1 Introduction

Nowadays, air pollution is one of the world's most serious environmental issues [1]. Every year, the presence of diverse sources of pollution and airborne particles kills millions of people around the world. According to the World Health Organization (WHO), about 92% of the world’s population lives in polluted areas, and of the total deaths in the world, 11.6% of them are due to air pollution exposure [2].

The amount of air pollution is increased by several causes, including a lack of strategic planning in urbanization concerns, the use of private cars instead of public transportation, and the development of diverse industrial sectors near cities [3, 4]. Furthermore, meteorological parameters such as temperature, wind speed and direction, relative humidity, and rainfall have an impact on air quality [4,5,6]. As a citation, several studies have reported the potential impact of meteorological factors on the ambient air quality [4, 7,8,9,10].

Iran, as a developing country, is no exception to the world’s present air pollution problems. As a result, hundreds of Iranians died of air pollution in 2017 [11]. According to the studies, Tabriz has been identified as one of Iran’s most polluted cities [2, 12, 13].

Tabriz’s Air Quality Monitoring Stations (AQMS), which is part of the Environmental Protection Organization, has established nine (spatial) stations in the city. These stations take hourly (temporal) measurements of air pollution concentrations, resulting in massive volumes of spatiotemporal data. This recorded information should be analyzed to monitor and control the pollution level and finally take appropriate actions. Traditional approaches are unable to process and interpret massive data with spatiotemporal dependencies due to their complexity. Appropriate data mining approaches are required to analyze this heterogeneous data and derive valuable information.

Clustering is one of the most important statistical approaches in data mining since it examines enormous datasets such as pollution levels in the air. Clustering is the process of grouping objects in datasets that have similar features together. In other words, it is a data function that divides observations into subgroups based on their similarity and identifies the data pattern. Clusters can be conveniently studied and interpreted due to clustering’s ability to reduce data size. As a result, in air pollution data with a large amount, clustering can recognize the pattern and provide more clear interpretation by lowering the amount of the data. The effects of meteorological factors on pollutant concentrations inside each cluster can be studied using obtained clusters of air pollution data.

Many studies have used clustering techniques to analyze spatial and temporal data. Some have used cluster analysis without considering the temporal dependency; they clustered the observed sites to identify the spatial pattern [14,15,16,17]. Some have considered the data as a set of time series in different places [14, 16, 18, 19], which did not consider the spatial nature of data. Some of them have also done clustering regardless of the spatial and temporal nature of data [20,21,22]. Models appeared to have insufficient accuracy and validity due to considering only one dimension (time or place) or none of them.

Recently, a new model-based clustering method for spatiotemporal data has been introduced by Cheam et al. [1], which considers both spatial and temporal dimensions for fitting a mixture model for spatiotemporal data. Changes in time and space were considered simultaneously in this method. Thus, we can cluster the days of the year according to the spatiotemporal information of pollutant concentrations.

Many studies have been conducted in Tabriz to investigate the level of air pollution concentration and its relationship with meteorological factors [23,24,25,26]. However, we did not find any study that clusters the pollutant concentrations according to the spatial and temporal dimensions and justifies extracted clusters with meteorological factors in Tabriz so far.

Therefore, the purpose of this study was to cluster the days of 2017 according to the hourly concentrations of O3 and PM10 collected from four stations of Tabriz using the mixture model–based clustering for spatiotemporal data and assessing the association of meteorological factors and concentration within detected clusters.

2 Method

2.1 Data

Tabriz is the capital of East Azerbaijan in Iran, and it is one of the country’s largest and most industrialized cities. Tabriz has a population of almost 1.8 million people and a land area of 320 square kilometers [24], according to the most recent census in 2017. Tabriz has four seasons and is semi-arid in terms of climate, so that it is rainy and moderate in spring, hot and dry in summer, rainy in autumn, and cold and snowy in winter. The ambient temperature reaches 30–35 °C on hot summer days and − 10 to − 20 °C on cold winter days [27]. Furthermore, the average yearly wind speed is 1.65 m per second [24, 28]. In 2017, the average temperature was around 13.32 °C, with the lowest and highest temperatures of (− 5, 39.40 °C), average relative humidity of 20.32%, and average wind speed of 7.02 km per hour, according to Tabriz Meteorological Organization reports.

Based on the information received from AQMS, the recorded hourly data on the pollutant concentrations in 2017 were more accurate and complete than in other recent years. As a result, the pollutant data collected in 2017 was selected to assess and analyze.

PM10 (particulate matter with an aerodynamic diameter less than 10 μm) and O3 (surface ozone) concentrations were measured at each station using beta attenuation and UV spectrophotometry, respectively. Table 1 shows the geographical coordinates of four quality monitoring stations in Tabriz (see Table 1).

Table 1 Geographical coordinates of air quality monitoring station
Full size table

Air quality information usually contains inaccurate, missing, and outlier data. As a result, the available information was examined for outliers and missing data. The outlier and inaccurate data were removed by the z-score method before inputing the missing data, estimating the parameters, and data mining [2, 29, 30]. As a result, the hourly concentrations of monitoring stations were compared to time trend data from the same stations and data from nearby monitoring stations. First, the original data were converted into z-scores (with mean = 0 and SD = 1) based on the following settings: (1) having \(|z|>4(|{z}_{t}>4|)\), (2) the difference from the prior value being greater than \(9\left({z}_{t}-{z}_{t-1 }>9\right)\), (3) the ratio of the z-score value to its centered moving average of order 3(MA3) being greater than \(2({z}_{t}/MA3\left({z}_{t}\right)>2)\), and (4) the difference between the singular monitoring air quality station and the prior value (\({z}_{t}-{z}_{t-1}\)) being at least twice greater than the difference between the city’s monitoring air quality station’s averaged increment (\(\mathrm{city}({z}_{t}-{z}_{t-1 })\)) and the prior value \({[(z}_{t}-{z}_{t-1 })/city({z}_{t}-{z}_{t-1 })]>2\), and then the outlier data were deleted from the original data.

According to WHO guidelines, there must be at least 50% valid information of 1 year and the desired station for statistical analysis, so the information for January to December 2017, which had at least 50% of the complete data, was selected (250 days). According to the instructions mentioned for the valid data in AQMS, PM10 and O3 data of four monitoring stations were considered in this research (Abrasan, Baghshamal, Rahahan, and Rastekoche). It should be noted that only the hourly information of 4 out of 9 stations was available. Of the total data available for pollutants (PM10, O3), the percentage of missing data was (11.88%, 0.6%) for Abrasan, (3.3%, 3.22%) for Baghshamal, (2.22%, 1.68%) for Rahahan, and (2.38%, 1.58%) for Rastekoche. The missing data were inputed using the linear interpolation method [31] with R (4.0.2) software (package: imputeTS version 3.1) to improve the accuracy and validity of the results. Hourly concentrations of each pollutant (PM10, O3) are used to compute the diurnal concentration. The acquired diurnal data are then utilized to assess the variations in air pollution in each cluster. Meteorological data for Tabriz was gathered from https://en.tutiempo.net in 2017.

2.2 Statistical Analysis

The spatiotemporal air pollution data were clustered by fitting a mixture model–based that took into account the spatiotemporal dimensions (STMC). To fit the model, spatial (geographical coordinates of data recording stations) and temporal (day and hour of recorded data) information were taken into account, an optimal number of clusters was determined using the Bayesian Information Criterion (BIC) [32] and Integrated Completed Likelihood criterion (ICL) [33], and the model parameters were estimated using the EM algorithm. Finally, each observation was categorized into an appropriate cluster.

STMC was compared to temporal mixture model–based clustering that only considered the temporal dimension (TMC) and mixture model–based clustering that did not consider the spatiotemporal dimensions (MC) to see how well it performed. BIC and ICL are the criteria used to compare the fitted models. The TMC and MC models were fitted in the same manner as the STMC models, except for a difference in considering the dimensions.

The nonparametric Mann–Kendall test (MK) [2, 34,35,36] was calculated to investigate the relationship between meteorological factors and pollutant concentrations (PM10 and O3) within the detected clusters.

R (4.0.2) software with “SpaTimeClus” (version 1.0) and “mclust” (version 5.4.6) packages were used for statistical analysis. The following is a brief explanation of STMC.

2.2.1 Mixture models for Spatiotemporal Data

Mixture model–based clustering is a broad family of algorithms designed for modeling an unknown distribution as a mixture of simpler distributions, sometimes called basis distribution [37, 38]. In this method, unlike other clustering methods that cluster the data based on some similarity measures, a Gaussian statistical distribution was considered for data. Thus, the purpose of model-based clustering is to estimate the parameters of the statistical distribution and hidden variables, considered as a cluster label of the data [37, 39]. Moreover, this model is applicable to a variety of data kinds such as continuous, ordinal, categorical, mixed, and functional.

Extension of mixture model–based clustering for spatiotemporal data was introduced in 2017 by Cheam et al. [1], which is a generalization for mixture model–based clustering considering only the temporal dimension [40]. Besides, the STMC method considers the variations in time and space simultaneously. Suppose each observation \({x}_{i}=\left\{{x}_{ijt}\right\}\begin{array}{c}t=1,\dots ,T\\ j=1,\dots ,J\end{array}\), as spatiotemporal data, including \(\mathrm{J}\times \mathrm{T}\) observations on the predetermined temporal framework \(m=\left({m}_{1},\dots ,{m}_{T}\right)\) and the spatial framework \(s=\left({s}_{1},\dots ,{s}_{J}\right)\). Therefore, \({\mathrm{x}}_{\mathrm{ijt}}\in {\mathbb{R}}\) is the observation i at location j and time t. For spatial coordinates, location j is a two-dimensional vector \({s}_{j}=\left({v}_{j},{w}_{j}\right)\).

The density function of the cth element (or cluster) for the spatiotemporal model is as follows:

$${f}_{c}\left({x}_{i}|{\vartheta }_{c}\right)=\prod \begin{array}{c}J\\ j=1\end{array}\prod \begin{array}{c}T\\ t=1\end{array}{\sum }_{h=1}^{H}{\Upsilon }_{cjth}\left({a}_{c}\right)\times \varphi \left({x}_{ijt}|{\left[{M}_{t}-{M}_{t-1}\right]}^{^{\prime}}{\beta }_{ch}+{x}_{ij\left(t-1\right)},\sigma \begin{array}{c}2\\ ch\end{array}\right)$$
(1)

so that \({\vartheta }_{c}=\left({a}_{c},{\beta }_{ch},\varrho \begin{array}{c}2\\ ch\end{array},h=1,\dots ,H\right)\) is a set of parameters of the element cth, \({M}_{t}=\left(1,{m}_{t},\dots ,m\begin{array}{c}\varrho \\ t\end{array}\right)\) the polynomial vector with Q-degree of \({M}_{t}\), \({\beta }_{ch}={\left({\beta }_{ch0},\dots ,{\beta }_{ch\varrho }\right)}^{^{\prime}}\) is the coefficient vector of the hth regression model for the cth element, and \(\phi \left(.|\mu ,{\varrho }^{2}\right)\) is the univariate Gaussian distribution density with mean \(\mu\) and variance \({\sigma }^{2}\). The weight of this mixture, \({\gamma }_{cjth}\), depends on both dimensions (time and space) via a logistic function. In the expectation and maximization steps of EM algorithm, the parameters of the model, which includes \({\vartheta }_{c}=\left({a}_{c},{\beta }_{ch},\sigma \begin{array}{c}2\\ ch\end{array},h=1,\dots ,5\right)\), were estimated. In this study, each \({x}_{ijt}\) represents the hourly concentration of the desired pollutant, i.e.,, observation i in time t (T = 24 h per day for n = 250 days) and location j (J = 4 stations).

2.2.2 Model Selection

Suppose \(M\) represents a set of selectable models with the optimal number of clusters. Each model \(\mathcal{M}\in \mathrm{M}\) is defined by three elements. C number of elements (number of clusters), H number of regressions of each element, and Q degree of polynomial regression. Therefore, for model M we have

$$\mathcal{M}=\left(\mathrm{C}.\mathrm{H}.\mathcal{Q}\right);\mathrm{ C}.\mathrm{H}.\mathcal{Q}\in {\mathbb{N}}^{*}.$$
(2)

so, M is obtained by applying the maximum number of each element. Thus, assuming \({C}_{\mathrm{min}}={H}_{\mathrm{min}}={\mathcal{Q}}_{\mathrm{min}}=1\), the number of selectable models is as follows:

$${\mathrm{Models}\left(M\right)=C}_{\mathrm{max}}\times {H}_{\mathrm{max}}\times {\mathcal{Q}}_{\mathrm{max}}$$
(3)

The selection of the best model based on BIC and ICL criteria, calculated for all models of M. Therefore, in this study, suitable elements for selecting the best models in STMC, TMC, and MC, based on the mentioned criteria, is obtained. Finally, each model with a suitable number of clusters is analyzed.

3 Results

Table 2 presents the information criteria and the number of free parameters for PM10 and O3 concentrations in all models (see Table 2). BIC for STMC was the highest in both pollutants. The number of free parameters of STMC in both pollutants was less than the other models. Since STMC had a better fitting on PM10 and O3 concentrations which is spatiotemporal data, it was selected as the final model. The number of optimal clusters in STMC was four for PM10 and three for O3.

Table 2 Goodness-of-fit criteria, number of free parameters, and number of clusters for the analyzed pollutants in each clustering model
Full size table

3.1 Description of Clusters Obtained from STMC

Figure 1 shows the scatter plot of the diurnal average concentrations of PM10 (right) and O3 (left), where each representative dot corresponds to the day profile of pollutant measurements for a cluster (see Fig. 1). Also, Table 3 presents descriptive information on the detected clusters and meteorological factors (see Table 3). It is important to note that the number of available days for analyzing the concentration variations of pollutants in each season was spring 93 days, summer 31 days, autumn 90 days, and winter 36 days.

Fig. 1
figure 1

Scatter plot of the diurnal average concentration of PM10 (right) and O3 (left) clusters in STMC model

Full size image
Table 3 Descriptive statistics for PM10 and O3 clusters obtained from STMC
Full size table

3.1.1 Description of Clusters Obtained for O3 Pollutant

In clustering O3 concentrations based on BIC and ICL criteria, STMC with three clusters was selected as the best model. O3 contains three clusters with average concentrations of low (cluster2), moderate (cluster1), and high (cluster3). O3-cluster1 encompasses 102 days which is the most member days.

Spring forms approximately 60% of this cluster, and the rest of the members are in O3-cluster3. The average concentration for spring in O3-cluster3 (65.51 \(\mu \mathrm{g}/{m}^{3}\)) is higher than O3-cluster1 (55.87 \(\mu \mathrm{g}/{m}^{3}\)). O3-cluster2 includes approximately 85% of autumn days (59 days), and the rest of the members are in O3-cluster1. Although O3-cluster1 encompasses less number of autumn days (22 days), the average concentration of O3-cluster1 (48.39 \(\mu \mathrm{g}/{m}^{3}\)) members is higher than O3-cluster2 (18.68 \(\mu \mathrm{g}/{m}^{3}\)) members. Additionally, most winter days (22 days) are in O3-cluster1.

3.1.2 Description of Clusters Obtained for PM10 Pollutant

In clustering PM10 concentrations based on BIC and ICL criteria, STMC with four clusters was selected as the best model. PM10-cluster4 has the highest number of days (90 days) and the highest average concentration (78.77 \(\mu \mathrm{g}/{m}^{3}\)). PM10-cluster1 and PM10-cluster2 were almost the same in member days (70 and 69 days, respectively), whereas the average concentration of PM10-cluster1 (49.69 \(\mu \mathrm{g}/{m}^{3}\)) was higher than the average concentration of PM10-cluster2 (43.81 \(\mu \mathrm{g}/{m}^{3}\)). PM10-cluster3 (41.87 \(\mu \mathrm{g}/{m}^{3}\)) contained the lowest member days (21 days), and its average concentration was close to PM10-cluster2.

Spring days were present in all clusters, and this season’s highest and lowest average concentrations belonged to PM10-cluster4 and PM10-cluster2, respectively. It is important to note that regarding the spring, PM10-cluster1 and PM10-cluster3 had moderate average concentrations and were close in values. However, PM10-cluster3 had a slightly lower average concentration than PM10-cluster1. Summer days were present in PM10-clusters 1, 2, and 4, but PM10-cluster1 included only 2 days which can be disregarded.

Additionally, PM10-cluster4 had a higher average concentration than PM10-cluster2. Autumn days were seen in PM10-cluster2 (approximately 60%) and PM10-cluster4 (approximately 37%). PM10-cluster4 containing low member days exhibited a higher average concentration than PM10-cluster2. Regarding winter, PM10-cluster1 included 50% of days and PM10-cluster4 30% of them. Furthermore, PM10-cluster4 had a higher average concentration than PM10-cluster1.

3.2 Analysis of the Correlation Between Pollutant Concentrations and Meteorological Factors Within Clusters in STMC

The MK test to assess the correlation between meteorological factors and the concentrations of PM10 and O3 was calculated, shown in Table 4 (see Table 4).

Table 4 Mann–Kendall correlation coefficient between meteorological parameters and PM10 and O3 clusters
Full size table

According to the MK’s test results, all clusters presented a statistically significant correlation between the average concentration of O3 and meteorological factors, except O3-cluster2 for rainfall. The relationship between the average concentration of O3 and temperature in all clusters was positive and weak to moderate with a very high statistical significance (p-value < 0.05); in respect of relative humidity, it was negative and moderate with a very high statistical significance (p-value < 0.05). Moreover, all clusters showed a positive and weak relationship between the average concentration of O3 and wind speed.

The MK’s test results for PM10 revealed a statistically significant correlation between average concentration and meteorological factors (temperature and relative humidity) in PM10-cluster2 and PM10-cluster3. The relationship between temperature and average concentration in PM10-cluster2 was positive and weak to moderate, with a very high statistical significance (p-value < 0.05), but it was negative and weak about relative humidity. The relationship between temperature and average concentration in PM10-cluster3 was positive and moderate to strong whereas the relationship between temperature and relative humidity was negative and moderate to strong. Furthermore, in PM10-cluster1, PM10-cluster2, and PM10-cluster3, the relationship between rainfall and average concentration was negative and weak. It is important to note that there was no statistically significant correlation between PM10 and wind speed.

4 Discussion

The goal of our study was to apply a mixture model–based clustering method for clustering the days according to hourly concentrations of PM10 and O3 over the 1 year in Tabriz. This method does not consider the type of data and can be used for various types such as continuous, ordinal, categorical, mixed, and functional. Spatial and temporal information is always available in the air pollution datasets, which is a type of functional data. Therefore, the mixture model–based clustering method is a proper choice for analyzing this functional data with spatial and temporal dependencies.

STMC with spatial and temporal dependencies, TMC with temporal dependency, and MC lacking in both dependencies were fitted to data to assess the applicability of statistical models for our data. BIC, ICL, and number of free parameters were calculated for these models. The results showed that the STMC compared to TMC and MC had a better fitting on spatiotemporal data of PM10 and O3. The following is a review of various pollutant clustering studies in which none of the dependencies (spatial and temporal) are simultaneously considered.

Jin et al. [41] demonstrated the application of the k-means clustering method to identify O3 spatial regimes (or clusters) over California’s San Joaquin Valley. Clusters show the days having the same O3 geographical distribution. In terms of concentration, of a total of six recognized regimes, two corresponded to low-O3-cluster, three to moderate-O3-cluster, and one to high-O3-cluster. Moreover, meteorological measurements were used to describe O3 spatial distributions and their correlation to those in San Francisco Basin.

Pandey et al. [42] demonstrated the spatial and temporal variability of PM1.0, PM2.5, PM10, NO2, and SO2 in India using the average linkage clustering approach. Clusters including monitoring sites represent similar behavior in terms of pollutant dispersions and spatial variations. According to the results, concentrations of all varieties of PM10 were highest in the winter and lowest in the wet season at all sites.

Huang et al. [43] investigated the characteristics of PM2.5 in China using a hierarchical clustering approach. According to the results, PM2.5 information, collected from 13 monitoring sites, was arranged in 3 clusters. One of the most important findings was that the temporal distribution of PM2.5 demonstrated that winter had the highest concentration, and fall has a higher concentration value than spring, and the lowest concentration belongs to summer. In terms of spatial distribution, three out of 13 monitoring sites exhibited the highest concentration of PM2.5.

In a study conducted in northern China [44], cluster analysis was used to reveal the spatial and temporal distribution of PM2.5, SO2, NO2, CO, and O3 pollutants. Pearson’s correlation coefficient was used to investigate the relationship between pollutant concentrations with each other. Hierarchical cluster analysis was used in this study to categorize nine cities into various groups based on a monthly average of the above pollutants in each city. Using cluster analysis, nine cities were divided into four clusters based on the monthly average of pollutants. According to the clustering analysis, air pollution was mainly associated with industrial city structures and geographical and socioeconomic factors.

Many studies on the relationship between pollutant concentrations and meteorological factors have been conducted. They indicate that the hourly, daily, monthly, and seasonal variations of air pollutants in a residential area can be caused by meteorological factors such as atmospheric temperature, relative humidity, wind speed, solar radiation intensity, and so on. The amount of pollutant concentration will be affected by changing the above elements during the seasons [2, 45, 46].

Among the meteorological factors, atmospheric stability and wind speed have the most influence on atmospheric dispersion and decreasing air pollution [30, 34]. The intensification of air pollution in Tabriz can be attributed to the temperature inversions and calm conditions.

Geographical characteristics, meteorological variables, and residential buildings in the direction of the wind entering the city all contribute to the city’s current situation. It should be mentioned that the data show a fluctuating variation in PM10 and O3 concentrations in Tabriz from 2006 to 2017 [25].

Based on the results of this study, days with similar average temperature, relative humidity, and rainfall are grouped in a cluster. O3-cluster3 with the highest temperature and lowest relative humidity has the hottest days of the year, while the coldest days belong to O3-cluster2 with the lowest temperature and highest relative humidity. Therefore, based on the average concentration of each cluster and the MK test, the effect of meteorological factors can be observed so that concentration has a direct and positive relationship with temperature and has an inverse and negative relationship with relative humidity and rain. In the following, the reasons for increasing and decreasing O3 concentration on hot and cold days of the year are briefly indicated.

Based on most conducted studies, there was a clear and logical trend in the monthly and seasonal variations of O3 concentration. The highest and lowest values of O3 concentration were reported in summer and winter, respectively, according to the variation pattern. This highest value may be due to increased atmospheric temperature, increased intensity of sunlight, long days, and long sunny hours in hot seasons, all of which increase photochemical reactions and O3 production, whereas the lowest value is related to reduced daylight (sunlight time), lower temperature, and sunshine duration [2, 45, 47]. O3 generation is also influenced by several processes and activities such as transport, deposition, and NOx titration. The increase in O3 concentration in the warm seasons is directly related to the increment degree of temperature that is one of the main parameters in controlling the O3 formation [48, 49].

According to the results of this study, PM10-cluster4 had the highest temperature. PM10-cluster3 and PM10-cluster2 had lower temperatures, higher rainfall, and higher wind speed. It is important to note that PM10-cluster3 and PM10-cluster2 were close in values with slight differences. In the following, the effect of meteorological factors on PM10 concentration in different seasons is summarized.

Because of the Asian dust phenomenon, PM10 concentrations may rise in the spring and summer [36, 50]. Furthermore, the concentration of PM10 rises during warmer seasons due to high ambient temperatures and low relative humidity. Moreover, the correlation between PM10 concentration with temperature and relative humidity is positive and negative, respectively [8].

The studies on the relationship between PM10 and rainfall show that the washout effects of rainfall reduce PM10 concentrations in the atmosphere. Rainfall and relative humidity eventually have a negative correlation with PM10 [51].

According to the results of the analysis of the average concentrations of O3 and PM10 in each cluster, the relationship between meteorological factors and average concentration in each cluster, and the member days of seasons in individual clusters, it is possible to conclude that the clustering method used in this study has good fitness, and the results were interpretable. Finally, model-based clustering provides a strong basis and shows excellent efficiency in analyzing data with spatiotemporal dimensions. Using the above approach on datasets with spatial and temporal information produces more reliable and accurate results [1].

5 Conclusion

Data mining methods are very applicable in evaluating air pollution data, and the results play a significant role in managing and preventing the accumulation of pollutants. Clustering, one of the data mining approaches, reduces the amount of data examined while revealing hidden information. As a result, analyzing individual clusters with a small amount of data decreases the errors in the results. In this study, the STMC using spatiotemporal dimensions classify the days in Tabriz, 2017, based on their PM10 and O3 concentrations. The results demonstrated that considering the dimensions of the data in the analysis could increase the efficacy of the clustering in our spatiotemporal air pollution data. The findings of STMC's examination of the obtained clusters in terms of available meteorological parameters suggested that the clustering was acceptable and meaningful.

6 Limitation

This study may have some limitations:

  • Over the year 2017, there were insufficient measurements of pollutant concentrations in the winter and summer. This insufficiency was evident in the data collected in January, February, July, and September.

  • Due to the different sources for pollutant production in each region, geographical conditions and meteorological factors that influence on the pollutant level variations, it was impossible to directly compare with the results of the other studies.