Use of statistical multivariate analysis to identify sources emitting particulate matter in Belo Horizonte, Brazil

Particulate matter is the mixture of solid and liquid particles. It may at times cause adverse effects to human health, in the climate, ecosystem and materials. The objective of this study was to identify the sources emitting particulate matter that contributes to environmental pollution at two sites in Belo Horizonte city. The metropolitan region of Belo Horizonte is characterized by numerous mining activities and other industries, high population and, consequently, a high number of vehicles. There are therefore several sources emitting pollutants into the atmosphere. Airborne particulate matter was collected in air filters during approximately one year using a high-volume sampler in strategic points. The concentration of particulate matter results showed that, in general, the values were 40% higher than the values recommended by legislation. During the summer rainy season, no sample exceeded the recommended limits at both sampling sites. The elemental composition was determined through neutron activation analysis and these values were considered variables in the multivariate statistical analysis for the identification of the main sources of particle emission and their contributions. The results confirmed a significant influence of the soil and motor vehicles in the degradation of the region’s air quality.


INTRODUCTION
Changes in the natural air constituents may negatively affect plant and animal species, as well as human health and well-being (CONAMA, 1990). The air quality standards established by the World Health Organization (WHO) and national councils such as the National Environment Council (CONAMA) and the State of Minas Gerais Environment Foundation (FEAM) are based on political, social and economic issues, as well as health risk factors and national capacity for air quality management (CONAMA, 1990;FEAM, 2017;WHO, 2006).
Industrial activities generate a large volume of waste that is emitted to the atmosphere, containing several compounds harmful to human health (Brasil et al., 2007). Automobiles are one of the greatest sources of pollution developed by man, and release gases and particles into the atmosphere that alter its natural composition (Albuquerque et al., 2012;Sharma and Khare, 2001). This activity has had a significant impact on air quality, since over the years the number of vehicles has been growing sharply (Nagendra and Khare, 2002;Tavares et al., 2014).
One of the major pollutants released into the atmosphere is particulate matter (PM), which is the mixture of solid and liquid particles that vary in shape and chemical composition, and cause adverse effects on human health, and to the climate, ecosystem and materials (Kappos et al., 2004).
Some elemental traits can cause biological effects, such as inflammation, allergies and DNA damage, generating free radicals in human body tissues (HEI, 2002). The PM composition in the atmosphere depends on emission factors that occur in each region, such as the occurrence of fires, mining activities, automotive vehicle flow or industrial activities (Calijuri and Cunha, 2013), as well as the processes of PM resuspension by natural or artificial agents. Further, PM can be emitted from several sources, fixed or mobile, presenting varied chemical composition (Adissi et al. 2013). Monitoring air quality is essential to the control of air pollution.
The use of monitoring networks associated with computational modeling resources is fundamental to monitoring the evolution of pollutant dispersions and thereby obtaining support for the elaboration of public environmental policies (CETESB, 2015;FEAM, 2017).
Techniques based on monitoring networks allow the investigation of local conditions of air quality. The equipment used is costly and requires operating conditions and adequate structure for sampling at specific points. The generated environmental data are used mainly for statistical analyses on which strategies to identify the main contributing sources of pollution at the monitored site are based (Tavares et al., 2014).
The main statistical tool used to aid and interpret air quality monitoring data (Rebelo et al., Rev. Ambient. Água vol. 13 n. 4, e2202 -Taubaté 2018 2015) is the Principal Component Analysis (PCA). Many studies, use PCA to assess the impact of pollutants in the atmosphere and in the environment (Almeida et al., 2017;Elbayoumi et al., 2014;Jacomino et al., 2009;Megido et al., 2017;Sánches-Rodas et al., 2017). This work therefore assessed air pollution in Belo Horizonte, the capital of the state of Minas Gerais, Brazil, at two sites located on the campus of Federal University of Minas Gerais, in order to identify sources emitting PM10. The elemental composition of airborne particulate matter was analyzed through neutron activation analysis in order that PCA could be used to identify the main sources of PM10 emission.

MATERIALS AND METHODS
Sampling of PM10 was performed in the northeast region of Belo Horizonte. The city has a territorial area of 331,401 km², and an estimated population for the year 2017 of 2,523,794 inhabitants (IBGE, 2016). It is 900 meters above sea level and its temperature varies between 16ºC to 31ºC. The climate is a humid subtropical climate, registering rainfall and high temperatures in summer, while winter is characterized by low rainfall and low temperatures (Belo Horizonte, 2016).
The metropolitan region of Belo Horizonte is characterized by numerous mining activities and other industries, and is highly populated. Because the high population density, there is a high number of vehicles and consequent traffic jams and high emissions of pollutants into the atmosphere.
The area studied, the campus of the Federal University of Minas Gerais (UFMG) is the second largest in terms of the circulation of people in the city of Belo Horizonte, with an estimated 55,000 people circulating daily. This movement causes transportation problems as described in the UFMG Technical Report (UFMG, 2010).
The first step of this work was the selection of PM sampling points. An assessment of the study area was performed in order to identify points exposed to various emission sources, both of natural origin (soil, vegetation, etc.), as well as anthropogenic sources (vehicle and industrial emissions). The proximity to access routes containing high vehicular flow was a factor, due to the urban characteristics of the study area. Based upon these considerations and the surroundings of the region and the main routes of the neighborhood, two points were selected as shown in Figure 1. Point 1 is in an area that includes an ecological preservation site and local transit routes (Coordinate Latitude 19°52'19.53" S and Longitude 43°58'15.14" W); Point 2 is on the margin of one of the main avenues of Belo Horizonte with intense vehicular flow (Coordinate Latitude 19°51'47.32" S and Longitude 43°57'30.27" W).
Several methods have been applied for PM monitoring. A technique widely used throughout the world is based on the High Volume Sampler (HVS), which uses a filter to retain particulate matter, sampling over a 24-hour period (USEPA, 1999). In this work, glass fiber filters were used to collect trace elements contained in suspended particulates in the air (Pöykiö et al., 2003). The HVS has an inertial impact separation head at a flow rate of 1.13 m 3 min -1 . The flow uncertainty is less than 1% (ENERGÉTICA, 2012). For this study, only one high sampler was available. The equipment was therefore placed for few days at Point 1 and on other days at Point 2.
At Point 1, due to the presence of animals in the area of ecological preservation and the inclination of the terrain, it was necessary to delimit and elevate the equipment by 1.2 meters ( Figure 1). The elevation of the equipment did not interfere with the sampling. It is important to mention that in both sampling points, in order to avoid the resuspension of dust, a bulkhead was placed in the inferior part of the equipment. Before each sampling procedure, the equipment was calibrated to ensure the required flow rate. Samplings were performed using a filter for the retention of particulate during a 24-hour period. The filter model used for monitoring and determining the concentration of PM was the Whatman GF/A fiberglass type, with dimensions of 203 x 254 mm. Before the samplings, the filter was prepared by heating it in a muffle at 400°C for a period of 4 hours to eliminate moisture and organic compounds, and it was later stored in a desiccator to avoid contact with air humidity.
Before the sampling, each entire filter was weighed in a precision balance, with a maximum capacity of 210 g and a standard deviation of 0.1 mg.
After the sampling, the concentration (C) of PM, given in μg m -3 , was determined by the initial mass (mi), final mass (mf) and volume of the sampled air (Vs) as shown in Equation 1.
The identification of elements present in the particulate matter was performed through Neutron Activation Analysis, which is a multielementar analytical technique to determine chemical elements in biological or geological material, effluent, oil, water and also in particulate matter (Menezes and Jacimovic, 2006). The sample is analyzed without being solubilized, a great advantage compared with other technical analytics that require that the sample is in a soluble form. In this study, the technique was applied via k0-standardization method. This method does not use standards of the involved element but neutron monitors, gamma spectroscopy system absolutely calibrated, good knowledge of spectral parameters in the irradiation channel and k0 constants available in literature. In this technique, the sample is bombarded by neutrons in a nuclear reactor and the radionuclides produced emit characteristic radiation formed in the (n,γ) reaction. The chemical elements are then identified and assayed after irradiation by measurement -gamma spectrometry.
Rev. Ambient. Água vol. 13 n. 4, e2202 -Taubaté 2018 The irradiation was carried out on the carousel of the TRIGA MARK I IPR-R1 reactor at CDTN (Nuclear Technology Development Centre) in Belo Horizonte, at 100 kW, under an average thermal neutron flux of 6.30 x 10 11 cm -2 s -1 . The averages parameters f and α are (21.7 ± 1.1) and (0.0026 ± 0.0011), respectively. The samples were irradiated simultaneously with neutron flux monitor Al-Au (0.1%) IRMM-530RA foil, cut into a 6 mm diameter 0.1 mm thick. The gamma spectroscopy was performed on an HPGe detector with 15% efficiency and for the spectra analysis -peak area evaluation -the HyperLab program was used. For the calculation of elemental concentrations, a software package called Kayzero for Windows ® was applied (Menezes and Jacimovic, 2006).
In this study, the sampled filter was cut into a fraction of 30 cm 2 and packed in a polyethylene tube, 2.3 cm high and 1.1 cm in diameter, and then irradiated. In order to discount the influence of the filter material, the blank was prepared in triplicate under the same conditions.
Subsequently, Multivariate statistics was used, which is commonly applied in environmental studies, through principal component analysis (PCA), which analyzes multiple measurements of variables, with the objective of investigating, measuring, explaining and predicting relationships among statistical variables (Hair et al., 2009).
In the same way as PCA, Factorial Analysis (FA) describes the original vectors in a smaller number of variables, where they are called common factors and related to the original vector through the linear model (Mingoti, 2005). Through some rules, the quantity of factors according to Kaiser's criterion (Elbayoumi et al., 2014;Mingoti, 2005) refers to eigenvalues greater or equal to 1, and many times it is necessary to carry out a rotation of these factors. One of the methods of rotation is the Varimax, which tries to find factors with great variability to better explain the PCA (Mingoti, 2005). Minitab Software was used for the execution and interpretation of statistical data.

Particulate Matter Concentration
According to legislation (USEPA, 2014), air is considered to be of Good Quality when the concentration based on PM10, is lesser than 50 μg m -3 ; Moderate, when it is between 51 and 100 μg m -3 ; Unhealthy for Sensitive Groups, between 101 and 150 μg m -3 ; Unhealthy, between 151 and 200 μg m -3 ; Very Unhealthy, between 201 and 300 μg m -3 ; and above 301 μg m -3 is assessed as Hazardous. Figure 2 shows the concentration values for particulate matter collected from Point 1 and Point 2, the central tendency and variability of the data and their classification according to the values of US-EPA. The samples from Point 1 presented a range of PM10 concentration of 9.4 μg m -3 occurring in the fall, and 123 μg m -3 in the spring. During the monitored period, the concentration of PM10 in samples showed abrupt changes over time. Two samples in the winter exceeded 50 μg m -3 , classified according to US-EPA as "Moderate". In the spring, three samples presented concentrations higher than 50 μg m -3 , two presented as "Moderate" and one as "Unhealthy for Sensitive Groups". In the summer rainy season, these limits were not exceeded, resulting in good air quality.
Point 2 samples presented a range of PM10 concentration of 29 μg m -3 in the summer season and 68 μg m -3 during spring time. The results showed a stable behavior throughout the sampling campaign, with only a slight decline in concentration in the summer. In winter, four samples exceeded 50 μg m -3 . In the spring, two samples exceeded 50 μg m -3 , and one of them had the highest value recorded in the whole campaign for this site, 68 μg m -3 . In the summer, no sample exceeded the recommended limits.
Related to elemental concentration results, Table 1 displays the range of the values determined for the elemental concentration in the particulate matter collected in the air filters. Elements that presented values lower than their limits of detection, that is, Ag, Cd and Hg, were not included in the calculations.
The Brazilian legislation does not foresee maximum values allowed in particulate matter for elemental concentration (Querol et al., 2001, except for lead (CETESB, 2017. Comparing the obtained values in Table 1 with the values of samples of PM10 collected in China (Song et al., 2017), for instance, the Chinese results are higher. One reason is due to poor environmental control. Comparing the same results with those from another Brazilian capital, São Paulo, there are similarities of pollution in terms of chemical elements and elemental concentrations (Albuquerque et al., 2012;Almeida et al., 2017). European elemental concentrations are below those found in this study. The report points out a tendency of reduction of pollution, due to the public policies in those countries (Leoni et al., 2018). < DL, lower than the value corresponding to the limit of detection.

Statistical analysis
PCA was used after normalization for statistical analysis. The variables used were the chemical elements of the samples of the PM10 and their elemental concentrations. As the variables were higher than the number of samples, it was necessary to reduce the number of variables, according to Mingoti (Mingoti, 2005). This was based on the concentration frequency of the variables, that is, concentration values with smaller occurrences in the samples were excluded from the analysis.
The PCA and the Varimax rotation were used for better data interpretation, providing the factors that contribute to the sample matrix. The results of PCA were highlighted that showed greater relevance among the factors of each variable (Mingoti, 2005).

Site: Point 1
The software Minitab reduced the original variables of sampling site Point 1 according to Mingoti (Mingoti, 2005), obtaining two factors, shown in Table 2. The Kaiser criterion (Rebelo et al., 2015;Stobinski et al., 2014) was taken into account, using values of the eigenvalues higher than 1. For this analysis, sixteen variables were used (Ba, Br, Ca, Ce, Co, Cr, Fe, Hf, K, La, Na, Sb, Sc, Tb, Th and Zn) of the thirty-three obtained. It was verified that the accumulated variance is 88.7% of the data. Factor 1 explained 70.2% of the total variation, where the elements Ba, Br, Ca, Co, Cr, Fe, Hf, Na, Sb, Sc, Th and Zn contributed to this first factor. From the characterization of the factor loadings, the main contributors can be identified. These elements indicate a strong influence of geogenic origin (Destefan and Holanda, 2011;Moreno et al., 2017;Noce et al., 1997), although it is difficult to differentiate its origin due to the processes of transport and resuspension of the soil. The elements Ba, Na and Zn strengthen even more this origin due to their elemental concentrations, averaging 165 ng m -3 , 328 ng m -3 and 128 ng m -3 , respectively. The elements in Factor 1 may come mainly from streets that are covered by paving stones and close to an ecological reserve.
Factor 2 explains 18.5% of the total variation, where the elements Ce and La are the main contributors for this factor. These elements indicate, according to the PM speciation profiles repository of air pollution sources of the United States Environmental Protection Agency -Igor Felipe Silva Moura et al.
Speciate (USEPA, 2015), influence from fuel combustion (gasoline, oil and diesel) of automotive supplies and paving stones of the municipal road in front of the sampled site (Silva et al., 2009;Xue et al., 2006). The elements La and Ce have average concentrations of 0.06 ng m -3 and 0.13 ng m -3 , respectively. There are no specific studies of isotopic ratios of rare earths present in PM in Brazil. In addition to these influences, the burning of organic material and construction can also contribute to the presence of these elements. The relationship between La and Ce of the analyzed samples showed an increase of approximately 10% in comparison with the crustal source of Belo Horizonte (Noce et al., 1997).

Site: Point 2
The original variables of the samplings carried out in Point 2 were reduced, according to Mingoti (Mingoti, 2005), based upon the concentration frequency of the variables. The Kaiser criterion (Rebelo et al., 2015;Stobinski et al., 2014) was also followed. Three factors were obtained as shown in Table 3, with a cumulative variance of 84.9% of the variability in the data. For this analysis, 17 variables were used (Ba, Ca, Ce, Co, Cr, Fe, Hf, K, La, Na, Sb, Sc, Sm, Ta, Tb, Th, Zn) from the thirty-three obtained, due to the removal of variables that were not found elementary concentrations representative for the calculation of PCA. Factor 1 explains 43.5% of the total variation, where 10 elements contributed to this factor (Ba, Ca, Co, Cr, Fe, Hf, Na, Sb, Sc and Zn). Through the factorial loads, the elements of this factor indicate a strong influence of geogenic origin (Noce et al., 1997).
Factor 2 explains 33.3% of the total variation, where Th, Tb, Ce and La contributed to this factor. Through factor loads, there is a strong influence on fuel combustion, similar to the second factor at Point 1 (Silva et al., 2009;Xue et al., 2006), which identified sources from automotive fuel burning, burning of organic material (leaves of trees), buildings, and coal combustion (wood burning) (Someshwar and Jain, 2003;USEPA, 2015).
Factor 3 explains 8.1% of the total variation, where Ta and K contributed to this factor. Factorial loads show the influence of oil combustion, cigarettes, fertilizers, Portland cement, diesel and furnace, due to the maintenance of the wooded sites at the university and the buildings in the vicinity of the sampling site (Guillén et al., 2017;USEPA, 2017

CONCLUSION
The first sampling site (Point 1) is near buildings and streets with a constant traffic of vehicles. The second sampling site (Point 2) is near an avenue with intense and constant traffic of vehicles (cars, buses and trucks). Concerning the concentration of airborne particulate matter, PM10, Point 2 presented higher range of values than Point 1. The highest value, in spring time, was considered "Unhealthy for Sensitive Groups" according to US-EPA. These results point out that this site near of the avenue presents more health risks. In general, the values were 40% higher than those recommended by legislation. In the summer rainy season, no sample exceeded the recommended limits.
The analysis of the elemental composition of PM10, resulted in the determination of thirtythree elements. This information was fundamental for the verification of possible PM10 emitters present in the sampling sites by means of the statistical technique of multivariate analysis, using the main component method.
Two factors were identified for the first sampled site and three factors for the second site. It was observed that in both locations the air quality is influenced by vehicular emissions, though more strongly at Point 2. At Point 1, the influence of the dust from the soil was confirmed. At this point, the lower contribution of the vehicular sources was expected due to its central position in relation to the avenues that surround the site. However, the results suggest that the pollutants are transported from the streets in direction of the study area.
This study suggests that air quality monitoring should be conducted periodically to identify and alert the population of the health risks related to air pollution. This monitoring should also provide data for a better statistic for the study of local air quality. Therefore, the methodologies used in this study to determine airborne particulate matter concentration, the chemical composition and a statistical tool for multivariate analysis, are useful in PM10 research.