source profile is determined while keeping source contribution constant. EPA PMF estimates uncertainties associated with its solutions using a bootstrapping method based on the base case solution that is selected from several random model runs. These uncertainties also include uncertainties originated from the rotational freedom.

[FIGURE 1 OMITTED]

Because before July 2003 the STN data were not accompanied by analytical uncertainties, a comprehensive set of analytical uncertainty structures estimated by Kim et al. (32) was used in this study. Based on the reported analytical uncertainties or estimated analytical uncertainties, the input data and associated uncertainty matrices were estimated. The measured concentrations below method detection limit (MDL) values were replaced by half of the MDL values, and their uncertainties were set at five sixths of the MDL values. Missing concentrations were replaced by the geometric mean of the concentrations, and their accompanying uncertainties were set at four times this geometric mean concentration.

In this study, samples for which [PM.sub.2.5] or OC data were not available or were below zero or for which [PM.sub.2.5] or OC data had an error flag were excluded from the datasets. The samples collected on July 4, 2002, and July 5, 2003, at Chicago and July 5, 2004, at Portland that were highly affected by fireworks displays were excluded in this study. Overall, 18% and 8% of the original data were not included at Chicago and Portland, respectively. IC S[O.sub.4.sup.2-] was excluded from the analyses to prevent double counting of mass concentrations, because XRF S and IC S[O.sub.4.sup.2-] showed good correlations (slope = 2.9, [r.sup.2] = 0.97, for Chicago data; slope = 2.6, [r.sup.2] = 0.95, for Portland data). Also, IC [Na.sup.+] and IC [K.sup.+] were chosen because of the higher analytical precision compared with XRF Na and XRF K. Chemical species that have values above 90% below MDL (Cd, Ce, Cs, Hf, In, Ir, La, Hg, Nb, P, Rb, Ag, Tb, Y, and Zr at Chicago; Sb, Cd, Ce, Cs, Co, Eu, Ga, Au, Hf, In, Ir, La, Hg, Mo, Nb, Rb, Sm, Ag, Ta, Tb, Sn, Y, and Zr at Portland) were excluded. Chemical species that have signal-to-noise (S/N) ratios below 0.2 (As, Ba, and Sc at Portland) were considered bad variables and were excluded as recommended by Paatero and Hopke. (35) Thus, a total of 210 samples and 36 species and 269 samples and 26 species including [PM.sub.2.5] mass concentrations collected between May 2001 and November 2003 and October 2002 and April 2005 were used for Chicago and Portland, respectively.