Source Apportionment of Air Particulate Matter by Chemical Mass Balance ( CMB ) and Comparison with Positive Matrix Factorization ( PMF ) Model

The Chemical Mass Balance (CMB) model is one of several receptor models that have been applied to air quality management. This model quantifies contributions using chemical signatures characterizing source-types rather than contributions from individual emitters. The CMB model uses the chemical composition of ambient pollution samples to estimate the contributions of different source types to the measured pollutant concentrations. The disadvantage of the model is that it cannot separate the sources having similar chemical compositions or for those sources for which source composition profiles are unavailable. Since CMB analysis is done on a sample-bysample basis, it is possible to estimate the daily contributions of individual sources, and can thereby provide useful information based on a limited number of samples to address air quality management issues. Samples of fine and coarse fractions of airborne particulate matter (PM) were collected using a ‘Gent’ stacked filter unit in two fractions of 0-2.2 μm and 2.2-10 μm sizes in a semi-residential (Atomic Energy Centre Dhaka, AECD) area from June 2001 to June 2002 of Dhaka. These samples were analyzed for elemental concentrations with PIXE. The chemical composition data set was analyzed by CMB using local source profiles obtained using a Principal Components Analysis (PCA) and regression analysis of data from this site and the source contributions are quantitatively estimated for each of the samples. The results of the CMB analysis were compared with the results obtained using positive matrix factorization (PMF) that had been done previously. It is observed that CMB provides comparable results except for limited discrepancies, especially for the PM2.2 fractions where sources have similar elemental signatures.


INTRODUCTION
Air pollution is a common concern in the large and growing metropolitan areas throughout the world, especially in South East Asian cities.There are six major pollutants termed criteria pollutants in the United States, namely, PM 10 and PM 2.5 , NO x , SO 2 , O 3 and CO (Akhter et al., 2004) that have known severe health implications (Dockery et al., 1994, Pope, 2000, Schwartz, 2001).Particulate matter (PM) is recognized as the most important air pollutant in the Dhaka, the capital city of Bangladesh (Begum et al., 2006a).Dhaka (latitude 23.77N, longitude 90.38E) is congested with a large number of motor vehicles, including both public and private transportation.The vehicles are believed to constitute the dominant source of air pollution in Dhaka (Begum et al., 2004).Moreover, construction of roads and buildings is taking place continuously throughout the city.Many small-scale factories are also located in and around the city.Brickfields that use coal and wood to fire the bricks, have grown up around Dhaka city because of the increased demand of construction materials and these brickfields are also major contributors to the air pollution in winter since they operate only winter when the meteorological conditions are sufficiently dry (Salam et al., 2003).
The identification of various sources of airborne particulate matter (APM) and assessment of their impact on the aerosol composition of an air shed are major goals of contemporary atmospheric research.The national air quality standards for the APM in many countries around the world have also created the need to identify the particle sources so that effective control strategies can be designed and implemented.The initial efforts at the identification of particle sources focused on dispersion models of point sources, and in most cases, this approach resulted in substantial reduction in APM levels.Thus, additional methods were required to identify and quantitatively apportion the particle mass to sources with better resolution.These methods are called receptor models.Measured properties of collected ambient samples are used in these models to infer the contributions of the sources to the pollutant concentrations.This approach requires that the samples be collected at the location of interest, the receptor site and that the samples so collected be analyzed for the properties that are characteristic of the pollutant sources.Therefore, measurements of the elemental composition of atmospheric aerosol particles are needed for compliance monitoring, studies of environmental deposition, and source attribution by receptor modeling techniques (Gordon, 1988).
Receptor models have been developed to identify and apportion the contributions of various sources to the airborne particulate matter concentrations.Factor analysis (FA) techniques are multivariate data analysis methods that are used in environmental studies to estimate the number and compositions of the sources as well as their contributions to the samples taken at the receptors.Principal Components Analysis (PCA) is one of the common forms of FA (Jollife, 1986).This method extracts the principal components, explaining the majority of variance of the data matrix that are then qualitatively interpreted as possible sources (Wolff et al., 1985).However, this method suffers from several drawbacks (Paatero et al., 1993).A new approach to FA called Positive Matrix Factorization (PMF) was developed (Paatero, 1997;1999).This model uses a weighted least-square approach and imposes non-negativity constraints for fitting the FA model.Fairly large data sets with reasonable uncertainties are needed both in PCA and PMF models in order to produce good results.However, obtaining such data sets is often difficult and expensive.Moreover, it may be required to differentiate source contributions in different seasons or during specific limited duration pollution episodes.In such cases, large data set may not be available.T P T he Chemical Mass Balance (CMB) model has the capability of analyzing individual samples if local source profiles or fingerprints are known (Abu et al., 2002).In order to obtain a source profile of a given particle source, two things must be determined: first, which chemical species (elements, compounds, or ions) characterize that source profile and second, the relative proportions of each of those elements that provide the signature for the source profile at a sampling site.These fingerprints or profiles can then be used in the CMB model (Watson et al., 1990) together with the sample's species concentrations and their associated analytical errors.In an earlier study, it was found that the source profiles obtained from PCA were comparable with PMF (Begum et al., 2006b), so the source profiles obtained from the PMF analysis could be used as input to the CMB model.Since the CMB analysis could be applied even on a single sample, it is possible to estimate the daily contribution from the individual source if the source profiles are properly estimated.It is also applicable when only a small data set is available (Chowdhury et al., 2003).Profiles obtained from a large set of data from one location could be applied to a small data set from another site if it is known that the same or similar sources will affect both sites.Particularly if these profiles can be derived easily from readily available data sets (Begum et al., 2006a).
articulate matter (PM) is found to be the major pollutant of concern in Dhaka and from the time series pattern, it is observed that PM concentration depends on the prevailing meteorological conditions (Begum et al., 2006a).Thus, the contributions of the sources should be estimated for the different seasons and these contributions could be useful for supporting specific policy interventions to reduce emissions.
he main objective of this study is to analyze elemental and BC concentration data of PM samples collected at a semi-residential site at Dhaka using the CMB model to apportion the source contributions.The local source profiles used in this analysis were obtained from a previous data analysis study (Begum et al., 2006b) and the data set used for the analysis is from July 2002 to June 2003 of the same site.The results of the CMB analyses have been compared with those obtained from the PMF analysis performed in an earlier study (Begum et al., 2004) to ascertain if such an approach of using derived profiles can be applied to a data set to obtain good mass apportionments.As a result of this apportionment, the seasonal patterns of various pollutant sources are also discussed.

Sampling
Sampling was done using a "Gent" type stacked filter sampler (Hopke et al., 1997) capable of collecting particulate matter in the PM 10-2.2 (coarse) and PM 2.2 (fine) size fraction samples collected between June 2001 and June 2002.The samples were collected on an 8 m pore filter (coarse fraction) and on a 0.4 m pore filter (fine fraction).The sampler was placed on the flat roof of the building in the AECD area.The roof height is 5 m and the intake nozzle of the sampler was located 1.8 m above the roof.The intake was about 80 m away from the roadside.
The sampler was placed so that the airflow was unobstructed.The effective sampling time was varied between 6-20 hours distributed uniformly over 24 hours in a day to avoid filter clogging and so that the flow rate remained within the prescribed limits of the sampler.This flow control ensures proper size fractionation and collection efficiency.

PM mass and BC determination
The masses of the coarse and fine fraction samples were determined by weighing the filters before and after the exposure.A Po-210 (alpha emitter) electrostatic charge eliminator (STATICMASTER) was used to eliminate the static charge accumulated on the filters before each weighing.
The concentration of black carbon (BC) in the fine fraction of the samples is determined by reflectance measurement using an EEL-type Smoke Stain Reflectometer.Secondary standards of known black carbon concentrations are used to calibrate the Reflectometer (Biswas et al., 2003).

Multielement analysis
Multielement analyses of the samples collected during above cited time intervals were made using Proton Induced X-ray Emission (PIXE) at the Institute of Geological and Nuclear Sciences (IGNS), New Zealand.X-ray spectra obtained from PIXE measurements were analyzed using the computer code GUPIX developed by Guelph University.Calibration of the PIXE system was performed by irradiating suitable Micrometer thin target standards in the similar experimental conditions.Data on the concentrations of eighteen elements, black carbon and mass were available for further analysis.

Meteorological conditions
In Bangladesh, the climate is characterized by high temperatures, high humidity most of the year, and distinctly marked seasonal variations in precipitation.Based on the meteorological conditions, the year can be divided into four seasons, pre-monsoon (March-May), monsoon (June-September), post-monsoon (October-November) and winter (December-February) (Salam et al., 2003).The meteorological data used in this study was obtained from a local meteorological station, located about 5 kilometers north of the semi-residential site.

Data analysis by Chemical Mass Balance
The CMB model uses the chemical composition of airborne particulate matter samples to estimate the contributions of different source types to the measured concentrations at the receptor site using previously estimated source fingerprints.The chemical composition of each sourcetype's emission (commonly known as source profile or source fingerprints) may be estimated from PCA or sometimes a reasonable estimate may be obtained from the United States Environmental Protection Agency (USEPA) library (SPECIATE).In the present study, the source profiles have been estimated from the results obtained from prior analysis (Begum et al., 2006b).This model quantifies contributions from chemically distinct source-types rather than contributions from individual pollution emitters.This model cannot separate the sources that have very similar chemical compositions.It does estimate the source contributions for each individual sample.The particle composition may differ from one sample to the next owing to differences in emission rates, wind directions, wind speed, and changes in emission compositions.However, the variations in the source compositions are not taken into account in these analyses.
The model consists of the following equations i = 1…I, j = 1…J (1)

Conditional Probability Function (CPF)
To analyze point source impacts from various wind directions, the conditional probability function (CPF) (Kim et al., 2003a;2003b) as calculated using source contribution estimates coupled with wind direction values measured on site.To minimize the effect of atmospheric dilution, daily fractional mass contribution from each source relative to the total of all sources was used rather than using the absolute source contributions.The same daily fractional contribution was assigned to each 3 hours period of a given day to match to the 3 hours average wind direction.Specifically, the CPF is defined as where m is the number of occurrence from wind sector that exceeded the threshold criterion, and n is the total number of data from the same wind sector.In this study, was set to be 45 degrees.The threshold was set at the upper 50 percentile of the fractional contribution from each source.The sources are likely to be located to the directions that have high conditional probability values.

RESULTS AND DISCUSSION
Table 1 shows the average mass black carbon and elemental concentrations along with median and standard deviations of fine and coarse particulate samples collected at this semi-residential area (SR) site in Dhaka, Bangladesh.Here all of the samples are collected on weekdays.The yearly average values for both PM 10 and PM 2.2 masses (Begum et al., 2006a) are much higher than the 1997 USEPA standards as well as the Bangladesh national air quality standard for PM 10 and PM 2.5 that were set at 50 µg/m 3 and 15 µg/m 3 , respectively.The PM 10 mass concentrations are also higher than the WHO guideline range (60-90 µg/m 3 ).
CMB analysis was performed using the CMB7 software obtained from the USEPA website (Watson et al., 1990).Local source profiles obtained from PCA (Begum et al., 2006b) method, except road dust, metal smelter and biomass burning source, are provided as the input of the model and the source contributions are quantitatively estimated for each of the samples.The profiles are different for coarse and fine particulate matter.PCA and PMF method can produce comparable source profiles, The PMF source profiles have been used for these sources.The source profiles for coarse and fine fractions are presented in Figs. 1 and 2. The CMB model diagnostics for PM are in the acceptable limits as prescribed in the CMB user guidelines (Watson et al., 1990).For each of the PM samples, the R 2 was between 0.8 and 0.9 and the chi-square varied from 0.89-4.Eight sources could be identified for both coarse and fine fraction of the PM samples collected from semi-residential area of Dhaka.Figs. 3 and 4 present the measured mass concentrations of the individual samples compared with the predicted mass concentrations by the CMB model.The R 2 values are 0.81 and 0.72 for coarse and fine particulate matter samples, respectively.In the case of the fine particle mass (Fig. 4), the CMB could account for about 86% of the measured mass concentrations.However, rather poorer reproduction of the mass concentrations was obtained for the coarse particle samples (64%).It is entirely clear why so much less mass was apportioned for the coarse particle samples as compared to the fine particle samples.There are no measurements of the carbonaceous constituents of the coarse PM that would include primary biological materials such as plant parts and broken pollen grains.The small amount of potassium they contain would be taken up by the biomass burning factor.Thus, some contribution from primary biological components would appear to be part of the low mass apportionment.et al., 2004).The CMB modeling could resolve two additional sources, fugitive Pb and a brick kiln source, that the previous PMF modeling was unable to derive.

Soil dust and road dust
The soil contains the characteristic elements Al, Si, K, Ca, Ti and Fe (Watson et al., 2001a;2001b).It has a few high peaks during winter season when rainfall is at a minimum.The CPF plots show the likely direction of the source.Wind blown dust could produce these particles.
The road dust is characterized by the Al, Si, Ca, K, Fe, Na, Mg, P, S, and BC and this source did not varied much during the study period.This profile contains S and BC as well as other soil components (Bhave et al., 2001, Ho et al., 2003).These species are attributed primarily to diesel vehicle emissions.The source is influenced by southerly winds.There is a busy road about a kilometer away on the eastern side of the sampling site with heavy traffic.This road probably explains the dependence of the source on wind direction.

Motor vehicles
Historically, lead has been used as the most reliable tracer of traffic.However, the gradual shift from leaded to unleaded petrol as fuel for automobiles has reduced traffic emissions of this element (Biswas et al., 2003).The motor vehicle (mixed diesel and gasoline engine exhaust) source profile is characterized by the high BC and S (Maricq et al., 2002;Kim et al., 2003b;Begum et al., 2004).This source is also mixed with the crustal elements, Mg, Al, Si, and Fe, suggesting the vehicular exhaust is mixed with resuspended road dust.This factor has a higher contribution during the dry season and is influenced by northwesterly wind.In Dhaka, two main traffic corridors extend in this direction.The sampling site lies between two large bus stations; one is on the northern side and the other station is to the west of the site.

Two-stroke engine emissions
The factor has the characteristic of having high Zn, BC and Fe (Chueinta et al., 2000).These elements are characteristic of lubricating oil used in two-stroke engine where engine oil is mixed with fuel.Zinc compounds are extensively employed as lubricants, antioxidants and as detergent/dispersant improvers for lubricating oils.Oxidation of lubricating oils upon exposure to air at high temperature results in the formation of metal oxide particles.The oil may also contain metal wear particles consisting of Zn, Cu, and Cd-bearing alloys (Drew, 1975).Zn could also be contributed from galvanized materials, tire wear and the use of zinc compounds in rubber production (Fergussion et al., 1991).This factor profile includes Na, S, and Mn suggesting that part of this factor may be attributed to refuse burning normally used in the slums for cooking purposes.The two-stroke engines emit large quantities of organic carbon (Snow and Crews, 2004).However, organic carbon was not measured in this study.These engines also emit black smoke and BC was part of the profiles.The time series plot shows relatively higher contributions in the wet season (Morales et al., 1996) and the CPF plot shows the influence of easterly and southeasterly winds (Begum et al., 2004).There are some galvanizing industries situated in the south and southeastern parts of Dhaka.Moreover, a relatively large, lower income population (slum dweller) lives in this area.

Sea salt
High Na and Cl characterize the sea salt component in the fine PM.The time series plot shows random peaks mostly during monsoon period between June and October and dominated by easterly winds.The sea salt component was expected to be seen during the monsoon season, when the wind is normally from the southeast.

Biomass burning/brick kiln
The third source may represent biomass burning/brick kiln characterized by K (Azad et al., 1998;Watson et al., 2001a;Begum et al., 2004).The source has seasonal variation as indicated in time series plot with high contributions in the dry season.The CPF plot shows a stronger influence of both southwesterly and northeasterly wind (Begum et al., 2004).There are many brick kilns around Dhaka and a large amount of wood and coal are burnt in those kilns.The kilns normally operate only between October and April and thus, explain the higher source contribution during the winter.In addition, slums in and around the city use biomass as a cooking fuel.

Metal smelters
The source characterized by high Al, Mg, Si, Ca and mixed with K, Fe, Mn, Zn, etc. could represent metal smelting sources (Harris et al., 2005).The time series plot shows random peaks mostly between November and February.This factor was influenced mostly by easterly and southeasterly wind.There are a significant number of small industrial facilities involved in metal working including foundries distributed around Dhaka.

Fugitive Pb
The fourth source shows a high value of Pb together with crustal elements (Na, Mg, Si, P, S and Cl) indicating fugitive/resuspended Pb (Harris et al., 2005).Pb has generally shown a tendency to be associated more with the fine fraction than the coarse fraction (Young et al., 2002).This tendency is because the Pb particles were formed at high temperature as part of the vehicle exhaust.However, handling of lead batteries in preparation for the reclamation of the lead produces coarse particle lead (Glover et al., 1991).The time series plot shows that this source contains several peaks all over the sampling period in case of fine particulate matter.The CPF plot shows the dependence on southeasterly winds and thus, battery recycling could again be a significant contributor to this factor.Although, Pb was eliminated from the gasoline in Bangladesh beginning in July 1999, there may be substantial accumulated lead in the dust near roadways because lead has a very long residence time in surface soil (Young et al., 2002).There are also Pb-battery recycling plants in old town of Dhaka in this direction.In addition, the use of lead weights to balance tires leads to their loss from the wheels into the roadways where they can be broken down over time to produce coarse lead particles.

Comparison of CMB and PMF results
Table 2 presents the comparison of estimated average source contribution obtained from CMB model calculations and those obtained from PMF modeling for the coarse and fine particulate matter using the same chemical composition data sets of all the samples.Both the modeling technique could produce comparable results, especially for coarse PM.In case of the CMB analysis, soil dust together with road dust accounted for about 47.6% of the total coarse PM whereas from PMF analysis sum of these factors were just over 50%.The second largest contributor was motor vehicle emissions including two-stroke engine exhaust that was about 43.3% and PMF also resolved that source as a 44% contribution to the mass.As CMB analysis is done on a sample-by-sample basis, it could be found that there are only few days where the contribution of fugitive Pb at the receptor site was high.Pb may come from the paints and the batteries industries.CMB resolved about 5% sea salt from the coarse fraction while PMF produced a somewhat smaller value.In case of fine PM, several differences could be noted.Although CMB and PMF calculation produced different source contribution for soil and road dust, the sum of the contributions of these sources remained the same in both sets of results.Similarly, the CMB analysis accounted for higher motor vehicle emissions and lower metal smelter, fugitive Pb and brick kiln contributions than the PMF model calculations.
In the CMB modeling, if sources have similar chemical signatures, it may not be able to differentiate between these sources.The collinearity between source profiles may account for the differences in calculated source contributions in the two modeling techniques.For example, sea salt and two-stroke sources have unique source fingerprints and thus, both models found similar results.However, in the case of the dust sources, more ambiguity in the results were obtained although the total dust contributions were quite similar.

CONCLUSIONS
From CMB calculations, eight sources have been identified from both coarse and fine particulate matter and this modeling technique produces comparable source contribution.Since the CMB analysis is done on a sample-by-sample basis, it is possible to determine the daily contribution from the individual source if the source profiles are properly estimated.This feature could result in the separation of more sources in the CMB results than in the PMF modeling.Although Pb was eliminated from the gasoline in Bangladesh in July 1999, source apportionment using CMB model calculation shows that about 3% of the contribution to FPM is from fugitive component containing Pb that is comparable with the PMF result of hot spot site, Dhaka (Begum et al., 2005).There may be substantial accumulated Pb in the dust near roadways.There could also be fugitive Pb emission from battery and other industries.
The sample-by-sample approach in CMB also can provide useful information based on limited number of samples collected at a receptor site to address air quality management issues where PMF cannot be used.As particulate matter (PM) is major pollutant of concern in Dhaka and from time series pattern, it is observed that PM concentration depends on prevailing meteorological conditions, the contributions of sources on monthly average could also be estimated (Figs. 9 and 10) and in turn could be useful for undertaking policy interventions.The method can also be helpful in identifying pollution episodes which occurs for a short period.
where = Concentration of species I measured at a receptor site i C ij F = Fraction of species i emissions from source j j S = Estimate of the contribution of source j I = Number of chemical species J = Number of source types These equations have a unique solution only when the number of species is equal to or greater than the number of sources.It has been generally found that the greater number of species, the more precise the apportionment.The method also requires precision estimates for the and i C ij F values as model input.These precision estimates result in realistic uncertainties associated with the source contribution estimate, that are calculated by the model.j S

Fig. 3 .
Fig. 3. Regression between measured and predicted mass concentrations for coarse particulate matter.

Fig. 4 .
Fig. 4. Regression between measured and predicted mass concentrations for fine particulate matter.

Fig. 5 .
Fig. 5. Time variations for coarse particulate matter at the semi-residential site in Dhaka.

Fig. 6 .
Fig. 6.Time variations for fine particulate matter at the semi-residential site in Dhaka.

Fig. 7 .
Fig. 7.The relationship of source contributions with wind direction for coarse particulate matter at the semi-residential site.

Fig. 8 .
Fig. 8.The relationship of source contributions with wind direction for fine particulate matter at the semi-residential site.

Fig. 9 .
Fig. 9. Monthly average contribution of sources obtained from PMF and CMB methods during the study period for coarse particulate matter at semi-residential site in Dhaka.

Fig. 10 .
Fig. 10.Monthly average contribution of sources obtained from PMF and CMB methods during the study period for fine particulate matter at semi-residential site in Dhaka.

Table 1 .
Mean, median and standard deviations of PM mass, BC and elemental concentration at a semiresidential area (concentrations in ng/m 3 ).

Table 2 .
The source contribution obtained from CMB.