PM_2.5 Source Apportionment: Reconciling Receptor Models for U.S. Nonurban and Urban Long-Term Networks

ABSTRACT

Chemical mass balance (CMB) and trajectory receptor models were applied to speciated particulate matter with aerodynamic diameter ≤2.5 μm (PM_2.5) measurements from Speciation Trends Network (STN; part of the Chemical Speciation Network [CSN]) and Interagency Monitoring of Protected Visual Environments (IMPROVE) monitoring network across the state of Minnesota as part of the Minnesota PM_2.5 Source Apportionment Study (MPSAS). CMB equations were solved by the Unmix, positive matrix factorization (PMF), and effective variance (EV) methods, giving collective source contribution and uncertainty estimates. Geological source profiles developed from local dust materials were either incorporated into the EV-CMB model or used to verify factors derived from Unmix and PMF. Common sources include soil dust, calcium (Ca)-rich dust, diesel and gasoline vehicle exhausts, biomass burning, secondary sulfate, and secondary nitrate. Secondary sulfate and nitrate aerosols dominate PM_2.5 mass (50–69%). Mobile sources outweigh area sources at urban sites, and vice versa at rural sites due to traffic emissions. Gasoline and diesel contributions can be separated using data from the STN, despite significant uncertainties. Major differences between MPSAS and earlier studies on similar environments appear to be the type and magnitude of stationary sources, but these sources are generally minor (<7%) in this and other studies. Ensemble back-trajectory analysis shows that the lower Midwestern states are the predominant source region for secondary ammoniated sulfate in Minnesota. It also suggests substantial contributions of biomass burning and soil dust from out-of-state on occasions, although a quantitative separation of local and regional contributions was not achieved in the current study.

Supplemental materials are available for this article. Go to the publisher's online edition of the Journal of the Air & Waste Management Association for a summary of input data, Unmix and PMF factor profiles, and additional maps.

IMPLICATIONS

The IMPROVE and STN data have been widely used for source apportionment by receptor models. Applying more than one model to available PM_2.5 data reveals similarities and differences in source contribution estimates. The inter-model discrepancies identify areas for more focused investigation. Together, the information could lead to better-defined emission reduction strategies.

INTRODUCTION

The state of Minnesota currently attains National Ambient Air Quality Standards (NAAQS)1,2 for carbon monoxide (CO), sulfur dioxide (SO₂), nitrogen dioxide (NO₂), ozone (O₃), PM₁₀ (particulate matter with aerodynamic diameters <10 μm), and PM_2.5 (particulate matter with aerodynamic diameters <2.5 μm). However, enhanced ambient monitoring and a more stringent Minnesota Air Quality Index (AQI) indicates an increase in days classified as “moderate” and “unhealthy” for sensitive groups in Minnesota since 2000.3 The AQI is affected mainly by O₃ and PM_2.5 concentrations. The 24-hr PM_2.5 NAAQS is 35 μg/m³ (98th percentile over 3 years) and the annual average NAAQS is 15 μg/m³ (spatial average over 3 years),4 based on public health criteria.5–7 Dominici et al.8 detected relationships between PM and mortality in the Twin Cities, despite the NAAQS attainment. The Minnesota PM_2.5 Source Apportionment Study (MPSAS) was undertaken to identify and quantify PM_2.5 source contributions via a weight-of-evidence (WOE) approach by applying receptor models to existing chemically speciated PM_2.5 measurements.

Receptor models infer source contributions and atmospheric processes from air quality and meteorological measurements.9–22 Receptor models complement, rather than replace, source-oriented dispersion and chemical transformation models that begin with source emission rates to estimate ambient concentrations.23,24 Both source and receptor models are mathematical representations of reality, requiring simplifying assumptions that create uncertainty. Receptor model assumptions include (1) source profiles are constant over the period of ambient and source sampling; (2) chemical species do not react with each other; (3) the number of sources or source categories is less than or equal to the number of species measured; and (4) measurement uncertainties are random, uncorrelated, and normally distributed. Source models assume a reasonable accuracy for source emissions, air movement, and physical and chemical processes. Source contribution estimates (SCEs) from different models may disagree. A WOE approach25 considers the results and uncertainties of individual methods and reconciles their differences to determine the most plausible range of SCEs. WOE adds confidence to source apportionment and regulatory strategies that are based on it.17,18

MPSAS used chemically speciated PM_2.5 measurements from eight urban and rural monitoring sites across the state, applied effective variance (EV), Unmix, and positive matrix factorization (PMF) solutions for the chemical mass balance (CMB) receptor model, estimated source regions using back trajectories, and evaluated uncertainties for local and regional PM_2.5 source contributions.26 Methods for reconciling the measurements among networks and the EV-CMB SCEs have been described by Chen et al.27 As part of the WOE approach, this paper documents Unmix, PMF, and trajectory-based analyses and reports reconciliation among the different model results.

RECEPTOR MODELING METHODS

Ambient Data and Source Profiles

The Speciation Trends Network (STN; part of the Chemical Speciation Network [CSN])28 was established in 1999 to monitor urban-scale community exposure to PM_2.5. STN has operated four sites in Minnesota, including two (Minneapolis—Phillips Neighborhood [MPN] and St. Paul—Harding High School [SPH]) in the Twin Cities, one in Rochester (ROC), and one at Mille Lacs (ML). MPN, SPH, and ROC are considered urban sites. The ML site is classified as nonurban, as it is located on nonurban tribal lands within 0.1 km of Mille Lacs Lake, a popular tourist attraction. Four Interagency Monitoring of Protected Visual Environments (IMPROVE)29,30 sites are located in remote areas in the north (Voyageurs National Park [VOYA]), northeast (Boundary Waters Canoe Area [BOWA]), southwest (Blue Mounds State Park [BLMO]), and southeast (Great River Bluffs State Park [GRRI]) corners of the state (Figure 1). Chen et al.27 described these sites in greater detail.

Figure 1. STN and IMPROVE PM_2.5 chemical speciation sites in Minnesota. The STN sites are (1) Mille Lacs (ML), (2) Minneapolis—Phillips Neighborhood (MPN), (3) St. Paul—Harding High School (SPH), and (4) Rochester (ROC). The IMPROVE sites are (1) Voyageurs National Park (VOYA), (2) Boundary Waters Canoe Area (BOWA), (3) Blue Mounds State Park (BLMO), and (4) Great River Bluffs State Park (GRRI).

STN and IMPROVE measure 24-hr PM_2.5 mass, elements, ions, organic carbon (OC), and elemental carbon (EC). The networks differ, however, in sampling frequency, measured species, analytical methods, and compensation for passive deposition and organic sampling artifacts.31,32 Particularly, IMPROVE OC and EC are quantified by the thermal optical reflectance (TOR) method, whereas STN OC and EC were measured by the thermal optical transmittance (TOT) method prior to 2007.33 For this study, field blank values were retrieved and subtracted from the STN data and errors were propagated to estimate STN uncertainties that were not available prior to July 15, 2003. A criterion based on material balance screened out IMPROVE and STN data with poor mass closure.27 MPSAS analyzed data acquired between 2000 and 2006, which had the largest number of simultaneous samples across the eight sites. Table S1 (Supporting Information) summarizes the average and signal-to-noise ratios (SNRs) for PM_2.5 mass and chemical concentrations by site.

Region-specific geological source profiles were measured by grab sampling and resuspension,34 as reported by Chen et al.,27 for (1) native soil dust; (2) paved road dust; (3) deicing material; (4) ready-mix concrete plant trackout; and (5) taconite ore and dust. Taconite35 is an iron ore mined primarily in northern Minnesota. The geological materials share common markers such as silicon (Si), aluminum (Al), calcium (Ca), and iron (Fe). Ca and Fe are particularly enriched in concrete plant trackout and taconite dust, respectively. Because in-state source profiles for gasoline vehicle, diesel vehicle, residential wood combustion (RWC), wildfire, coal-fired utility plant, and other emissions were not available, those from other studies36–38 and archived by U.S. Environmental Protection Agency (EPA)39 were used. Secondary sulfate, secondary nitrate, and secondary organic carbon (SOC) were represented by their components in proportion to the stoichiometric ratios, that is, ammonium nitrate (NH₄NO₃), ammonium sulfate ([NH₄]₂SO₄), ammonium bisulfate ([NH₄]HSO₄), and organic matter (OM), in EV-CMB modeling.27

Unmix and PMF Receptor Modeling

Unmix40,41 and PMF42,43 solve the CMB equations44 by identifying source-related “factors” in the data set. Practically, non-negative factor loadings and scores are derived and interpreted as source profiles and source contributions, respectively. Although measured source profiles are not used as inputs to Unmix and PMF, measured profiles are needed to justify the assignment of a source factor to a source type. There is no guarantee that mathematically derived Unmix and PMF factors represent real sources; they often appear to be mixtures of source types that have correlated SCEs in space and or time.22 A needed demonstration of any Unmix or PMF application is that at least one measured source profile can be reasonably associated with each derived source factor.16,21 Unmix and PMF typically require large ambient data sets, for example, more than 100 samples distributed across time and space, for which the underlying source profiles are relatively constant and for which there is large independent variability among actual source contributions. Comparison of source contributions from the EV, Unmix, and PMF CMB solutions provides evidence of the extent to which these assumptions are complied with for a given application.

Ambient data sets used here are identical to those of Chen et al.,27 but they were divided into five groups for Unmix and PMF receptor modeling (Table 1). STN and IMPROVE data were analyzed separately due to the different variables measured and precision estimates. Data from the four STN sites were analyzed together owing to their proximity (<250 km apart) and similar source types. Transport and mixing from one site to another could occur within several hours where PM_2.5 source profiles are likely to remain the same. STN data were separated into post- and pre-July 15, 2003, periods to minimize the effects of changes in STN sampling and analysis protocols that occurred on July 15, 2003.45,46 It provides an opportunity to examine the robustness of Unmix and PMF solutions under changing measurement methods.

Table 1. Data Groupings for Unmix and PMF model inputs. *

Modeling Group^a	Location Code ^b (Network)	Period	Number of Samples Reported	Number of Valid Samples ^c	Fraction of Data Used for Receptor Modeling	PM2.5 Concentration Range (Mean) in μg/m^3
I	ML (STN)	July 14, 2003 ^d –July 30, 2005	234	209	89.3%	0.8–59 (7.4)
	MPN (STN)	July 17, 2003–December 13, 2006	367	334	91%	1.3–31.1 (9.9)
	SPH (STN)	July 20, 2003–March 29, 2005	100	98	98%	1.0–28.8 (9.9)
	ROC (STN)	July 20, 2003–December 20, 2006	189	179	94.7%	1.0–34.6 (10.1)
	Total		890	820	92.1%
II	ML (STN)	September 8, 2002–July 11, 2003	92	90	97.8%	1.2–28.3 (7.7)
	MPN (STN)	January 11, 2001–July 11, 2003	238	222	93.2%	1.3–36.8 (11.8)
	SPH (STN)	September 22, 2001–July 14, 2003	104	97	93.2%	1.5–30.2 (11.2)
	ROC (STN)	September 22, 2001–July 14, 2003	97	87	89.7%	1.9–29.7 (10.7)
	Total		531	496	93.4%
III	BLMO (IMPROVE)	July 25, 2002–December 30, 2005	399	229	57.4%	0.9–30.7 (7.7)
IV	GRRI (IMPROVE)	July 25, 2002–December 30, 2005	392	358	91.3%	1.0–58.2 (8.4)
V	BOWA (IMPROVE)	January 5, 2000–December 30, 2005	641	394	61.5%	0.2–22.2 (4.4)
	VOYA (IMPROVE)	January 5, 2000–December 30, 2005	673	478	71%	0.1–21.5 (4.1)
	Total		1314	872	66.4%
Notes: ^aTotals are provided for groups including more than one site. Modeling group I: STN sites after July 15, 2003; modeling group II: STN sites before July 15, 2003. ^bSee site locations in Figure 1. ^cValidation includes removing incomplete data and data with poor material balance.27 ^dAlthough overall the AQS database shows uncertainties only after July 15, 2003, in this one instance at ML uncertainties were reported on July 14, 2003. *Although the same data are used in this study and in Chen et al.,27 a mislabeling of sampling sites was discovered after final proofing of that paper. Chen et al.27 labeled GRRI data as BLMO and BLMO data as GRRI, and a corrigendum had been issued. This paper presents the correct information.

IMPROVE BLMO and GRRI sites are distant from each other and from BOWA and VOYA; they probably have different source influences, and so their data were modeled separately. Data from BOWA and VOYA were pooled because of their proximity and common exposure to regional sources in the northern part of the state. The number of valid samples ranges from 229 to 872 among the five groups of data. Sample elimination from BLMO, BOWA, and VOYA resulted mostly from incomplete measurements where either Teflon-membrane or quartz-fiber filters were missing. The missing data are random in nature, as PM_2.5 concentrations and EV-CMB source contributions are consistent between the overall and annual periods.27 PM_2.5 loadings ranged from 0.1 to 59 μg/m³ with higher PM_2.5 concentrations found at the urban sites.

Uncertainty weighting coefficients are not required by Unmix software, but they are required inputs for PMF software.47 Among all combinations of species and number of factors that provided Unmix solutions, results were sought that (1) maximized the number of species; (2) maximized the number of factors; and (3) contained the fewest low-SNR species. These combinations were also analyzed by PMF using the inverse analytical precisions of ambient concentrations as weighting coefficients. In subsequent tests, additional species were included and the number of factors was adjusted (−2 to +2) in PMF, as suggested by Reff et al.,48 to achieve source factors that are best associated with measured source profiles. Other criteria for selecting the optimal solutions included consistencies of overall, seasonal, and daily SCEs among Unmix, PMF, and EV-CMB results and of source types in the Minnesota emissions inventory and emission-based conceptual model.27 Potential source regions were further verified by the ensemble air parcel back trajectories (EBTs).

Back-Trajectory Analysis

EBT methods infer source regions associated with different SCE levels based on air mass history represented by backward air mass trajectories.49,50 Receptor–source relationships are represented by the time spent in the respective source grid cell (i, j) by air parcels numerically released from the receptor location and advected backwards in time. For air parcels associated with a particular sample period t, the residence time probability is calculated by

(1)

where m_ij,t is the number of trajectory segment endpoints located in cell i,j divided by the total number of endpoints (M_t ) over the sample period. P_ij,t may be weighted by a species' concentration or factor score for the sample period. The difference (DP _ij ) between weighted and unweighted probabilities,

(2)

(N: total number of samples) reveals areas that are likely to contain sources contributing to the species or factor.51

The Hybrid Single Particle Lagrangian Integrated Trajectory (HYSPLIT) model52 calculated EBTs for each monitoring site using the National Oceanic and Atmospheric Administration (NOAA) Eta (now called North American Mesoscale) Data Assimilation System (EDAS) meteorological data (http://www.arl.noaa.gov/ss/transport/archives.html) with a spatial resolution of 80 km. For each 24-hr sample, 8-day back trajectories were calculated every 3 hr with a starting height of 500 m above ground level. Hourly trajectory endpoints, up to 1536 per day, were reported. Residence times are calculated for every 1° by 1° cell in the modeling domain by summing the number of endpoints that falls within the grid cell.20,53

RESULTS AND DISCUSSION

Multiple Unmix and PMF solutions for each of the five data groups were obtained26 with the most interpretable results summarized in Table S2 (Supporting Information). Unmix and PMF resolved seven common factors (i.e., soil dust, calcium [Ca]-rich dust, diesel and gasoline vehicle exhausts, biomass burning, secondary sulfate, and secondary nitrate) for the Group I and II STN data. A bromine (Br) factor that could not be associated with any source type came out of Group I data. A more reasonable factor emerged from the Group II data that was associated with a road salt deicing profile. The common factors for IMPROVE sites included soil dust, Ca-rich dust, mobile emission, biomass burning, and secondary nitrate. Unmix and PMF resolved separate coal-fired boiler and secondary sulfate factors for the BLMO data set but a mixed factor in other IMPROVE data sets. Factors that could be associated with road salt and taconite dust profiles only emerged at the BOWA and VOYA receptors.

Comparability of Source and Factor Profiles

Minnesota soil/dust profiles are pretty consistent with respect to important markers such as OC, Si, potassium (K), Ca, titanium (Ti), and Fe (Figure 2a). The MPN soil dust profile (MPNSoil), a composite of soil material collected around MPN, contains the lowest carbon abundance (7.6% ± 0.6%) and the highest Si (14.4% ± 1.0%) abundance among all the Minnesota dust profiles, corroborating with soil that has minimal impacts from deposited engine exhaust. High S, NO₃ ⁻, and NH₄ ⁺ fractions in a freeway dust profile (HwayDust) might result in part from deposition of background aerosol. The Ca abundance of 9.4% ± 0.7% for the ROC road dust profile (ROCDust) is much higher than the 3.9–5.9% abundances in the other profiles, consistent with a limestone-rich soil in southeastern Minnesota. Trackout from a ready-mix concrete plant near SPH (MNCal) shows an even higher Ca abundance of 19.9% ± 1.4% (Figure 2b), similar to the abundances found in cement kiln emissions.37 Fe is the dominant component (33–51%) in the taconite dust profiles (MNTac1 and MnTac2), followed by OC, Si, Ca, and manganese (Mn) (Figure 2c). The Mn fraction is higher in taconite (0.3–0.7%) than in other geological materials (<0.2%).

Figure 2. Comparison of PMF geological factors with measured geological source profiles: (a) soil dust, (b) Ca-rich dust, and (c) taconite dust. Error bars in part c (not shown in parts a and b) represent the boot-strapping uncertainties for the Group I taconite factor calculated by PMF. See Table 1 for modeling groups I–V. OCR and ECR indicate OC and EC measured by TOR, whereas OCT and ECT indicate OC and EC measured by TOT.

The PMF geological factors, that is, soil dust, Ca-rich dust, and taconite, derived from the post-July 15, 2003, STN data show the best agreement with the MPNSoil, MNCal, and MNTac2 profiles in terms of major marker species (i.e., Si for soil dust, Ca for Ca-rich dust, and Fe for taconite; Figure 2a–c). Corresponding factors from the pre-July 15, 2003, data also agree for these markers within reported uncertainties. The geological source factors and measured source profiles differ with respect to several trace elements such as Ni, Cu, As and Se that have minor abundances. These elements are often variable within a fugitive dust source category.54,55 The BOWA-VOYA data yield the only IMPROVE taconite factor, consistent with the Fe and Mn abundances in MNTac2 and the sites' proximity to taconite processing facilities (e.g., U.S. Steel Corp. Minntac at Mt. Iron; see Watson et al.46). There is more variability among the three IMPROVE Ca-dust factors, with their Ca fractions ranging from 7.5% to 12.2%, lower than those of STN Ca factors (13.9–18.0%). The BOWA-VOYA Ca factor contains higher abundances of Ti, V, Cu, Zn, Se, and Sr than those of the MNCal profile, implying that (1) the mixture of crustal contributions differs from those measured and/or (2) several source types are collinear for these elements (Figure 2b).

Combustion sources such as gasoline exhaust, diesel exhaust, and biomass smoke contain higher OC and EC abundances than those in the geological profiles (Figure 3a–c). OC and EC by themselves are insufficient to distinguish among the combustion sources.56,57 Both the post- and pre-July 15, 2003, STN data yield, via PMF, a high OC (40.3–64.9%) and low EC (0–1.6%) factors that also contains K abundances of 6.4–6.8% (Figure 3c). This K abundance is similar to that of the Big Bend Regional Aerosol Visibility and Observational (BRAVO) study open burning (BVBURN; 5.7%) profile. The EC/OC ratio in the BVBURN profile (measured by the IMPROVE TOR method) appears to be higher (i.e., 0.24 compared to <0.03 for the biomass burning factors), though this in part reflects different OC–EC splits made by TOR and TOT. Two other factors dominated by OC and EC have lower K contents (<1.6%) and are assigned to gasoline and diesel engine exhaust (Figure 3a–b). The diesel factor has an EC/OC ratio (0.27–0.54 by TOT) higher than that of the gasoline factor (0.13–0.14 by TOT). Higher Fe, Zn, Br, and Pb abundances are associated with the diesel factor, despite their low abundances in measured diesel profiles (LVOnRDIE, NWHDc, and PHDIES; Figure 3a–b). The diesel factor may be influenced by resuspended road dust and tire/brake wear that are not included in engine exhaust samples. The IMPROVE data yield only gasoline and biomass burning factors. The EC/OC ratio (by TOR) ranges from 0.34 to 0.48 for gasoline vehicle exhaust and from 0.16 to 0.20 for biomass burning smoke. Higher Zn and Pb fractions in these IMPROVE gasoline factors may indicate a blend of gasoline and diesel exhaust contributions.

Figure 3. Comparison of PMF combustion factors with measured combustion source profiles: (a) gasoline vehicle, (b) diesel vehicle, and (c) biomass burning (BB). See Table 1 for modeling groups I–V.

Road salt factors were identified by water-soluble Na⁺ (STN) and Cl⁻ (IMPROVE) markers. The Na⁺ and Cl⁻ abundances in the two salt factors (Groups II and V) are 32.9% and 7.4%, respectively. The only standalone coal factor resolved by PMF was at BLMO, and it features Se and S abundances of 0.1% and 12.0%, respectively. The Se/S ratios in the secondary sulfate factors for BLMO, GRRI, and BOWA-VOYA are much lower. Coal-burning contributions were not separated for GRRI, BOWA-VOYA, and STN sites, likely due to insufficient SNR of the marker species for PMF analysis, not because the coal combustion contribution was absent. Ideally, secondary sulfate and secondary nitrate factors contain only the secondary inorganic species, such as sulfate, bisulfate, nitrate, and ammonium, but this is seldom achieved for PMF source factors.22 In this study, the secondary inorganic aerosol factors also contain OC, and EC as well as trace elements at low levels (see Supporting Information).

Unmix factors are presented in the Supporting Information. These factors contain fewer species than the PMF factors and some differ from the PMF factors with respect to common species. However, reasonable temporal correlations (r ² > 0.7) were achieved between the Unmix factors and their counterpart PMF factors.

Source Contribution Estimates

Average SCEs from PMF, Unmix, and EV-CMB solutions are compared in Tables 2 and 3 for the Group I and Group II data, respectively. Individual factors and source types are grouped into categories of (1) secondary, (2) stationary, (3) area, and (4) mobile sources according to U.S. EPA definitions.58 The secondary sulfate and nitrate contributions dominate in most cases, followed in sequence by the mobile-source (a close one at MPN and SPH by EV-CMB), area-source, and stationary-source contributions.

Table 2. SCEs (in μg/m³) by source type, site, and receptor model for STN sites after July 15, 2003 (modeling group I)

Sources/Site, Model		ML			MPN			SPH			ROC
		PMF	Unmix	EV-CMB	PMF	Unmix	EV-CMB	PMF	Unmix	EV-CMB	PMF	Unmix	EV-CMB
Secondary	Sulfate	1.96	2.19	1.93	2.78	2.74	2.65	2.57	2.56	2.44	3.01	2.96	2.83
	Nitrate	1.70	2.32	2.00	2.29	3.03	2.43	2.67	3.42	2.91	2.70	3.41	3.03
	SOC			0.37			0.12			0.13			0.18
	Sum	3.66	4.51	4.30	5.07	5.77	5.20	5.24	5.98	5.48	5.71	6.37	6.04
Stationary	Taconite (dust)	0.05		0.05	0.08		0.09	0.07		0.09	0.04		0.04
	Sum	0.05	0.00	0.05	0.08	0.00	0.09	0.07	0.00	0.09	0.04	0.00	0.04
Area	Soil dust	0.94	0.46	0.48	0.58	0.31	0.33	0.51	0.28	0.29	0.55	0.25	0.25
	Ca-rich dust	0.16	0.52	0.07	0.29	0.59	0.15	0.23	0.48	0.07	0.48	0.92	0.34
	Salt			0.08			0.13			0.12			0.11
	Biomass burning	0.07	0.05	0.46	0.20	0.10	0.83	0.19	0.10	0.71	0.13	0.07	0.77
	Sum	1.17	1.03	1.09	1.07	1.00	1.44	0.93	0.86	1.19	1.16	1.24	1.47
Mobile	Gasoline	2.36	1.25	0.92	3.32	2.72	1.83	3.10	2.25	0.83	2.75	1.85	0.76
	Diesel	0.07	0.34	1.16	0.19	0.68	3.17	0.31	0.76	4.84	0.12	0.55	2.30
	Sum	2.43	1.59	2.08	3.51	3.40	5.00	3.41	3.00	5.67	2.87	2.39	3.06
Total		7.3	7.1	7.5	9.7	10.2	11.7	9.7	9.8	12.4	9.8	10.0	10.6
Measured PM2.5			7.4			9.9			9.9			10.1

Table 3. SCEs (in μg/m³) by source type, site, and receptor model for STN sites before July 15, 2003 (modeling group II)

Sources/Site, Model		ML			MPN			SPH			ROC
		PMF	Unmix	EV-CMB	PMF	Unmix	EV-CMB	PMF	Unmix	EV-CMB	PMF	Unmix	EV-CMB
Secondary	Sulfate	2.67	2.69	2.21	3.66	3.72	3.01	2.86	3.21	2.58	3.18	2.92	2.72
	Nitrate	2.04	2.92	2.33	2.76	4.07	3.35	2.82	4.00	3.43	2.97	4.13	3.55
	SOC			0.29			0.11			0.22			0.22
	Sum	4.71	5.61	4.83	6.42	7.79	6.47	5.68	7.21	6.23	6.15	7.05	6.49
Stationary	Taconite (dust)			0.02			0.06			0.08			0.02
	Sum	0.00	0.00	0.02	0.00	0.00	0.06	0.00	0.00	0.08	0.00	0.00	0.02
Area	Soil dust	0.31	0.33	0.37	0.50	0.64	0.63	0.42	0.59	0.45	0.44	0.44	0.49
	Ca-rich dust	0.16	0.28	0.03	0.41	0.63	0.13	0.32	0.54	0.06	0.76	1.09	0.38
	Salt	0.37		0.25	0.60		0.45	0.54		0.37	0.49		0.39
	Biomass burning	0.04	0.08	0.51	0.10	0.17	0.79	0.09	0.17	0.89	0.03	0.07	0.71
	Sum	0.88	0.69	1.16	1.61	1.44	2.00	1.37	1.30	1.77	1.72	1.60	1.97
Mobile	Gasoline	1.57	0.82	1.43	2.06	1.18	2.91	3.03	1.18	1.63	2.05	1.11	1.31
	Diesel	0.36	0.52	0.90	1.19	1.32	2.53	1.24	1.99	3.96	0.49	0.75	1.90
	Sum	1.93	1.34	2.33	3.25	2.50	5.44	4.27	3.17	5.59	2.54	1.86	3.21
Total		7.5	7.6	8.3	11.3	11.7	13.9	11.3	11.7	13.7	10.4	10.5	11.7
Measured PM2.5			7.7			11.8			11.2			10.7

PMF and Unmix estimate the lowest and highest secondary source contributions, respectively, mainly due to differences in secondary nitrate. Taconite dust is the only stationary source identified by PMF and Unmix. PMF and EV-CMB estimate similar, though negligible, taconite contributions (0.04–0.09 μg/m³) for the post-July 15, 2003, period. Before July 15, 2003, only EV-CMB resolved a taconite contribution at 0.02–0.08 μg/m³. The PMF taconite factors could contain both dust and stack emissions, but they are still comparable to the measured taconite dust profile (MNTac2).

All three solutions estimate similar area source contributions post July 15, 2003 (Table 2), but the dust and biomass burning fractions differ among the solutions. EV-CMB estimates higher biomass burning contributions than the other two solutions for the pre-July 15, 2003, period, whereas the soil dust SCEs appear to be consistent among the three models (Table 3). Much higher salt SCEs are reported by PMF and EV-CMB for the pre-July 15, 2003, period (i.e., 0.25–0.60 μg/m³ compared to 0.08–0.13 μg/m³ post July 15, 2003). A known contamination during STN sample analysis might have resulted in biased Na⁺ concentrations and salt SCEs prior to July 2003.45

For the mobile source category, PMF and Unmix solutions yield higher gasoline exhaust SCEs than diesel exhaust SCEs, whereas the EV-CMB finds the opposite. This disparity is most apparent at MPN, SPH, and ROC after July 15, 2003 (Table 2). The sum of all EV-CMB source contributions exceeds 115% of measured PM_2.5 mass at MPN and SPH. When compared source by source with the PMF/Unmix solutions, most of the excess mass in EV-CMB SCEs is attributed to larger mobile source contributions. Although the STN data were blank-corrected, the correction could be insufficient to compensate for the organic sampling artifacts at the urban sites.31 EV-CMB possibly attributes the organic sampling artifact to mobile, biomass burning, and/or SOC sources that contain large OC abundances. In addition, the mismatch between OC–EC splits in ambient data (by the STN TOT protocol) and in source profiles (by the IMPROVE TOR protocol) influences the diesel SCEs, as EC is the key marker for diesel emissions.27 The PMF and Unmix achieve better mass closure in terms of factor contributions (Table 2 and 3) because PM_2.5 mass is fitted directly by these solutions, whereas PM_2.5 is a floating species in the EV-CMB solution. As a result, reproduction of the PM_2.5 mass is almost guaranteed by PMF and Unmix solutions and is not a stringent performance measure.

Table 4 shows the SCEs for IMPROVE sites (Groups III–V). All three solutions show larger area source contributions than mobile source contributions, consistent with minimal on-road traffic near the sites. The area source SCEs are consistent across solutions (within ±20%) for BOWA, VOYA, and BLMO, though PMF and Unmix attribute less of the area source PM_2.5 to biomass burning than does the EV-CMB solution. At GRRI, the Unmix area source contribution (0.58 μg/m³) is much lower than the PMF or EV-CMB SCEs, whereas PMF and EV-CMB results differ only in the Ca-rich dust contribution (Table 4). EV-CMB reports negligible Ca-rich dust (<0.06 μg/m³) SCEs at the IMPROVE sites and attributes Ca mostly to the coal-fired boiler source. PMF and Unmix resolve the coal-fired boiler source at BLMO; for other sites, however, the coal combustion contribution might be embedded in the secondary sulfate factor. There are also inconsistencies with respect to the mobile source contributions. Except for GRRI, PMF reports mobile SCEs of ∼0.1 μg/m³, compared to SCEs of ∼0.3 μg/m³ by EV-CMB. PMF and EV-CMB mobile SCEs agree better at GRRI (0.53 vs. 0.62 μg/m³). Unmix mobile SCEs are similar to PMF values for BOWA and VOYA but are similar to EV-CMB values for BLMO and GRRI (Table 4).

Table 4. SCEs (in μg/m³) by source type, site, and receptor model for IMPROVE sites (modeling groups V, III, and IV)

Sources/Site, Model		BOWA			VOYA			BLMO			GRRI
		PMF	Unmix	EV-CMB	PMF	Unmix	EV-CMB	PMF	Unmix	EV-CMB	PMF	Unmix	EV-CMB
Secondary	Sulfate	2.23	1.99	1.91	1.97	1.88	1.68	2.68	2.22	2.49	3.31	4.07	3.32
	Nitrate	0.64	0.98	0.85	0.70	1.04	0.87	2.48	3.06	3.00	2.43	3.34	2.77
	SOC			0.03			0.02			0.01			0.04
	Sum	2.87	2.97	2.79	2.67	2.92	2.57	5.16	5.28	5.50	5.74	7.41	6.13
Stationary	Taconite (dust)	0.06	0.17	0.02	0.04	0.13	0.01			0.00			0.01
	Coal utility			0.22			0.20	0.68	0.53	0.36			0.43
	Sum	0.06	0.17	0.24	0.04	0.13	0.21	0.68	0.53	0.36	0.00	0.00	0.44
Area	Soil dust	0.46	0.01	0.22	0.34	0.01	0.15	0.47	0.81	0.38	0.15	0.22	0.16
	Ca-rich dust	0.05	0.50	0.02	0.04	0.38	0.00	0.64	0.92	0.04	0.62	0.35	0.06
	Salt	0.20	0.39	0.04	0.21	0.41	0.04			0.10			0.05
	Biomass burning	0.69	0.34	0.89	0.67	0.27	0.93	0.60		1.20	1.28	0.01	1.23
	Sum	1.40	1.24	1.17	1.26	1.07	1.12	1.71	1.73	1.72	2.05	0.58	1.50
Mobile	Gasoline	0.09	0.05	0.33	0.09	0.05	0.29	0.11	0.24	0.31	0.53	0.47	0.62
	Sum	0.09	0.05	0.33	0.09	0.05	0.29	0.11	0.24	0.31	0.53	0.47	0.62
Total		4.4	4.4	4.5	4.1	4.2	4.2	7.7	7.8	7.9	8.3	8.5	8.7
Measured PM2.5			4.4			4.1			7.7			8.4

Unlike those at urban STN sites, the sum of all IMPROVE EV-CMB source contributions agrees with measured PM_2.5 concentration within ±3%. For IMPROVE sites, ambient and source profile OC and EC were measured by the same IMPROVE TOR protocol. Diesel contributions, which might be overestimated by EV-CMB for the STN sites, are not detected at the IMPROVE sites. The IMPROVE network corrects the organic sampling artifacts with backup filters31 that show larger OC areal densities than the 15-min duration field blanks used for STN data. SOC is a minor (0.01–0.04 μg/m³) contributor at the IMPROVE sites compared to SOC contributions at the STN sites (0.12–0.37 μg/m³), which might be due to an uncompensated OC artifact.

Reconciled Source Attribution for Urban and Rural Areas

From the WOE viewpoint, fractional source contributions (FSCs) for “urban” and “rural” areas of Minnesota may be best estimated by the mean or median FSCs across STN and IMPROVE sites, respectively, and among PMF, Unmix, and EV-CMB results (Figure 4). Uncertainties of the urban and rural FSCs are then determined from the 10th and 90th percentiles of all (4–12) estimates. These mean FSCs sum to a good mass closure, that is, 105–109% (see Table 5). Figure 4a–b confirms a similar source apportionment between post- and pre-July 15, 2003, STN data, with larger differences for taconite dust (1–3%), salt (1–4%), and gasoline (16–21%) contributions. Chow et al.56 suggested difficulties in determining the gasoline-diesel split accurately without organic marker compounds. Overall mobile FSCs (gasoline + diesel) are found more consistent at 29–34%. The average mobile FSC is only 4% of PM_2.5 at the nonurban IMPROVE sites where secondary sulfate and biomass burning fractions are higher (Figure 4c). The higher biomass burning contribution at IMPROVE sites (i.e., averaged 13%) may be due to prolonged smoldering emissions from wild or prescribed fires.59,60

Table 5. FSCs (% of PM_2.5) from this study compared with those from similar environments

		STN						IMPROVE
Source Category	Source	This Study Post-July 15, 2003	This Study Pre-July 15, 2003	Detroit I,^a Michigan (PMF)	Detroit II, ^b Michigan (PMF)	Chicago, ^c Illinois (PMF)	Chicago, ^c Illinois (CMB)	This Study	BOND. ^d Illinois (PMF)	MACA, ^e Kentucky (PMF)
PM2.5 (μg/m^3)		7.4–10.1	7.7–11.8	16.0	15.8	15.0	15.0	4.1–8.4	10.9	12.7
Secondary	Sulfate	27	29	31	27	37	31	41	27	49
	OP-rich Sulfate									16
	Nitrate	28	31	28	25	21	20	28	24	3
	SOC	2	2					0
Stationary	Taconite (dust)	1	3					1
	Iron/Steel/Mn			3		4	2
	Copper smelter					1			2
	Lead smelter								4
	Metal plating				3				5
	Industrial zinc				5
	Industrial					2
	Refinery						0
	CFPP						1	6		1
Area	Soil dust	5	4	8	4	6	2	5	6	5
	Ca-rich dust	4	4				1	5	3
	Salt	1	4	4		3		4		3
	BB	3	3	3	2	5	11	13	12	14
Mobile	Gasoline	21	16	15	21 ^f			4	17 ^f	7
	Diesel	13	13	4	14					3
	Mobile					23	31
	Sum	105	109	96	101	102	99	107	100	101
Notes: ^aDetroit Allen Park, November 1, 2000–May 1, 2005.62 ^bDetroit Allen Park, January 1, 2002–March 31, 2005.61 ^cChicago Lawndale, 2001–2003.63 ^dBondville, IL, 2002–2004.61 ^eMammoth Cave, KY, 2002–2004.64 ^fMixture of gasoline and SOC contribution.

Figure 4. Model-collective estimates (i.e., square and number: average; bar: median; box: 10th and 90th percentiles) of fractional source contributions (FSCs) for (a) STN sites after July 15, 2003; (b) STN sites prior to July 15, 2003; and (c) IMPROVE sites.

This is the first use of available Minnesota data for statewide PM_2.5 source apportionment, so opportunities for comparison with other studies are limited. Receptor modeling, especially PMF, has been applied to several Midwestern STN sites, including Detroit and Chicago,61–63 and rural IMPROVE sites at Bondville (BOND), Illinois, and Mammoth Cave National Park (MACA) in Kentucky61,64 (see Figure S1 in Supporting Information), and these are compared with the Minnesota SCEs in Table 5. The Detroit and Chicago sites are located in the upper Midwest and likely influenced by sources similar to those affecting MPN and SPH. Buzcu-Guven et al.61 and Gildemeister et al.62 resolved eight factors for Detroit, including regional sources such as secondary sulfate, secondary nitrate, soil dust, biomass burning, and mobile sources (gasoline, diesel). Gildemeister et al.,62 based on STN data mostly acquired prior to July 15, 2003, reported a salt factor responsible for ∼4% of PM_2.5 mass. Detroit-specific industrial factors such as steel, metal plating, and zinc were not found for Minnesota. As for Chicago, PMF and EV-CMB both identified nine factors/sources,63 though the two solutions differ notably in industrial and biomass burning sources (Table 5). Gasoline and diesel contributions were not separated at Chicago by either PMF or EV-CMB analysis.

SCEs at BOND61 differ from those at Minnesota IMPROVE sites. Despite its rural nature, industrial sources such as copper, lead smelting, and metal plating accounted for 11% of BOND PM_2.5 and mobile source contributions constituted as much as 17% of PM_2.5. Buzcu-Guven et al.61 attributed these contributions to long-range transport from Indiana and Ohio. Such strong industrial impacts were not found at MACA. Zhao and Hopke64 reported a mobile FSC (10%) higher than those found at Minnesota rural sites. Secondary nitrate contributes less at MACA, possibly due to elevated sulfate concentrations and warmer weather in winter.65 About 16% of MACA PM_2.5 mass was associated with an optical pyrolysis (OP)66,67-rich sulfate factor that Zhao and Hopke64 attributed to acidic catalyzed SOC.

Potential Source Regions

PMF-factor-score-weighted probability maps (DP _ij ) were generated for each STN and IMPROVE site. These maps are more useful for estimating locations of regional sources, often including secondary sulfate, secondary nitrate, windblown dust, and large-scale wildfires or prescribed burns. Spatial resolution of the trajectories is too coarse to resolve urban-scale sources such as vehicle exhaust and localized industrial emissions.16 Figure 5 shows these maps for ROC and GRRI, a STN-IMPROVE pair that are only ∼70 km apart despite differing in their urban-rural setting. Factor probability maps are expected to be similar between ROC and GRRI if the factor is dominated by regional-scale sources and/or long-range transport.

Figure 5. EBT probability maps for (a, b) secondary sulfate, (c, d) soil dust, (e, f) biomass burning, and (g, h) secondary nitrate factors resolved by PMF for the ROC and GRRI sites.

Occurrences of high secondary sulfate factor scores correspond with a region in Central Plains stretching from Great Lakes to northern Texas (Figure 5a–b). This region, particularly the lower Midwest, contains the highest densities of coal-fired power plants in the country.68 Coal combustion is a major source of SO₂, of which conversion rate to sulfate varies with season and weather conditions. Trajectories weighted by the “soil dust” factor suggest both local (for ROC only) and potential intrastate sources (Figure 5c–d). Major dust sources are located along a corridor from the south of Minnesota into neighboring states of Iowa, Missouri, and Arkansas (Figure S2.)

Wild and prescribed fire locations determined from the Moderate Resolution Imaging Sepctroradiometer (MODIS) satellite confirms large areas of agricultural burning in the Central Plains and most of the southeastern states (see Supporting Information). Another region with biomass burning appears northwest of Minnesota and runs across the U.S.-Canadian border into the provinces of Manitoba and Saskatchewan. Biomass burning probability maps (Figure 5e–f) do not reveal source regions as clearly as those for secondary sulfate and dust factors, though the Central Plain–southeastern states and northwestern Minnesota are indeed highlighted for ROC and GRRI, respectively. Transport of wildfire smoke from the northern Rockies also shows up in both maps. The secondary nitrate weighted trajectories show a common source region to the west of ROC/GRRI (Figure 5g–h). This is consistent with ammonia (NH₃) emissions in the vast rural areas of Minnesota, North Dakota, South Dakota, Nebraska, and Iowa, as ammonium is the counterpart of nitrate in the secondary nitrate formation.69 Another source region to the east of Minnesota also shows up in the ROC map.

CONCLUSIONS

This study represents the first effort to apply multiple receptor models to both urban and rural PM_2.5 monitoring data across the state of Minnesota. PMF and Unmix factors were associated with source types based on comparison with measured source profiles. There was evidence of some collinearity among source factors, but not so large that it would influence the source attribution. PMF, Unmix, and EV-CMB reached consensus on (1) secondary sources (sulfate, nitrate, and organic carbon) dominate PM_2.5 mass, particularly at rural sites; (2) mobile sources outweigh area sources including fugitive dust and biomass burning at urban sites, and vice versa at rural sites; (3) taconite production and coal-fired boilers were the only stationary sources, with minor primary PM_2.5 contributions when compared to the other source categories; (4) gasoline and diesel exhaust contributions were separated only at the urban STN sites, but the large variability among the SCEs from different CMB solutions indicate that the split is not accurate; and (5) secondary sulfate and nitrate were similar for all of solution methods. SOC was resolved only by EV-CMB.

Overall urban and rural PM_2.5 source contributions were estimated by averaging fractional source contributions (FSCs) from PMF, Unmix, and EV-CMB for STN (urban) and IMPROVE (rural) sites. This led to a good mass closure with respect to resolved source/factor contributions, that is, 105–109%. Urban FSCs (post July 15, 2003), by the decreasing order, are secondary nitrate (28%), secondary sulfate (27%), gasoline engine exhaust (21%), diesel engine exhaust (13%), soil dust (5%), Ca-rich dust (4%), biomass burning smoke (3%), SOC (2%), salt (1%), and taconite dust (1%). Compared with source apportionment studies for Detroit and Chicago, major differences appear on the type and magnitude of stationary sources that may be location-specific. FSCs for the Minnesota rural IMPROVE sites are secondary sulfate (41%), secondary nitrate (28%), biomass burning (13%), coal utility (6%), soil dust (5%), Ca-rich dust (5%), gasoline vehicle (4%), salt (4%), SOC (<1%), and taconite dust (1%). Coal utility contributions were not commonly resolved for Midwestern IMPROVE sites, but a few other studies (e.g., Buzcu-Guven et al.61 and Zhao and Hopke64) reported multiple sulfate factors implying fresh and aged aerosols from coal combustion.

Ensemble back-trajectory analysis was consistent with results from similar studies that identified emitters in lower Midwestern states as the largest contributors to secondary ammonium sulfate. Transport from the north tended to produce levels that were below sulfate annual averages. Biomass burning caused by wildfires and prescribed burns may have large contributions from out-of-state on occasions, as indicated by the satellite observations and trajectory analysis. Regional transport is likely important for biomass burning and dust particles concentrations at Minnesota sampling sites, in addition to substantial local RWC and road dust emissions. Further separation of regional and local burning or dust contributions is not achieved in this study but may be accomplished in the future with more specific organic markers, which will also help narrow down the uncertainties in mobile source contributions.

ACKNOWLEDGMENTS

This study was supported, in part, by the Minnesota Pollution Control Agency (MPCA). The information expressed in this paper does not necessarily reflect the MPCA's policies. The authors thank D. Trimble, S. Kohl, and S. Chancellor for their assistance in the laboratory and data analysis and H. Wiegner for helping collect the Minnesota geological dust samples.