Fine Particulate Matter Source Apportionment for the Chemical Speciation Trends Network Site at Birmingham, Alabama, Using Positive Matrix Factorization

By Baumann, Karsten Jayanty, R K M; Flanagan, James B

ABSTRACT The Positive Matrix Factorization (PMF) receptor model version 1.1 was used with data from the fine particulate matter (PM^sub 2.5^) Chemical Speciation Trends Network (STN) to estimate source contributions to ambient PM^sub 2.5^ in a highly industrialized urban setting in the southeastern United States. Model results consistently resolved 10 factors that are interpreted as two secondary, five industrial, one motor vehicle, one road dust, and one biomass burning sources. The STN dataset is generally not corrected for field blank levels, which are significant in the case of organic carbon (OC). Estimation of primary OC using the elemental carbon (EC) tracer method applied on a seasonal basis significantly improved the model’s performance. Uniform increase of input data uncertainty and exclusion of a few outlier samples (associated with high potassium) further improved the model results. However, it was found that most PMF factors did not cleanly represent single source types and instead are “contaminated” by other sources, a situation that might be improved by controlling rotational ambiguity within the model. Secondary particulate matter formed by atmospheric processes, such as sulfate and secondary OC, contribute the majority of ambient PM^sub 2.5^ and exhibit strong seasonality (37 +- 10% winter vs. 55 +- 16% summer average). Motor vehicle emissions constitute the biggest primary PM^sub 2.5^ mass contribution with almost 25 +- 2% long-term average and winter maximum of 29 +- 11%. PM^sub 2.5^ contributions from the five identified industrial sources vary little with season and average 14 +- 1.3%. In summary, this study demonstrates the utility of the EC tracer method to effectively blank-correct the OC concentrations in the STN dataset. In addition, examination of the effect of input uncertainty estimates on model results indicates that the estimated uncertainties currently being provided with the STN data may be somewhat lower than the levels needed for optimum modeling results.

(ProQuest: … denotes formulae omitted.)

INTRODUCTION

Epidemiological studies suggest that ambient particulate matter (PM) has significant associations with adverse respiratory and cardiovascular health effects,1-5 which prompted the U.S. Environmental Protection Agency (EPA) to promulgate National Ambient Air Quality Standards (NAAQS) in July 1997. The majority of the earlier epidemiological studies focused on linking human exposures to the mass, chemical components, and size of PM. More recent studies have been conducted to understand associations between PM emission sources and human exposure.6-9 To provide nationally consistent data for the assessment of trends, EPA also established, in association with the PM NAAQS, the fine PM (PM^sub 2.5^) Chemical Speciation Trends Network (STN) program.10,11 In this program, 24- hr integrated filter-based samples are collected every 3 or 6 days at each monitoring site and analyzed to determine gravimetric mass and chemical composition, including ions, trace elements, and carbonaceous fractions (i.e., organic carbon [OC] and elemental carbon [EC]).

Various PM^sub 2.5^ source apportionment studies have been conducted to understand the sources and contributions of PM^sub 2.5^ in the southeastern United States.12-20 These studies demonstrate the applicability of different source apportionment methods and models to testing spatial and temporal representativeness of model results, which supports efforts to develop PM^sub 2.5^ management strategies with greater promise of success for different locations in the southeastern United States. Most of these studies applied the Chemical Mass Balance (CMB) and Positive Matrix Factorization (PMF) models, supported by other observation-driven statistical manipulations (species ratios, normalizations, specific marker additions, and trajectories) to assess or improve the mathematical performance of the models. A few recent studies also compared source apportionment results from different receptor models applied to the same dataset, demonstrating that the resulting differences are due to the inherent uncertainty of the multivariate approach.21-23 Hence, the estimation of the absolute uncertainty of the receptor model results remains difficult, especially with respect to the limited spatial representativeness of the results across large urban areas (possibly due to the influence of different local sources). A recent focus on the use of PMF for PM^sub 2.5^ source apportionment called for more thorough and transparent documentation to help develop standard protocols and procedures necessary to guide the user along the various decisions in the process, to create results that are consistent and reproducible, and to indicate where future research is needed to improve the model’s general applicability.24 In this paper, we follow this call closely by providing detailed information on the reasoning, sensitivities, and effects of certain decisions made.

The main goal of this study was to conduct source apportionment of PM^sub 2.5^ using a large dataset from a highly industrialized urban STN site. The North Birmingham site (designated NBHM from here on) has a large amount of heavy producing industry, which is not typical for most other urban agglomerations in the southeastern United States (which are dominated by transportation and business services), and it was chosen partly for that reason. The NBHM site was also chosen because (1) it is in an EPA-designated nonattainment area for PM^sub 2.5^; (2) it was part of some of the above- mentioned investigations, and so provides a certain history of PM^sub 2.5^ characterization; (3) it was the subject of a detailed “evidencebuilding” study that investigated source impacts from strictly observational concentration ratio analyses;25 (4) it is sampled at a higher frequency (every third day instead of every sixth), which produces a larger dataset resulting in increased statistical robustness; and (5) it employs the MetOne SASS sampler, which is the most widely used sampler in the STN.

The 3-yr average (2002-2004) of the annual average PM^sub 2.5^ mass concentration is 17.5 [mu]g.m^sup -3^ for NHBM,25 which clearly exceeds the current NAAQS for PM^sub 2.5^ of 15 [mu]g.m^sup -3^ and therefore forces the state of Alabama to develop a state implementation plan (SIP) for effective reduction of ambient PM^sub 2.5^ concentrations. The SIP must be submitted to EPA by April 2008. Demonstrating the feasibility and utility of STN data, the PMF receptor model26 is applied here in a systematic, comparative way to examine issues such as OC blank correction (here indirect and retroactive), uncertainty, and outlier treatment, as well as to build confidence in its results, help identify local primary sources, and estimate their contributions to the observed ambient PM^sub 2.5^ with acceptable uncertainty.

METHOD AND APPROACH

PM^sub 2.5^ filter samples integrated over 24-hr (12:00 p.m. to 12:00 p.m.) have been collected at the NBHM site every third day since January 2001. The NBHM site (Aerometric Information Retrieval System [AIRS] code 010730023, latitude 33.553056, longitude 86.814900) is located in North Birmingham at 3009 28th Street, Birmingham, AL 35207, which is in Jefferson County, and is categorized as being in an urban and center city setting. Figure 1 and Table A-1 (Appendixes Table A-1 through A-4 and Figure 1-A are provided as supplemental material available online only at http:// secure.awma.org/journal/pdfs/2008/1/10.3155-1047- 3289.58.1.27_supplmaterial.pdf) provide a combined overview of the site’s location and the different emission sources surrounding it. Three steel pipe manufacturers; two large coking operations; a mineral wool plant; an asphalt batching plant; and fugitive dust sources from coal and coke storage yards, limestone quarries, and metal fabricating operations dominate the industrial activities within 10 km of the site.25 The site is also near three major interstate highways, two large state highways, local roads with heavy traffic, significant railroad traffic with diesel locomotives serving the nearby manufacturing plants, and the Birmingham International Airport (BHM) approximately 5 km to the east- northeast. BHM is Alabama’s largest airport, serving more than 3 million travelers annually with more than 160 arrivals and departures daily, most between 6:00 a.m. and 6:00 p.m.

The PM^sub 2.5^ samples were collected on Teflon, nylon, and quartz fiber filters at 6.7 L/min downstream from individual sharp cut cyclones (SCCs) of a SASS aerosol speciation sampler (MetOne Instruments). Gravimetric and elemental mass concentrations were determined from the Teflon filters, with a total of 48 elements measured via energy-dispersive X-ray fluorescence (XRF). The nylon filters were used for the analysis of cations (ammonium [NH^sub 4^^sup +^], sodium [Na^sup +^], and potassium [K^sup +^]) and anions (sulfate [SO^sub 4^^sup 2-^] and nitrate [NO^sub 3^^sup -^]) by ion chromatography. The quartz fiber filters were used for OC and EC applying the thermal optical transmittance (TOT) method.27 Initial Data Screening

The original dataset contained 602 data records, from January 13, 2001 to February 13, 2006. Potential outliers were identified by application of a simple plausibility test using the mass balance approach, where the sum of all identified species concentrations (Sum ID) is treated as “reconstructed mass” and compared with the measured gravimetric mass. For this test, major metal oxides (MetOx) mass concentration was calculated from the XRF-detected concentrations of aluminum, silicon, calcium, iron, and titanium for silicate minerals according to Malm et al.,28 as shown in eq 1.

[MetOx] = 2.2[Al] + 2.49[Si] + 1.63[Ca] + 2.42[Fe] + 1.94[Ti] (1)

The [MetOx] was then added to the other identified mass concentrations, including the organic mass (OM), which was assumed to be equal to 1.4 x [OC]. Figure 2 shows Sum ID as a percentage of the measured total PM^sub 2.5^ gravimetric mass (reported as TMAC) versus TMAC.

The data were screened for outliers by comparing the Sum ID mass fraction to the measured gravimetric mass: points where the Sum ID lay outside the range of 80-120% of the measured gravimetric mass were considered outliers. Of the 602 initial data points, 87 were eliminated as outliers at this stage, leaving 515 data points, a reduction of almost 15%. The remaining dataset was still large enough to allow robust uncertainty analysis via PMF bootstrapping, which is described below.

After screening for outliers, the concentration of each species was normalized to its long-term average, and the resulting time series for each species was investigated for potential additional outliers. Figure A-1 in the online supplemental material shows the normalized concentrations time series of the 12 most prominent species (i.e., the ones with the highest signal-to-noise ratio [S/ N]). Seasonal variations become apparent with opposing amplitudes for SO^sub 4^^sup 2-^ versus NO^sub 3^^sup -^ (i.e., summer highs and winter lows for SO^sub 4^^sup 2-^ versus summer lows and winter highs for NO^sub 3^^sup -^). Intermittent spikes in Mn, Fe, and Zn appear simultaneously, possibly indicating same-source influence, whereas K^sup +^ shows sporadic spikes around July 4 and January 1, pointing to the influence of celebrations involving fireworks. The exclusion of those samples had relatively little influence on the PMF model results described later.

Data Uncertainties

For each reported species concentration value, a corresponding uncertainty (UNC) was calculated according to eq 2:

… (2)

where MDL is the annually averaged method detection limit; P is the concentration-dependent precision, combining analytical and field sampling uncertainties; and CONC is the individually measured species concentration.

Species MDL values were resolved annually, and they varied slightly because of variations in the individual MDL for each laboratory instrument, particularly for the XRF instruments used to analyze the trace metals on the Teflon filters. EPA’s Air Quality System database system (http://www.epa.gov/ttn/airs/airsaqs/) contains pointwise MDL and uncertainty values for individual samples/ days taken in July 2003 or later, but not for samples taken earlier. Annual average MDL values were computed for NBHM and used in eq 2 to minimize irrelevant point-topoint variations in the input dataset.

For the screened data, a P value of 7.1% was chosen for most species, with the following exceptions: Na, 9.8%; Mg, 8.5%; Al, 8.6%; and Si, 8.4%. The P values were selected to be similar to those that have been used to estimate the uncertainty values in EPA’s AQS database system. Analysis of data from sites with side-by- side STN samplers29 indicates that 7.1% is reasonable for many species, with a few very well behaved species (e.g., sulfate) having slightly lower P values. Other species, which are present at lower levels than SO^sub 4^^sup 2-^ and are prone to filter contamination, exhibit significantly higher P values. The P values selected for Na, Mg, Al, and Si were increased because of known difficulties associated with analysis of very light elements by the XRF method.30- 32

Table A-2 (see online supplemental material) summarizes the resulting relative uncertainty averages and standard deviations for each species. Table 1 provides the arithmetic and geometric means, standard deviations, minimum and maximum values, fraction of values below detection limit (BDL), and MDL averages from the five annual means for the species measured. Only species with S/N values greater 0.5 are listed. S/N ratios are automatically computed for each species as shown in eq 3.

… (3)

If the measured value of any species was below the MDL, then its CONC value was replaced by MDL/2 (i.e., the center of the MDL increment >0), and its corresponding UNC value was set to the MDL, so that regression results for low CONC values were weighted equally. From the original 602 records, only 18 were missing because of partially missing species data; therefore, no additional missing data records occurred for any of the species.

Values for primary OC (POC) estimated from application of the EC tracer method (described below), and for the new total mass resulting from the difference of TMAC and secondary OC (SOC = OC – POC) are listed in both tables as well. Because POC replaced the original OC, TMAC was adjusted accordingly (i.e., by subtracting SOC from it) to maintain the original proportionality for the PMF. TMAC is typically an underestimate of the actual mass because of the loss of semi-volatiles in the equilibration step. The unreported mass of other organic elements (OM – OC) correlates differently with POC and SOC. Separating SOC from POC reduces the effect of the OM/OC multiplier, because the PMF analysis can scale the results to account for the amounts of TMAC that correlate with POC and SOC. That is also one reason TMAC enters the PMF analysis with very low weight (i.e., TMAC is weak; see Estimating Uncertainties of Modeled Results section). Determination of POC concentration and its associated uncertainty is described in the following section.

Estimating POC

In the STN program, aerosol carbon is measured as OC and EC using a thermal-optical method. OC aerosol can be directly emitted to the atmosphere in the particulate form (primary) or can be produced by gas-to-particle conversion processes (secondary). EC is emitted mainly from combustion sources. Because POC and EC are mostly emitted from the same sources, EC can be used as a tracer for primary combustion-generated OC.33-35 The calculated POC values then replace the original OC values in the dataset that is submitted to PMF.

The atmospheric formation of secondary organic aerosol (SOA) increases the ambient concentration of OC and the ambient OC/EC ratio. OC/EC ratios exceeding the expected primary emission ratio are an indication of SOA formation. For southern and central California, 30-80% of the total OC has been identified as secondary in summer. 33,34,36-38 Primary ratios of OC/EC vary from source to source and show temporal and diurnal patterns,33 but because EC is emitted only by combustion sources, gaseous tracers of combustion (carbon monoxide [CO], nitric oxide [NO], oxides of nitrogen [NO^sub x^]) can be used to determine periods dominated by primary aerosol emissions. Ozone, its production efficiency estimated from regressions with NO^sub z^ (NO^sub y^ – NO^sub x^), and the sulfate- tototal-oxidized-S ratio are possible indicators of photochemical activity and can be used as indicators for periods when SOA production should be expected. In this case, increases in the OC/EC ratio correlated to more general ozone episodes are indicative of SOA production.39

The EC tracer method takes advantage of the fact that POC and EC are both emitted by most combustion sources. Primary ratios of OC to EC, (OC/EC)^sub p^, can be determined from a subset of ambient measurements if a large dataset is available and if conditions for SOA production are unfavorable.34,35 Alternatively, (OC/EC)^sub p^ can be derived from an emissions inventory of the principal sources for an area of interest.33,40 In this paper, the EC tracer method was applied using data from filters having the lowest 20% of OC/EC ratios measured in each season, in which the four seasons are defined as follows: spring (March, April, and May), summer (June, July, and August), fall (September, October, and November), and winter (December, January, and February).

Assuming that POC can be defined by

POC = (OC/EC)^sub p^ * EC + b (4)

the contribution of SOC can then be estimated as

SOC = OC – POC (5)

where (OC/EC)^sub p^ is the ratio of OC to EC for the primary sources affecting the site of interest, and b is the noncombustion contribution to the POC.34,35 A negative value of b can be interpreted as POC correction because of the presence of noncombustion or other EC (without associated OC, e.g., mechanical abrasion in industrial processes involving grinding of polymorphic carbon-containing minerals). EC is the measured EC concentration, SOC is the SOA contribution to the total OC, and OC is the measured nonblank-corrected OC concentration reported in the STN dataset. All of these parameters are time-dependent because of the temporal variations in anthropogenic emissions and meteorology. The application of the EC tracer method requires the determination of the (OC/EC)p ratio and b for the area and period of interest from the linearly regressed slope and intercept of the measured OC versus EC. A recent study showed that the Deming linear least-squares regression is the superior method among several ambient EC tracer methods.41

Table A-3 (see online supplemental material) summarizes the slope (OC/EC)^sub p^ and intercepts (b) of the Deming linear OC versus EC regressions for the different seasons. In a recent study by Yu et al.,42 (OC/EC)^sub p^ values were obtained from an emission/ transport model (EPA’s Models-3/Community Multiscale Air Quality [CMAQ] model). The EC tracer method was then applied to IMPROVE (Interagency Monitoring of PROtected Visual Environments) and SEARCH (South-Eastern Aerosol Research and CHaracterization) network data for a 10-week summer period in 1999 to estimate the temporal and spatial distributions of primary and secondary OC aerosols over the continental United States. The modeled (OC/EC)^sub p^ for the NBHM site averaged 1.59 +- 0.36, although the authors acknowledged significant temporal variability in (OC/EC)^sub p^ even over a relatively small period of time (10 weeks). The 3-month summer values (June, July, August) determined from our more recent observations (2001-2006), average 3.21 +- 0.78, indicating significant year-to-year variability. For each of the 515 samples remaining (after outliers were removed based on the reconstructed mass Sum ID), concentration values (CONC) for POC were calculated based on eq 4, except that the intercept b was ignored, because the original OC values for STN data (unlike IMPROVE) are not blank- corrected, leaving the intercept highly susceptible to potential sampling artifacts. Omitting the intercepts builds upon procedures suggested by Solomon et al.43 and applied by Tolocka et al.,44 Kim et al.,45,46 and Qin et al.47 It assumes here that the positive and negative OC artifacts outweigh the noncombustion contributions of OC and EC, with the latter driving the occurrence of negative intercepts. All 21 intercepts listed in Table A-3 average 0.25 [mu]g.m^sup -3^, but at a relatively high standard error (SE) of +- 1.09 [mu]g.m^sup -3^. The smallest and largest seasonal average is – 0.01 [mu]g.m^sup -3^ in fall and 0.37 [mu]g.m^sup -3^ in summer. The POC uncertainties were determined from propagation of errors of the slope, EC, and OC. A new total gravimetric mass, NTM, was calculated by subtracting SOC from TMAC, and its uncertainty was estimated as the RMSE of the individual uncertainties. The POC CONC averaged 4.55 +- 3.21 [mu]g.m^sup -3^ and UNC averaged 1.76 +- 0.82 [mu]g.m^sup – 3^, whereas the POC fraction of OC was 66 +- 24% and its fraction relative to the total original gravimetric mass, TMAC, was 22 +- 10%. The relative contributions of SOC to OC and TMAC are correspondingly 34 +- 24% and 11 +- 8%, respectively.

Figure 3 shows the seasonal trend of SOC/OC fraction versus the OM/OC ratio with OM determined from mass closure. Despite the large uncertainty in both values from individual error propagation, a trend becomes apparent for an increased fraction of SOC with an increased OM/OC in the summer relative to the winter months, possibly indicating the presence of more highly oxygenated OC in summer, although different levels of particlebound water will likely contribute to the observed seasonal difference. The overall mean for OM/OC appears to be approximately 1.4 +- 0.01 (SE of total mean), with a summer average of 1.54 +- 0.03 and a winter average of 1.27 +- 0.03. Coincidentally, the factor 1.4 has been widely used in the past, originating from very limited theoretical and laboratory studies from more than 20 yr ago that suggested this value to be the lowest reasonable estimate for urban aerosols.48-50 More recent investigations, 51,52 however, suggest a factor of 1.6 +- 0.2 and 2.1 +- 0.2 to be more accurate in urban and nonurban environments during the summer, respectively. Our summer value is slightly lower and possibly indicative of this particular site being more heavily influenced by primary emissions from nearby industrial sources, making the influence from secondary atmospheric sources less important.

Estimating Uncertainties of Modeled Results

Receptor models apply factor analysis to concentrations of particles (and gases) that are (1) emitted from sources of potential influence to the receptor, and (2) measured in ambient air at the receptor location, to both identify the presence of and quantify source contributions to the ambient pollution level. However, all models trying to apportion PM^sub 2.5^ mass concentrations observed in different environments suffer from inherent difficulties of accurately estimating uncertainties of the models’ results. This problem has been recently addressed with mixed success in numerous applications to study urban air quality and human exposure issues.13,22,45,53-60

Three fundamental assumptions underlie all the receptor models: (1) PM compounds are present in different proportions in different source emissions; (2) these proportions remain relatively constant for each source type over the evaluated measurement period; and (3) changes in these proportions between source and receptor, i.e., the chemical transformation during atmospheric transport and dispersion of emissions, are negligible or can be approximated. Here, we employed EPA version 1.1 of the PMF receptor model, solving the multilinear regressions using constrained, weighted, least- squares.26,61,62 The general model assumes that “p sources, source types, or source regions (factors) impact a receptor, and linear combinations of the impacts from the p factors give rise to the observed concentrations of the various species, so that

… (6)

where x^sub ij^ is the concentration at a receptor for the jth species on the ith day, g^sub ik^ is the contribution of the kth factor to the receptor on the ith day, f^sub kj^ is the fraction of the kth factor that is species j, and e^sub ij^ is the residual for the jth species on the ith day”.26 In EPA PMF1.1, it is assumed that only the receptor concentrations (x^sub ij^) are known, and the goal is to estimate the contributions g^sub ik^ and the fractions (or profiles) f^sub kj^. It is furthermore assumed that the contributions and mass fractions are all non-negative, hence constraining the least squares. Additionally, EPA PMF1.1 allows the user to set the uncertainty of each x^sub ij^. Species days with large uncertainty are not allowed to influence the estimation of the contributions and profiles as much as those with small uncertainty, hence weighting the least squares. The task of EPA PMF1.1 is to minimize the sum of squares, Q (goodness of fit parameter), as follows:

… (7)

where s^sub ij^ is the uncertainty in the jth species for day i. EPA PMF1.1 operates in a robust mode, meaning that outliers are not allowed to overly influence the fitting of the contributions and profiles. PMF does this in the robust mode by dynamically reducing the weights (s^sub ij^) for points that fit poorly through an iterative process until each such point’s residual falls within the critical limit of 2s^sub ij^. If the model is appropriate for the data and if the uncertainties specified are truly reflective of the uncertainties in the data, then Q should be approximately equal to the number of sample records in the concentration dataset. If all species are perfectly accounted for by the computed factors and if all input uncertainties fully account for all true uncertainty, then the computed robust Q should be equal to or at least within a factor of 2 of the theoretically ideal Q (Q^sub 0^) on the basis of the number of samples (n), sources (factors) (p), and the numbers of “strong” and “weak” species (m^sub strong^ and m^sub weak^, respectively), such that63

Q^sub 0^ * n +- (m^sub strong^ [mu] m^sub weak^/3 + p) (8)

Because Q heavily depends on the uncertainties of the input concentrations, we conducted three different EPA PMF1.1 model runs comparing the effects of different input uncertainties on model performance and results. In all three comparison runs, the total gravimetric mass was categorized as weak, therefore its input uncertainty values were downweighted by a factor of 3, as were all species with 0.5 < S/N < 2 (As, Br, Mg, Ni, Ti, and V; see Table 1), to reduce their influence on the fitting process. All species with S/ N < 0.5, including Na, were entirely excluded from the fit (labeled "bad"), leaving a total of 16 strong species with S/N > 2. (OC, EC, SO^sub 4^^sup 2-^, NO3-, NH^sub 4^^sup +^, AL, CA, Cl, Cr, Cu, Fe, K, Mn, Pb, Si, and Zn). For all three comparison runs, the linear regressions of the 515 samples with 16 strong and 7 weak species were fit to p = 10 sources, hence Q^sub 0^ consistently equaled 4292.

Model results were generated for four cases of different input data uncertainty estimates and treatments summarized in Table 2. Note that the sodium ion was excluded, because it significantly worsened the quality of the fitted results of all cases. This negative impact may be due to the mixed influences of primary and secondary Na sources, which are subject to seasonal and higher frequency variations and sensitivities. Total XRF-sodium (NAX) was excluded because of its poor S/N ratio and high BDL fraction of 84%. The three main cases are:

(1) Use of original data with non-blank-corrected OC and uncertainties estimated based on eq 2 with P values equal to 7.1% for all species except Mg (8.5%), Al (8.6%), and Si (8.4%).

(2) Same as case 1, except that POC replaced OC, NTM = TMAC – OC + POC replaced TMAC, and their uncertainties increased correspondingly due to above-error propagation.

(3a) Same as case 2, except that uncertainties for all species in the input uncertainty matrix were increased by 5% (i.e., 0.05 x CONC added to UNC from eq 2) using the C3 option during the PMF model run.

One additional variation (case 3b) was investigated, in which 15 samples were removed before doing a second model run, which had exceedingly high residuals for potassium around July 4 and January 1; additional dates with similar K-governed outliers indicating unusual biomass combustion or other activities causing PM emissions high in potassium were removed; these included March 14, 2001 (near St. Patrick’s Day), November 24, 2001 (Saturday after Thanksgiving), May 14, 2002, February 11, 2003, September 21, 2003 (Sunday, fall equinox and harvest festival), February 27, 2004 (Friday after Ash Wednesday and beginning of Lent), October 9, 2004 (Saturday, Blountsville Harvest Festival celebrated ~50 km northeast of the NBHM site; winds were from 58 [degrees] at 1.8 m/sec daily average), June 15, 2005 (possibly associated with cleanup work after Tropical Storm Arlene had made landfall on June 11, 2005 just west of Pensacola, FL, moving northward and weakening the following days), and August 2, 2005. The computed Q-robust (Q^sub r^) goodness-of- fit values for cases 1, 2, 3a, and 3b are 16,276, 13,917, 7,574, and 6,915, yielding Q^sub r^/Q^sub 0^ ratios of 3.8, 3.2, 1.8, and 1.7, respectively. Note that Q^sub r^/Q^sub 0^ < 2 indicates an overall acceptable fit.26 Table 3 summarizes the values of Q for which the impact of outliers has been capped (Q^sub r^) and not capped (Q- true, or Q^sub t^), and compares the closeness of Q^sub r^ to Q^sub 0^ from eq 8. Generally, Q^sub r^ and Q^sub t^ are within +-50% for all cases, but are closer for larger input uncertainties, indicating an improved fit of the outliers on average for each random solution. The range of Q^sub r^ was small for all cases, indicating that the model is generally finding the global minimum.

In PMF, modeled result uncertainties are estimated using a bootstrapping technique, which involves choosing random samples with replacement from the dataset and analyzing each sample the same way. Because PMF, like other multivariate receptor models, requires only the ambient measurement data to estimate the source contribution, bootstrap is an appropriate (and possibly the only) way to estimate model uncertainties. The use of a larger dataset like ours, containing over 500 samples (records), is advantageous, because it reduces the probability of taking the same records during the bootstrap sampling process. PMF uses a fixed block size of three consecutive sample records, covering a 7-day period for the 1-every- 3-days sampling rate here, and the blocks are chosen randomly. Unlike the block size, the number of bootstrap samples has an effect on the distribution of the statistical estimator such as the model uncertainty; hence at least a few hundred of bootstraps are desirable.

Uncertainties of our factor profiles here were determined consistently from 300 bootstrap runs, which all converged for all 10 factors of each of the four cases. Each run’s factors were mapped to the corresponding base case factors and retained when all the run’s source contribution estimates (SCE) correlated with the base case’s SCEs at r^sup 2^ > 0.6. All of the case 3b solutions fulfilled this condition, and case 3a had only one factor, the one representing the biomass burning (BB) source, miss this condition; this source is governed by K and hence this lack of correlation is caused by the included K outliers. For comparison, case 2 had 9 factors miss this condition, and case 1 had 107. For both cases, most of the factors that missed the condition were also associated with the BB source; however, unmapped factors of case 1 also included the primary metal processing source (43 factors) and coking source (24 factors).

In the Results and Discussion section, we will describe sensitivities of the PMF model output, in particular the scaled residuals and factor profiles resulting from the different input uncertainties; identify and interpret the factor profiles by assigning and associating realistic source classes; calculate absolute SCEs; and relate the SCEs to the daily measured total mass for both temporal investigations on seasonal trends and spatial investigations on correlations with air mass transport (i.e., wind direction). Differences in results obtained from application of the CMB approach will be investigated and discussed later in a companion paper.

RESULTS AND DISCUSSION

PMF Residuals

The model provides scaled residuals, which are the measured concentrations minus the modeled concentrations divided by the provided input uncertainty (per eq 2, increased by 0.05 x CONC for case 3) for each species and sample, and which should lie between – 3 and +3. Table 4 lists the number of scaled residuals of each species outside that range for the four different cases. Finding many residuals for a species or date lying outside this range may indicate that the species uncertainty or S/N is too low, its data need to be blank-corrected, or a particular sample needs to be downweighted or identified as an outlier and eliminated entirely. The Pearson’s coefficient of determination (r^sup 2^), which is a measure of how well the modeled concentrations match the measured concentrations, is also shown for each species.

Scaled residuals of the downweighted weak species showed the lowest number of exceedances of the nominal value 3, and NO^sub 3^^sup -^, Ca, and Cl had no exceedances for any of the four cases. As expected, the biggest improvements from case 1 to case 2 were achieved by reducing exceedances of OC, because of a combination of an inherent blank correction and increased uncertainty. Even EC exceedances were significantly reduced (from 20 to 1), probably because of the introduction of its linear relationship with POC via application of the EC tracer method. Sulfate might have suffered an increase of exceedances (all positive) for a few coincident high SO^sub 4^^sup 2-^ and high POC cases that left most of the measured SO^sub 4^^sup 2-^ unexplained. More significant shifts toward larger unexplained mass fractions were seen for Mn and Zn, with Mn data likely having not enough signal (S/N = 4.8) and being more often below detection in comparison with Zn.

As expected, all species clearly benefited from a uniform 5% increase (adding 0.05 x CONC) of the input uncertainties via PMF’s C3 option (case 3a), both in terms of number of scaled residuals exceeding the threshold and how well the modeled values matched the observations. Except for the weak species, only the correlation coefficients for Pb and K remained below 0.6. Removing the 15 samples with the identified outliers for K mentioned above helped improve both species’ match, but more significantly so for K, whose r^sup 2^ improved from 0.32 to 0.95.

All scaled residuals were regressed among the different cases, and these results are summarized in Table A-4, confirming the above in more detail, in that these results show that the first improvement step (replacing original non-blank-corrected OC with POC and an increased-more realistic-uncertainty from error propagation) reduced the amount of the residuals, both in number of threshold exceedances and magnitude; i.e., the observed species were explained more fully by the fitted values. A more significant improvement in those two indicators was achieved by the second step of uniformly adding 5% of CONC to the uncertainties of all input species. The quality of this improvement was further enhanced by excluding 15 K outliers from the input dataset (case 3b), as indicated by the slope values and correlation coefficients of the linearly regressed scaled residuals of all species, but especially of Cl, Ca, NH^sub 4^^sup +^, and Si, which were further reduced by 32, 13, 3, and 1.5%, respectively.

PMF Factor Profiles

PMF, like other multivariate linear regression analyses, yields factors that in most cases and settings do not represent clean single source types. Therefore, the following discussion characterizes the resulting factors as emission source profiles, with the understanding that some profiles may include or be “contaminated” by multiple source types, potentially affecting the source apportionment. Figure 4 compares the profiles obtained for each of the 10 factors from the 3 main cases (1, 2, and 3a), which were all based on 515 input data records. Case 3b, which excluded the 15 potassium outliers yielded very similar profiles to case 3a, with the main difference being the apportionment of less sulfate to the road dust (2.5% less), primary metal processing (5.5%), and pulp- paper-wood production (7.1%) factors, and more SO^sub 4^^sup 2-^ to the coking industry (2.7% more), secondary metal (3.3%), and mineral processing (4.3%) factors. Table 5 summarizes each species’ mass fractions of each factor’s apportioned mass for case 3b.

Focusing on the least uncertain and best fitted results of case 3b, the secondary SO^sub 4^^sup 2-^ factor (sSO4) was identified by its large SO^sub 4^^sup 2-^ mass fraction of 58% and its molar NH^sub 4^^sup +^/SO^sub 4^^sup 2-^ ratio of 1.66, indicating that approximately 34% of all SO^sub 4^^sup 2-^ is best explained as bisulfate. Its seasonal contribution aggregates shown in Figures 5 and 6 also indicate a distinct summer high versus a winter low. This factor’s POC mass fraction is 1.1% (+5.9/-1.1% upper/lower SE), whereas the original (non-blank-corrected) OC dataset yields a 14% (+2.4/-0.5%) mass fraction. This much higher OC fraction must be interpreted as a significant SOC contribution, which seemed to have been effectively eliminated by the applied EC tracer method described earlier, and in fact compares well with the method’s total average SOC contribution of 11.9% (see Table 9).

The factor labeled sNO3 represents the secondary nitrate source, showing distinct winter highs and summer lows as expected, with an average molar ratio of its apportioned NH^sub 4^^sup +^/NO^sub 3^^sup -^ of 0.9. In addition to model uncertainty, this ratio of less than 1.00 may be in part due to the potential contributions from motor vehicle emissions indicated by a 22% (+7.4/-12%) POC mass fraction and an OC/EC ratio of 2.4 +- 0.88 (root mean square error [RMSE]). About a 1% contribution from mineral dust (i.e., crustal elements calculated according to eq 1) further supports a mobile source contribution to this factor.

The MVEH factor (motor vehicle source) is characterized by distinctly higher PM^sub 2.5^ mass contributions on weekdays relative to weekends. This factor is also characterized by the relatively large POC mass fraction (66 + 5.4/-6.6%) and a POC/EC ratio of 3.1 +- 0.19. However, MVEH cannot be distinctly isolated from the RDUST factor (road dust source), which indicates a similar POC/EC ratio, although at a much lower POC mass fraction of approximately 13 + 9.3/-13% pointing to vehicle tires’ transfer and abrasion mechanisms as possible link between these two factors. RDUST contributions are also characterized by a summer high, potentially driven by drier conditions, higher winds, and more heavy construction activities than during other seasons. Another distinctive RDUST feature is the 43 +- 4% mass fraction of the crustal materials (using Al, Si, Fe, Ti, and the relationship in eq 1), and a K/Si ratio of approximately 0.1, resembling the soil- derived ratio of 0.15 from Lewis et al.64

The BB factor is clearly determined by the large POC (~11 + 17/- 7%) and near zero EC fraction along with the largest K fraction of all factors, plus a K/Si ratio of 1.9 that is significantly larger than the soil-derived 0.15. This factor’s contributions are also the lowest during the summer, indicating combined influences from residential wood burning in fall and winter and prescribed burning conducted mostly in spring (see, for example, Lee et al.).65

The remaining five factors are identified as industrial sources and all show trends toward lower weekend contributions, except for the factor labeled COKE. Table 6 provides a summary of the associations made for tentatively interpreting these five factors to resemble potential actual sources on the basis of the following relationships of indicator species prominently represented in each factor profile. The PULP/WOOD factor is characterized by relatively large POC/EC ratio and POC, Cl, and K mass fractions, plus some mineral components (Ca, Si) that are typically part of the emissions from pulp-paper-wood manufacturing and processing.66 However, there are no major paper mills in the area, and only a few small local wood products manufacturers (standard industrial classification [SIC] code 2499 to the northeast and SIC code 3995 to the southwest). Given that the PULP/WOOD factor shows an OC/EC ratio of approximately 8 and an EC loading significantly greater than that of the BB factor (which is zero), both factors seem to resemble sources of biomass burning and wood combustion, as indicated by their similar K/Si ratios and seasonality (Figure 6). Controlling the rotational ambiguity between the EC, chloride (both statistically zero for BB), and nitrate (statistically zero for PULP/WOOD) regressions could improve the factor profiles and help better separate the factor contributions. Model runs with one less factor (i.e., 9 factors instead of 10) did not yield any improvements; on the contrary, that worsened overall model performance, as indicated by larger deviations of Q^sub r^ from Q^sub t^ and increased number of scaled residuals outside the tolerable range. Additional evidence for somewhat confounding factor separation becomes apparent when modern/fossil carbon split estimates are compared with splits determined from carbon isotope measurements by Zheng et al.67 Assuming here that 80% of SOC plus all TC (POC + EC) from BB and PULP/WOOD together originates from modern carbon, constituting an upper limit estimate of 26 + 7/-6% from error propagation, the carbon isotope method yielded 30% modern carbon (and 70% fossil), indicating a discrepancy within the uncertainties of the PMF analysis and the SOC estimation.

The COKE factor shows a significant EC mass fraction (~8.5 + 5.8/ -7.1%) and a POC/EC ratio of 2.6 +- 1.7, pointing to its combustion- related processes utilizing coke ovens and blast furnaces. Because coking plant fenceline measurements yielded OC/EC ratios between 0.6 and 1.4,25 this factor’s higher average value of 2.6 indicates the PMF model’s inability to cleanly resolve any single specific source type by instead including other source types of similar profile in more than one factor (as, for example, the METL1 factor described below). The COKE factor also displays a more significant (primary) sulfate mass fraction of approximately 25 + 13/-17%, which may be attributed to one of the two coking facilities in the NBHM area (ABC Drummond, 4.5 km to the northeast), where ammonium sulfate fertilizer is produced as a recovery byproduct.68 However, the primary sulfate contribution of this coking source is less than 3% of the average total mass apportioned to the regional sSO4 and only approximately 0.9% of the total average apportioned mass (including SOC) of 20.3 +- 2 [mu]g.m^sup -3^; see Table 9. Among all factors with industrial source characteristics, this factor shows the least statistical difference between weekend and weekday contributions, supporting its identity, because both Sloss (smaller than ABC Drummond but only 1.3 km away to the northeast) and ABC Drummond coking plants operate several oven batteries 24 hr a day, 7 days a week, 365 days a year. The factor is also characterized by relatively high Fe, Cr, Mg, and Ca mass fractions, which point to the involvement of significant coal grinding and coke handling activities.25 It therefore also matches the characteristics of facilities like the large historic U.S. Steel plant in Fairfield, approximately 15 km to the southwest, which hosts a multitude of operations more closely resembling the METL1 and METL2 factor profiles. Similar to the confounding PULP/WOOD and BB factor relationships mentioned above, the COKE factor may carry signatures from the factor labeled METL1, which is very similar except for almost opposite loadings on NO^sub 3^^sup -^ (high) and NH^sub 4^^sup +^ (low).

The factor labeled METL1 is identified as a primary metal processing source on the basis of its extremely high POC/EC at relatively high POC mass fraction (13 + 26/-13%), particularly high Cu, but also significant Fe, Pb, Zn, and Mg contributions. Also, the trend toward smaller weekend contributions was relatively small, indicating mainly 24/7 operations. Sources that fit this profile are from primary steel and alloy production using miscellaneous furnaces year round, such as at American Cast Iron Pipe (ACIPCO) 2.2 km to the southwest (see Table A-1) and the U.S. Pipe and Foundry Co. with several stacks within 1 km to the northeast, east, and southeast. Regular operations in these gray iron foundries and blast furnaces also include coke transfer, railroad operation, and hot/cold sheet metal production. The proximity of these sources may be why the METL1 factor is high in NO^sub 3^^sup -^ and near zero in NH^sub 4^^sup +^.

The last two factors carry features of fugitive dust emissions that are typical for secondary metal and mineral processing sources (labeled METL2 and MINEP, respectively). They also show the most distinct differences between weekend low and weekday high contributions. Particularly high contributions from Zn but also elevated Fe, Mg, Mn, and Pb are characteristic components of fugitive dust emissions associated with secondary metal processing. There are several iron pipe facilities in the neighborhood of NBHM that fit this profile. The MINEP factor was characterized by both fugitive dust and combustion emissions with particularly high EC and Ca mass fractions of approximately 12% each, and POC/EC of 1.6+- 1.4, pointing to a variety of companies in the area (see Figure 1) that apply hot and cold processes, including limestone quarries, other stone, cement, asphalt paving, and mineral wool production. Table 6 summarizes the above factor-source profile associations.

SCEs

Daily SCEs in units of mass concentration ([mu]g.m^sup -3^) were calculated for all four cases and each of the 515 sample dates from the correspondingly modeled species’ absolute mass fractions of each factor’s apportioned average mass. Similar to the analysis of the residuals, the effects of the different input data on the SCE results relative to the four cases were investigated and are summarized in Table 7. As the quality of the model results improved systematically from case 1 to case 3b as described earlier, the POC inherent blank-correction of the OC data reveal a most significant shift of SCE from the biomass burning factor to the pulppaper-wood factor, as well as a shift from the coking to the primary metal processing factor as indicated by the case 2/case 1 ratios (slopes) of 0.25 and 1.66 for BB and PULP/WOOD on the one hand, and 0.35 and 1.85 for COKE and METL1 on the other hand. Although SCE shifts from BB to PULP/WOOD were traceable via their main indicators (POC/EC, Cl, K), some of the case 1 BB SCEs may have shifted to RDUST and MINEP factors on the basis of the mineral dust indicator species (especially Ca, Fe, Si) and Mg. Similarly, the increase in METL1 contributions might have been in part caused by an increased association of metals from the secondary metal processing METL2, showing a slope of 0.53. The additional 5% uncertainty increase applied uniformly to all input data caused another but much smaller shift in the same direction for the BB and primary and secondary metal processing factors, and for the motor vehicle factor, whose contributions had been originally underestimated by 19%. The case 3b model run, which excluded the 15 K outliers, further consolidated the SCEs relative to case 3a for RDUST, BB, and most industrial factors except PULP/WOOD and METL1. The secondary, more regional atmospheric factors (SO^sub 4^^sup 2-^ and NO^sub 3^^sup -^) and traffic factors were not affected by the exclusion of these outliers.

The predicted PM^sub 2.5^ contributions from all sources (all 10 PMF-derived factors plus the SOC from the EC tracer method) were compared by linear regression with the 515 gravimetrically measured PM^sub 2.5^ concentrations for the four different cases. Table 8 summarizes the slope, its SE, intercept, and correlation coefficient (r^sup 2^) for each case’s regression result. Although the case 1 results reveal significant uncertainty in the slope and a large intercept (1.76 [mu]g.m^sup -3^), reflecting artifacts from using non-blank-corrected OC data, the PMF-resolved factors from all other cases seem to effectively reproduce the measured values and account for most of the variation in the PM^sub 2.5^ concentrations at an r^sup 2^ of approximately 0.96. Excluding the 15 outlier records from the analysis seems to mainly reduce the uncertainty in the average 97% of the apportioned mass. Table 8 also includes a summary of how well the sSO4 contributions matched the measured ammonium sulfate mass (excluding nitrateammonium) overall. For all cases, PMF overapportioned the sSO4-SCE but the POC replacement significantly improved this agreement from a factor (regression slope) of 1.4 to less than 1.3. As with total mass, the exclusion of outliers also reduced the uncertainty in the slope but not the slope itself. A negative intercept here means that PMF yields a maximum primary sulfate SCE of approximately 1 [mu]g.m^sup -3^ for samples that resulted in zero contributions from sSO4. If the regression is forced through zero (i.e., if no sSO4 factor contributions are allowed for zero SO^sub 4^^sup 2-^ observations), the overapportionment is reduced from 21% to less than 15%. Figures 5 and 6 and Table 9 provide an overview of the SCE results obtained for the best performing case (3b), presenting seasonally incremented time series of absolute contributions, relative seasonal contributions, and tabulated absolute/relative contributions including mean SEs containing propagated upper and lower errors of individual factor profiles. On global average, the regional atmospheric (secondary) factors (sSO4, sNO3, and SOC) combined make up 51 +- 3.8% or 10.5 +- 0.77 [mu]g.m^sup -3^ of the total PM^sub 2.5^ mass, followed by traffic related factors of approximately 30 +- 2%, assuming the MVEH (24.6 +- 1.8%) and RDUST (5.9 +- 0.9%) factors are uniquely related. All five industrial factors combined follow at 14 +- 1.3%, corresponding to 2.8 +- 0.3 [mu]g.m^sup -3^. The regional plus local traffic (MVEH only) factors alone already slightly exceed the annual PM^sub 2.5^ NAAQS of 15 [mu]g.m^sup -3^, with an average 15.5 +- 0.85 [mu]g.m^sup -3^. The regional source contributions vary the most seasonally, with a summer maximum of 57 +- 16%, despite the winter maximum of secondary NO3 – of 11 +- 6%. The motor vehicle source contributions are also highest in winter at almost 29 +- 11%, likely as a combined result of increased vehicular cold-start emissions and meteorological conditions (shallower and more stratified boundary layer, preventing downmixing of more buoyant stack emissions). Relative source contributions from all five industrial factors showed comparatively little seasonal variation.

The remaining text describes the sensitivities of local winds, source locations, and PM^sub 2.5^ contributions from local versus regional sources, which are limited to non-calm conditions with clearly detectable wind directions. Sensitivities are being investigated in a combined manner to help identify local sources that are most important for the development of emissions control strategies and SIPs and that have the potential to most effectively mitigate critical source impacts, reduce the annual PM^sub 2.5^ average at the NBHM site, and re-establish compliance with the NAAQS.

Associations with Air Mass Transport

This section discusses the transport-related site characteristics to identify predominant directions of local sources relative to surface wind direction. Daily prevailing wind directions were calculated based on vector additions for each of the 515 sample days from hourly wind speed and direction data collected at the collocated SEARCH site (see http://www.atmosphericresearch. com/ studies/SEARCH/index.html for more information on SEARCH). The hourly wind data had been reduced from 1-min raw data applying vector averaging for wind speeds greater 0.2 m/sec.69 The wind data coverage for the PM sample days was nearly 100%, and all hourly wind speed records were accompanied by a valid wind direction value, indicating that calm conditions did not persist beyond any 1-hr period, so that hourly wind directions and thus daily aggregates seem to represent actual conditions fairly well. To minimize the effect of atmospheric dilution, daily fractional mass contribution from each source relative to the total of all 11 identified sources (i.e., all 10 PMF-derived factors plus SOC from the EC tracer method) was used rather than the absolute SCE. Wind rose plots (Figure 7) were used in conjunction with the map in Figure 1 to associate local emission source types and locations with the average fractional PM^sub 2.5^ mass contributions in air masses arriving from different directions.

Figure 7 shows correlations with wind direction plots (wind roses) of certain gas and particle-phase species concentrations (CO, NO^sub y^, sulfur dioxide [SO2], PM^sub 2.5^, EC, SO^sub 4^^sup 2- ^) and the 11 individually identified factors resembling certain source categories. Not shown is the wind frequency distribution, which indicates that southeasterly, southwesterly, and northerly component winds are the most frequent, and northeasterly winds least frequent on the 515 sample days of the 5-yr period. Winds blew strongest from the south and weakest from the northeast, which explains the relatively low average concentrations, especially PM^sub 2.5^, EC, SO^sub 4^^sup 2-^, and SO2 for southerly flow. Higher SO2 and SO^sub 4^^sup 2-^ averages for air masses arriving from northerly and especially easterly directions point to the more regional influence of high-stack emissions from coal-fired power plants. Indeed, the sSO4 factor wind rose shows about a 6-8% enhancement in PM^sub 2.5^ contributions for air masses arriving from the northwest and east, where coal-fired power plants are located 24 and 25 km away, respectively.

The fact that SO2 is highest relative to SO4 2-in air masses arriving from the east-northeast as compared with other directions may be an effect of SO2 emissions from (1) nearby primary metal producing facilities (U.S. Pipe) totaling more than18 t/yr annually,70 and (2) aircraft taking off from BHM International Airport, whose runway is less than 4 km away, considering that kerosene-based jet fuel contains significantly more sulfur (>500 ppm) than gasoline or even diesel fuel (<15 ppm). If the OC/EC ratio served as indicator for SOA, then the airshed with the highest SOA transport to the site would have to be south-southwest from the site. Indeed, the SOC rose confirms that southerly winds carry the highest SOC contributions (almost 20% of the total PM^sub 2.5^ mass) on average, although for winds slightly shifted to the south- southeast.

Similarities between the CO,NO^sub y^, and EC roses point to same source origins, one of which is likely the MVEH source, with maximum PM^sub 2.5^ contributions of 20-30% in air masses arriving from all directions except east and southeast. Different CO/NO^sub y^ ratios, however, indicate a different degree of influence from other contributing sources, such as pulp-paper-wood, mineral processing, and coking from the NE. Coking emissions seem to severely impact the site from the southeast and west. The largest and only coke producing plants in the vicinity of the site are ABC Drummond (4.3 km) and Sloss (1.3 km) to the northeast; SMI Steel (5.6 km) to the east-southeast, and U.S. Steel Fairfield (15 km) to the west- southwest are listed under the same SIC code, but their emissions carry signatures that likely also fall under other categories than COKE; see Figure 1. With more than 700 t/yr of PM^sub 2.5^ emissions, Fairfield is the largest industrial plant in the area with a multitude of operations in addition to coking. 70 It operates a large blast furnace, oxygen-induction furnaces, a seamless tube manufacturing operation, and a hot and cold rolled sheet metal operation, including galvanized sheet lines. In addition, there is substantial potential for fugitive dust emissions from trucking and railroad activity on-site, including coke storage and transfer, as well as other raw material and slag and debris storage,25 which make this plant fall under coking, primary and secondary metal processing, but also mineral dust source categories. It is therefore not surprising that despite its distance, its emissions contribute to the PM^sub 2.5^ in air masses arriving from the southwest on average in the COKE, METL2, and RDUST factors. Several mineral processing plants are lined up to the northeast from the site and average a combined (MINEP) contribution of more than 6% from that direction.

The primary and secondary metal processing contributions are channeled from the SE, contributing up to 3 and 6%, respectively. Figure 1 and Table A-1 confirm the highest density of metal processing companies in this wind sector, with the largest METL2 plants, each listed with 10-20 t/yr PM^sub 2.5^ emissions only 3-4 km away at 111-134[degrees] from the site, and several METL1 facilities further away. The situation is similar for the southwest sector, where in addition to ACIPCO, several other large sources matching the METL1 characteristics are approximately 20 km or more away, for example, Citation Castings, Griffin Wheel, and U.S. Pipe- Bessemer, with annual PM^sub 2.5^ emissions of 58, 200, and 69 t/ yr, respectively, whereas a few smaller METL2-matching facilities (besides Fairfield at 15 km) are closer in, for example, Consolidated Pipe and Supply at 4.2 km and LB Foster at 3.9 km. The west-southwest contribution may also reflect emissions from ACIPCO, 2.2 km away with annual PM^sub 2.5^ emissions of 392 t/yr.

The last three source categories road dust (RDUST), PULP/WOODand BB are contrasted in the wind rose plot of Figure 7, d and e. Whereas the BB factor reflects the character of a diffuse regional area source, and PULP/WOOD that of more distinct local sources, indicating impacts from individual industrial and possibly residential locations surrounding the site nearby, the RDUST factor points to specific directions of higher and lower impacts. Higher impacts point to line sources of major highways and roads intersecting at the site, such as I-65 (northsouth axis), I-20/59 (east-west axis), and Highways 78 and 280 (northwest and southeast), with winds parallel to the line sources generally causing accumulative effects. Interstates I-20 and I-59 merge approximately 5 kmeast of the site, only a few meters away from the BHM International Airport’s runway, which may explain the more than 7% road dust contribution determined for this wind direction. Other contributions point to industrial yards, and public parks in the immediate vicinity of the site whose RDUST contributions may vary with wind speed and season. As illustrated in Figure 6 and Table 9, road dust contributes significantly more in summer than in winter, reflecting generally drier and windier summer conditions. A sensitivity test comparing this source’s contributions on rainy days versus dry days would corroborate this interpretation. In contrast, the BB factor shows relatively smooth transitions between sectors, indicating a spatially wider spanning character with contributions from regional land management activities utilizing prescribed burning, whereby more local contributions from spread-out residential areas to the northeast (Tarrant City) and broadly to the south toward downtown Birmingham may be harder to distinguish from the PULP/WOOD factor. Our seasonal analysis found the minimum of both factors in summer, a clear PULP/WOOD maximum in winter, and more broadly elevated contributions for BB outside the summer months, suggesting that confined (industrial and residential) wood burning is the main driver in winter, whereas open biomass burning under controlled and uncontrolled conditions (i.e., prescribed burning) spans winter and spring. However, spring and fall contributions from other directions point to possible influences from distant wild fires, which would require case-by-case trajectory analyses to pinpoint their individual locations of origin, but are impossible to control from a regulatory and SIP development point of view. CONCLUSIONS

Estimation of POC from seasonal application of the EC tracer method significantly improved the performance of the PMF receptor model version 1.1 applied to a 5-yr dataset to estimate source contributions to ambient PM^sub 2.5^ in the highly industrialized urban setting of the northern part of Birmingham in Jefferson County, AL, which suffers nonattainment status for the PM^sub 2.5^ NAAQS. Use of estimated POC rather than “raw” OC obtained from the STN database may be advantageous because the STN OC data are not blank corrected. Because the EC tracer method indirectly applies an OC blank correction, its incorporation into the data reporting process would significantly improve the direct utility of the PMF modeling tool.

The PMF model was able to consistently resolve 10 factors resembling two secondary, five industrial, one motor vehicle, one road dust, and one biomass burning sources using three different uncertainty estimates. An additional variation that added special identification and treatment of outliers strengthened the factor loadings of all 10 factors but especially the ones representing biomass burning (BB), pulppaper-wood production (PULP/WOOD), and mineral dust sources related to traffic (RDUST) and mineral processing (MINEP). These factors indicated potential inclusion of multiple source types into one factor’s profile, where individual control of rotational ambiguity of certain regressions might have improved the source apportionment result. For example, the PULP/ WOOD factor carried signatures from wood combustion similar to BB. Similarly confounding were the RDUST and MINEP as well as the COKE and METL1 factor profiles and contributions.Arecent model comparison with an advanced version (PMF2)71 showed that for sites in less industrialized settings, differences in the least-squares search algorithm, non-negativity constraints, and rotations contribute to differences in factor profiles and SCE of individual samples; however, average results are similar within the models’ uncertainty bounds and worth investigating for the NBHM dataset.

Model runs with fewer (8 and 9) and more (11) factors yielded unsatisfactory results with significantly worse model performance as indicated by larger deviations of Q^sub r^ from Q^sub t^ with Q^sub r^/Q^sub 0^ > 2 and increased number of scaled residuals outside the tolerable range. In addition to the introduction of POC, the uniform increase of input data uncertainty and exclusion of a few outlier samples identified by spikes in K further improved the model results, explaining 97% of the measured total PM^sub 2.5^ mass at a minimal intercept of 0.25 [mu]g.m^sup -3^ and an r^sup 2^ of 0.96.

Secondary particles including SO^sub 4^^sup 2-^ and SOC (SOC determined from the difference OC – POC) combined contribute the majority of ambient PM^sub 2.5^ in summer with 55 +- 16% compared with 37 +- 10% in winter. SOC associated with the sSO4 from the model run with the original OC data compares well with the average SOC contribution estimated by the EC tracer method. A small primary SO^sub 4^^sup 2-^ source was found associated with the major coking plants in the area, contributing less than 3% of the average total mass apportioned to the regional sSO4 and only approximately 0.9% to the total average apportioned PM^sub 2.5^ mass.

Motor vehicle emissions constitute the largest primary PM^sub 2.5^ mass contribution with a long-term average of almost 25 +- 2% and a winter maximum of 29 +- 11%, likely due to the combined effects from increased cold-start emissions and meteorological conditions. The annual PM^sub 2.5^ NAAQS of 15 [mu]g.m^sup -3^ is already slightly exceeded by the regional secondary plus local primary traffic sources alone at a combined average of 15.5 +- 0.85 [mu]g.m^sup -3^, so that SIP development efforts should focus on traffic and certa