Exploring the background features of acidic and basic air pollutants around
an industrial complex using data mining approach
Ho-Wen Chen
a,1, Ching-Tsan Tsai
b,1, Chin-Wen She
a, Yo-Chen Lin
b, Chow-Feng Chiang
b,⇑ aDepartment of Environmental Science and Engineering, Tunghai University, 181 Section 3, Tauchung Port Road, Taichung 407, Taiwan
b
Department of Public Health, China Medical University, 91 Hsueh-Shih Road, Taichung 404, Taiwan
a r t i c l e
i n f o
Article history: Received 9 March 2010
Received in revised form 27 July 2010 Accepted 8 August 2010
Available online xxxx Keywords:
Cluster analysis High-tech industrial park Inorganic acid and base Spatial and temporal background distribution
a b s t r a c t
Air pollution data around a monitored site are normally difficult to analyze due to highly inter-related meteorological and topographical factors on top of many complicated atmospheric chemical interac-tions occurred in local and regional wind fields. The challenge prompts this study to develop a compre-hensive data-mining algorithm of cluster analysis followed by meteorological and interspecies correlations to mitigate the inherent data complexity and dissimilarity. This study investigated the background features of acidic and basic air pollutants around a high-tech industrial park in Taiwan. Monthly samplings were taken at 10 sites around the park in a year. The temporal distribution plots show a baseline with two characteristic groups of high and low peaks. Hierarchical cluster analysis con-firms that high peaks were primarily associated with low speed south wind in summer for all the chem-ical species, except for F, Cl, NH
3and HF. Crosschecking with the topographical map identifies several
major external sources in south and southwest. Further meteorological correlation suggests that HCl is highly positively associated with humidity, while Clis highly negatively associated with temperature,
both for most stations. Interestingly, HNO3is highly negatively associated with wind speed for most
stations and the hotspot was found in summer and around the foothill of Da-Tu Mountain in the north-west, a stagnant pocket on the study site. However, Fis highly positively associated with wind speed
at downwind stations to the prevailing north wind in winter, indicating an internal source from the north. The presence of NHþ
4 stimulates the formation of NO 3, SO
2
4 (R = 0.7), and HNO3, H2SO4, NH3
(R = 0.3–0.4). As H2SO4could be elevated to a level as high as 40% of the regulated standard, species
interactions may be a dominate mechanism responsible for the substantial increase in summer from external sources.
Ó 2010 Elsevier Ltd. All rights reserved.
1. Introduction
As of 1998, 85 industrial parks have been officially built, oper-ated or planned in Taiwan to ensure the island’s continual eco-nomic growth. Due to the limited land area in Taiwan, most industrial complexes are inevitably situated adjacent to urban development areas where people reside and work. The Taichung site of Central Taiwan Science Park (CTSP) is one of the most important bases worldwide for the high-tech industries of semi-conductors, electronics and electrical peripherals with a total investment of USD 25 billions. The South part of Taichung site is surrounded by the Greater Taichung area residing about 1.5 mil-lions populations. Inorganic acidic (HF, HCl, HNO3, H2SO4) and
ba-sic (NH4OH) chemicals are the concerned pollutants emitted from
the cleaning and etching processes (Tsai et al., 2003). However, some of these pollutants may be attributed to external sources such as urban activity, power plant or other combustion processes (Orel and Seinfeld, 1997). In addition, ocean is another potential source, which may compound the complexity in source identifica-tion. Accordingly, understanding the temporal and spatial back-ground features of pollutants around the site is critical to addressing the public concerns and searching for the responsible parties in environmental forensics (Gargava and Aggarwal, 1996; Neves, 1996). Many early studies covered a variety of aspects, including formation mechanisms (Baek et al., 2004; Tsai et al., 2004), health effects (Spengler et al., 1990), concentration con-tours, hot spots and source identification (Danalatos and Glavas, 1999; Bari et al., 2003; Walker et al., 2004; Chittaro et al., 2006; Horng and Cheng, 2008; Zhu et al., 2008; Hosseinlou and Sohrabi, 2009), and meteorological effects (Webb and Vincent, 1999; Jäggi et al., 2006; Pejman et al., 2009). However, little effort was made to develop a comprehensive data analysis approach for compli-cated air pollution systems like the Taichung site.
0045-6535/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.chemosphere.2010.08.019
⇑ Corresponding author. Tel.: +886 4 22053366; fax: +886 4 22037717. E-mail address:amur.chiang@gmail.com(C.-F. Chiang).
1 H.W. Chen and C.T. Tsai have equal contribution to this study.
Contents lists available atScienceDirect
Chemosphere
j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / c h e m o s p h e r e
Prior to the full operation, the site was monitored every month for five gaseous species (HCl, HF, HNO3, H2SO4and NH3) and five
fine particles ionic species (Cl, F, NO 3, SO
2
4 and NH
+4
) at 10 stations around the site for 1 year. The primary objective of this study was to develop a comprehensive data-mining algorithm to explore the background features of acidic and basic pollutants around the Taichung site. Special effort was made to identify po-tential external sources to the study site under the meteorological and topographical constraints. Accordingly, an algorithm employ-ing hierarchical cluster analysis (HCA) followed by meteorological and interspecies correlation analysis was developed in this study so that monitored data with specified similarity could be simplified into several groups for better insight.
2. Materials and methods 2.1. Study site and sampling stations
The Taichung site of the Central Taiwan Science Park (CTSP) launched its first test run in middle 2005, but was not fully oper-ated until early 2007.Fig. 1a shows a site map of the Taichung site, having a total area of 413 hectares. The east and south of the site are immediately connected to the Greater Taichung Metropolitan area with 1.5 millions of population. About 5 km to the south is the Taichung Industrial Park, a conventional machinery industry complex being heavily operated for about 30 years. About 3 km to the southwest is the Taichung Refuse Incinerator with a daily
Fig. 1. The Taichung site of the Central Taiwan Scientific Park investigated in this study, showing: (a) the surrounding potential external pollution sources and 10 sampling stations around the site, (b) three-dimensional topographical map of the study site.
throughput of 900 wet tones. About 2 km to the east is a national highway system. The coast line of the Taiwan Strait and Taichung coal-burned power plant are about 15 km to the east of the study site. The topography of the site is characterized by a low altitude terrain of 50–300 m above the sea level, declining from northwest
to southeast as shown inFig. 1b.
As shown in Fig. 1a, a total of 10 sampling sites distribute
around the boundary of the factory lots and evenly covers the en-tire area on the Taichung site. Stations 6–8 locates in the north, Stations 1, 5, 9 in the south, and 2, 4, 10 in the east. Station 3 is the only site located within the factory complex.
2.2. Sampling device and schedule
In each sampling event, 10 chemical species were monitored, including five ionic species (NHþ
4, F
, Cl, NO
3and SO
2
4 ) on
partic-ulate matter and five gaseous species (HF, HCl, NH3, HNO3 and
H2SO4). The air samples were collected by a porous metal denuder
which was well reported byTsai et al. (2003, 2004). The denuder
contains three primary structured units: a single-stage impactor
with a cutoff aerodynamic diameter of 2.5
l
m, two porous metaldiscs with a pore size of 100
l
m with a disc diameter of 4.7 cm,and a three-stage filter pack. As air samples enter into the first
stage impactor, coarse particles are removed. The remaining gas-eous pollutants are then absorbed by a layer of sodium carbonate (5% w/v) and citric acid (4% w/v) coated on each of the porous me-tal discs for acidic and basic species. The filter packs consist of a Teflon filter (Gelman Science, 2
l
m pore size) to collect fine parti-cles, a 4.7-cm nylon filter (Gelman Science, 1l
m pore size) tocol-lect HNO3and HCl, and a 4.7-cm glass fiber filter (AP40, Millipore
Inc.) coated with citric acid to collect NH3that volatilizes from the
collected particles on the first layer of Teflon filter. Particulate NO 3
and Clconcentrations were determined as the sum of those
col-lected on the Teflon filter and the nylon filter, respectively. The
flow rate was maintained at 2 L min1. Samplers were placed at
an average height of 1.5–2 m from the ground level. The samplings were taken simultaneously at all 10 sampling sites once every month for a period of 18 h in a day throughout the entire study from March 2005 to February 2006, prior to the full-scaled opera-tion of the site.
2.3. The algorithm of data analysis
A total of 1120 (eight stations 12 months + two
sta-tions 8 months) valid air data were obtained, each monitoring 10 chemical species for a period of 12 months from March 2005
Draw clustered radar chart of
average meteorological
measurements
Classify all the monthly wind rose diagrams into the selected
clusters
Perform meteorological
correlation for each species Normalize the averaged concentrations with the
overall average for each species
Perform hierarchical cluster analysis
Draw clustered radar chart of average concentrations for ionic species
Draw clustered radar chart of average concentrations for gaseous species Select the one with a higher temporal or spatial
variation as the basis for clustering
Calculate average concentrations for cluster analysis for each species
Perform ANOVA among temporal concentrations for each species
Perform ANOVA among spatial concentrations for each species
Perform species correlation for
each species
Fig. 2. The algorithm of data mining approach developed in this study, incorporating a preliminary ANOVA and hierarchical cluster analysis followed by a confirmatory meteorological and interspecies correlation.
to February 2006. The sampling was collected once every month. On each sampling date, four meteorological parameters (wind speed, wind direction, ambient air moisture, and ambient air tem-perature) were also obtained each hour for 14 h (5 am–7 pm) from a local meteorological station. This gives a total of 762 (four param-eters 14 h 12 months) meteorological data. The air pollutant and meteorological data were analyzed using a novel hierarchical cluster analysis (HCA) approach developed in this study. For the type of air dispersion concentration data inherent with high dis-similarity like this study, it is advantageous to aggregating mea-surements into different groups with spatial or temporal similarity (Lee et al., 2001; Reghunath et al., 2002; Banerjee et al., 2009). The HCA algorithm shown inFig. 2consists of eight steps and are briefly stated.
1. For each chemical species, use one way ANOVA to determine whether the monitored concentrations are significantly dif-ferent among the sampling sites and/or among sampling times during the entire course of study.
2. Select the one with a higher temporal or spatial variation as the basis for clustering.
3. If temporal basis is selected, calculate the spatial-averaged concentration for each sampling time and each chemical species, and vice versa. For easy understanding, the temporal basis is assumed in the following steps.
4. Calculate the yearly arithmetic mean of all the station-aver-aged concentration for each chemical species obtained in Step 3.
5. Divide the monthly station-averaged concentration of each chemical species by the corresponding yearly mean concen-tration to obtain the corresponding standardized unitless concentration for all the sampling months.
6. Calculate the Euclidean distance (dij) between the two 10
species-dimensioned points of any two sampling months, dij¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P10
p¼1ðxip xjpÞ2
q
i, j denoting any two sampling months (1–12), p denoting the type of pollutant (1–10). It can be realized that dij= djiand dij= 0 as i = j. A total of 72
Fig. 3. Results of the temporal distributions of concentrations monitored in each month for the 10 sampling sites in this study, illustrating a pattern of two groups of peaks for: (a) HNO3and, (b) Cl.
(12 122) dijare calculated and used as the basis for
clus-tering in this study.
7. Plot the radar chart of average concentration for each clus-tered group determined in Step 6 by superimposing all the clusters on the same chart for comparison. The radar chart is useful for visual inspection as to whether there is a dis-tinct separation in concentration among the clustered groups for each chemical species.
8. Plot wind roses for each month using the hourly wind data and categorize each month of wind roses in the correspond-ing cluster as determined in Step 6. The clustered wind rose diagrams are useful in evaluating the effect of clustered wind data on the clustered radar chart of concentrations for each chemical species.
9. Perform correlation analysis between concentration and each meteorological parameter for each species and each station.
10. Perform correlation analysis between the station-averaged concentrations of any two species.
3. Results and discussion
3.1. Background temporal and spatial profiles
To examine whether a long-term temporal or spatial distribu-tion pattern existed for the gaseous and particulate species moni-tored in this study, their levels versus sampling months were superimposed on the same plot for all the 10 sampling stations for each chemical species.Fig. 3illustrates the distribution profiles
of HNO3and Clfrom March 2005 to February 2006 (12 months).
In general, these plots display varied temporal and spatial varia-tions for different chemical species. However, a general pattern of baseline with two characteristic groups of high and low peaks appearing in summer and winter could be distinctly identified for each species. For instance, as shown inFig. 3a, HNO3had a
base-line of 1–2 ppb with a summer peak of 4–7 ppb and a winter peak of 3–4 ppb. Nevertheless, an exceptional high summer peak of 7 ppb appeared at Station 8 and a winter peak of 8.5 ppb at Station
7 for HNO3. Interestingly, the two hotspots were located around
the foothill of Du-Tu Mountain to the northwest of the study site. In another instance shown inFig. 3b, particulate Clalso displayed
the pattern of two major groups, but with much wider and larger variations among different sampling sites. The other four ionic spe-cies also demonstrated the type of distribution. Ionic spespe-cies ab-sorbed on fine particles was likely associated with long-range transport and hence affected by regional sources. Two major regio-nal sources must be noted in this study. One is the Taichung coal-burned power plant and Taiwan Strait, about 15 km to the west of the study site. Another is the Taichung city and Taichung Industrial Park, about 5 km to the south of the study site. The variation could be also due to the ever changing surface winds (1.5–2 m above ground) over a typical complex terrain like the Taichung site. As the denuder sampling took a total of 14 h, it would be difficult to assure the actual upwind and downwind relationship.
Table 1shows the mean concentration and percent variation coefficient (CV) of the 10 sampling stations for each sampling month and each species. For such a relatively small site of 413 hectares (2800 m 1500 m), it would be justified to simplify the analysis by taking an arithmetic mean of the concentrations collected at all the sampling stations each month for each species. For the five gaseous species, the average concentration and the percent of the regulated standard (in parenthesis) was
0.5–2.0
l
g m3 (5–10%) for HF, 1–4 ppb (1–4%) for HCl, and0.5–3 ppb (1–8%) for HNO3, 5–20
l
g m3 (10–40%) for H2SO4,and 5–20 ppb (0.5–2%) for NH3, using the factor-surrounding air
quality standard (SAQS), 10
l
g m3 for total F including HF,100 ppb for HCl, 40 ppb for NHO3, 50 for
l
g m3for H2SO4, and1000 ppb for NH3, currently regulated by Taiwan Environmental
Protection Agency (Taiwan EPA, 2007). The above analysis shows that the average background gaseous concentrations in general
Table 1
Results of mean concentrations and variation coefficients (CV) of the 10 sampling sites for each chemical species monitored in this study.
HF H2SO4 HCl HNO3 NH3
Meana
CV% Mean CV% Mean CV% Mean CV% Mean CV%
March-2005 0.6 10.0 16.1 17.0 3.1 207.4 0.9 12.2 20.4 33.1 April-2005 0.7 1.4 10.6 50.9 2.8 60.4 0.4 42.5 15.8 43.9 May-2005 0.6 5.0 11.3 34.5 1.7 20.0 1.0 17.0 9.6 31.9 June-2005 2.2 30.5 22.4 33.7 4.3 38.1 1.7 33.5 9.1 19.9 July-2005 1.1 33.6 19.8 42.7 1.2 40.0 2.9 52.1 15.5 73.9 August-2005 0.6 5.0 5.9 49.2 0.8 38.8 1.4 49.3 4.9 63.3 September-2005 1.2 56.7 6.5 38.5 1.4 13.6 1.0 32.0 16.3 29.1 October-2005 0.7 1.4 6.1 36.2 1.3 5.4 0.7 15.7 7.4 30.4 November-2005 1.2 45.8 12.1 51.3 1.2 18.3 0.4 17.5 9.5 22.5 December-2005 1.7 58.8 9.2 49.1 3.0 139.7 1.5 157.7 14.6 46.6 January-06 1.2 38.3 5.4 21.7 1.0 38.0 0.3 16.7 6.7 36.0 February-2006 0.9 40.0 4.0 41.5 1.3 23.9 0.8 30.0 6.8 33.7 F SO2 4 Cl NO 3 NHþ4
Mean CV% Mean CV% Mean CV% Mean CV% Mean CV%
March-2005 0.5 10.0 7.0 33.3 0.8 91.3 3.3 23.3 7.0 27.4 April-2005 0.6 11.7 5.4 21.1 0.6 5.0 0.6 38.3 3.7 15.1 May-2005 0.5 4.0 6.6 27.0 0.9 45.6 1.0 13.0 3.4 30.6 June-2005 0.6 21.7 3.3 25.8 0.9 53.3 0.8 30.0 3.1 23.6 July-2005 0.5 20.0 10.6 17.6 0.5 60.0 2.0 29.5 8.3 22.2 August-2005 0.5 6.0 10.9 48.6 0.4 20.0 0.9 47.8 4.4 35.5 September-2005 0.5 10.0 2.2 24.6 0.6 11.7 0.9 50.0 2.8 40.4 October-2005 0.5 2.0 2.4 36.3 0.7 17.1 0.5 38.0 2.6 6.9 November-2005 0.5 2.0 5.3 21.5 0.9 17.8 1.0 24.0 4.8 15.0 December-2005 0.7 8.6 2.3 32.2 1.2 38.3 1.6 26.9 3.9 9.0 January-2006 0.5 20.0 1.9 55.3 1.4 18.6 0.5 36.0 1.6 11.3 February-2006 0.5 30.0 7.5 47.6 0.7 25.7 2.2 64.6 4.4 56.8
a Unit is ppb for HCl, HNO
3and HN3and islg m3for the rest parameters.
contributed less than 10% of the regulated standard, except for the H2SO4. For ionic species, their mean levels of concentrations
ap-peared to have less spatial variations than the ones for gaseous
species. As also shown inTable 1, the CV among the 10 sampling
stations for each sampling month, disregarding a few outliers (>100%), varied from 10% to 50% for all the species. Nevertheless, the temporal distribution plot and spatial Cv calculation was inad-equate to generate more concrete findings.
Tsai et al. (2003, 2004)reported the monitoring results of a fully operated Scientific Park in northern Taiwan, having the manufac-turing processes similar to this study site. Their monitoring results
showed that HF and H2SO4 had a concentration as high as 5–
10
l
g m3 (5–10%) and 30–35l
g m3 (60–70%) around 13sam-pling stations in summer at a wind speed of 1.7 m s1. It should
be noted that HF is considered an indicator species to the
semicon-ductor manufacturing process, while H2SO4could be generated in
the photochemical reaction associated with the emission of SO2
from combustion, and favorably occurred in summer in the sub-tropical area like Taiwan. It is likely, for the study site, that
H2SO4is the chemical species most difficult to comply with the
regulated standard, particularly in summer, due to the potential transport from Taichung Industrial Park and Taichung City. 3.2. Hierarchical cluster analysis
To begin with the data-mining algorithm proposed in this study (Fig. 2), spatial and temporal variations were evaluated using one
way ANOVA approach for each chemical species for 10 sampling sites during 12 sampling months. The dependent variables were the concentrations of each chemical species, and the independent variables were either the sampling month or the sampling station, respectively. The results show that temporal variation in average concentration was significantly different among all the sampling stations (p < 0.001), while spatial variation was not among all the sampling months (p < 0.001). The relatively small area (413 hect-ares) of the study site may be part of the reason for the insignifi-cant spatial variation.
To explore features of the temporal variation, all the sampling months were grouped into three statistically meaningful clusters based on the hierarchical cluster analysis (HCA) proposed in this study.Fig. 4a shows the results of three clusters at a Euclidean dis-tance of 5–7, with Cluster I (October, January, February, and Au-gust), Cluster II (April, May, November, and December) and Cluster III (March, June, and July). The arithmetic mean of wind speed, relative humidity, and temperature of each clustered month was calculated and plotted in a triangle chart as shown inFig. 4b. It clearly shows the separation power of the cluster analysis pro-posed in this study, that the mean wind speed is noticeably lower (about 0.7 m s1) for Cluster III than that (about 1.1 m s1) for
Clus-ters I and II. However, the mean of relative humidity and temper-ature among the three clusters are not so visually different.
By applying the same segregation rule, the 12 months of wind
roses were grouped into three clusters inFig. 5. On the basis of
most frequent wind direction, Cluster III of wind roses primarily
(a) clustered categories
(b) meteorological conditions
(c) particulate acidic and basic pollutants
(d) gaseous acidic and basic pollutants
Fig. 4. Results of hierarchical cluster analysis showing: (a) three clusters of sampling months, (b) clustered triangle chart in average number for three meteorological parameters, (c) clustered radar charts in average concentration for five fine particle ionic species, (d) clustered radar charts in average concentration for five gaseous species.
selects for south and southeast wind in summer at a low speed of
2–4 m s1(June and July), with an exception for March at a high
speed of 6–7 m s1(March). High speed (4–7 m s1) of northeast
(October, January, February) and northwest (February and August) winter winds dominated for Cluster I. However, Cluster II shows no obvious pattern in wind speed and direction. Surprisingly, it was
Fig. 5. Results of three clusters of wind rose diagrams, showing high northern winds (>3 m s1) prevailed in winter for Cluster I, low southern winds (<3 m s1) dominated in
summer for Cluster III, and other types of winds for Cluster II.
also noted that the high speed (6–7 m s1) south wind of April was
grouped in Cluster II, while the low speed (2–4 m s1) northeast
wind of July was classified in Cluster III. As the cluster analysis is based on the concentrations monitored around the site for each species, the imprecise separation for wind roses reflects the com-plicate nature of an air dispersion system like the study site. Many local and regional factors other than wind roses may also affect the background concentrations of the study site.
To further explore the association between clustered meteoro-logical parameters and pollutant concentrations, two concentra-tion-clustered radar charts were made, one for the five ionic
species in Fig. 4c and another for the five gaseous species in
Fig. 4d. Interestingly, Cluster III was clearly separated from the
other two clusters for NO
3, SO 2
4 , and NH
þ
4 in Fig. 4c, and for
HNO3, H2SO4and HCl inFig. 4d. Several interesting points can be
noted after crosschecking between the clustered radar charts and
wind roses. The elevated levels of NO
3, SO 2
4 , and NH
þ
4 inFig. 4c
and HNO3, H2SO4and HCl inFig. 3d were primarily associated with
southern summer low wind in cluster III (Fig. 5), suggesting exter-nal sources existed in the open southern area to the study site as
displayed in the 3-D topography map ofFig. 1b. The average levels
of NO
3, SO24 , and NHþ4in Cluster III were significantly increased by
50–100% from those in the other two clusters. However, the levels of Fand ClinFig. 4c and HF and NH
3inFig. 4d remained
rela-tively stable in all clusters, suggesting that they were little influ-enced by wind speed and direction ands likely associated with internal sources. These analyses lead to a hypothesis that Taichung Industrial Park and Taichung City southern to the study site could be a significant external source contributing to the background
concentrations of NO
3, SO 2 4 , NH
þ
4, HNO3, H2SO4and HCl. On the
other hand, the levels of NH3, HCl and H2SO4were distinctly lower
in Cluster I than those in the other two clusters shown inFig. 4d, and are associated with the dispersion driven by northern high speed winds of Cluster I inFig. 5.
Although the use of clustering analysis may oversimplify the complicated nature of the air system, it does help to look into
the dominate factors affecting the levels of chemical species under study when a clustering basis is properly used. Theoretically the separating power of the proposed HCA can be enhanced as the monitoring frequency increases in a year, i.e., from once to twice a month (Lee et al., 2001; Reghunath et al., 2002; Banerjee et al., 2009). With the aid of clustered radar charts, clustered wind roses and three-dimensional topography map, the HCA algorithm pro-posed in this study offers a comprehensive approach to explore the features of a complicate air dispersion system like the Taichung site, to which external pollutant sources may play an important role in regional transport.
3.3. Meteorological and interspecies correlation
Taiwan is in a subtropical area, the relative humidity and ambi-ent temperature monitored were as high as 60–80% and 29–31 °C in summer and as low as 40–50% and 16–18 °C in winter. In order to examine the association between the monthly monitored con-centrations for each station and each of the three metrological parameters (relative humidity, temperature, and wind speed) col-lected on the same sampling date, a univariate correlation analysis was performed. The results of correlation coefficients (R) were gi-ven inTable 2for the five gaseous species (HF, H2SO4, HCl, HNO3,
and NH3) and the five corresponding ionic species (F, SO24 , Cl ,
NO
3, NHþ4). It can be seen that HCl were highly positively
associ-ated (R > 0.6) with relative humidity (RH) at a significance level (
a
) of 0.05 or 0.01 for seven of the 10 sampling stations. Interest-ingly, the three excluded stations were located at the foothill of Da-Tu Mountain in the northeast of the study site, a stagnant area favorable for the development of a local hotspot. HF also demon-strated the similar correlation but only for three sampling stations. It was noted that Station 3, inside the factory complex, showed a very strong correlation of 0.8 with relative humidity for both of HF and HCl.On the other hand, the concentration of Fis positively (R = 0.6)
associated with the wind speed at Stations 1, 3, 4, 5 and 10. It
Table 2
Results of correlation coefficients between pollutants concentrations and each of the three meteorological data measured in the corresponding month for each sampling site and each chemical species.
No. HF H2SO4 HCl HNO3 NH3
RHa
Tempa
W.Sa
RH Temp W.S RH Temp W.S RH Temp W.S RH Temp W.S
S1 0.505 0.489 0.458 0.428 0.366 0.074 0.637* 0.039 0.262 0.494 0.447 0.755** 0.096 0.368 0.237 S2 0.341 0.227 0.566 0.316 0.271 0.407 0.704* 0.125 0.252 0.139 0.466 0.52 0.187 0.004 0.143 S3 0.837** 0.145 0.451 0.234 0.219 0.572 0.792** 0.099 0.207 0.197 0.476 0.726** 0.128 0.226 0.362 S4 0.446 0.083 0.544 0.401 0.344 0.49 0.763** 0.117 0.326 0.187 0.594* 0.642* 0.164 0.073 0.275 S5 0.501 0.131 0.465 0.032 0.192 0.318 0.700* 0.219 0.101 0.042 0.652* 0.588* 0.152 0.178 0.243 S6 0.483 0.132 0.244 0.186 0.275 0.387 0.082 0.021 0.12 0.015 0.687* 0.601* 0.35 0.376 0.056 S7 0.752** 0.508 0.33 0.531 0.187 0.309 0.528 0.456 0.267 0.492 0.334 0.328 0.043 0.180 0.416 S8 0.214 0.403 0.359 0.389 0.065 0.274 0.39 0.327 0.253 0.103 0.387 0.414 0.197 0.046 0.197 S9 0.649 0.122 0.114 0.53 0.323 0.662 0.858** 0.049 0.384 0.475 0.519 0.818** 0.163 0.198 0.205 S10 0.671* 0.209 0.311 0.105 0.409 0.596 0.751* 0.020 0.598 0.009 0.801** 0.651 0.295 0.126 0.394 No. F SO2 4 Cl NO 3 NHþ4
RH Temp W.S RH Temp W.S RH Temp W.S RH Temp W.S RH Temp W.S
S1 0.132 0.402 0.813** 0.261 0.550 0.271 0.141 0.301 0.641* 0.131 0.011 0.061 0.091 0.233 0.161 S2 0.051 0.272 0.391 0.022 0.371 0.101 0.161 0.801* 0.122 0.242 0.051 0.084 0.124 0.213 0.061 S3 0.334 0.075 0.912** 0.432 0.481 0.222 0.252 0.832* 0.031 0.174 0.032 0.121 0.531 0.321 0.112 S4 0.035 0.302 0.771** 0.273 0.591* 0.312 0.382 -0.511 0.061 0.124 0.011 0.171 0.302 0.282 0.252 S5 0.262 0.384 0.801** 0.422 0.431 0.112 0.201 0.951* 0.032 0.141 0.204 0.022 0.491 0.271 0.151 S6 0.071 0.412 0.572 0.392 0.482 0.384 0.282 0.442 0.252 0.344 0.142 0.213 0.574 0.171 0.252 S7 0.471 0.535 0.343 0.291 0.432 0.341 0.343 0.612* 0.071 0.222 0.024 0.371 0.282 0.332 0.404 S8 0.425 0.012 0.334 0.291 0.442 0.323 0.131 0.571 0.092 0.113 0.023 0.012 0.184 0.301 0.144 S9 0.224 0.531 0.444 0.491 0.551 0.341 0.711* 0.791* 0.081 0.144 0.151 0.204 0.513 0.432 0.243 S10 0.381 0.273 0.811** 0.401 0.482 0.223 0.111 0.731* 0.502 0.561 0.282 0.161 0.604 0.501 0.131 a
RH denotes relative humidity, Temp denotes temperature, W.S denotes wind speed.
*
significance levela= 0.05.
** significance levela= 0.01.
should be noted that Stations 3 and 4 are situated at the upwind to
the prevailing north wind in winter. For this reason, Fcould be
very likely associated with internal sources as Ta-Du Mountain might block external sources in winter. However Clis little
asso-ciated with the wind velocity and negatively assoasso-ciated (R = 0.6) with the ambient temperature for most stations. The concentration of HNO3was positively correlated (R = 0.6) with temperature while
negatively correlated (R = 0.6) with the wind speed, both at a significance level of 0.05 or 0.01. Further analysis shows that
HNO3 levels increased when the wind velocity was less than
2 m s1and/or the temperature is higher than 28 °C.
To further investigate the chemical speciation as often occurred in the ambient environment of a urban city, a correlation was also
performed among different chemical species as shown inTable 3.
Ammonium (NHþ
4) was highly positively associated (R = 0.7) with
the two oxidized form of acidic anions (SO2
4 and NO
3) at a
signif-icance level of 0.01. In a theoretical simulation study, Orel and
Seinfeld (1997) pointed that the formation H2SO4and HNO3from
SO3and NO2, respectively, was involved in two subsequent steps,
formation of acids in gaseous phase followed by heterogeneous condensation with water vapor on the existing particles as a sur-face site for catalysis (Danalatos and Glavas, 1999; Baek et al., 2004). They also concluded that NHþ
4 plays an important role in
the formation of SO2
4 and NO3 from their gaseous forms. This
demonstrates that particle concentration and species competition may play an important part to contribute the background concen-tration of the elated species. As many factors may affect the con-centration profiles monitored in this study, the results of the statistical analysis poses a challenge for developing a more ad-vanced data-mining technique.
4. Conclusions
To explore the background features of acidic and basic air pol-lutants around a new high-tech industrial park, this paper pre-sents a novel analytical algorithm of data mining followed by meteorological and interspecies correlation. By clustering air-monitoring data and applying the segregation rule to the corre-sponding meteorological data, inherent data complexity and dis-similarity can be mitigated and key influencing factors can be identified. The proposed algorithm is also valuable in dealing with the problem of source identification for a complicate air dis-persion system like the Taichung site in this study. This study
concludes that, at the presence of NHþ
4, several background
peaks, such as NO
3, SO 2
4 , HNO3, H2SO4 and HN3, occurred in
summer when the prevailing wind velocity is slow (<2 m s1),
are likely associated with the external sources from Taichung ur-ban city and Taichung Industrial Park. The results of this study
will serve as a basis for the air quality planning and management for the Taichung site.
Acknowledgement
The authors would like to thank the National Science Council of the Republic of China for financially supporting this study under Contract No. NSC95-2221-E-324-011.
References
Baek, B.H., Aneja, V.P., Tong, Q., 2004. Chemical coupling between ammonia, acid gases, and fine particles. Environ. Pollut. 129, 89–98.
Banerjee, D., Chakraborty, S.K., Bhattacharyya, S., Gangopadhyay, A., 2009. Appraisal and mapping the spatial-temporal distribution of urban road traffic noise. Int. J. Environ. Sci. Technol. 6 (2), 325–335.
Bari, A., Ferraro, V., Wilson, L.R., Luttinger, D., Husain, L., 2003. Measurements of gaseous HONO, HNO3, SO2, HCl, NH3, particulate sulfate and PM2.5in New York,
NY. Atmos. Environ. 37, 2825–2835.
Chittaro, P.M., Usseglio, P., Fryer, B.J., Sale, P.F., 2006. Spatial variation in otolith chemistry of Lutjanus apodus at Turneffe Atoll, Belize. Estuar. Coast. Shelf Sci. 67, 673–680.
Danalatos, D., Glavas, S., 1999. Gas phase nitric acid, ammonia and related particulate matter at Miditerranean costal site, Patras, Greece. Atmos. Environ. 33, 3417–3425.
Gargava, P., Aggarwal, A.L., 1996. Industrial emissions in a coastal region of India: prediction of impact on air environment. Environ. Int. 22, 361–367. Hosseinlou, H.M., Sohrabi, M., 2009. Predicting and identifying traffic hot spots
applying neuro-fuzzy systems in intercity roads. Int. J. Environ. Sci. Technol. 6 (2), 309–314.
Horng, C.L., Cheng, M.T., 2008. Distribution of PM2.5, acidic and basic gases near
highway in central Taiwan. Atmos. Res. 88, 1–12.
Jäggi, M., Ammann, C., Neftel, A., Fuhrer, J., 2006. Environmental control of profiles of ozone concentration in a grassland canopy. Atmos. Environ. 40 (28), 5496– 5507.
Lee, J.Y., Chen, J.Y., Lee, K.K., Lee, S.Y., Lee, M.H., 2001. Statistical evaluation of geochemical parameter distribution in a ground water system contaminated with petroleum hydrocarbons. J. Environ. Qual. 30, 1548–1563.
Neves, N.M.S., 1996. Air monitoring at Camacari petrochemical complex. Water Sci. Technol. 33 (3), 9–16.
Orel, A.E., Seinfeld, J.H., 1997. Nitrate formation in atmospheric aerosols. Environ. Sci. Technol. 11 (10), 1000–1007.
Pejman, A.H., Nabi Bidhendi, G.R., Karbassi, A.R., Mehrdadi, N., Esmaeili Bidhendi, M., 2009. Evaluation of spatial and seasonal variations in surface water quality using multivariate statistical techniques. Int. J. Environ. Sci. Technol. 6 (3), 467– 476.
Reghunath, R., Murthy, T.R.S., Raghavan, B.R., 2002. The utility of multivariate statistical techniques in hydrogeochemical studies: an example from Karnataka, India. Water Res. 36 (10), 2437–2442.
Spengler, J.D., Brauer, M., Koutrakis, P., 1990. Acid air and health. Environ. Sci. Technol. 24, 946–956.
EPA Taiwan, 2007. Emission Standard for Stationary Airborne Pollutants Sources. Taipei, Taiwan.
Tsai, C.J., Aggarwal, S.G., Chang, C.T., Hung, I.F., 2003. Concentration profiles of acidic and basic air pollutants around industrial park of Taiwan. Water Air Soil Pollut. 151, 287–304.
Tsai, C.J., Chang, C.T., Liu, T.W., Huang, C.C., Chien, C.L., Chein, H.M., 2004. Emission characteristics and control efficiency of acidic and basic gases and aerosols from packed towers. Atmos. Environ. 38 (4), 643–646.
Table 3
Results of linear correlation between two mean concentrations of all 10 sampling stations for any two species during the entire sampling period of 12 months.
HF H2SO4 HCl HNO3 NH3 F SO2 4 Cl NO 3 NHþ4 HF 1 H2SO4 0.356** 1 HCl 0.385** 0.416** 1 HNO3 0.387** 0.443** 0.391** 1 NH3 0.112 0.187* 0.172 0.107 1 F 0.445** 0.040 0.193* 0.135 0.074 1 SO2 4 0.284 ** 0.195* 0.090 0.301** 0.013 0.248** 1 Cl 0.301** 0.059 0.260** 0.168 0.057 0.309** 0.318** 1 NO 3 0.089 0.230* 0.057 0.154 0.356** 0.037 0.438** 0.009 1 NHþ 4 0.140 0.403 ** 0.052 0.379** 0.324** 0.088 0.679** 0.218* 0.716** 1 *Significance levela= 0.05. **Significance levela= 0.01.
Walker, J.T., Whitall, D.R., Robarge, W., Paerl, H.W., 2004. Ambient ammonia and ammonium aerosol across a region of variable ammonia emission density. Atmos. Environ. 38, 1235–1246.
Webb, M.P., Vincent, C.E., 1999. Comparison of time-averaged acoustic backscatter concentration profile measurements with existing predictive models. Mar. Geol. 162 (1), 71–90.
Zhu, X., Zhihua, F., Wu, X., Meng, Q., Wang, S.W., Tang, X., Ohman-Strickland, P., Georgopoulos, P., Zhang, J., Bonanno, L., Held, J., Lioy, P., 2008. Spatial variation of volatile organic compounds in a ‘‘Hot Spot” for air pollution. Atmos. Environ. 42 (32), 7329–7338.
![TraditionalMLCalgorithmsmainlytacklethebatchMLCproblem,wheretheinputdataarepresentedinabatch[24,28].Nevertheless,inmanyMLCapplicationssuchase-mailcategorization[22],multi-labelexamplesarriveasastream.Onlineanalysisistherefore dimensionreducermotivatedbyma](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACH5BAEAAAAALAAAAAABAAEAAAICRAEAOw==)