Upload
others
View
11
Download
0
Embed Size (px)
Citation preview
Queensland University of Technology
Science and Engineering Faculty
School of Chemistry, Physics and Mechanical Engineering
Characteristics of spatial variation, size distribution, formation and growth of particles in urban environments
Farhad Salimi (n7411545)
A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS
OF THE DEGREE OF DOCTOR OF PHILOSOPHY
2014
I
Abstract
Ultrafine particles (UFP; particles smaller than 100 nm) have attracted attention as
there have been many toxicological studies showing their adverse health effects. UFPs also
affect our climate significantly, as they can act as cloud condensation nuclei (CCN).
Therefore, studying the physical and chemical characteristics of UFPs will enhance our
knowledge of their effects on climate and human health.
The overall aim of this PhD project was to further investigate the physical/chemical
dynamics and transformations of UFPs. This study was conducted as part of a major study
called ‘Ultrafine Particles from Traffic Emissions and Children’s Health’ (UPTECH), which
sought to determine the relationship between children’s exposure to UFPs in schools and
their health. This project included comprehensive measurements of air quality parameters
at each of 25 primary schools within the Brisbane metropolitan area. While measurements
were conducted in schools, the results of this thesis are not limited to school environment
and have urban implications. Indoor measurements have been conducted as a part of the
UPTECH project as well, however, they are not the focus of this thesis and will not be
presented here.
This PhD project investigated three important parameters which relate to an
aerosols impact on climate and human health, namely spatial variation, particle size
distribution and new particle formation. The first part of the project included an
investigation of the spatial variation of particle number concentration (PNC) across
microscale environments, such as schools, and the effect of this spatial variation on
exposure assessment. Three outdoor sites were used to measure PNC at each school, which
provided sufficient data to pursue this aim. Coefficients of variation (CV) were calculated to
quantify the spatial variation and these were adjusted to account for an instrument
uncertainty of CV = 0.3. The study found a positive relationship between traffic density and
spatial variation, and wind speed had negative correlation with spatial variation. Spatial
variation increased while the wind was blowing from the road towards school. A generalised
additive model (GAM) was used to establish the relationship between spatial variation and
its affecting parameters (wind speed and traffic density). Wind speed had a decreasing
effect on spatial variation as it increased beyond 1.5 m/s and traffic density had an
II
increasing effect on spatial variation until it reached 70 vehicles/5min. The relationship
between the spatial variation and traffic density was not strong, suggesting the presence of
other sources and/or factors that could influence this variation. Therefore, it was not
possible to draw quantitative conclusions about the conditions under which the microscale
environment was expected to display spatial variation. The findings from the part of the
study, and the importance of particle size in determining particle transformation, dynamics
and the identification of particle sources/origins, led us to the next step of the study, which
investigated particle size distribution.
In this part of the thesis, particle size data were analysed using sophisticated
mathematical and statistical techniques, with the aim of identifying the sources/origins of
urban UFPs. Particle number size distribution (PNSD) within the size range 9-414 nm was
measured at each school using a Scanning Mobility Particle Sizer (SMPS). Around 84,000
SMPS scans were collected during the whole project, which made the processing and
analysis of the data a large and complex task and therefore, a clustering technique was used
to analyse the data. Four clustering techniques were compared and the K-mean technique
was found to perform best on the PNSD data. The Dunn index and silhouette width
determined five clusters to the optimum number and therefore, the K-means clustering
technique was applied to the whole PNSD dataset, in order to find five clusters.
A new parameterisation method, based on GAM, was also developed to characterise
the PNSD data. The new method was able to parameterise the PNSD data, with advantages
over the traditional methods. Afterwards, each cluster was attributed to its origin/source
using the diurnal variation of clusters, PNSD parameterisation results and several other air
quality parameters. The five clusters were found to be attributed to three main
sources/origins: 1) Photochemically induced nucleated particles; 2) Urban background; and
3) Vehicle generated particles. It was found that a significant number of particles were due
to new particle formation (NPF) events, which was much higher than in previous studies
which employed the traditional method of identifying NPF events in the same environment.
In this study, NPF was found to be an important contributor to total PNC that can potentially
have effects on climate and human health.
III
Based on the importance of the NPF phenomenon and the significant gaps in
knowledge which exist in this area, the next part of this study aimed to determine the
characteristics of NPF events, particularly in relation to their chemistry. A Compact Time of
Flight Aerosol Mass Spectrometer measured particle composition in five schools. Firstly, the
PNSD measured at these schools were analysed and five NPF events were detected. Growth
rates (GR) of the identified new particle formation events ranged between 3.3-4.6 nm/hr.
These events happened on days with high temperature, together with high solar radiation
intensity and low humidity. Mass composition of the chemical species (i.e. nitrate,
ammonium, sulphate and organics) increased rapidly after the start of nucleation, and the
size distribution of organics, nitrate and sulphate followed the same trend as PNSD during
the nucleation events. The mass fraction percentage of each chemical species was
calculated and its trend was determined using the GAM model. The mass fraction of
organics increased after the start of NPF event, while the mass fraction of other species
decreased, indicating the important role of organics in the growth of newly formed
particles. f44 and f43 were calculated in order to further study the role of organics. f44
decreased when the particles were mainly vehicle generated, while NPF generated particles
had a lower f43 and f44. In other words, NPF and vehicle generated particles clustered in
different locations in the f44 vs f43 plot, indicating the importance of this plot for source
identification and particle transformation characteristics.
In summary, this thesis has provided insight into the transformation and dynamics of
urban aerosols by undertaking an advanced analysis of comprehensive data measured
within an urban environment. The results of this thesis have considerably broadened
scientific knowledge in this field and can help researchers to further understand aerosol
dynamics and transformation.
IV
Keywords
Aerosols, ultrafine particles, urban, particle number concentration, spatial variation, school,
microscale, GAM, exposure assessment, particle number size distribution, SMPS, clustering,
source apportionment, new particle formation, AMS, growth rate.
Statement of Original Authorship
The work contained in this thesis has not been previously submitted to meet
requirements for an award at this or any other higher education institution. To the best of
my knowledge and belief, the thesis contains no material previously published or written by
another person except where due reference is made.
Signature:
Date: March 2014
V
QUT Verified Signature
VI
Table of Contents Abstract .................................................................................................................................................... I
Keywords ................................................................................................................................................ IV
Statement of Original Authorship ........................................................................................................... V
List of Publications .................................................................................................................................. X
List of Tables .......................................................................................................................................... XI
List of Figures ......................................................................................................................................... XI
Chapter 1:........................................................................................................................................... XI
Chapter 2:........................................................................................................................................... XI
Chapter 3:........................................................................................................................................... XI
Chapter 3S: ........................................................................................................................................ XII
Chapter 4:.......................................................................................................................................... XII
Chapter 5:.......................................................................................................................................... XII
Acknowledgments ................................................................................................................................ XIV
Acronyms .............................................................................................................................................. XV
Chapter 1: Introduction .......................................................................................................................... 1
1.1. Description of Scientific Problems Investigated ..................................................................... 1
1.2. Overall aims of the study ........................................................................................................ 2
1.3. Specific objectives of the study............................................................................................... 3
1.4. Account of Scientific Progress Linking the Scientific Papers ................................................... 4
Chapter 2: Literature Review .................................................................................................................. 7
2.1. Introduction ............................................................................................................................ 7
2.2. Particle size ............................................................................................................................. 9
2.3. Particle concentration ........................................................................................................... 10
2.4. Particle Chemical Composition ............................................................................................. 12
2.4.1. Offline methods ............................................................................................................ 13
2.4.2. Online methods ............................................................................................................. 14
2.5. Spatial variation of particle concentration ........................................................................... 15
2.5.1. Spatial variation of particle mass concentration .......................................................... 15
2.5.2. Spatial variation of particle number concentration ...................................................... 16
2.5.3. Effects of spatial variation on exposure estimation in microscale environments such as
schools 17
2.5.4. Quantification of the spatial variation .......................................................................... 18
2.6. Particle number size distribution .......................................................................................... 19
VII
2.6.1. PNSD modes .................................................................................................................. 21
2.6.2. Cluster analysis of PNSD data ....................................................................................... 23
2.7. New Particle Formation ........................................................................................................ 24
2.7.1. Mechanisms .................................................................................................................. 25
2.7.2. NPF detection ................................................................................................................ 27
2.7.3. Formation, growth rate and condensation sink ........................................................... 29
2.7.4. Affecting parameters .................................................................................................... 31
2.7.5. Chemistry of NPF ........................................................................................................... 32
2.8. Gaps in knowledge ................................................................................................................ 33
2.9. References ............................................................................................................................ 33
Chapter 3: Spatial variation of Particle Number Concentration in School Microscale Environments
and its Impact on Exposure Assessment ............................................................................................... 42
Statement of Joint Authorship .......................................................................................................... 43
Abstract ............................................................................................................................................. 44
TOC Art .............................................................................................................................................. 45
Keywords ........................................................................................................................................... 45
3.1. Introduction .......................................................................................................................... 45
3.2. Material and methods .......................................................................................................... 47
3.2.1. Background ................................................................................................................... 47
3.2.2. Instrumentation ............................................................................................................ 47
3.2.3. Data analysis ................................................................................................................. 48
3.3. Results and discussion .......................................................................................................... 50
3.3.1. Data availability ............................................................................................................. 50
3.3.2. Traffic density ................................................................................................................ 50
3.3.3. Particle number concentration ..................................................................................... 50
3.3.4. Spatial variation ............................................................................................................ 52
Coefficient of variation: ................................................................................................................ 52
Number of monitoring sites required: .......................................................................................... 52
Correlation between CVc and traffic density, wind speed/ direction, inter-site distances: ......... 53
Non-parametric regression model: ............................................................................................... 54
3.4. Acknowledgments ................................................................................................................. 55
3.5. Supporting information ........................................................................................................ 56
3.6. References ............................................................................................................................ 56
3.7. Figures: .................................................................................................................................. 57
VIII
Chapter 3S: Supporting Information ..................................................................................................... 65
3S.1. Measurements .......................................................................................................................... 66
3S.2. Quality assurance and data processing .................................................................................... 66
3S.2.1. Particle number concentration: ........................................................................................ 66
3S.2.2. Meteorological parameters: ............................................................................................. 68
3S.3. References: ........................................................................................................................... 78
Chapter 4: Assessment and Application of Clustering Techniques to Atmospheric Particle Number
Size Distribution for the Purpose of Source Apportionment ................................................................ 79
Statement of Joint Authorship .......................................................................................................... 80
Abstract ............................................................................................................................................. 81
TOC Art .............................................................................................................................................. 82
Keywords ........................................................................................................................................... 82
4.1. Introduction .......................................................................................................................... 82
4.2. Materials and Methods ......................................................................................................... 85
Background ................................................................................................................................... 85
Instrumentation, quality assurance, and data processing ............................................................ 85
Clustering particle number size distribution data......................................................................... 86
Non-parametric estimation of particle number size distribution and temporal trends ............... 87
4.3. Results and discussion .......................................................................................................... 88
Particle number size distribution .................................................................................................. 88
Optimum clustering technique and number of clusters ............................................................... 88
Clustering PNSD data .................................................................................................................... 89
Temporal trend of clusters............................................................................................................ 92
4.4. Acknowledgment .................................................................................................................. 93
4.5. References ............................................................................................................................ 94
4.6. Figures: .................................................................................................................................. 94
Chapter 5: Investigation into Chemistry of New Particle Formation and Growth in Subtropical Urban
Environment ....................................................................................................................................... 101
Statement of Joint Authorship ........................................................................................................ 102
5.1. Abstract ............................................................................................................................... 103
5.2. Introduction ........................................................................................................................ 103
5.3. Materials and Methods ....................................................................................................... 105
5.3.1. Background ..................................................................................................................... 105
5.3.2. Instrumentation .............................................................................................................. 105
IX
5.3.3. Data analysis ................................................................................................................... 106
5.4. Results and Discussion ........................................................................................................ 108
5.4.1. Identified New Particle Formations ............................................................................ 108
5.4.2. Evolution of Chemical Compositions .......................................................................... 110
5.4.3. Role of organics ........................................................................................................... 114
5.5. Acknowledgments ............................................................................................................... 119
5.6. References .......................................................................................................................... 120
Chapter 6: Conclusions ....................................................................................................................... 123
6.1. Principal Significance of the Findings .................................................................................. 123
6.2. Directions for future research............................................................................................. 130
X
List of Publications
1) Farhad Salimi, Mandana Mazaheri, Sam Clifford, Leigh R. Crilley, Rusdin
Laiman, and Lidia Morawska. “Spatial Variation of Particle Number
Concentration in School Microscale Environments and Its Impact on
Exposure Assessment”, Environmental Science & Technology 2013 47
(10), 5251-5258.
2) Farhad Salimi, Zoran Ristovski, Mandana Mazaheri, Rusdin Laiman, Leigh
R. Crilley, Sam Clifford and Lidia Morawska. “Assessment and Application
of Clustering Techniques to Atmospheric Particle Number Size
Distribution for the Purpose of Source Apportionment”, under review in
Environmental Science & Technology.
3) Farhad Salimi, Leigh R. Crilley, Svetlana Stevanovic, Zoran Ristovski,
Mandana Mazaheri, and Lidia Morawska. “Investigation into Chemistry of
New Particle Formation and Growth in Subtropical Urban Environment”,
submitted to Environmental Science and Technology.
XI
List of Tables
Table3S. 1 Characteristics of the monitored schools. ........................................................................... 73
Table3S. 2: Schematic diagrams and location of monitoring sites at each school. The large dashed
line surrounds the school area and the solid line across the road shows the location of automatic
traffic counters. ..................................................................................................................................... 75
List of Figures
Chapter 1: Figure 1. 1: Visualisation of the issues investigated in this thesis. ......................................................... 3
Chapter 2: Figure 2.1: Main sources and impacts of atmospheric aerosol [18]. ...................................................... 9
Figure 2.2: Schematic diagram of TSI 3781 water-based CPC [54]. ...................................................... 12
Figure 2.3: Schematic of an Aerodyne AMS [80]. ................................................................................. 15
Figure 2.4: Global map of observation of PNSD. The dots indicate observation sites, dashed and
rectangles indicate regions with airborne and ship observations respectively [43]. ... Error! Bookmark
not defined.
Figure 2. 5: Schematic diagram of a TSI EC [108]. ................................................................................. 20
Figure 2.6: An example of an SMPS scan. ............................................................................................. 21
Figure 2.7: Idealised schematic of an atmospheric particle number size distribution [109]. .............. 22
Figure 2.8: Illustration of processes included in NPF [124]. ................................................................. 25
Figure 2.9: Schematic diagram of main size regimes of atmospheric clusters and their relation to NPF
[137]. ..................................................................................................................................................... 27
Figure 2.10: Evolution of PNSD in a day with NPF event [2, 138]. ........................................................ 28
Figure 2.11: Flow chart of the procedure for NPF events detection [140]. .......................................... 28
Figure 2.12: Growth rate of particles during NPF events [145]. ........................................................... 31
Chapter 3: Figure 3. 1: Traffic density box plot for each school (a), and diurnal traffic pattern during weekdays
(b) and weekends (c). Solid line represents the mean value and dashed lines show the 95%
confidence interval. .............................................................................................................................. 57
Figure 3. 2: Box plot of inter-site average PNC at each school (a), together with its weekday (b) and
weekend (c) diurnal pattern. Solid line represents the mean value and dashed lines show the 95%
confidence interval. .............................................................................................................................. 58
Figure 3. 3: Corrected CV box plot for each school (a) during school hours (shaded boxes) and 24hrs
(white boxes). Weekday (b) and weekend (c) diurnal variations of corrected CV. Solid line represents
the mean value and dashed lines show the 95% confidence interval. ................................................. 59
Figure 3. 4: Corrected CV between each pair of sites during school hours (9am-3pm). ..................... 59
XII
Figure 3. 5: Inter-site average concentration ratio at schools which displayed spatial variability of
PNC. The left, middle and right circles correspond to sites A, B, and C, respectively and the three sites
for each school are connected via a dashed line. ................................................................................. 60
Figure 3. 6: Effects of (a) traffic density for the regression model (2), with their 95% CIs (dashed
lines). The histogram below (b) shows the frequency of the values of traffic density, which explains
the width of the 95% CI for the fitted smooth function in (a). ............................................................. 61
Figure 3. 7: Partial effects of (a) wind speed and (b) traffic density for the regression model (3) with
their 95% CIs (dashed lines). The histograms below show the frequency of the values of (c) wind
speed and (d) traffic density, which explain the width of the 95% CI for the fitted smooth functions in
(a) and (b). ............................................................................................................................................. 62
Chapter 3S: Figure 3S. 1: Schematic diagram of measurement sites and traffic counter location. Arrow shows the
direction of the prevailing wind direction (based on the previous years’ statistics). ........................... 66
Figure 3S. 2: Comparison between wind data measured at S01 and the nearest monitoring station to
the school. ............................................................................................................................................. 69
Figure 3S. 3: PNC measured by three CPCs placed side by side and measuring the same aerosol (a),
together with a box plot of the corresponding calculated CV (b), and regression plot of CV vs. mean
PNC for the side by side measurements. .............................................................................................. 69
Figure 3S. 4: CV variation by time at S04. ............................................................................................. 70
Figure 3S. 5: CV by wind direction at S14. ............................................................................................ 70
Figure 3S. 6: Polarplot of CV at S20. ..................................................................................................... 71
Figure 3S. 7: CV vs. traffic for different wind directions. ...................................................................... 71
Figure 3S. 8: Available data (days) at each school. ............................................................................... 72
Figure 3S. 9: Spearman’s correlation coefficient between CV and traffic/wind direction. .................. 73
Chapter 4: Figure 4. 1: An example of a fitted smooth function (solid line) on PNSD data (circles) and the
identified local peaks (solid circles)...................................................................................................... 94
Figure 4. 2: Dunn index and Silhouette width for Kmeans, SOM, PAM, and CLARA clustering
techniques for different cluster numbers. .............................................................................................. 95
Figure 4. 3: Clustering results using SMPS data. Graphs on the left show the particle size distributions
with 95% confidence intervals, the associated graphs on the middle show the diurnal cycle of the
hourly percentage of occurrence for each cluster, and the graphs on right show the associated solar
radiation, PM2.5,, PM10, NOx, CO, SO2, total PNC. ............................................................................... 96
Figure 4. 4: Density of peaks in particle number size data at each cluster. ........................................... 97
Figure 4. 5: Temporal trend of occurrence of each cluster, with their 95% confidence intervals. ........ 98
Chapter 5: Figure 5. 1: Average diurnal pattern of solar radiation (SR), humidity, temperature and condensation
sink (CS) on nucleation and non-nucleation day at S12 and S25. Vertical dotted lines show the start
of nucleation. ...................................................................................................................................... 109
Figure 5.2: Average diurnal pattern of organics, nitrate, sulphate and ammonium concentrations on
nucleation and non-nucleation days at S12 and S25. ......................................................................... 110
Figure 5.3: Time series of the mass concentration of particle species and particle number size
distribution at S12. .............................................................................................................................. 112
XIII
Figure 5.4: Time series of the mass concentration of particle species and particle number size
distribution at S25. .............................................................................................................................. 113
Figure 5.5: Mass fraction of each chemical species (sulphate, nitrate, ammonium and organics)
during the nucleation events at s12 and s25. Vertical dashed lines represent the point that the newly
formed particles reached the AMS’s lowest detection limit. ............................................................. 114
Figure 5. 6: f44vs f43 at each hour of the day for all data measured at S12 during the non-nucleation
days. The triangle from Ng et al. [41] is drawn as a visual aid. ........................................................... 117
Figure 5. 7: f44 vs f43 at each hour of the day for all data measured at S12 during the nucleation days.
The triangle from Ng et al. [41] is drawn as a visual aid. .................................................................... 118
Figure 5.8: The pattern observed in f44 vs f43 plot after the aerosol particles were dominated by either
traffic or NPF events. .......................................................................................................................... 119
Figure 5. 9: Diurnal variation of the f57 at S12 during the nucleation and non-nucleation days. ....... 119
XIV
Acknowledgments
I would not have been able to complete the PhD journey without the help of so
many people, in so many ways, over the past three years. I am grateful to my principal
supervisor, Professor Lidia Morawska, for her guidance and support right from the early
stages of my PhD until its completion. I am also thankful to my associate supervisors,
Professor Zoran Ristovski and Dr Mandana Mazaheri, for their invaluable assistance.
I would like to thank all of my colleagues and friends at the ILAQH, particularly my
fellow PhD students, Rusdin Laiman, Megat Mokhtar and Leigh Crilley, because without
them, the long field measurement campaigns would have been impossible. I also thank Dr
Sam Clifford for his immense help in performing the data analysis and programming
required for this project, and I am thankful to Dr Svetlana Stevanovic for her help with
conducting the analysis for, and the writing of my third paper.
I would also like to thank all of my school teachers, university lecturers and tutors
who introduced me to different aspects of science. Finally, I would like to thank my parents,
brother and sister for instilling in me the confidence and courage to pass the challenges I
have faced, and am still yet to face in life.
XV
Acronyms
BOM — Bureau of Meteorology
CCN — Cloud Condensation Nuclei
CLARA — Clustering of Large Applications
CPC — Condensation particle Counter
DERM — Department of Environment and Resource Management
DMA — Differential Mobility Analyser
GAM — Generalised Additive Model
HEPA — High Efficiency Particulate Air
NAIS — Nano Air Ion Spectrometer
NPF — New Particle Formation
PAM —Partitioning Around Medoids
PCA — Principal Component Analysis
PMF — Positive Matrix Factorisation
PNC — Particle Number Concentration
PNSD — Particle Number Size Distribution
PSL — Polystyrene latex
SMPS — Scanning Mobility Particle Sizer
SOM — Self Organising Map
TOF — Time Of Flight
UFP — Ultrafine Particle
XVI
UPTECH — Ultrafine Particles from Traffic Emissions and Children’s Health
1
Chapter 1: Introduction
1.1. Description of Scientific Problems Investigated
The effects of ultrafine particles (UFPs; particles with diameter <100 nm) on human
health have attracted much attention recently, as there have been a significant number of
toxicological studies showing their adverse health effects. Since UFPs can penetrate deeper
into the lung, they can be potentially more harmful compared to larger particles. Therefore,
there is a definite need to investigate the physical and chemical characteristics of UFPs.
Particle number concentration (PNC) can be used as an indicator of the concentration of
UFPs, since the contribution of UFPs to PNC is the highest compared to larger particles.
Spatial variation of PNC is of importance as it can potentially affect human exposure.
PNCs normally display higher spatial variation compared to particle mass concentration,
however its spatial variation has been studied less. All of the studies related to the spatial
variation of PNC have focused on macroscale environments with site distances in the order
of kms, however the spatial variation of PNC in microscale environments (e.g. factories,
schools etc.) has yet to be studied. It is important to know the level and characteristics of
spatial variation in microscale environments, in order to accurately assess human exposure
to UFPs within that microenvironment. Therefore, there is a need to investigate the spatial
variation of PNC in microscale environments.
Particle size is one of the most important parameter characterising particles and
their health effects, and it is possible to measure the number size distribution of particles
using instruments such as the Scanning Mobility Particle Sizer (SMPS). The processing and
analysis of measured particle number size distribution (PNSD) data is often a challenging
task, as very large datasets are generally extracted from the long term measurement of
PNSD. Therefore, there is a need to develop a sophisticated approach for the processing,
analysis and interpretation of PNSD data. Clustering techniques have recently been
employed in PNSD analysis and interpretation, as they are able to categorise the whole
measured dataset into similar groups (clusters). However, there are several clustering
techniques available whose performances have not been assessed on PNSD data so far. In
addition, parameterising each particle size spectra needs further investigation, as the
2
current method (fitting multilognormal functions) has drawbacks, such as introducing a
predefined shape for the spectra. Therefore, there is a need to investigate potential
clustering techniques and assess their functionality for source identification, in addition to
developing more sophisticated methods for PNSD parameterisation.
Atmospheric new particle formation (NPF) is an important source of UFPs which can
potentially affect our climate. Several studies have investigated NPF events in different
types of environments and most of them focused on the physical characteristics of NPF
events. Studies have shown differences in NPF events in urban compared to remote areas,
however there are less studies focusing on NPF events in urban areas and even fewer which
consider the chemistry of NPF events. Chemical composition plays an important role in
determination of the adverse health effects posed by exposure to aerosols. No study has
ever investigated the chemistry of NPF events in an urban subtropical environment, which
makes it really important to conduct research in this area.
In summary, this thesis reports on the investigation of several characteristics of UFPs
in urban environments, particularly those which determine their transformation, dynamics
and effect on climate and humans health. This is not only important but also novel work, as
similar studies have never been conducted in some environments and there are a very
limited number available in total.
1.2. Overall aims of the study
Taking into account the research problems defined in the section above, the broad
aim of this research program was to give further insight into physical/chemical dynamics
and transformation of UFPs. The pursuit of this aim also identified other more specific aims,
which can be summarised as follows (Figure 1. 1):
1) To investigate the spatial variation of particle number concentration in
urban microscale environments and its impact on human exposure
assessment.
2) To identify and apply the best analysis techniques and methods on PNSD
data for source apportionment.
3) To conduct source identification of urban aerosols.
3
4) To investigate the chemistry of NPF events in urban environments.
Figure 1. 1: Visualisation of the issues investigated in this thesis.
1.3. Specific objectives of the study
In order to achieve the abovementioned aims of the study, the objectives were to:
Conduct measurements at each of 25 state primary schools across the
Brisbane metropolitan area, within the framework of UPTECH project.
Particularly, to measure the physical and chemical parameters (size
distribution, concentration and chemical composition) of ambient UFPs.
Identify a scientific method to quantify the spatial variation of PNC in a
microscale environment, such as a school.
Apply the method found above to PNCs measured at the 25 schools which
were investigated during the UPTECH project.
Determine an accurate scientific technique to quantify the level of spatial
variation related to instrumental uncertainty.
4
Apply the method found above to PNCs measured at each of the 25 schools.
Determine the relationship between the spatial variation of PNC and its
potential affecting parameters (e.g. traffic density and meteorological
conditions).
Determine the effect of spatial variation of PNC on variations in human
exposure.
Investigate the application of different clustering techniques on PNSD.
Apply the optimum technique to the PNSD data collected during the UPTECH
project.
Develop a new method to parameterise the particle size spectra.
Use the combination of the clustering technique and parameterisation
technique found above for source apportionment purposes.
Analyse PNSD to detect NPF events.
Investigate the chemistry of NPF events in urban subtropical environments
using the AMS measurements conducted during the UPTECH project.
Determine the impact of different chemical species on particle growth.
1.4. Account of Scientific Progress Linking the Scientific Papers
Three papers frame the body of this thesis (Figure 1. 2). In the first paper, spatial
variations of PNC in microscale environments were investigated using the data collected at
three monitoring sites at 25 schools. Coefficient of variation (CV) was used to quantify the
spatial variation. Variations related to the instruments measurement uncertainty were
calculated by employing the side by side test. A CV = 0.3 was found to be related to
measurement uncertainty and all of the calculated CVs were corrected taking this into
account. Spatial variation was observed in 12 out of 25 schools and due to their level of
spatial variation, three were found to require three monitoring sites to provide accurate
PNC measurements, while the remaining nine needed only two. Spatial variations of PNC
were calculated at all schools and their relationship with meteorological parameters and
traffic density was investigated. Higher spatial variations were observed when the wind was
blowing from the adjacent road toward the schools and less spatial variation was observed
at higher wind speeds. Ratios of mean concentration at each site to the site with lowest
mean concentration were also calculated, which varied from one to about six. These
5
findings show that exposure estimation depends on the monitoring site. The impact of
traffic density and wind speed on spatial variation were further investigated and wind speed
was found to have no effect below 1.5 m/s. Above this level, an increase in wind speed
correlated with a decrease in spatial variation. Traffic density was found to have a significant
effect, as spatial variation increased linearly with increases in traffic density, until it reached
70 (5min-1). This part of the study enhanced our understanding of the nature of the spatial
variation of PNC and the impact of its affecting parameters. However, due to the particles
coming from other sources than vehicles, it was not possible to draw quantitative
conclusions on the conditions under which spatial variation can be expected in a microscale
environment.
The results of the first paper indicated the potential role of sources, other than
vehicle emissions, contributing to the aerosols measured at the 25 sites. Therefore, the
main aim of the second paper was to analyse PNSD with the purpose of source
apportionment. More than 84,000 SMPS scans were collected, therefore requiring a
sophisticated method to analyse such a large data set. Several clustering techniques
recently employed for analysing PNSD in the literature (i.e. K-means, PAM, CLARA, SOM and
AP) were selected and their performances compared using the Dunn index and Silhouette
width. K-means performed the best on the PNSD data and five was found to be the
optimum number of clusters. Therefore, whole dataset were categorised into five clusters
using the K-means clustering technique. A Generalised Additive Model (GAM), with a basis
of penalised B-splines, was used to develop a new method to parameterise the PNSD data.
Each cluster was attributed to its source/origin based on the characteristics of each cluster,
together with the results of the GAM parameterisation. The five clusters were attributed to
three main sources/origins: 1) Photochemically induced nucleated particles; 2) Vehicle
generated particles; and 3) Regional backgrounds. It was found that the photochemically
induced nucleated particles contributed significantly to overall aerosol concentrations,
which were previously underestimated using the traditional methods of finding nucleated
particles performed in the same area.
The abovementioned results revealed the high contribution of newly formed
particles to overall aerosol concentrations in the studied area. Therefore, the third paper
was focused on new particle formation (NPF; also known as nucleation) events. There have
6
been several studies related to the physical aspects of NPF events, but no study has ever
focused on the chemistry of NPF on the studied area. The main aim of the third paper was
to investigate the chemical species that contributed to NPF events and particle growth,
together with the nature of their contribution. Aerosol Mass Spectrometers (AMS’s) were
used to measure the chemical composition of particles in 5 out of 25 schools and the results
of these measurements were used, together with the other measurements, in the analysis.
Five NPF events, with a growth rate range of 3.3-4.6 (nm/hr), were observed at two out of
the five sites. Nucleation events occurred on days with higher solar radiation intensity,
higher temperatures and less humidity. Condensation sinks (CS), as a measure of the surface
area of pre-existing particles, were calculated and it was found that, on days when NPF
events occurred, there was a smaller CS before the start of nucleation. Analysis of the AMS
data revealed the significant increase in mass concentration of sulphate, ammonium, nitrate
and organics as a consequence of nucleation events. Time series of the mass concentration
of particle species showed an increase in the concentration of small particles, followed by
the subsequent growth of particle diameter, except for ammonium, for which no clear trend
was observed. To further investigate the role of each chemical species, the mass fraction
percentage of each chemical species (i.e. sulphate, nitrate, ammonium and organics) was
calculated versus time. A GAM with a basis of penalised B-spline was applied to determine
the trend of each chemical species. The organics fraction increased following the start of
nucleation, while ammonium, sulphate and nitrate decreased, indicating the importance of
organics in the growth of particles after the start of nucleation. On the other hand, the
fraction of sulphate and ammonium peaked before the start of nucleation, which is an
indication of their involvement in triggering NPF. f43 and f44 were analysed in order to
further study of the role of organics in NPF events. f44 decreased when the particles were
generated from vehicle emissions, while both f44 and f43 decreased after the start of NPF
events. Vehicle and NPF generated particles clustered in different locations on the f44 vs f43
plot, indicating the application of this plot for identification of the source and
transformation characteristics of organic aerosols. The results of the third paper revealed
the importance of sulphates and ammonium in NPF events, possibly through ternary water-
sulphuric acid-ammonia nucleation, and the key role of organics in the growth of newly
formed particles.
6
1st Paper
Main Aims:
To
Quantify the spatial variation of PNC.
Identify and quantify main parameters
affecting spatial variation of PNC.
Determine the adequate number of
measurement sites required for PNC
monitoring at each school.
Establish the level of deviation in exposure
estimation resulting from the location of the
monitoring site.
2nd Paper
Main Aims:
To
Identify the optimum clustering technique and
number of clusters by comparing the performance of
three clustering techniques (i.e. PAM, CLARA, and
SOM) on PNSD measured at 25 sites across Brisbane.
Apply the identified method on the whole PNSD.
Develop a new method to parameterise particle size
spectra using GAM.
Attribute each cluster to its source/origin using the
characteristics of each cluster, parameterisation of
the results and several other air quality parameters.
3rd Paper
Main Aims:
To
Investigate the characteristics of NPF events in a
subtropical urban area.
Identify the main chemical species contributing to NPF
events and particles growth.
Characterise the nature of the contribution of each
chemical species compared to results found in the
literature.
Figure 1. 2: Structure of the scientific continuity of the research papers
7
Chapter 2: Literature Review
2.1. Introduction
The word ‘aerosol’ refers to two-phase systems which include particulate matter
suspended in the air/gas as a mixture of solid and/or liquid particles. Scientifically, it refers
to particulate suspension in gaseous medium and it is in this context which it used
throughout this thesis. Aerosol particles originate from natural or anthropogenic sources.
Natural sources include windblown dust, sea spray, photochemically formed particles, forest
fires and volcanic emissions, while combustion activities (e.g. burning fossil fuels and
biomass) are the main anthropogenic sources [1, 2]. Aerosol particles are divided into two
main groups, being primary and secondary, depending on their formation mechanisms.
Primary particles are directly emitted from their source into the atmosphere, whereas
secondary particles are formed in the atmosphere by gas to particle conversion resulting
from the chemical reaction of gaseous components in the atmosphere [3-5].
Aerosol particles play a key role in several important atmospheric and climate
conditions, including atmospheric visibility [6, 7], climate change [8], acid deposition [9] and
stratospheric ozone depletion [10]. Aerosols can significantly affect the atmosphere in three
important ways. Firstly, they can act as cloud condensation nuclei (CCN), enabling the
formation and growth of water droplets and ice crystal. Secondly, they can scatter and
absorb sunlight, which can affect the thermal structure of the atmosphere. Thirdly, through
heterogeneous interactions which can modify the trace gas composition of the atmosphere,
the chemical composition of aerosol particles and gas-phase reaction pathways [11-18]. An
increase in the number of aerosols can increase the number of particles which can act as
CCN, resulting in a decrease in droplet radius, lower precipitation efficiency in clouds, higher
cloud lifetime and an increase in the cloud’s reflectivity [16, 17]. The effect of an aerosol on
climate depends on the physical and chemical properties of the particles it contains, such as
their size, shape and chemical composition [12].
Aerosol particles can also have an adverse effect on human health by penetrating
into the lung through inhalation, and numerous epidemiological studies have shown
associations between ambient aerosol levels and adverse health outcomes, including
premature death, respiratory and cardiovascular disease, and neurodegenerative disorders
8
[19-28]. The level of the adverse effect of aerosol particles on human health is a complex
issue but particles size is one of the main contributors, as this governs the site of particle
deposition along the respiratory tract. Smaller particles, such as UFPs, that can penetrate
deeply into the pulmonary system have serious adverse effects on human health, leading to
pulmonary and cardiovascular diseases [6, 19, 29-32]. UFPs have even been observed in the
brain and central nervous system of rats [33].
Major sources of aerosol particles, their health and environmental effects, are
illustrated in Figure 2.1. While the detailed health effects of some aerosol particles are not
yet known, there are significant epidemiological studies showing the adverse health effects
of airborne particles. For instance, the World Health Organisation (WHO) has estimated that
exposure to aerosol particles causes around 348,000 premature deaths each year, which
reduces the average life expectancy by 8.6 months [34]. Toxicity of particles also relates to
their chemical composition [35] and there is a significant amount of evidence showing an
association between the chemical composition of aerosol particles and their health
outcomes [35-38].
The properties (e.g. size and chemical composition) of aerosol particles emitted in
the atmosphere undergo transportation and transformation as a result of various physical
and chemical processes [1]. Therefore, it is not only vital to understand the physical and
chemical properties of aerosol particles, but also their dynamics, transformation and
transportation, in order to evaluate their effects on climate and human health, and establish
effective control strategies. Despite significant progress in aerosol science, there are still
gaps in knowledge which need to be addressed in relation to the physical/chemical
properties, dynamics and transformation of aerosol particles.
9
Figure 2.1: Main sources and impacts of atmospheric aerosol [18].
2.2. Particle size
Particle size is an important physical property which can be used to understand the
characteristics of aerosols, such as their possible source/origin, formation and removal
mechanisms, site of deposition in the human respiratory tract, atmospheric lifetime, and
chemical composition [1]. Aerosol particles cover a wide range of sizes, from a few
nanometres to 100 µm in diameter, and the nature of physical laws governing particles
dynamics depends strongly on particles size. Size can be simply defined as the geometric
diameter of spherical particles with unit density. However, almost all aerosol particles have
irregular shapes and different densities. Therefore, several equivalent (e.g. aerodynamic,
diffusive and mobility) diameters have been defined, based on the different measurement
techniques.
Aerodynamic diameter is defined as the diameter of a spherical particle with unit
density which has the same settling velocity as the particle in question. Aerodynamic
diameter is useful for characterising particles with significant inertia, which are typically
particles larger than 0.5 µm. Smaller particles undergo Brownian motion and diffusion
diameter is used for them. Diffusion diameter is defined as the diameter of a spherical
10
particle with density of unity which has the same diffusion as the particle in question.
However, electrical mobility diameter is usually used for UFPs, because the most commonly
used instruments measure this diameter. Electrical mobility diameter is defined as the
diameter of a spherical particle with unit density which has the same electrical mobility as
the particle in question [39]. Particles with diameters less than 2.5 µm, 0.1 µm and 50 nm
are called fine, ultrafine and nano- particles, respectively.
2.3. Particle concentration
Mass concentration is the most commonly measured aerosol property and is defined
as the mass of particles in a unit volume of air, which is normally expressed in μgm-3. Air
pollution and workplace exposure standards are mainly stated in terms of particle mass
concentration. Particle number concentration (PNC) is another important aerosol
characteristic which has emerged in the field of aerosol science following that of particle
mass concentration. PNC is defined as the number of particles in a unit volume of air and is
commonly expressed in number/cm3.
UFPs make the highest contribution to ambient PNC, while making a very low
contribution to particle mass concentration, therefore, PNC can be used as a measure of
UFPs. A study in the US showed that 25% and 75% of aerosol particles in an urban area were
smaller than 10nm and 50nm, respectively [40]. Another study showed that 26% and 89% of
particles in the studied urban area were smaller than 10nm and 100nm, respectively [41].
UFPs have both natural and anthropogenic sources, with new particle formation (NPF) and
vehicular emissions being the main source of natural and anthropogenic particles,
respectively [42-46]. Particles emitted from petrol and diesel engines are in range of 20-
60nm and 20-130nm, respectively [3, 47, 48], and therefore, they contribute mainly UFPs.
While significant toxicological studies indicate the adverse effects of UFPs on human health
[49-52], PNC are not currently used for air quality regulations and particle number emission
limits have only recently been introduced in the upcoming Euro5/6 European vehicle
emissions standards [53].
Mass concentration can either be measured directly by gravimetric methods or by
the measurement of another property (e.g. light scattering), which can then be related to
particle mass. PNC can either be measured by sampling particles using a collection method
11
(e.g. filter or impinger) and then counting the number of particles collected, or by the direct
measurement of PNC employing real time methods (e.g. Condensation Particle Counter
(CPC)) [1, 39].
CPCs use an optical detector to count the number of particles in the sampled flow
through a viewing volume and are commonly used for the measurement of PNC [54]. The
CPC works with a condensation system which condenses a working vapour on the particles
to enlarge their size, in order to become detectable by a conventional optical detection
system [54, 55]. A supersaturated condition is required inside the condensation system to
make the working fluid condense on particles at a high rate, for which there are currently
four techniques available - expansion, conductive cooling, differential diffusion and mixing
[56].
CPCs are usually either water-based or butanol-based, however, in indoor
experiments, it is better to use water-based CPCs, in order to avoid issues such as exposure
to butanol vapour and accidental spills. The measurement of PNC inside the CPC includes
three processes: 1) the supersaturation of a working fluid (e.g. water or butanol); 2) the
growth of particle by condensation of the vapour on the surface of them; and 3) the
detection of particles using optical methods [39]. The different parts of a TSI 3781 CPC are
presented in Figure 2.2 (all types of water-based CPCs have similar parts). In water-based
CPCs, cold aerosol flow is introduced into a warm wet walled tube. Since the mass diffusion
of water vapour from the wall to the centre of the tube is faster than the warming of the
aerosol stream, the supersaturated condition is achieved inside the tube.
12
Figure 2.2: Schematic diagram of TSI 3781 water-based CPC [54].
CPCs are able to detect particles with a size range of a few nanometres up to several
micrometres. For instance, the TSI 3781 CPC can detect particles with size range of 6nm to
more than 6μm, with a concentration range of 0-5 × 105 particles/cm3. The manufacturer
claims that it has an uncertainty of 10% [54], and therefore, CPCs are deemed to be a
reliable and accurate instrument for the measurement of PNC. According to the literature
review, no studies are available regarding the performance of this type of CPC, therefore,
the performance claimed by the manufacturer has been stated here and the performance
assessment conducted in this thesis can be found in chapter 3.
2.4. Particle Chemical Composition
The chemical composition of atmospheric aerosol particles includes sulphates,
nitrates, ammonium, organic material, elemental carbon, crustal species, sea salt, metal
oxides, hydrogen ions and water. Fine particles predominantly include sulphate,
ammonium, organic and elemental carbon [2]. The impact of aerosol particles on human
health and the environment relates to both their chemical composition and physical
properties [12, 35-38].
As mentioned above, atmospheric aerosols can contain many different compounds,
which makes measuring the chemical composition of aerosol particles more complicated
13
than the measurement of their size or concentration [57]. Offline measurement methods
have been used conventionally for measuring the chemical composition of aerosol particles,
however, online measurements methods have recently become common due to advances in
instrument development and the advantages of online measurements.
2.4.1. Offline methods
Offline measurements methods involve the collection of aerosol particles on
substrates (e.g. filters, impactors, and denuders) and the transfer of the collected samples
to a laboratory for extraction and chemical analysis. Atomic absorption, scanning electron
microscopy (SEM), high performance liquid chromatography (HPLC), gas chromatography
(GC), ion chromatography (IC), proton nuclear magnetic resonance (PNMR), secondary ion
mass spectrometry (SIMS), inductively coupled plasma mass spectrometry (ICP-MS),
Inductively Coupled Plasma Atomic Emission Spectrometry (ICP-AES) and laser microprobe
mass spectrometry are the main analytical techniques available for the offline chemical
analysis of collected particles. More details on the application of the abovementioned
techniques can be found in [58-60].
Filters are the most common medium used for the collection of aerosol particles for
chemical analysis because filter samplers are inexpensive and simple to use. However, they
require extensive manual work and a large number of filters for long term field
measurements. Moreover, size information can only be provided by use of filters with
different pore sizes in series, which results in a very coarse resolution. Impactors employ a
range of film and foil substrates in order to sample aerosol particles classified according to
their aerodynamic diameters. Impactors which collect aerosol particles down to 0.05µm,
including the Berner impactor [61] and the Micro Orifice Uniform Deposit Impacotr
(MOUDI), are commonly used for analysing the chemical composition of aerosol particles.
However, aerosol particles can bounce off the impactor substrate, which leads to incorrect
sizing of the particles [62], and while coated substrates have been used to address this
issue, coating usually interferes with chemical analysis, particularly when particulate organic
carbon contents are present.
Using filters or impactors can introduce positive or negative errors in relation to the
measured composition of the collected aerosol particles. Volatilisation of semivolatile
14
compounds [63, 64] and adsorption of gas-phase compounds onto filters [62] are the main
source of negative and positive errors, respectively. The diffusion denuder-based sampling
technique has been developed to address the artefacts associated with the measurement of
semivolatile compounds [65-67] and while denuder-based samplers do not have the
limitations of the filter and impactor based methods, they still have the common
disadvantages of all the offline methods, such as contamination during sample collection
and analysis, long time requirements and most importantly, long sampling periods, which
can result in missing the short-term chemical changes that can occur in air masses and are
vital for investigating the transformation and variation of chemical species.
Therefore, due to the range of issues related to offline measurements mentioned
above, online measurement of the chemical composition of aerosol particles should be
considered when investigating the chemistry related to aerosol dynamics.
2.4.2. Online methods
Online (also called real-time and in-situ) chemical measurements of aerosol particles
increase the time resolution of samples and reduce errors due to volatilisation and
adsorption. Online instruments include the Aethalometer [68], Multiangle Absorption
Photometre (MAAP), Online XRF, OC&EC, flame photometric detectors (FPD) [69],
inductively coupled plasma [59] and the particle into liquid sampler (PILS) [70, 71]. PILS is
the only one of these techniques that provides size-resolved data, however, it does not
provide size discrimination of particle composition and only detects particles that can act as
water condensation nuclei. Aerosol mass spectrometers (AMS’s) are the most significant
development in the past 20 years, which have been shown to be capable of providing
valuable information about the sources, chemical composition and transformations of
aerosol particles [60].
The AMS has attracted considerable attention recently as a means of measuring the
size discriminated chemical composition of individual particles in real-time, since sampling
errors are significantly reduced compared to other methods and the characterisation
process is very quick. A comprehensive history of the development of the AMS can be found
in [72-76]. In the AMS, particles are sampled and focused into a narrow beam through a
supersonic expansion inlet nozzle [77-79], which then enters a particle sizing chamber. The
15
sizing chamber also conducts time-of-flight particle velocity measurements that can be
correlated with particle density and size. After passing through the chamber, particles are
vaporised and ionised before being analysed by a mass spectrometer (Figure 2.3) [80, 81].
Figure 2.3: Schematic of an Aerodyne AMS [80].
2.5. Spatial variation of particle concentration
There is considerable spatial variation in aerosol particle concentration, due to the
transient nature of aerosols, spatial variation in particle sources, atmospheric processes and
meteorological conditions. For instance, the annual average concentration of PM2.5 varies by
more than an order of magnitude in the relatively clean air of remote areas compared to the
polluted air in urban areas [82]. Understanding the spatial variation of particles helps in
understanding aerosol dynamics, as well as human exposure.
2.5.1. Spatial variation of particle mass concentration
Spatial and temporal variations of particle mass concentration in large macroscale
environments have been investigated in several studies [83-93]. Wongphatarakul et al.
investigated the variation in chemical component data collected at 21 locations around the
world and found strong similarities in the average concentration at close sites, with a
correlation coefficient of higher than 0.7 [83]. Adgate et al. investigated the spatial variation
in outdoor and indoor, as well as personal exposure to, particle mass concentration across
Minneapolis, USA. They found lower spatial variation in outdoor compared to indoor
16
concentrations and concluded that a centrally located ambient particle mass concentration
measurement site was sufficient in order to estimate average outdoor PM2.5 in the
Minneapolis-St Paul metropolitan area [84]. Morawska et al. investigated the variation in
ozone, NOx and PM10 across three locations in Brisbane, Australia and found PM10 to be
more spatially variable compared to ozone and NOx. Based on this, they suggested using
more sites for the assessment of exposure to particulate mass within an urban environment
[85]. DeGaetano et al. investigated the distribution in measurements of PM2.5 concentration
at 20 stations across New York City, USA and they found high diurnal, weekly and seasonal
variation, while at the same time observing very small spatial variation in 90% of the
observations [86]. Similarly, Martuzevicius et al. and Sarnat et al. found low spatial variation
in PM2.5 concentration across 11 locations in the Cincinnati metropolitan area and Israeli,
Jordanian and Palestinian cities, respectively [87, 93].
In summary, while the abovementioned studies established that variations in particle
mass concentration were higher compared to gaseous pollutants, the spatial variation of
particle mass concentration was still found to be relatively low across different urban
environments. However, PNC as a measure of UFP concentration, can potentially exhibit
higher spatial variation due to its more transient nature and several studies have
investigated this important issue [94-102].
2.5.2. Spatial variation of particle number concentration
Spatial variation of PNC across Helsinki was investigated by Buzorius et al. [94] and
they found traffic density to be the main factor affecting the spatial variation of PNC.
Puustinen et al. [98] investigated the spatial variation of PNC and particle mass
concentration over four European cities (i.e. Amsterdam, Athens, Birmingham, and
Helsinki). They found that single measurement sites were not representative of PNC over a
wider urban area, however, a single central site could represent the overall PM10 and PM2.5
in the same urban area. In another study, spatial and temporal variation of PNC was
investigated across four sites in Augsburg, Germany [99]. Significant differences in absolute
PNC levels at those sites were observed and they concluded that using a single site for long
term epidemiological studies was not adequate. PNC measurements at 14 sites in Los
Angeles were conducted and significant variability in total PNC was observed [100]. Hudda
17
et al. [102] quantified the spatial variation of PNC using coefficient of divergence across six
sites within an urban area and found the highest variation during the morning rush hour. In
addition to the abovementioned studies, other studies have found similar results in other
environments as well [103, 104].
In summary, spatial variation of PNC was found to be higher than the particle mass
concentration, and most studies focused on the spatial variation of PNC in macroscale
environments or in environments with high exposure (e.g. near busy roads and highways).
However, humans spend most of their time in environments with lower exposure levels
[105].
Some studies were conducted in microscale environments, however, they were
mostly focused on environments with high levels of PNC, such as in the vicinity of a major
highway. For instance, Zhu et al. [95] investigated the variation of PNC downwind and
upwind from a highway and found traffic density and wind speed/direction to be the main
factors influencing the PNC level near highways. They found PNC to decrease exponentially
with distance from the highway along the same vector as wind direction. The effects of
traffic and meteorological parameters on the spatial variation of PNC were investigated in
another study which was conducted in Brisbane, Australia [97]. Considerable heterogeneity
in PNC in a busy urban environment and the exponential decrease of PNC with increasing
wind speed was also observed in that study.
2.5.3. Effects of spatial variation on exposure estimation in microscale
environments such as schools
Children are particularly vulnerable to exposure to particulate matter and they spend
a significant amount of their time in schools [106]. Children are exposed to particles emitted
from vehicles coming to pick up/drop off children, as well as from the adjacent roads.
Therefore, it is important to assess children’s exposure to aerosol particles and it has been
the subject of recent studies in the literature.
Hochstetler et al. [107] measured indoor and outdoor PNC at four urban elementary
schools and assessed children’s exposure to the diesel exhaust particles generated by school
buses. They found that the presence of school buses significantly affected outdoor UFPs,
which were strongly associated with vehicle counts. Zhang et al. [108] measured PNC and
18
other air quality parameters inside and outside five schools in both urban and rural areas in
southern Texas. A greater variation was found for indoor versus outdoor PNC and indoor
sources, such as gas fan heaters, were also found to be important contributors to indoor
PNC and consequently, to indoor to outdoor ratios. In another study, indoor and outdoor
PNC were measured at 7 primary schools to assess the children’s exposure to PNC [109].
They observed a maximum outdoor PNC of about 3.7 × 104 cm-3 at two schools and
concluded that this was due to heavy traffic on adjacent roads.
One monitoring site was employed in order to measure outdoor PNC in the school
grounds for all of the abovementioned studies. A single monitoring site has also always been
used for studying PNC in microscale environments such as schools. However, the spatial
variation of PNC can potentially affect study results, as the location of the single outdoor
monitoring site can influence the indoor to outdoor ratio results, as well as estimations of
exposure and consequently, the conclusions drawn from the study. Therefore, the spatial
variation of PNC in microscale environments needs to be assessed in order to address the
effect of spatial variation on exposure estimation results. No study has ever investigated the
level of spatial variability of PNC across microscale environments and its possible impact on
human exposure estimation.
2.5.4. Quantification of the spatial variation
Coefficient of variation (CV) and coefficient of divergence are methods which have
been employed in the literature for quantification of the spatial variation of air quality
parameters, including PNC. Coefficient of variation is defined as:
1)
where STD and AVG are standard deviation and average of the observations
respectively. CD is defined as [88] :
2) √
∑ (
)
where f and h represent the two monitoring sites, and n is the number of
observations. CD approaches zero if the measured value of the two sites are similar and if
19
they are very different, it approaches unity. CV is equal to zero when the measured values
at all the sites are the same and it increases as the variation increases. CV can exceed 1 in
contrast with CD which has the maximum value of 1.
CV and CD are both good methods for quantification of spatial variation. CD can only
be used when spatial variation between two monitoring sites is of interest, however, CV can
be applied to more than two sites as well. Therefore, CV should be employed if the spatial
variations of more than two sites are of interest. Instrumental uncertainty can affect the
quantification of variation and it should be taken into account for interpretation of the
spatial variation. Surprisingly, to our knowledge, no study has considered the instrumental
uncertainty associated with PNC measurements in their analysis.
2.6. Particle number size distribution
Particle size distribution is one of the most important parameters characterising
aerosols, as it represents the distribution of a specific property of an aerosol within a
particular size range [39]. Particle size distribution measurements have been performed in
different locations around the world, mostly in Europe and North America [43]. Particle size
is mostly weighted by number to construct a particle number size distribution (PNSD). PNSD
in the submicrometre size range is usually measured using a SMPS, which includes an
electrostatic classifier (EC) that selects a narrow size range (bin) of particles and a CPC to
measure the concentration of particles at each bin [55]. A schematic diagram of a TSI EC is
presented in Figure 2. 4.
20
Figure 2. 4: Schematic diagram of a TSI EC [110].
PNSD’s are represented by plotting particle number concentration on the y-axis
versus particle size on the x-axis (Figure 2.5). The particle concentrations related to each bin
are normalised to remove the effect of bin width. The concentration is usually divided by
the bin width to give a normalised value independent of the bin width as follows:
3)
⁄
where dN is particle concentration, Dp is midpoint diameter, Dp,u is upper channel
diameter and Dp,l is lower channel diameter.
21
Figure 2.5: An example of an SMPS scan.
2.6.1. PNSD modes
PNSD consists of nucleation, Aitken, accumulation and coarse modes (Figure 2.6).
Different modes of particles can refer to different sources and particle dynamics. Particles
smaller than 20 nm or so are called nucleation mode particles, or particles in the nucleation
mode. Particles in nucleation mode are mainly the result of gas to particle conversion of
vapours with low volatilities. Particle in the range of ~ 20- ~ 100 nm are in the Aitken mode
and they can be the result of anthropogenic activities, such as combustion or growth by
condensation, and coagulation processes from aerosols in the nucleation mode. Particles in
the range ~100nm - ~ 1 μm are said to be in the accumulation mode and they are usually
result of combustion processes. Particles larger than ~ 1 µm are in the coarse mode and
mechanical processes, windblown dust, sea spray, volcanos and plants are the main source
of these particles. Dust, pollen and ground materials are in the coarse range, smoke and
fumes are in the submicrometer range and newly formed particles are just a few
nanometres in diameter and they approach the size of gas molecules.
22
Figure 2.6: Idealised schematic of an atmospheric particle number size distribution [111].
Each mode of the PNSD has been found to be lognormally distributed, therefore,
multi-lognormal distribution functions are commonly used to fit the PNSD data [112]. The
following multi-lognormal distribution function has been commonly used for this purpose
[111, 113, 114]:
4) ( ̅ ) ∑
√ ( )
[
[ ( ) ( ̅ )]
( )]
where is the diameter of an aerosol particle. Three parameters characterise an
individual lognormal mode : the mode number concentration , geometric variance
and geometric mean diameter [111]. It is beneficial to use this technique for long time
averaged data in particular, when a general trend is of interest. However, it should be noted
that some local peaks can be concealed as a result of using this technique, as particle
number size data is not always multi-lognormal and there are some cases that the multi-
lognormal function is not the best fit for the data. Therefore, a new approach to
parameterise PNSD data needs to be developed which can be applied to particle size data
without the drawbacks of the multi-lognormal approach.
23
2.6.2. Cluster analysis of PNSD data
In order to study temporal atmospheric transformations, PNSD data are normally
collected at short intervals (e.g. five minutes) for long periods of up to several years. This
results in a large amount of data collected by the SMPS, which makes handling the data a
challenging task. After performing pre-processing and quality control of the whole dataset,
the next step is analysing and grouping the PNSD data based on their properties. This
process used to be done manually, which was very time consuming and subject to human
error, and the results were subjective and not repeatable. Automatic clustering techniques
have recently attracted attention for their high performance in relation to PNSD data
analysis. Clustering techniques are sophisticated methods which can be applied to the SMPS
data, to group them automatically. Cluster analysis or clustering refers to the task of
grouping a set of objects into similar groups (clusters). Different algorithms have been
developed for different purposes, which results in different clustering outcomes for
different datasets.
The performance of four clustering algorithms (Fuzzy, K-means, K-median and model
based) were compared using the Dunn index and Silhouette validation values, as outlined by
Beddows et al [115]. The Dunn index [116] and Silhouette width [117] are scores resulting
from the nonlinear combination of compactness and separation. The K-means technique,
which is an iterative algorithm that minimises the within clusters sum of squares to find a
predefined number of clusters [118], was found to be the optimum method for clustering
PNSD data. In this work, 10 and 15 clusters were found in the datasets collected in rural and
urban areas, respectively. Then, each cluster was related to its origin using other air quality
parameters and the temporal profile of each cluster. The main sources/origins were traffic
emissions and regional background.
Clustering was also employed to analyse three years’ worth of PNSD data at an urban
station in Helsinki, Finland [119]. Seven clusters were extracted from the whole dataset and
they were analysed using the PNSD physical properties, clusters temporal frequencies and
their relation to local meteorological parameters. The clusters were found to be related to
three sources and origins, where 69% of the PNSD was found to be the result of
anthropogenic activities, 29% were maritime-type particles and 2% were newly formed
24
particles. In another study, source and time variability of aerosol particles were examined
using clustering and positive matrix factorisation (PMF) techniques [120]. Nine clusters were
identified and attributed to traffic (69%), dilution (15%), summer background (4%) and
regional pollution (12%) using directional association, diurnal variation, meteorological and
pollution variables. A combination of the abovementioned techniques was found to be
capable of separating the components and quantifying their contribution to total particles.
In a similar study, PNSD were separated into eight clusters and the source/origins of each
cluster were found to be: 1) Traffic, regional nucleation; 2) Winter traffic; 3) Traffic with low
regional background, long-range transport; 4) Night traffic; 5) Night traffic, long-range
transport; 6) Summer regional with high traffic; 7) Summer regional with low traffic; and 8)
Summer long-range transport with low traffic [121].
There are several other clustering methods (e.g. PAM, CLARA, SOM, and AP)
available, however their performance in terms of clustering PNSD data is yet to be
evaluated. The partitioning around medoids (PAM) technique is an iterative algorithm
similar to K-means, which builds up the clusters around a set of representative objects using
the sum of pair wise dissimilarities to assign each data to the nearest representative object
[122]. Clustering for large applications (CLARA) has been developed to run the PAM
algorithm quicker by performing it on a number of subgroups of data [122], while the self-
organising map (SOM) method is an algorithm based on artificial neural networks, which is
able to map high dimensional data in two dimensions and is extensively used for data
mining purposes [123]. Affinity propagation (AP) is a more recent method compared to the
aforementioned, which has been used in several different areas since its introduction. The
AP algorithm considers all data as potential exemplars and then searches for the optimum
set and their corresponding clusters by communicating between the data points [124]. All of
the above mentioned algorithms can be implemented using different packages in R [125].
2.7. New Particle Formation
John Aitken made the first apparatus to measure the number of fog and dust
particles in the late 19th century, which subsequently led to the first evidence of particle
formation in atmosphere. However, very little progress was made in understanding the
underlying mechanisms of NPF events before the development of the first instruments
capable of measuring the PNSD down to a few nanometres [2, 126].
25
2.7.1. Mechanisms
NPF is initially triggered by the nucleation of vapour molecules which results in the
formation of thermodynamically-stable charged or neutral clusters. Afterwards,
organic/inorganic vapour condenses on the clusters to form aerosol particles, which grow
further to reach a diameter that can act as a CCN [126]. The aforementioned processes are
illustrated in Figure 2.7.
Figure 2.7: Illustration of processes included in NPF [126].
Several mechanisms for NPF have been proposed and the most common ones are: 1)
homogeneous binary water-sulphuric acid nucleation; 2) homogenous ternary water-
sulphuric acid-ammonia nucleation; 3) ion-induced nucleation of binary or ternary vapours;
and 4) organics involved nucleation [127].
Homogeneous binary water-sulphuric acid nucleation: SO2 oxidises in the
atmosphere and forms sulphuric acid (H2SO4) molecules, which reaches different states of
hydration, so that not only individual molecules of water and sulphuric acid, but individual
water and hydrated sulphuric acid molecules start clustering due to low vapour pressure [2].
Nucleation rates estimated based this mechanism are lower than observations in the
atmosphere, indicating the presence of another factor that affects nucleation in the
atmosphere.
26
Homogenous ternary water-sulphuric acid-ammonia nucleation: This
mechanism involves sulphuric acid, water and ammonia, and the nucleation of sulphuric
acid particles is enhanced in this process by the presence of ammonia [128-130].
Ion-induced nucleation: Ions can enhance nucleation by stabilising the newly
formed clusters and it also can increase the growth rate of particles [131-133]. Ion-induced
nucleation has been observed in the atmosphere [134, 135].
Organics involved nucleation: This mechanism involves organic species, as well as
ammonium and sulphuric acid. There has been evidence of the involvement of organics,
such as cis-pinonic, pinic acid and amines, in the process of nucleation, as they enhance the
nucleation and particle growth rates [136-138].
In the atmosphere, ion or neutral clusters are ubiquitous and activation of these
clusters can cause new particle formation events as well [139-141]. Concept of activation
probability explains the activation mechanism. Employing this mechanism resulted in a
model which followed the observed NPF events trends closely [142].
In a recent study, state of the art instrumentations capable of measuring particle size
and chemistry down to 1 nm were used to investigate NPF events. Three main size regimes
in atmospheric neutral clusters were found. Regime 1 (particles with a mobility diameter
less than 1.1 - 1.3 nm) included clusters with very small net growth as an equivalent lost and
the formation of clusters was a result of evaporation, vapour uptake and chemical reaction.
Regime 2 (particles with a mobility diameter between 1.1 - 1.9 nm) included clusters with
medium growth due to the condensation of sulphuric acid and its stabilisation by ammonia,
amines, or other organic vapours. Regime 3 (particles with a mobility diameter range of 1.5 -
3 nm) included clusters with a fast growth rate, which cannot be explained by the
condensation of sulphuric acid. This fast growth rate was due to enhanced vapour uptake,
which was likely related to the presence of oxidised organic vapours [143].
27
Figure 2.8: Schematic diagram of main size regimes of atmospheric clusters and their relation to NPF [143].
2.7.2. NPF detection
NPF events have been observed in many types of environments around the world,
including the free troposphere, subarctic Lapland, remote boreal forests, coastal and
mountainous environments, and urban and suburban areas [43]. A typical NPF event is
illustrated in Figure 2.9, which is based on an event observed on 11 August 2001, in
Pittsburgh, US [144]. A significant increase in PNC can be observed at around 10 am and
then the particles grew smoothly versus time for several hours, even as different air parcels
passed the measurement site. This shows that this NPF is a regional event and has taken
place over a distance of several hundred kilometres [2]. A banana shape, similar to the one
illustrated in Figure 2.9, is usually observed in PNSD contour plots as an indicator of the NPF
event [145].
28
Figure 2.9: Evolution of PNSD in a day with NPF event [2, 144].
The most commonly used criteria for the detection of NPF events in PNSD contour
plots were proposed by Dal Maso et al. [146]. The measurements days were classified into
class 1 & 2 event, non-event and undefined days. A day was classified as an event day if a
distinct nucleation mode was observed in the PNSD and the particles grew for more than an
hour. A flow chart of their proposed criteria is illustrated in Figure 2.10.
Figure 2.10: Flow chart of the procedure for NPF events detection [146].
29
2.7.3. Formation, growth rate and condensation sink
Particle formation rate is defined as the amount of critical clusters formed during the
NPF and can be used to estimate the activation coefficient. Particle growth rate is defined as
the growth of particle diameter vs. time due to the condensation of the vapour on the
newly formed clusters/particles and can be used to calculate the source rate of
condensation vapour [142]. Neutral clusters can be measured by almost none of the
available instruments, therefore, the formation rate of larger particles is usually calculated
using the following formula [43]:
5) ( ) ( )
where JD is the flux of particles passed the size D, t is the time and ( ) is the
PNSD. Equation (5) requires a PNSD function, as well as the particle growth rate, and this
kind of information is usually not available. Therefore, instead of using the instantaneous
particle formation rate (JD), average formation can be calculated using the following formula
[43]:
6)
where is the duration of NPF event, ND,Dmax is the total PNC in the size range [D,Dmax] and
Dmax is the maximum diameter that clusters reach during their growth in . The first term in
equation (6) can be calculated using the PNSD data, the second and third term on the right
hand side of the equation represent loss due to self-coagulation and scavenging by pre-
existing particles, and the fourth term is the loss due to transport of the particles. The last
three terms on the right hand side of the equation are normally small and assumed to be
zero and therefore, the equation can be simplified to [43]:
7)
The growth rate of the particles during NPF event can be calculated based on the
evolution of the mean diameter of the newly formed particles using the formula below:
30
8)
where GR is the growth rate, and Dm belongs to the [D, Dmax] size range. To calculate
the growth rate, linear regression line is fitted to the mean diameter versus time [145].
Condensation sink (CS) is another important parameter representing the available
surface area on particles for condensation. CS determines the rate of condensation of
vapours on pre-existing particles and can be calculated using PNSD [147, 148]. CS can be
calculated using the following formula [147]:
9)
i
iipMpppMp NdDdddnddDCSi ,
0
2)()(2
where D is the diffusion coefficient, dp is the particle diameter, Ni is the distribution
of particles and M , which is the coefficient correcting for free molecular effects [149], is:
10)
KnKnKn
KnM
121
3
4
3
41377.0
1
where is the sticking coefficient and can be assumed to be one [150], and Knudsen
number (Kn) is:
11) p
v
dKn
2
where v which is the free mean path is defined as:
12) )
)110
1(
)2.296
1101(
)(2.296
)(101
(
T
T
pov
where P is the pressure in Kpa, T is the temperature in K and o can be assumed to
be 0.039 µm which is the free mean path of sulphuric acid at standard conditions [150].
31
2.7.4. Affecting parameters
Solar radiation: Several factors have been found to affect NPF events, including solar
radiation, temperature, relative humidity and sulphuric acid concentration. A strong
correlation between solar radiation intensity and NPF events has been observed and most
nucleation events were found to occur during day time hours [43]. Hydroxyl can be formed
in the presence of high solar radiation, which can then react with SO2 to form sulphuric acid
in the air. These photochemical activities can enhance the photoixidation reactions, often
leading to higher particle growth rates as well [151] (Figure 2.11).
Figure 2.11: Growth rate of particles during NPF events [151].
Temperature: Temperature can affect the condensation/evaporation of molecules
to/from particles, as the low temperature enhances condensation, whereas high
temperatures enhance the evaporation process. Therefore, in theory, low temperatures are
favourable for NPF and the growth of newly formed particles. However, since solar radiation
correlates with the temperature, then the unfavourable effect of temperature on NPF
events may not be observed in filed measurements.
Humidity: inverse correlation between relative humidity and NPF occurrence and
particle formation rates has been observed in different continental locations [43, 152, 153].
These observations are surprising, as water and sulphuric acid have been found to co-
nucleate very efficiently in laboratory experiments. Several reasons for this somewhat odd
observation have been suggested, such as the water uptake of pre-existing particles and
32
their consequent increase in CS and coagulation. However, it seems that this relates simply
to the inverse correlation of humidity with solar radiation [152].
Sulphuric acid: Sulphuric acid is an important factor in binary and ternary nucleation
and is formed by the oxidation of SO2 in the atmosphere [154]. The measurement of
sulphuric acid is quite difficult and rarely available in field measurements, therefore, a proxy
often needs to be defined. SO2 concentration and solar radiation are the main driving force
for sulphuric acid production, therefore, a sulphuric acid concentration proxy has been
defined as follows [151]:
13) [ ] [ ]
where [H2SO4] and [SO2] are the sulphuric acid and SO2 concentrations, Rad is the
solar radiation intensity and CS is the condensation sink of pre-existing particles.
2.7.5. Chemistry of NPF
Most of the studies on NPF events have focused on physical aspects of this
phenomenon, while a few studies investigated the chemistry behind it. Zhang et al. [153]
investigated the chemistry of particles during the growth phase of NPF events in an urban
environment and found the distinctive growth of sulphate, ammonium, organics and nitrate
in the ultrafine mode during NPF events. Sulphate was found to be the first and fastest
species to grow, ammonium increased next with a 10-40 minute delay and secondary
organics contributed significantly to the growth of particles at a relatively later time during
the event. Nitrate was found to be a minor contributor to UFPs and the growth of newly
formed particles.
Creamean et al. [138] measured the CCN activity of newly formed ambient particles
directly at remote site during two nucleation events using an ultrafine aerosol time of flight
mass spectrometer (UF-ATOFMS). They detected amines and sulphate in aerosol particles
during NPF events and confirmed the important role of sulphuric acid in the formation of
particles, while amines were involved in the growth of particles. Bzdek at al. [154] employed
a nano aerosol mass spectrometer (NAMS) at a coastal site to measure the composition of
particles with a mobility diameter of 18 nm. Ammonium nitrate was found to contribute
significantly to aerosol particle growth and they also find a significant fraction of organics in
33
aerosol particles. In one recent study, Kulmala et al [143] used a particle size magnifier
(PSM), nano air ion spectrometer (NAIS), atmospheric pressure interface time of flight mass
spectrometer (APi-TOF) and chemical ionization mass spectrometer (CIMS) to measure the
particle size and chemical composition down to 1 nm mobility diameter during NPF events.
They detected the presence of amines in the gas phase and sulphuric acid-amine clusters in
the very early stages of the nucleation process.
There have been important findings regarding the chemistry and physics of NPF
events during the last few years, however, the scale of these studies and their geographical
spread is still not enough to draw any comprehensive conclusions. Therefore, more research
on NPF events, particularly in relation to their chemical aspects, needs to be undertaken,
especially in less studied geographical areas.
2.8. Gaps in knowledge
The literature review identified the following gaps in knowledge:
There is no data available regarding the level of spatial variation of PNC in
microscale environments such as schools. (Section 2.5)
The relationships between the spatial variation of the PNC in microscale
environments and its affecting parameters are yet to be determined.
(Section 2.5)
The impact of spatial variation of particle number concentration on exposure
assessments has not been studied. (Section 2.5)
There are no data available regarding the performance of some clustering
algorithms (such as CLARA, SOM, PAM, AP) on PNSD data. (Section 2.6.2)
Multilognormal function is the only method used for PNSD parameterisation,
despite its drawbacks. Application of more sophisticated methods, such as
GAM for parameterisation of PNSD data, has not yet been assessed.
(Section 2.6.1)
There is little information available on the chemistry of NPF events,
particularly in urban environments. (Section 2.7.5)
There is no data available regarding the chemistry of NPF events in urban
subtropical environments. (Section 2.7.5)
2.9. References 1. Hinds, W.C., Aerosol technology: properties, behavior, and measurement of airborne
particles. 1999, New York: Wiley.
34
2. Seinfeld, J.H. and S.N. Pandis, Atmospheric chemistry and physics: from air pollution to climate change. 2006, Hoboken, N.J: Wiley.
3. Morawska, L., et al., Ambient nano and ultrafine particles from motor vehicle emissions: Characteristics, ambient processing and implications on human exposure. Atmospheric Environment, 2008. 42(35): p. 8113-8138.
4. Kittelson, D.B., Engines and nanoparticles: a review. Journal of Aerosol Science, 1998. 29(5–6): p. 575-588.
5. Ristovski, Z.D., et al., PARTICLE EMISSIONS FROM COMPRESSED NATURAL GAS ENGINES. Journal of Aerosol Science, 2000. 31(4): p. 403-413.
6. Davidson, C.I., R.F. Phalen, and P.A. Solomon, Airborne particulate matter and human health: a review. Aerosol Science and Technology, 2005. 39(8): p. 737-749.
7. McMurry, P.H., M.F. Shepherd, and J.S. Vickery, Particulate matter science for policy makers: A NARSTO assessment. 2004: Cambridge University Press.
8. Solomon, S., et al., Climate Change 2007. The Physical Science basis: Summary for policymakers. Contribution of Working Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change. Climate Change 2007: The Physical Science basis: Summary for policymakers. Contribution of Working Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change, 2007.
9. NAPAP, 1990 Integrated Assessment Report. National Acid Assessment Program, Washington DC. 1990.
10. Scientific assessment of ozone depletion: 2010, W.M. Org, Editor. 2010. 11. Stevens, B. and G. Feingold, Untangling aerosol effects on clouds and precipitation in a
buffered system. Nature, 2009. 461(7264): p. 607-613. 12. Yu, H., et al., A review of measurement-based assessments of the aerosol direct radiative
effect and forcing. Atmospheric Chemistry and Physics, 2006. 6(3): p. 613-666. 13. Randles, C. and V. Ramaswamy, Absorbing aerosols over Asia: A Geophysical Fluid Dynamics
Laboratory general circulation model sensitivity study of model response to aerosol optical depth and aerosol absorption. Journal of Geophysical Research: Atmospheres (1984–2012), 2008. 113(D21).
14. Ramanathan, V., et al., Atmospheric brown clouds: Impacts on South Asian climate and hydrological cycle. Proceedings of the National Academy of Sciences of the United States of America, 2005. 102(15): p. 5326-5333.
15. Lohmann, U. and J. Feichter, Global indirect aerosol effects: a review. Atmospheric Chemistry and Physics, 2005. 5(3): p. 715-737.
16. Ramanathan, V., et al., Aerosols, climate, and the hydrological cycle. science, 2001. 294(5549): p. 2119-2124.
17. Kuang, Z. and Y.L. Yung, Reflectivity variations off the Peru Coast: Evidence for indirect effect of anthropogenic sulfate aerosols on clouds. Geophysical research letters, 2000. 27(16): p. 2501-2504.
18. Kolb, C.E. and D.R. Worsnop, Chemistry and Composition of Atmospheric Aerosol Particles. Annual Review of Physical Chemistry, 2012. 63(1): p. 471-491.
19. Pope II, C.A. and D.W. Dockery, Health effects of fine particulate air pollution: lines that connect. Air & Waste Management Association, 2006(56): p. 709-742.
20. Dockery, D.W., Health effects of particulate air pollution. Annals of epidemiology, 2009. 19(4): p. 257-263.
21. Dockery, D.W. and C.A. Pope, Acute respiratory effects of particulate air pollution. Annual review of public health, 1994. 15(1): p. 107-132.
22. Dockery, D.W., et al., An association between air pollution and mortality in six US cities. New England journal of medicine, 1993. 329(24): p. 1753-1759.
23. McConnell, R., et al., Air pollution and bronchitic symptoms in Southern California children with asthma. Environmental health perspectives, 1999. 107(9): p. 757.
35
24. Šrám, R.J., et al., Adverse reproductive outcomes from exposure to environmental mutagens. Mutation Research/Fundamental and Molecular Mechanisms of Mutagenesis, 1999. 428(1): p. 203-215.
25. Stayner, L., et al., Predicted lung cancer risk among miners exposed to diesel exhaust particles. American journal of industrial medicine, 1998. 34(3): p. 207-219.
26. Hoek, G., et al., Association between mortality and indicators of traffic-related air pollution in the Netherlands: a cohort study. The lancet, 2002. 360(9341): p. 1203-1209.
27. Gauderman, W.J., et al., Effect of exposure to traffic on lung development from 10 to 18 years of age: a cohort study. The Lancet, 2007. 369(9561): p. 571-577.
28. Peters, A., et al., Translocation and potential neurological effects of fine and ultrafine particles a critical update. Part Fibre Toxicol, 2006. 3(13): p. 1-13.
29. Knibbs, L.D., T. Cole-Hunter, and L. Morawska, A review of commuter exposure to ultrafine particles and its health effects. Atmospheric Environment, 2011. 45(16): p. 2611-2622.
30. Seaton, A., et al., Particulate air pollution and acute health effects. The Lancet, 1995. 345(8943): p. 176-178.
31. Li, N., et al., Ultrafine particulate pollutants induce oxidative stress and mitochondrial damage. Environmental health perspectives, 2003. 111(4): p. 455.
32. Delfino, R.J., C. Sioutas, and S. Malik, Potential role of ultrafine particles in associations between airborne particle mass and cardiovascular health. Environmental health perspectives, 2005. 113(8): p. 934.
33. Oberdörster, G., et al., Translocation of inhaled ultrafine particles to the brain. Inhalation toxicology, 2004. 16(6-7): p. 437-445.
34. WHO, Health risks of particulate matter from long-range transboundary air pollution. Rep., WHO reg. Off. Eur, Copenhagen, Denmark, 2006.
35. Yang, H., et al., Comparative study of cytotoxicity, oxidative stress and genotoxicity induced by four typical nanomaterials: the role of particle size, shape and composition. Journal of applied Toxicology, 2009. 29(1): p. 69-78.
36. Schlesinger, R., et al., The health relevance of ambient particulate matter characteristics: coherence of toxicological and epidemiological inferences. Inhalation toxicology, 2006. 18(2): p. 95-125.
37. Hahn, F.F., et al., Particle size and composition related to adverse health effects in aged, sensitive rats. Research report (Health Effects Institute), 2005(129): p. 1.
38. Valavanidis, A., K. Fiotakis, and T. Vlachogianni, Airborne particulate matter and human health: toxicological assessment and importance of size and composition of particles for oxidative damage and carcinogenic mechanisms. Journal of Environmental Science and Health, Part C, 2008. 26(4): p. 339-362.
39. Baron, P.A., P. Kulkarni, and K. Willeke, Aerosol measurement: principles, techniques, and applications. 2011, Hoboken, N.J: Wiley.
40. Stanier, C.O., A.Y. Khlystov, and S.N. Pandis, Ambient aerosol size distributions and number concentrations measured during the Pittsburgh Air Quality Study (PAQS). Atmospheric Environment, 2004. 38(20): p. 3275-3284.
41. Woo, K.-S., et al., Use of continuous measurements of integral aerosol parameters to estimate particle surface area. Aerosol Science & Technology, 2001. 34(1): p. 57-65.
42. Fine, P.M., S. Shen, and C. Sioutas, Inferring the Sources of Fine and Ultrafine Particulate Matter at Downwind Receptor Sites in the Los Angeles Basin Using Multiple Continuous Measurements Special Issue of Aerosol Science and Technology on Findings from the Fine Particulate Matter Supersites Program. Aerosol Science and Technology, 2004. 38(S1): p. 182-195.
43. Kulmala, M., et al., Formation and growth rates of ultrafine atmospheric particles: a review of observations. Journal of Aerosol Science, 2004. 35(2): p. 143-176.
36
44. Wåhlin, P., et al., Pronounced decrease of ambient particle number emissions from diesel traffic in Denmark after reduction of the sulphur content in diesel fuel. Atmospheric Environment, 2001. 35(21): p. 3549-3552.
45. Shi, J.P., et al., Sources and concentration of nanoparticles (<10 nm diameter) in the urban atmosphere. Atmospheric Environment, 2001. 35(7): p. 1193-1202.
46. Harrison, R.M., M. Jones, and G. Collins, Measurements of the physical properties of particles in the urban atmosphere. Atmospheric Environment, 1999. 33(2): p. 309-321.
47. Ristovski, Z.D., et al., Influence of Diesel Fuel Sulfur on Nanoparticle Emissions from City Buses. Environmental Science & Technology, 2006. 40(4): p. 1314-1320.
48. Harris, S.J. and M.M. Maricq, Signature size distributions for diesel and gasoline engine exhaust particulate matter. Journal of Aerosol Science, 2001. 32(6): p. 749-764.
49. WHO, W.H.O., Air Quality Guidelines: Global Update 2005: Particulate Matter, Ozone, Nitrogen Dioxide, and Sulfur Dioxide. 2005: World Health Organization Europe.
50. Wichmann, H.E. and A. Peters, Epidemiological Evidence of the Effects of Ultrafine Particle Exposure. Philosophical Transactions: Mathematical, Physical and Engineering Sciences, 2000. 358(1775): p. 2751-2769.
51. Oberdörster, G., Pulmonary effects of inhaled ultrafine particles. International Archives of Occupational and Environmental Health, 2000. 74(1): p. 1-8.
52. Ning, L., et al., Ultrafine Particulate Pollutants Induce Oxidative Stress and Mitochonrial Damage. Environmental Health Perspectives, 2003. 111(4): p. 455.
53. http://www.dieselnet.com/standards/eu/ld.php. 54. TSI, Model 3781 Water-based Condensation Particle Counter, Operation and Service Manual.
2006. P/N 1930111, Revision A, April 2006. 55. Morawska, L., et al., JEM Spotlight: Environmental monitoring of airborne nanoparticles. J.
Environ. Monit., 2009. 11(10): p. 1758-1773. 56. Kulkarni, P., P.A. Baron, and K. Willeke, Aerosol measurement: principles, techniques, and
applications. 2011: Wiley. 57. Saxena, P. and L.M. Hildemann, Water-soluble organics in atmospheric particles: A critical
review of the literature and application of thermodynamics to identify candidate compounds. Journal of atmospheric chemistry, 1996. 24(1): p. 57-109.
58. Chow, J.C., Measurement methods to determine compliance with ambient air quality standards for suspended particles. Journal of the Air & Waste Management Association, 1995. 45(5): p. 320-382.
59. Spurný, K., Analytical chemistry of aerosols. 1999: CRC Press. 60. McMurry, P.H., A review of atmospheric aerosol measurements. Atmospheric Environment,
2000. 34(12): p. 1959-1999. 61. Berner, A., et al., The size distribution of the urban aerosol in Vienna. Science of the Total
Environment, 1979. 13(3): p. 245-261. 62. Turpin, B.J., J.J. Huntzicker, and S.V. Hering, Investigation of organic aerosol sampling
artifacts in the Los Angeles Basin. Atmospheric Environment, 1994. 28(19): p. 3061-3071. 63. Katz, M. and C. Chan, Comparative distribution of eight polycyclic aromatic hydrocarbons in
airborne particulates collected by conventional high-volume sampling and by size fractionation. Environmental Science & Technology, 1980. 14(7): p. 838-843.
64. Wang, H.-C. and W. John, Characteristics of the Berner impactor for sampling inorganic ions. Aerosol Science and Technology, 1988. 8(2): p. 157-172.
65. Possanzini, M., A. Febo, and A. Liberti, New design of a high-performance denuder for the sampling of atmospheric pollutants. Atmospheric Environment (1967), 1983. 17(12): p. 2605-2610.
66. Eatough, D., et al., Determination of the size distribution and chemical composition of fine particulate semivolatile organic material in urban environments using diffusion denuder technology. Inhalation Toxicology, 1995. 7(5): p. 691-710.
37
67. Lane, D.A. and L. Gundel, Gas and particle sampling of airborne polycyclic aromatic compounds. Polycyclic Aromatic Compounds, 1996. 9(1-4): p. 67-73.
68. Hansen, A., H. Rosen, and T. Novakov, The aethalometer—an instrument for the real-time measurement of optical absorption by aerosol particles. Science of the Total Environment, 1984. 36: p. 191-196.
69. Kittelson, D., et al., Total sulfur aerosol concentration with an electrostatically pulsed flame photometric detector system. Atmospheric Environment (1967), 1978. 12(1): p. 105-111.
70. Weber, R., et al., A particle-into-liquid collector for rapid measurement of aerosol bulk chemical composition. Aerosol Science & Technology, 2001. 35(3): p. 718-727.
71. Orsini, D.A., et al., Refinements to the particle-into-liquid sampler (PILS) for ground and airborne measurements of water soluble aerosol composition. Atmospheric Environment, 2003. 37(9): p. 1243-1259.
72. Suess, D.T. and K.A. Prather, Mass spectrometry of aerosols. Chemical Reviews, 1999. 99(10): p. 3007-3036.
73. Johnston, M.V., Sampling and analysis of individual particles by aerosol mass spectrometry. Journal of Mass Spectrometry, 2000. 35(5): p. 585-595.
74. Noble, C.A. and K.A. Prather, Real‐time single particle mass spectrometry: A historical review of a quarter century of the chemical analysis of aerosols. Mass Spectrometry Reviews, 2000. 19(4): p. 248-274.
75. Sipin, M.F., S.A. Guazzotti, and K.A. Prather, Recent advances and some remaining challenges in analytical chemistry of the atmosphere. Analytical Chemistry, 2003. 75(12): p. 2929-2940.
76. Coe, H. and J.D. Allan, Mass spectrometric methods for aerosol composition measurements. analytical techniques for atmospheric measurement, 2006: p. 265.
77. Liu, P., et al., Generating particle beams of controlled dimensions and divergence: II. Experimental evaluation of particle motion in aerodynamic lenses and nozzle expansions. Aerosol Science and Technology, 1995. 22(3): p. 314-324.
78. Liu, P., et al., Generating particle beams of controlled dimensions and divergence: I. Theory of particle motion in aerodynamic lenses and nozzle expansions. Aerosol Science and Technology, 1995. 22(3): p. 293-313.
79. Zhang, X., et al., Numerical characterization of particle beam collimation: Part II integrated aerodynamic-lens–nozzle system. Aerosol Science and Technology, 2004. 38(6): p. 619-638.
80. Canagaratna, M., et al., Chemical and microphysical characterization of ambient aerosols with the aerodyne aerosol mass spectrometer. Mass Spectrometry Reviews, 2007. 26(2): p. 185-222.
81. Drewnick, F., et al., A new time-of-flight aerosol mass spectrometer (TOF-AMS)—Instrument description and first field deployment. Aerosol Science and Technology, 2005. 39(7): p. 637-658.
82. NARSTO, Particulate Matter Science for Policy Makers. Electric Power Research Institute, Palo Alta, CA, 2003.
83. Wongphatarakul, V., S.K. Friedlander, and J.P. Pinto, A Comparative Study of PM2.5 Ambient Aerosol Chemical Databases. Environmental Science & Technology, 1998. 32(24): p. 3926-3934.
84. Adgate, J.L., et al., Spatial and temporal variability in outdoor, indoor, and personal PM2.5 exposure. Atmospheric Environment, 2002. 36(20): p. 3255-3265.
85. Morawska, L., et al., Spatial variation of airborne pollutant concentrations in Brisbane, Australia and its potential impact on population exposure assessment. Atmospheric Environment, 2002. 36(21): p. 3545-3555.
86. DeGaetano, A.T. and O.M. Doherty, Temporal, spatial and meteorological variations in hourly PM2.5 concentration extremes in New York City. Atmospheric Environment, 2004. 38(11): p. 1547-1558.
38
87. Martuzevicius, D., et al., Spatial and temporal variations of PM2.5 concentration and composition throughout an urban area with high freeway density--the Greater Cincinnati study. Atmospheric Environment, 2004. 38(8): p. 1091-1105.
88. Kim, E., et al., Spatial Variability of Fine Particle Mass, Components, and Source Contributions during the Regional Air Pollution Study in St. Louis. Environmental Science & Technology, 2005. 39(11): p. 4172-4179.
89. Wilson, J.G., et al., A review of intraurban variations in particulate air pollution: Implications for epidemiological research. Atmospheric Environment, 2005. 39(34): p. 6444-6462.
90. Johansson, C., M. Norman, and L. Gidhagen, Spatial & temporal variations of PM10 and particle number concentrations in urban air. Environmental Monitoring and Assessment, 2007. 127(1): p. 477-487.
91. Querol, X., et al., Spatial and temporal variations in airborne particulate matter (PM10 and PM2.5) across Spain 1999-2005. Atmospheric Environment, 2008. 42(17): p. 3964-3979.
92. Patel, M.M., et al., Spatial and temporal variations in traffic-related particulate matter at New York City high schools. Atmospheric Environment, 2009. 43(32): p. 4975-4981.
93. Sarnat, J.A., et al., Assessing the spatial and temporal variability of fine particulate matter components in Israeli, Jordanian, and Palestinian cities. Atmospheric Environment, 2010. 44(20): p. 2383-2392.
94. Buzorius, G., et al., Spatial variation of aerosol number concentration in Helsinki city. Atmospheric Environment, 1999. 33(4): p. 553-565.
95. Zhu, Y., et al., Concentration and size distribution of ultrafine particles near a major highway. Journal of the Air & Waste Management Association (1995), 2002. 52(9): p. 1032.
96. Zhu, Y., et al., Study of ultrafine particles near a major highway with heavy-duty diesel traffic. Atmospheric Environment, 2002. 36(27): p. 4323-4335.
97. Holmes, N.S., et al., Spatial distribution of submicrometre particles and CO in an urban microscale environment. Atmospheric Environment, 2005. 39(22): p. 3977-3988.
98. Puustinen, A., et al., Spatial variation of particle number and mass over four European cities. Atmospheric Environment, 2007. 41(31): p. 6622-6636.
99. Cyrys, J., et al., Spatial and temporal variation of particle number concentration in Augsburg, Germany. Science of The Total Environment, 2008. 401(1-3): p. 168-175.
100. Moore, K., et al., Intra-Community Variability in Total Particle Number Concentrations in the San Pedro Harbor Area (Los Angeles, California). Aerosol Science and Technology, 2009. 43(6): p. 587-603.
101. Morawska, L., et al., Variation in indoor particle number and PM2.5 concentrations in a radio station surrounded by busy roads before and after an upgrade of the HVAC system. Building and Environment, 2009. 44(1): p. 76-84.
102. Hudda, N., et al., Inter-community variability in total particle number concentrations in the eastern Los Angeles air basin. Atmospheric Chemistry and Physics, 2010. 10(23): p. 11385-11399.
103. Dall'Osto, M., et al., On the spatial distribution and evolution of ultrafine particles in Barcelona. Atmospheric Chemistry and Physics, 2013. 13(2): p. 741-759.
104. Gomišček, B., et al., Spatial and temporal variations of PM1, PM2.5, PM10 and particle number concentration during the AUPHEP—project. Atmospheric Environment, 2004. 38(24): p. 3917-3934.
105. Kornartit, C., et al., Activity pattern and personal exposure to nitrogen dioxide in indoor and outdoor microenvironments. Environment International, 2010. 36(1): p. 36-45.
106. Burtscher, H. and K. Schüepp, The occurrence of ultrafine particles in the specific environment of children. Paediatric Respiratory Reviews, 2012. 13(2): p. 89-94.
107. Hochstetler, H.A., et al., Aerosol particles generated by diesel-powered school buses at urban schools as a source of children’s exposure. Atmospheric Environment, 2011. 45(7): p. 1444-1453.
39
108. Zhang, Q. and Y. Zhu, Characterizing ultrafine particles and other air pollutants at five schools in South Texas. Indoor Air, 2012. 22(1): p. 33-42.
109. Diapouli, E., A. Chaloulakou, and N. Spyrellis, Levels of ultrafine particles in different microenvironments — Implications to children exposure. Science of The Total Environment, 2007. 388(1–3): p. 128-136.
110. TSI, Model 3936 Scannin Moblity Particle Sizer (SMPS) Spectrometer. 2005. 111. Hussein, T., et al., Evaluation of an automatic algorithm for fitting the particle number size
distributions. Boreal Environment Research, 2005. 10(5): p. 337-355. 112. Whitby, K.T., The physical characteristics of sulfur aerosols. Atmospheric Environment
(1967), 1978. 12(1-3): p. 135-159. 113. Hussein, T., et al., Urban aerosol number size distributions. Atmospheric Chemistry and
Physics, 2004. 4(2): p. 391-411. 114. Bae, M.S., et al., Characteristics of size distributions at urban and rural locations in New York.
Atmospheric Chemistry and Physics, 2010. 10(10): p. 4521-4535. 115. Beddows, D.C.S., M. Dall’Osto, and R.M. Harrison, Cluster Analysis of Rural, Urban, and
Curbside Atmospheric Particle Size Data. Environmental Science & Technology, 2009. 43(13): p. 4694-4700.
116. Dunn, J.C., Well-separated clusters and optimal fuzzy partitions. Journal of cybernetics, 1974. 4(1): p. 95-104.
117. Rousseeuw, P.J., Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 1987. 20: p. 53-65.
118. Hartigan, J.A. and M.A. Wong, A k-means clustering algorithm. Applied statistics, 1979: p. 100-108.
119. Wegner, T., et al., Properties of aerosol signature size distributions in the urban environment as derived by cluster analysis. Atmospheric Environment, 2012. 61(0): p. 350-360.
120. Dall'Osto, M., et al., Urban aerosol size distributions over the Mediterranean city of Barcelona, NE Spain. Atmos. Chem. Phys. Discuss, 2012. 12: p. 16457-16492.
121. Sabaliauskas, K., et al., Cluster analysis of roadside ultrafine particle size distributions. Atmospheric Environment, 2012.
122. Kaufman, L. and P.J. Rousseeuw, Finding Groups in Data: An Introduction to Cluster Analysis. 2009, New York: John Wiley & Sons, Inc.
123. Kohonen, T., Self-organizing maps. Vol. 30. 2001, New York: Springer. 124. Frey, B.J. and D. Dueck, Clustering by Passing Messages Between Data Points. Science, 2007.
315(5814): p. 972-976. 125. R Development Core Team, R: A language and environment for statistical computing. 2010,
R foundation for Statistical Computing: Vienna, Austria. 126. Kulmala, M., How Particles Nucleate and Grow. Science, 2003. 302(5647): p. 1000-1001. 127. Kulmala, M. and V.-M. Kerminen, On the formation and growth of atmospheric
nanoparticles. Atmospheric Research, 2008. 90(2–4): p. 132-150. 128. Ziereis, H. and F. Arnold, Gaseous ammonia and ammonium ions in the free troposphere.
Nature, 1986. 321(6069): p. 503-505. 129. Coffman, D.J. and D.A. Hegg, A preliminary study of the effect of ammonia on particle
nucleation in the marine boundary layer. Journal of geophysical research, 1995. 100(D4): p. 7147-7160.
130. Ball, S., et al., Laboratory studies of particle nucleation: Initial results for H2SO4, H2O, and NH3 vapors. Journal of Geophysical Research: Atmospheres (1984–2012), 1999. 104(D19): p. 23709-23718.
131. Lovejoy, E., J. Curtius, and K. Froyd, Atmospheric ion-induced nucleation of sulfuric acid and water. Journal of Geophysical Research: Atmospheres (1984–2012), 2004. 109(D8).
40
132. Sorokin, A. and F. Arnold, Laboratory study of cluster ions formation in H2SO4-H2O system: Implications for threshold concentration of gaseous H2SO4 and ion-induced nucleation kinetics. Atmospheric environment, 2007. 41(18): p. 3740-3747.
133. Kirkby, J., et al., Role of sulphuric acid, ammonia and galactic cosmic rays in atmospheric aerosol nucleation. Nature, 2011. 476(7361): p. 429-433.
134. Kulmala, M., et al., Toward direct measurement of atmospheric nucleation. Science (New York, N.Y.), 2007. 318(5847): p. 89-92.
135. Lee, S.-H., et al., Particle formation by ion nucleation in the upper troposphere and lower stratosphere. Science, 2003. 301(5641): p. 1886-1889.
136. O'Dowd, C.D., et al., Aerosol formation: Atmospheric particles from organic vapours. Nature, 2002. 416(6880): p. 497-498.
137. Wang, L., et al., Atmospheric nanoparticles formed from heterogeneous reactions of organics. Nature Geoscience, 2010. 3(4): p. 238-242.
138. Creamean, J.M., et al., Measurements of Aerosol Chemistry during New Particle Formation Events at a Remote Rural Mountain Site. Environmental Science & Technology, 2011. 45(19): p. 8208-8216.
139. Laakso, L., et al., Kinetic nucleation and ions in boreal forest particle formation events. Atmospheric Chemistry and Physics, 2004. 4(9/10): p. 2353-2366.
140. Kulmala, M., L. Pirjola, and J.M. Makela, Stable sulphate clusters as a source of new atmospheric particles. Nature, 2000. 404(6773): p. 66-69.
141. Kulmala, M., et al., On the existence of neutral atmospheric clusters. Boreal Environ. Res, 2005. 10(2): p. 79-87.
142. Kulmala, M., K. Lehtinen, and A. Laaksonen, Cluster activation theory as an explanation of the linear dependence between formation rate of 3nm particles and sulphuric acid concentration. Atmospheric Chemistry and Physics, 2006. 6(3): p. 787-793.
143. Kulmala, M., et al., Direct observations of atmospheric aerosol nucleation. Science, 2013. 339(6122): p. 943-946.
144. Stanier, C.O., A.Y. Khlystov, and S.N. Pandis, Nucleation Events During the Pittsburgh Air Quality Study: Description and Relation to Key Meteorological, Gas Phase, and Aerosol Parameters Special Issue of Aerosol Science and Technology on Findings from the Fine Particulate Matter Supersites Program. Aerosol Science and Technology, 2004. 38(S1): p. 253-264.
145. Cheung, H.C., Z.D. Ristovski, and L. Morawska, Observation of new particle formation in subtropical urban environment. Atmospheric Chemistry and Physics Discussions, 2010. 10(10): p. 22623-22652.
146. Dal Maso, M., et al., Formation and growth of fresh atmospheric aerosols: eight years of aerosol size distribution data from SMEAR II, Hyytiälä, Finland. Boreal Environment Research, 2005. 10(5): p. 323-336.
147. Lehtinen, K.E.J., et al., On the concept of condensation sink diameter. Boreal Environment Research, 2003. 8(4): p. 405-411.
148. Pirjola, L., et al., FORMATION OF SULPHURIC ACID AEROSOLS AND CLOUD CONDENSATION NUCLEI: AN EXPRESSION FOR SIGNIFICANT NUCLEATION AND MODEL COMPRARISON. Journal of Aerosol Science, 1999. 30(8): p. 1079-1094.
149. Fuchs, N.A. and A.G. Sutugin, HIGH-DISPERSED AEROSOLS, in Topics in Current Aerosol Research, G.M. Hidy and J.R. Brock, Editors. 1971, Pergamon. p. 1.
150. Clement, C.F., M. Kulmala, and T. Vesala, Theoretical consideration on sticking probabilities. Journal of Aerosol Science, 1996. 27(6): p. 869-882.
151. Salma, I., et al., Production, growth and properties of ultrafine atmospheric aerosol particles in an urban environment. Atmospheric Chemistry and Physics Discussions, 2010. 10(6): p. 13689-13721.
41
152. Hamed, A., et al., The role of relative humidity in continental new particle formation. Journal of Geophysical Research: Atmospheres, 2011. 116(D3): p. D03202.
153. Zhang, Q., et al., Insights into the Chemistry of New Particle Formation and Growth Events in Pittsburgh Based on Aerosol Mass Spectrometry. Environmental Science & Technology, 2004. 38(18): p. 4797-4809.
154. Bzdek, B.R., et al., Nanoparticle Chemical Composition During New Particle Formation. Aerosol Science and Technology, 2011. 45(8): p. 1041-1048.
42
Chapter 3: Spatial variation of Particle Number
Concentration in School Microscale Environments and its Impact
on Exposure Assessment
Farhad Salimi, Mandana Mazaheri, Sam Clifford, Leigh R. Crilley, Rusdin
Laiman, and Lidia Morawska
International Laboratory for Air Quality and Health, Queensland University of
Technology, GPO Box 2434, Brisbane QLD, 4001, Australia
Farhad Salimi, Mandana Mazaheri, Sam Clifford, Leigh R. Crilley, Rusdin Laiman, and Lidia
Morawska. “Spatial Variation of Particle Number Concentration in School Microscale
Environments and Its Impact on Exposure Assessment”, Environmental Science &
Technology 2013 47 (10), 5251-5258.
43
Statement of Joint Authorship
Title: Spatial Variation of Particle Number Concentration in School Microscale
Environments and Its Impact on Exposure Assessment
Authors: Farhad Salimi, Mandana Mazaheri, Sam Clifford, Leigh R. Crilley, Rusdin
Laiman, and Lidia Morawska.
Farhad Salimi (candidate): Contributed to the measurements, performed data analysis and
wrote the manuscript.
Mandana Mazaheri: Supervised the project and reviewed the manuscript.
Sam Clifford: Assisted with data analysis, wrote part of the manuscript and reviewed it.
Leigh R. Crilley: Contributed to the measurements.
Rusdin laiman: Contributed to the measurements.
Lidia Morawska: Supervised the project, assisted with data interpretation and reviewed the
manuscript.
44
Abstract
It has not yet been established whether the spatial variation of particle number
concentration (PNC) within a microscale environment can have an effect on exposure
estimation results. In general, the degree of spatial variation within microscale
environments remains unclear, since previous studies have only focused on spatial variation
within macroscale environments. The aims of this study were to determine the spatial
variation of PNC within microscale school environments, in order to assess the importance
of the number of monitoring sites on exposure estimation. Furthermore, this paper aims to
identify which parameters have the largest influence on spatial variation, as well as the
relationship between those parameters and spatial variation. Air quality measurements
were conducted for two consecutive weeks at each of the 25 schools across Brisbane,
Australia. PNC was measured at three sites within the grounds of each school, along with
the measurement of meteorological and several other air quality parameters. Traffic density
was recorded for the busiest road adjacent to the school. Spatial variation at each school
was quantified using coefficient of variation (CV). The portion of CV associated with
instrument uncertainty was found to be 0.3 and therefore, CV was corrected so that only
non-instrument uncertainty was analysed in the data. The median corrected CV (CVc) ranged
from 0 to 0.35 across the schools, with 12 schools found to exhibit spatial variation. The
study determined the number of required monitoring sites at schools with spatial variability
and tested the deviation in exposure estimation arising from using only a single site. Nine
schools required two measurement sites and three schools required three sites. Overall, the
deviation in exposure estimation from using only one monitoring site was as much as one
order of magnitude. The study also tested the association of spatial variation with wind
speed/direction and traffic density, using partial correlation coefficients to identify sources
of variation and non-parametric function estimation to quantify the level of variability.
Traffic density and road to school wind direction were found to have a positive effect on
CVc, and therefore, also on spatial variation. Wind speed was found to have a decreasing
effect on spatial variation when it exceeded a threshold of 1.5 (m/s), while it had no effect
below this threshold. Traffic density had a positive effect on spatial variation and its effect
increased until it reached a density of 70 vehicles per five minutes, at which point its effect
plateaued and did not increase further as a result of increasing traffic density.
45
TOC Art
Keywords
Spatial variation, microscale environment, particle number concentration, exposure, schools.
3.1. Introduction
There is significant toxicological evidence indicating the adverse human health
effects of ultrafine particles (UFPs, particles with diameter <100nm) [1-4]. UFPs are the main
contributor to particle number concentration (PNC) in most typical urban environments [5].
Therefore, the study of exposure to PNC has received increasing attention in recent years,
due to its significant impact on human health. There is a considerable amount of literature
on exposure to PNC in microscale environments and most of these studies used a single
monitoring site. However, it is not yet known whether spatial variation of PNC can cause
biases on exposure estimation, and with this in mind, we undertook this study to determine
the spatial variation of PNC in microscale environments and investigate its effect on human
exposure estimation.
Exposure to PNC in microscale environments, particularly schools, has been widely
studied. Schools have been the focus of much attention because children are the group
most vulnerable to particulate exposure and they spend significant amount of their time at
46
school [6]. Children are exposed to vehicle emitted UFPs from the roads adjacent to their
schools, as well as vehicles coming to pick up/drop off children (including school buses).
Hochstetler et al [7] measured PNC at one indoor and one outdoor site in each of four urban
elementary schools, in order to assess the children’s level of exposure to diesel exhaust
particles generated by school buses, and to determine if there was an association between
indoor and outdoor aerosol characteristics. Zhang et al [8] measured PNC, as well as some
other air pollutants, simultaneously, inside and outside five schools in both urban and rural
areas, in order to find the main factors effecting indoor UFP concentration. In another study,
the indoor and outdoor PNC for 7 primary schools, as well as PNC inside vehicles, were
measured using a single outdoor site to assess the level of exposure of children to PNC [9].
The aforementioned studies all used a single outdoor site to measure PNC, in order
to determine exposure and/or indoor to outdoor concentration ratios. However, the
presence of spatial variation in PNC within any of the studied environments has the
potential to affect exposure calculations. In other words, it is very likely that the location of
the outdoor site can influence outdoor to indoor analysis, as well as the results of exposure
calculations, and consequently the conclusions of the study. To the best of our knowledge,
there are no studies in the literature regarding how the level of spatial variability of PNC can
impact on human exposure estimation.
In order to investigate this, we firstly need to quantify the spatial variation of PNC
within a microscale environment. Many studies have been published on the spatial variation
of PNC, however most of them focused on macroscale environments, with site separations
of up to several kilometres apart [10-13]. A small number of studies have also assessed the
spatial variation of PNC in smaller scale environments, however these focused on
environments with high PNC (e.g. near highways) [14-17], while humans tend to spend most
of their outdoor time in microscale environments with a relatively lower PNC [18]. In
particular, there is no available data related to the spatial variability of PNC within the
microscale school environment. In addition, none of the aforementioned studies considered
variation due to instrumental uncertainty, which can potentially cause bias in the results
and may result in the authors drawing false conclusions.
47
The main aims of the present study were to: (1) quantify the spatial variation of PNC,
(2) determine the level of variation due to instrumental uncertainty, (3) identify and
quantify main parameters affecting spatial variation of PNC, (4) determine the adequate
number of measurement sites required for PNC monitoring at each school, and (5) establish
the level of deviation in exposure estimation resulting from the location of the monitoring
site.
3.2. Material and methods
3.2.1. Background
This study is part of the 'Ultrafine Particles from Traffic Emissions and Children’s
Health' (UPTECH) project, which seeks to determine the effects of exposure to traffic related
UFPs on the health of primary school-aged children. Overall, 25 state primary schools, coded
S1 to S25, were randomly selected across the Brisbane Metropolitan Area, in Australia. Air
quality measurements were conducted for two consecutive weeks at each school, during
the period October 2010 to August 2012. Further details and information on the study
design can be found on the project website [19]. The average temperature variation
between the coldest and the hottest months in Brisbane’s subtropical environment is about
10 deg (http://www.bom.gov.au). Seasonality in PNC has not previously been observed in
this area [20] and therefore, the effect of seasonal variation on PNC was assumed to be
negligible in this work. The monitored schools were located between 1.5 km to 30 km from
Brisbane City and Table S1 summarises the characteristics of the 25 monitored schools and
their monitoring dates. Schematic diagrams of the schools, including the location of traffic
counters and measurement sites within the school grounds are illustrated in Table S2.
3.2.2. Instrumentation
PNC, down to 6nm in diameter, was measured using a TSI 3781 water-based
condensation particle counter (CPC) at each site. A cyclone, with a theoretical cut size of
3.3μm, was connected to each CPC. The CPC logged data at a frequency of 1 Hz and was set
to record with an averaging interval of 30 seconds. The TSI 3781 water-based CPC is a small,
light and easy to use unit, with low power consumption, which was found to be a good
choice for environmental monitoring, particularly in outdoor sites where batteries are the
main power supply. The CPCs at sites A and C were placed inside a Stevenson screen and
powered by batteries, while the CPC at site B housed inside a powered trailer, along with
48
several other instruments. A “Monitor Sensors” weather station was mounted on top of the
trailer, positioned about five metres above the ground. It measured meteorological
parameters at site B and data was recorded at 30 sec intervals using a specific data logger.
Traffic density was also measured by a “MetroCount 5600” automatic traffic counter, which
counted the number of passing vehicles, with a resolution time of five minutes, and
classified them into different categories.
3.2.3. Data analysis
3.2.3.1. Quantification of spatial variation of PNC
Coefficient of variation (CV) was chosen to characterise the spatial variation of PNC
and was calculated based on the measured PNC for the three sites at each school. The CV
for each time interval (CVi ) was defined as:
i
ABC
iAVG
STDCV i)(
where “iABCSTD )( ” and “ iAVG ” are the standard deviation and mean of the
measured PNCs for each time interval, i , respectively. The CV will be zero when all values
were exactly the same at all three sites and it will increase as the difference between the
values also increases. However, given that CPCs never read exactly the same concentrations
while they are sampling the same aerosol (due to instrumental uncertainties), the CV had to
be corrected to account for this uncertainty. In order to achieve this, the three CPCs were
set up to sample the same ambient air through a splitter, with the same sampling line
length, over a three day period. These data were then combined with the data collected
during the side by side quality assurance measurements and averaged according to five
minute intervals, prior to calculating the corresponding CVs (Figure 3S. 3a). More than 95% of
the calculated CVs were lower than 0.3 (Figure 3S. 3b) and therefore, a CV = 0.3 was defined
as the CPCs variation due to measurement uncertainty.
Variations lower than CV = 0.3 were considered within the range of variation due to
instrumental uncertainty, and were consequently deemed insignificant. No correlation
between the average PNC and CV was found (Figure 3S. 3c), and thus, CPC variation due to
measurement uncertainty was assumed to be independent of the value of PNC. Therefore, it
49
was concluded that a CV = 0.3 corresponds to the variation due to instrumental uncertainty,
regardless of the concentration value.
3.2.3.2. General analysis
Prior to analysis, the raw data for each school were imported into R [21]and
averaged according to five minute intervals. Following this, the CV was calculated for each
five minute period and further analysis was conducted for the sub-group which displayed
significant variation. All of the analyses described in this section were conducted using
functions implemented inside the “openair” package [22].
Diurnal variation of CV was plotted using the “TimeVariation” function, which reveals
information about the nature of the temporal variation (Figure 3S. 4). The CV value was then
plotted by wind direction using the “pollutionRose” function. This approach was found to be
an informative tool for air pollutant species analysis [23]. For instance, Figure 3S. 5 shows the
variation of CV by wind direction at S14, revealing an association between CV and the
Easterly wind direction. The “polarPlot” function was then used to illustrate the dependence
of CV on joint wind speed/direction dependence, using a bivariate polar plot of CV to be
varied by wind speed/direction. These plots are illustrated as a fitted continuous surface,
which is calculated using the Generalised Additive Model (Figure 3S. 6). Finally, linear
regression was done in order to model the relationship between CV and wind speed, PNC
and traffic density. In addition, CV was plotted against the aforementioned parameters and
a linear regression model was fitted using the least squares approach. The CV was also split
for different wind directions to assess the directional variability of the linear model (Figure
3S. 7).
3.2.3.3. Correlation between CV and traffic density/ meteorological parameters
In order to further investigate the association of CV with other parameters (i.e. wind
speed/direction and traffic density), the partial correlation coefficient method was used to
calculate correlations between CV and those selected parameters, by calculating the
correlation between a pair of variables, while all other variables remain constant [24]. Prior
to doing so, the corrected CV (CVc) was determined by deducting 0.3 from the calculatedCV.
Negative values of the corrected CV were then set to zero, which corresponded to no spatial
variation. In the end, Spearman’s partial correlation coefficient was employed to investigate
the association of CV with wind speed/direction and traffic density.
50
3.2.3.4. Non-parametric function estimation
While the partial correlation between variables gives an indication that there is a
relationship between them, it does not specify the nature of that relationship. Therefore,
the relationship between CV and traffic and meteorological parameters can be further
analysed using the Generalised Additive Model, with a basis of penalised B-splines [25]. This
approach allows for the flexible estimation of non-linear effects without assuming, a priori,
the functional form of the non-linearity. The resulting fitted smooth functions, and their
95% confidence intervals, indicate the contribution of each regression term to CV.
3.3. Results and discussion
3.3.1. Data availability
Erroneous measured concentration data were detected (as described in the
Supporting Information (SI)) and omitted when at least one of the three CPCs recorded an
erroneous value. Data availability varied from about 2.5 days to more than 11 days at each
school (Figure 3S. 8). Almost all of the missing data was due to the malfunction of the CPCs
(as described in SI), including the continued failure of one of the CPCs at schools S21 and
S22. However, adequate data was still available to address the aims of this project. About
150 days of data, over almost two years, were included in the analysis and data availability
at each school varied from 2.1 days at S21 to 10.2 days at S8, which were 16% and 78% of
what was expected, respectively. Based on this experience, these instruments were found
to be unreliable and their use is not recommended for long measurement campaigns.
3.3.2. Traffic density
Traffic density varied largely between schools, with a maximum of more than 250 (5
min-1) at S7 and S20 (Figure 3. 1a). Diurnal variation patterns were found to be very similar
between schools, with two distinct peaks during weekdays, at around 8am and 5pm,
corresponding to morning and afternoon rush hours (Figure 3. 1b). Weekend traffic tended
to follow a bell shaped pattern, peaking at around noon (Figure 3. 1c). This is in agreement
with the traffic pattern observed in previous studies conducted in the Brisbane
Metropolitan Area, as well as in Helsinki (Finland), Lecce (Italy) and Somerville (USA) [26-29].
3.3.3. Particle number concentration
Continuously measured PNC was averaged across the three sites within each school
(summarised in Figure 3. 2a). PNC varied across the schools, with the median value at each
51
school ranging from 3×103 to 1.4×104 cm-3, and maximum of 3.5×104 cm-3. The highest
values for PNC were observed at three schools, S3, S15 and S23, where the concentration
peaked at a level higher than or equal to 3×104 cm-3. In particular, S15 was located adjacent
to a road with high traffic density with around 180 (5 min-1) during rush hours. The median
PNC at this school was around 6×103 cm-3, which was 5.3 times lower than the peak value.
The wind heading at this site was mostly from the school to the road and high PNC was
associated with times when the wind blew from the road towards the school. Similarly, S23
was located adjacent to a road with traffic density of 150 (5 min-1) downwind of the
prevailing wind direction. The median PNC was about 1.3×104 cm-3, which was 2.3 times less
than the peak value. In contrast, S3 was located near a road with low traffic density and the
highest PNC was associated with a school event during which significant particles were
emitted due to the use of a fog machine. The median value observed at this school was
about 9×103 cm-3, which was 3.9 times less than the peak value.
The lowest median concentrations were observed at S1, S13, S14 and S16, all being
less than 5×103 cm-3. Among them, S1 had the lowest median concentration, at around
3×103 cm-3. This school was located in a very quiet residential area, 1.3 km from the ocean,
and less than 50 vehicles per 5 minutes passed on the adjacent road during rush hours. Both
S14 and S16 were located in similar areas, with a low traffic density. However, S13 was
located close to a road whose traffic density peaked at 140 (5 min-1) during rush hours, but
the median PNC was around 3×103 cm-3 and the highest concentration did not exceed 1×104
cm-3. The effect of the adjacent road on PNC at this school was very small, since during rush
hour, the prevailing wind direction was from the school towards the road. These findings
show that PNC was associated with traffic density and wind direction, which is in agreement
with previous findings [20, 30].
The average diurnal variation of PNC during weekdays and weekends was also
calculated and plotted (Figure 3. 2b,c). Weekday PNC was found to be the lowest during the
early morning hours followed by an increase to a peak around 7am, which was associated
with the morning traffic rush hour. A smaller peak, associated with afternoon rush hour,
was observed around 6pm, while the highest peak was observed at around noon and was
associated with new particle formation events [31]. The variation of PNC during the
weekends was different, in that no morning peak was observed and the afternoon peak was
52
delayed by two hours. However, the same late morning peak was observed, which could be
associated with a broad traffic density peak and/or new particle formation events during the
middle of the day. These observations are in line with several studies that have been
conducted around the world [7, 10, 32-34].
3.3.4. Spatial variation
Coefficient of variation: During weekdays, CVc was found to have the lowest value
in the early morning, which increased as a result of morning rush hour traffic (Figure 3. 3b). It
also increased during the evening, in association with the evening rush hour. However, CVc
was low during the the middle of the day, in contrast with the significant increase in PNC
described in Section 3.3. This can be explained by the fact that increased PNC during the
middle of the day was associated with new particle formation events, which occur over a
large scale area and do not tend to cause any spatial gradient in PNC. During the weekends,
CVc peaked at around 10am and 3pm, in association with the broad traffic peak observed
during those hours. However, in line with weekday observations, the CVc dropped around
the middle of the day, despite the new particle formation related increase in PNC at around
the same time.
CVc was calculated for both school hour (9am-3pm) and total 24 hour periods (Figure
3. 3a). Its median value ranged from 0 to about 0.35, with a maximum value exceeding 1.2.
Similar work by Kearney et al. [32], in houses in Ontario, Canada, found that the median
hourly CV ranged from 28 to 34% across 6 sites, with distances ranging from 0.2 to 18.8km,
which corresponds to a CVc of 0 to 0.04. CVc was also calculated according to hourly
intervals, in order to determine if using a larger time interval could affect the results,
however no significant differences were found.
In general, spatial variation of PNC was higher during school hours for all schools,
with the exception of S7, and significant spatial variation was observed in 12 schools,
indicating the need for using more than one site for exposure monitoring purposes. Further
analysis was conducted in order to establish the minimum number of monitoring sites
required.
Number of monitoring sites required: The significance threshold of CV, due to
the uncertainty between two CPCs, was calculated according to the same procedure
53
outlines in Section 2.3.1, with the amount of variation due to instrument uncertainty
between two CPCs found to be the same as that of three CPCs (i.e. CV = 0.3). This threshold
was then used to correct the calculated CV between each pair of sites (summarised in Figure
3. 4). This analysis showed that only one monitoring site was needed in the 13 schools in
which no spatial variability was observed. This is because data collected at all of the
measurement sites showed the same level of PNC within the measurements uncertainties
and therefore, adding an additional monitoring site would not result in any more
information. Two monitoring sites were adequate for nine of the schools, because spatial
variability was not observed between one pair of sites at these schools. Three of the schools
(i.e. S7, S12 and S14) required three sites, because spatial variability was observed between
all three pairs of sites.
Therefore, using only a single monitoring site at schools with significant spatial
variability in PNC will result in inaccurate exposure estimation. Ratios of mean concentration
at each site versus the site with the lowest mean concentration reveal the level of over or
underestimation in exposure estimation based on the location of the monitoring site (Figure
3. 5). Overall, the level of over or underestimation differed according to the spatial variation
of the microenvironment in question. For instance, at S2, the estimation of exposure
differed by as much as 7 times, depending on the location of the monitoring site. The
average difference in exposure estimation due to the location of the monitoring site was
found to be around 100%. This shows the significant effect of spatial variation of PNC within
a microscale environment on exposure assessment results, together with the importance of
using more than a single monitoring site for environments with significant spatial variability.
It was not possible to provide quantitative guidelines as to the conditions under
which schools would need one, two or three monitoring sites. However, schools which were
located downwind from a road with a traffic density of above 150 (5min-1) tended to need
more monitoring sites due to their higher spatial variability. Therefore, it is recommended
that spatial variation be assessed in microenvironments that display similar conditions, prior
to conducting measurements for exposure estimation.
Correlation between CVc and traffic density, wind speed/ direction, inter-
site distances: The correlation between CVc and traffic density and wind speed/direction
54
was tested using Spearman’s partial correlation coefficient for four averaging intervals.
Wind direction was coded to have a higher value when it was blowing from the direction of
the main adjacent road towards the school, in order to allow us to test the association
between wind direction and spatial variation. Overall, CVc was found to be positively
correlated with traffic density and wind direction, and larger correlation coefficients were
found to correspond to larger averaging intervals (Figure 3S. 9). The correlation coefficient
for wind speed and CVc was less than 0.05, and therefore, further analysis is needed in
order to determine their relationship. Mean inter-site distances across all schools ranged
between 50 to 180 metres, depending on the size of the school, and Spearman’s partial
correlation coefficients were calculated for CVc versus each mean inter-site distance. This
was done to test the relationship between inter-site distance and spatial variation, however
no significant relationship was found.
CVc was found to be positively correlated with traffic density and wind direction
using Spearman’s partial correlation coefficient and higher spatial variation was observed
while the wind was blowing from the direction of the adjacent road towards the school.
Non-parametric regression model: In order to understand the impact of traffic
density on spatial variation, a non-parametric regression model was fit to the data with a
time resolution of five minutes, according to the following equation:
)()( 1 ijtrjij trafficfCVLog
The regression model quantifies the coefficient of variation as the sum of a smooth
function of traffic density and a random effect mean, j1 , for each school. The random
effect mean is included as an assumption that each school has its own base level of spatial
variation. The model explained 35.1% of the variation in the data and the smooth function
of traffic density and random mean terms were significant with p < 0.001.
The smooth function of traffic density (Figure 3. 6a) increased super-linearly until
approximately 80 vehicles per five minutes, at which point it plateaued until approximately
150 vehicles per five minutes. Above 150 vehicles per five minutes, the 95% confidence
interval of the smooth function of traffic density became much wider, indicating a much
55
more variable effect of traffic density on the coefficient of variation due to the small
number of observations at such values (Figure 3. 6b).
The above regression model was refit with the five minute average wind speeds and
modelled as a smooth function, with a basis of thin plate regression splines, as follows:
)()()( 1 ijwsijtrjij windspeedftrafficfCVLog
This regression model explained 24.4% of the variation in the data set.
The smooth function of wind speed (Figure 3. 7a) stayed flat from 0 to 1.5 m/s and
decreases from 1.5 to 3 m/s. At values above 3 m/s, the mean effect of wind speed on the
CVc increased with wind speed, but the 95% confidence interval got progressively wider due
to the small number of observations in that range (Figure 3. 7c). Therefore, wind speed was
found to have no effect on the CVc until it reached about 1.5 m/s. In the range 1.5-3 (m/s),
the amount of spatial variation was found to decrease with increasing wind speed (as
measured by CVc). In other words, as wind speed increased beyond a certain threshold value
(1.5 m/s), the difference in PNC at each location decreased. Beyond 3 m/s, the confidence
interval became too wide to make any meaningful inferences. The smooth function of traffic
density (Figure 3. 7b) was much the same as in the model without wind.
In summary, it was found that wind speed had no effect on the spatial variation of
PNC until it reached about 1.5 m/s. Above this threshold, an increase in wind speed was
found to reduce spatial variation. Traffic density had a significant effect on spatial variation
and the amount of effect increased linearly as the density increased from 0 to 70 (5min-1).
The effect of traffic plateaued after exceeding a traffic density of 70 vehicles per five
minutes.
3.4. Acknowledgments
This work was supported by the Australian Research Council (ARC), Department of
Transport and Main Roads (DTMR) and Department of Education, Training and Employment
(DETE) through Linkage Grant LP0990134. Our particular thanks go to R. Fletcher (DTMR)
and B. Robertson (DETE) for their vision regarding the importance of this work. We would
also like to thank all members of the UPTECH project, including Prof G. Marks, Dr P.
56
Robinson, Prof K. Mengersen, Prof Z. Ristovski, Prof G. Ayoko, Dr C. He, Dr G. Johnson, Dr R.
Jayaratne, Dr S. Low Choy, Prof G. Williams, W. Ezz, M. Mokhtar, N. Mishra, L. Guo, Prof C.
Duchaine, Dr H. Salonen, Dr X. Ling, Dr J. Davies, Dr L. Leontjew Toms, F. Fuoco, Dr A. Cortes,
Dr B. Toelle, A. Quinones and P. Kidd for their contribution to this work. We thank former
ILAQH, QUT members Dr M. Falk, Dr F. Fatokun, Dr J. Mejia and Dr D. Keogh for their
contributions at the early stage of the project. Thanks also to Prof T. Salthammer from
Fraunhofer WKI in Germany for VOC analysis, R. Appleby and C. Labbe for their
administrative assistance.
3.5. Supporting information
Supporting information includes details of the locations of the measurement sites,
quality assurance and data processing procedure eight figures and two tables. This material
is available free of charge via the Internet at http://pubs.acs.org.
3.6. References
57
3.7. Figures:
Figure 3. 1: Traffic density box plot for each school (a), and diurnal traffic pattern during weekdays (b) and weekends (c). Solid line represents the mean value and dashed lines show the 95% confidence interval.
58
Figure 3. 2: Box plot of inter-site average PNC at each school (a), together with its weekday (b) and weekend (c) diurnal pattern. Solid line represents the mean value and dashed lines show the 95% confidence interval.
59
Figure 3. 3: Corrected CV box plot for each school (a) during school hours (shaded boxes) and 24hrs (white boxes). Weekday (b) and weekend (c) diurnal variations of corrected CV. Solid line represents the mean value and
dashed lines show the 95% confidence interval.
Figure 3. 4: Corrected CV between each pair of sites during school hours (9am-3pm).
60
Figure 3. 5: Inter-site average concentration ratio at schools which displayed spatial variability of PNC. The left, middle and right circles correspond to sites A, B, and C, respectively and the three sites for each school are connected via
a dashed line.
61
Figure 3. 6: Effects of (a) traffic density for the regression model (2), with their 95% CIs (dashed lines). The histogram below (b) shows the frequency of the values of traffic density, which explains the width of the 95% CI for the
fitted smooth function in (a).
62
Figure 3. 7: Partial effects of (a) wind speed and (b) traffic density for the regression model (3) with their 95% CIs (dashed lines). The histograms below show the frequency of the values of (c) wind speed and (d) traffic density,
which explain the width of the 95% CI for the fitted smooth functions in (a) and (b).
1. WHO, W.H.O., Air Quality Guidelines: Global Update 2005: Particulate Matter, Ozone, Nitrogen Dioxide, and Sulfur Dioxide. 2005: World Health Organization Europe.
2. Wichmann, H.E. and A. Peters, Epidemiological Evidence of the Effects of Ultrafine Particle Exposure. Philosophical Transactions: Mathematical, Physical and Engineering Sciences, 2000. 358(1775): p. 2751-2769.
3. Oberdörster, G., Pulmonary effects of inhaled ultrafine particles. International Archives of Occupational and Environmental Health, 2000. 74(1): p. 1-8.
4. Ning, L., et al., Ultrafine Particulate Pollutants Induce Oxidative Stress and Mitochonrial Damage. Environmental Health Perspectives, 2003. 111(4): p. 455.
5. Morawska, L., et al., JEM Spotlight: Environmental monitoring of airborne nanoparticles. J. Environ. Monit., 2009. 11(10): p. 1758-1773.
6. Burtscher, H. and K. Schüepp, The occurrence of ultrafine particles in the specific environment of children. Paediatric Respiratory Reviews, 2012. 13(2): p. 89-94.
7. Hochstetler, H.A., et al., Aerosol particles generated by diesel-powered school buses at urban schools as a source of children’s exposure. Atmospheric Environment, 2011. 45(7): p. 1444-1453.
8. Zhang, Q. and Y. Zhu, Characterizing ultrafine particles and other air pollutants at five schools in South Texas. Indoor Air, 2012. 22(1): p. 33-42.
63
9. Diapouli, E., A. Chaloulakou, and N. Spyrellis, Levels of ultrafine particles in different microenvironments — Implications to children exposure. Science of The Total Environment, 2007. 388(1–3): p. 128-136.
10. Buzorius, G., et al., Spatial variation of aerosol number concentration in Helsinki city. Atmospheric Environment, 1999. 33(4): p. 553-565.
11. Larsen, M.L., Spatial distributions of aerosol particles: Investigation of the Poisson assumption. Journal of Aerosol Science, 2007. 38(8): p. 807-822.
12. Cyrys, J., et al., Spatial and temporal variation of particle number concentration in Augsburg, Germany. Science of The Total Environment, 2008. 401(1-3): p. 168-175.
13. Hudda, N., et al., Inter-community variability in total particle number concentrations in the eastern Los Angeles air basin. Atmospheric Chemistry and Physics, 2010. 10(23): p. 11385-11399.
14. Zhu, Y., et al., Study of ultrafine particles near a major highway with heavy-duty diesel traffic. Atmospheric Environment, 2002. 36(27): p. 4323-4335.
15. Zhu, Y., et al., Concentration and size distribution of ultrafine particles near a major highway. Journal of the Air & Waste Management Association (1995), 2002. 52(9): p. 1032.
16. Hitchins, J., et al., Concentrations of submicrometre particles from vehicle emissions near a major road. Atmospheric Environment, 2000. 34(1): p. 51-59.
17. Durant, J.L., et al., Short-term variation in near-highway air pollutant gradients on a winter morning. Atmospheric Chemistry and Physics, 2010. 10(17): p. 8341-8352.
18. Kornartit, C., et al., Activity pattern and personal exposure to nitrogen dioxide in indoor and outdoor microenvironments. Environment International, 2010. 36(1): p. 36-45.
19. UPTECH. Available from: http://www.ilaqh.qut.edu.au/Misc/UPTECH%20Home.htm. 20. Mejía, J.F., et al., Trends in size classified particle number concentration in subtropical
Brisbane, Australia, based on a 5 year study. Atmospheric Environment, 2007. 41(5): p. 1064-1079.
21. R Development Core Team, R: A language and environment for statistical computing. 2010, R foundation for Statistical Computing: Vienna, Austria.
22. Carslaw, D.C. and K. Ropkins, openair — An R package for air quality data analysis. Environmental Modelling & Software, 2012. 27–28(0): p. 52-61.
23. Henry, R., et al., Source Region Identification Using Kernel Smoothing. Environmental Science & Technology, 2009. 43(11): p. 4090-4097.
24. Zar, J.H., Biostatistical analysis. 4th. Upper Saddle River, NJ: Prentice Hall, 1999. 1: p. 389-94. 25. Wood, S.N., Thin plate regression splines. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 2003. 65(1): p. 95-114. 26. Hussein, T., et al., Urban aerosol number size distributions. Atmospheric Chemistry and
Physics, 2004. 4(2): p. 391-411. 27. Contini, D., et al., Analysis of particles and carbon dioxide concentrations and fluxes in an
urban area: Correlation with traffic rate and local micrometeorology. Atmospheric Environment, 2012. 46(0): p. 25-35.
28. Padró-Martínez, L.T., et al., Mobile monitoring of particle number concentration and other traffic-related air pollutants in a near-highway neighborhood over the course of a year. Atmospheric Environment, 2012. 61(0): p. 253-264.
29. Morawska, L., et al., Differences in airborne particle and gaseous concentrations in urban air between weekdays and weekends. Atmospheric Environment, 2002. 36(27): p. 4375-4383.
30. Pérez, N., et al., Variability of Particle Number, Black Carbon, and PM10, PM2.5, and PM1 Levels and Speciation: Influence of Road Traffic Emissions on Urban Air Quality. Aerosol Science and Technology, 2010. 44(7): p. 487-499.
31. Morawska, L., et al., Ambient nano and ultrafine particles from motor vehicle emissions: Characteristics, ambient processing and implications on human exposure. Atmospheric Environment, 2008. 42(35): p. 8113-8138.
64
32. Kearney, J., et al., Residential indoor and outdoor ultrafine particles in Windsor, Ontario. Atmospheric Environment, 2011. 45(40): p. 7583-7593.
33. Moore, K., et al., Intra-Community Variability in Total Particle Number Concentrations in the San Pedro Harbor Area (Los Angeles, California). Aerosol Science and Technology, 2009. 43(6): p. 587-603.
34. Bhangar, S., et al., Ultrafine particle concentrations and exposures in seven residences in northern California. Indoor Air, 2011. 21(2): p. 132-144.
65
Chapter 3S: Supporting Information
Farhad Salimi, Mandana Mazaheri, Sam Clifford, Leigh R. Crilley, Rusdin
Laiman, and Lidia Morawska
International Laboratory for Air Quality and Health, Queensland University of
Technology, GPO Box 2434, Brisbane QLD, 4001, Australia
Number of Pages: 12
Number of Tables: 2
Number of Figures: 8
66
3S.1. Measurements Three outdoor sites, named A, B and C, were selected within the grounds of each
school, to cover the whole school area. The prevailing wind direction in the schools region
was determined by examining the previous 10 years of wind data from the Queensland
Bureau of Meteorology (BOM) and Department of Environment and Resource Management
(DERM) for the month of interest. Sites were chosen to be approximately collinear along the
prevailing wind direction and in accordance with the schools Health and Safety
requirements, in order to avoid disturbing school activities (Figure 3S. 1). Sites A and C were
located at opposite ends of the school and were usually close to the adjacent roads. Site B
was located in a central location within the school grounds, to represent the general air
quality as best as possible. An automatic traffic counter was placed on the busiest road
adjacent to the school, preferably upwind of the prevailing wind direction. Vehicles
travelling on the roads that had no traffic counter could have affected PNC at the school,
however their effects were much lower than those travelling on the road with the traffic
counter, as they were less busy and/or downwind of the prevailing wind direction.
Figure 3S. 1: Schematic diagram of measurement sites and traffic counter location. Arrow shows the direction of the prevailing wind direction (based on the previous years’ statistics).
3S.2. Quality assurance and data processing
3S.2.1. Particle number concentration:
CPC performance was checked three times a week during the monitoring period.
Total and aerosol flow rates were measured and recorded by a Sensidyne Gilibrator-2
flowmeter and a zero reading was taken through a particle filter to ensure that there was no
leakage of air into the CPC. The logger’s internal time reading was also recorded against the
local time to ensure identification of the corrected time stamps. In addition, side by side
comparisons were conducted by sampling five times from the same aerosol to make sure
67
that the CPC measurements were similar, within a tolerance of 20%. All observations,
including the times when an instrument fault was detected, were entered into a meta data
sheet after each performance check.
Erroneous measured data were identified by inspecting the PNC time series and
meta data, and any data collected during the malfunction of a CPC were omitted. Following
this, the PNC time series for each site were observed visually and any rapid and/or un-
characteristic changes in concentration were identified and the corresponding data were
selected. It was then compared with the measured concentrations at the other two sites.
Data were omitted if the same trend was not observed at the other sites and/or it was
coincident with a change in the instrumental condition. The quality assured data were
imported into the data base using Microsoft Access 2007. The presence of erroneous and
missing data was mainly due to the malfunction of the CPCs, which was usually due to a
design fault in the water handling system, which resulted in excessive water entering the
sampler tube and a subsequent reduction in aerosol flow rate. This, in turn, caused an
underestimation of particle concentration, making the data unusable. We attempted to
rectify the water handling fault by using an external dryer to stop the condensation of water
in the sampling tube. This improved the function of the instrument, however frequent
failure of the CPC pump also resulted in erroneous and missing data. These issues were
discusses extensively with the manufacturer and the production of this instrument was later
discontinued, possibly as a result of the aforementioned issues.
The TSI 3781 CPC uses a nominal aerosol flow rate (120 cm3/sec) to calculate
concentration [1]. However, the aerosol flow rate was lower than the nominal flow rate for
a considerable fraction of the measurement period, resulting in the underestimation of
measured PNC. Therefore, the concentration was corrected using the following formula:
3S.1) Corrected concentration = (Nominal flow rate/Measured flow rate)×Measured
concentration
This correction was applied when the measured flow rate was at least 80% of the
nominal flow rate, otherwise, the data were discarded.
68
3S.2.2. Meteorological parameters:
Vector and scalar averaging was carried out on wind direction and wind speed data,
respectively. Any wind speed below 0.5 ms-1 was assumed to be indicative of calm
conditions and was set to zero. At many schools, site B was located near a building, which
could potentially affect the quality of data collected by the weather station. The measured
wind data at each school were compared to the data collected from the nearest weather
monitoring station (operated either by the Queensland Department of Environment and
Resource Management of the Australian Bureau of Meteorology), which were located
between 1 and 6 km from the schools.
The data were imported into R [2] and compared using the “pollutionRose” function
in the “openair” package [3]. This function subtracts the reference wind speed/direction
value from the second data set and plots the difference. The nearest air monitoring station
was used as the reference site and comparisons were made with the data measured at the
schools (Figure S2). A difference of 0 degrees corresponds to no difference between the
wind headings.
As an example, Figure S2 shows a slight mean negative displacement from the north,
which means that the measured wind direction was lower at the reference station by only -
3.4 degrees. In addition, the mean wind speed was higher than the reference value by 2
m/sec. A non-zero difference indicated a possible error in the measured data. This
comparison was done for all of the schools and if the weather station location was
surrounded by buildings and a discrepancy of more than 40 degrees (when compared to the
nearest DERM/BOM station) was observed for more than 20% of the data, then the data of
that station was used instead of the data from the weather station located at the school.
69
Figure 3S. 2: Comparison between wind data measured at S01 and the nearest monitoring station to the school.
Figure 3S. 3: PNC measured by three CPCs placed side by side and measuring the same aerosol (a), together with a box plot of the corresponding calculated CV (b), and regression plot of CV vs. mean PNC for the side by side
measurements.
70
Figure 3S. 4: CV variation by time at S04.
Figure 3S. 5: CV by wind direction at S14.
71
Figure 3S. 6: Polarplot of CV at S20.
Figure 3S. 7: CV vs. traffic for different wind directions.
72
Figure 3S. 8: Available data (days) at each school.
73
Figure 3S. 9: Spearman’s correlation coefficient between CV and traffic/wind direction.
Table3S. 1 Characteristics of the monitored schools.
School
code
Monitoring period School location
S1 15/11/2010 –
29/11/2010
Close to ocean 1.3 km to E; elevation = 21 m
S2 18/10/2010 –
01/11/2010
Elevation = 16 m
S3 01/11/2010 –
15/11/2010
Close to river 500 m to E and 800 m to SW;
elevation = 22 m;
S4 28/02/2011 –
21/03/2011
Close to Airport about 5 km to NE; Close to
racecourse 300 m S; elevation = 13 m
74
S5 21/03/2011 –
11/04/2011
Elevation = 44 m
S6 16/05/2011 –
30/05/2011
Adjacent to a road with heavy traffic on E
S7 30/05/2011 –
14/06/2011
Elevation = 26 m,
S8 14/06/2011 –
27/06/2011
Elevation = 36 m,
S9 11/07/2011 –
25/07/2011
Close to swamp 300 m to E; Elevation = 36 m,
S10 25/07/2011 –
05/08/2011
Elevation = 50 m,
S11 23/08/2011 –
05/09/2011
Elevation = 51 m,
S12 08/08/2011 –
22/08/2011
Elevation = 46 m,
S13 03/10/2011 –
17/10/2011
Close to swamp 600 m to NE; Elevation = 12 m
S14 17/10/2011 –
31/10/2011
Elevation = 67 m
S15 31/10/2011 –
14/11/2011
Close to swamp 1 km to E; Elevation = 16 m
S16 14/11/2011 –
28/11/2011
Elevation = 38 m
S17 28/11/2011 –
12/12/2011
Elevation = 16 m
75
S18 19/03/2012 –
02/04/2012
Elevation = 41 m
S19 05/03/2012 –
19/03/2012
Elevation = 64 m
S20 16/04/2012 –
30/04/2012
Elevation = 53 m
S21 28/05/2012 –
12/06/2012
Elevation = 42 m
S22 12/06/2012 –
25/06/2012
Close to Airport about 5 km to E; Elevation = 8
m
S23 16/07/2012 –
30/07/2012
Close to city centre about 1 km to S; Elevation
= 39 m
S24 30/07/2012 –
13/08/2012
Elevation = 50 m
S25 13/08/2012 –
27/08/2012
Close to Airport about 4 km to SE; Close to
ocean about 4 km to NE; Elevation = 20 m
Table3S. 2: Schematic diagrams and location of monitoring sites at each school. The large dashed line
surrounds the school area and the solid line across the road shows the location of automatic traffic counters.
S1)
S2)
S3)
76
S4)
S5)
S6)
S7)
S8)
S9)
S10)
S11)
S12)
77
S13)
S14)
S15)
S16)
S17)
S18)
S19)
S20)
S21)
78
S22)
S23)
S24)
S25)
3S.3. References:
1. TSI, Model 3781 Water-based Condensation Particle Counter, Operation and Service
Manual. 2006. P/N 1930111, Revision A, April 2006.
2. R Development Core Team, R: A language and environment for statistical computing,
2010, R foundation for Statistical Computing: Vienna, Austria.
3. Carslaw, D.C. and K. Ropkins, openair — An R package for air quality data analysis.
Environmental Modelling & amp; Software, 2012. 27–28(0): p. 52-61.
79
Chapter 4: Assessment and Application of Clustering
Techniques to Atmospheric Particle Number Size Distribution for
the Purpose of Source Apportionment
Farhad Salimi, Zoran Ristovski, Mandana Mazaheri, Rusdin Laiman, Leigh R.
Crilley, Congrong He, Sam Clifford and Lidia Morawska
International Laboratory for Air Quality and Health, Queensland University of
Technology, GPO Box 2434, Brisbane QLD, 4001, Australia
Farhad Salimi, Zoran Ristovski, Mandana Mazaheri, Rusdin Laiman, Leigh R. Crilley,
Sam Clifford and Lidia Morawska. “Assessment and Application of Clustering Techniques to
Atmospheric Particle Number Size Distribution for the Purpose of Source Apportionment”,
under review in Environmental Science and Technology
80
Statement of Joint Authorship
Title: Assessment and Application of Clustering Techniques to Atmospheric
Particle Number Size Distribution for the Purpose of Source Apportionment Spatial
Variation
Authors: Farhad Salimi, Zoran Ristovski, Mandana Mazaheri, Rusdin Laiman, Leigh R.
Crilley, Congrong He, Sam Clifford and Lidia Morawska
Farhad Salimi (candidate): Contributed to the measurements, performed data analysis and
wrote the manuscript.
Zoran Ristovski: Helped with data interpretation and reviewed the manuscript.
Mandana Mazaheri: Supervised the project and reviewed the manuscript.
Rusdin laiman: Contributed to the measurements.
Leigh R. Crilley: Contributed to the measurements.
Sam Clifford: Helped with programming in R and reviewed the manuscript.
Congrong He: Contributed to the measurements and reviewed the manuscript.
Lidia Morawska: Supervised the project and reviewed the manuscript.
81
Abstract
Long-term measurements of particle number size distribution (PNSD) produce a very
large number of observations and their analysis requires an efficient approach in order to
produce results in the least possible time and with maximum accuracy. Clustering
techniques are a family of sophisticated methods which have been recently employed to
analyse PNSD data, however, very little information is available comparing the performance
of different clustering techniques on PNSD data. This study aims to apply several clustering
techniques (i.e. K-means, PAM, CLARA and SOM) to PNSD data, in order to identify and
apply the optimum technique to PNSD data measured at 25 sites across Brisbane, Australia.
A new method, based on the Generalised Additive Model (GAM) with a basis of penalised B-
splines, was proposed to parameterise the PNSD data and the temporal weight of each
cluster was also estimated using the GAM. In addition, each cluster was associated with its
possible source based on the results of this parameterisation, together with the
characteristics of each cluster. The performances of four clustering techniques were
compared using the Dunn index and Silhouette width validation values and the K-means
technique was found to have the highest performance, with five clusters being the
optimum. Therefore, five clusters were found within the data using the K-means technique.
The diurnal occurrence of each cluster was used together with other air quality parameters,
temporal trends and the physical properties of each cluster, in order to attribute each
cluster to its source and origin. The five clusters were attributed to three major sources and
origins, including regional background particles, photochemically induced nucleated
particles and vehicle generated particles. Overall, clustering was found to be an effective
technique for attributing each particle size spectra to its source and the GAM was suitable
to parameterise the PNSD data. These two techniques can help researchers immensely in
analysing PNSD data for characterisation and source apportionment purposes.
82
TOC Art
Keywords
Particle number size distribution, clustering, source apportionment
4.1. Introduction
Atmospheric aerosols affect climate, air quality and subsequently human health [1-
3]. Despite their small contribution to particle volume and mass, ultrafine particles (particles
83
with diameter <100nm) make a significant contribution to particle number concentration
(PNC) [4, 5] and toxicological studies show evidence of their adverse effects on human
health [6]. Therefore, measurements of the chemical and physical properties of aerosol
particles are crucial in order to understand their effects on climate and human health. One
of the most important properties of particles is their size distribution, which helps in
understanding aerosol dynamics, as well as determining their sources [7, 8]. Long-term
particle number size distribution (PNSD) measurements have been conducted in a number
of different environments and the measured size range can extend from less than 10nm up
to more than 10μm. In addition, these long-term measurements generally result in a large
number of observations and analysing such a massive data set often requires sophisticated
techniques. The clustering technique has recently been used to divide particle size data into
groups with similar characteristics and then relate each group to its sources and/or to
investigate aerosol particle formation and evolution [9-14].
Several clustering algorithms currently exist, which makes the selection of an
appropriate clustering technique a daunting task. Determining the most appropriate
number of clusters can be an additional challenge for researchers. Clusters should ideally be
compact, well-separated and scientifically relevant. Beddows et al [9] assessed the
performance of four clustering techniques (Fuzzy, K-means, K-median and model based
clustering) on PNSD data using different validation indices, particularly the Dunn index, and
found the K-means technique capable of finding clusters with smallest size, furthest
separation and highest degree of inner cluster similarity compared to others. Throughout
their work, four techniques were evaluated while several other methods (e.g. Partitioning
around Medoids (PAM), Clustering of Large Applications (CLARA), and Self Organizing Map
(SOM), and Affinity Propagation (AP)) are available which their performance on PNSD data
has not been assessed so far. Therefore, these techniques were selected to be compared
with K-means using two validation measures.
K-means is an iterative algorithm minimising the within cluster sum of squares to
find a given number of clusters [15]. PAM is an iterative algorithm similar to K-means which
constructs clusters around a set of representative objects by assigning each data to the
nearest representative object using sum of pair wise dissimilarities [16]. CLARA performs
PAM on a number of subgroups of data, allowing faster performance for a large number of
84
observations [16]. SOM is a neural networks based method, with the ability to map high
dimensional data to two dimensions and has been widely used in data mining researches
[17]. The AP algorithm is a relatively new clustering technique which has been employed in
different fields since its introduction in 2007. AP considers all data as potential exemplars
and finds the best set of exemplars and corresponding clusters by exchanging messages
between the data points [18].
In a recent study, three years of PNSD data were clustered using the K-means
technique to produce seven clusters. Those clusters were found to form three main groups,
anthropogenic (69%), maritime (29%) and nucleation (2%), which characterised the whole
data set [11]. In another study, Dall’Osto et al. found nine clusters within the PNSD data
collected over a one year period in an urban area and found four typical PNSD groups using
diurnal variation, directional and pollution association [10]. The authors called those groups
traffic, dilution, summer background and regional pollution, which included 69%, 15%, 4%
and 12% of the total data respectively. In the above and all other previous studies, PNSD
data were averaged to decrease the number of data and consequently, reduce
computational cost and complexity. However, averaging can encumber the transient
characterisation of PNSD data and the larger the averaging interval, the more transient
characteristics will be lost.
Parameterisation of PNSD data in terms of a mixture of few log-Normal components
is common and beneficial, particularly for data averaged over a longer interval. Multi-log-
Normal function with predefined number of peaks where the means of each log-Normal
distribution are constrained to vary around some initial estimates in the nucleation, Aitken,
accumulation and coarse modes have been used in literature [19-22]. However, not all of
the measured particle size data are able to be expressed in this way and this imposing a
predefined number of separated peaks may not accurately represent all of the variation in
the collected data. In addition, this method can result in losing the transient trends if
applied to each single particle size spectra.
This study aimed to identify the optimum clustering technique and number of
clusters by comparing the performance of three clustering techniques (i.e. PAM, CLARA, and
SOM) with the K-means technique for several numbers of clusters and to associate each
85
cluster with its possible sources using the cluster characteristics, PNSD parameterisation
results, diurnal variation, temporal variation and several air quality parameters.
4.2. Materials and Methods
Background
This study was performed within the framework of the Ultrafine Particles from
Traffic Emissions and Children’s Health (UPTECH) project, which aimed to determine the
effects of exposure to traffic related ultrafine particles (UFPs) on the health of primary
school-aged children. Air quality measurements were conducted for two consecutive weeks
at each of the 25 randomly selected state primary schools across the Brisbane Metropolitan
Area, in Australia, during the period October 2010 to August 2012. Further details regarding
the UPTECH project can be found in [23] and the study design is available online [24].
Instrumentation, quality assurance, and data processing
PNSD within the size range 9 - 414 nm was measured every 5 minutes using a TSI
Scanning Mobility Particle Sizer (SMPS). The SMPS system included a TSI 3071 Differential
Mobility Analyser (DMA) connected to a TSI 3782 water-based Condensation Particle
Counter (CPC). A combination of a diaphragm pump and a critical orifice was used to supply
a sheathe flow of 6.4 lpm. A zero particle filter and silica gel dryer were used to supply a dry,
particle free air stream. PNC was measured using a TSI 3781 water-based CPC, particle mass
concentration (PM2.5, and PM10) measurements were conducted using a TSI DustTrak,
solar radiation and other meteorological parameters were measured by a Monitor Sensors
weather station, and EcoTech gas analysers measured gaseous emission (i.e. CO, NOx and
SO2) concentrations. These data were averaged according to five minute intervals prior to
data analysis.
The sheathe and aerosol flow rate of the SMPS system was checked three times a
week using a bubble flow meter. A zero check of the system was done at the start of the
measurements at each school using a high efficiency particle (HEPA) filter connected to the
inlet of the system. Size accuracy of the SMPS was calibrated using monodisperse
polystyrene latex (PSL) particles with a nominal diameter of 100 nm. Size accuracy
calibrations were conducted five times throughout the whole measurement campaign and
86
all instruments passed the test with a maximum error of 3.5% from the nominal diameter,
as recommended by Wiedensohler et al [25]. Particle losses inside the tube were corrected
using the formula derived for the laminar flow regime [26]. Equivalent tube length was used
to correct for particle loss inside the bipolar charger and DMA [27, 28]. On-site calibration
checks (span and zero) of the gas analysers were conducted on the second day of the two
week measurement campaigns at each school when the analysers reached the stable
running conditions. The DustTrak zero check was performed every second day and
calibrated if it was not showing zero within the uncertainty of the instrument. Information
regarding the quality assurance and data processing procedures for the CPC can be found in
[23].
Clustering particle number size distribution data
Principal Component Analysis (PCA) was performed on the whole data set, in order
to reduce its high dimensionality and remove the correlation between features, so as to
increase the performance of the clustering. Ideally, clusters should minimise intra-cluster
variation (compactness) and maximise the distance between clusters (separation) [29, 30],
resulting in small, homogenous clusters which are clearly separated from each other.
Validation measures reflecting the compactness and separation of the clusters were used to
find the optimal method and number of clusters using the “clValid” package in R [29, 31]. As
the number of clusters increases, improving compactness, the separation decreases due to
multiple clusters being created which could be described by a single cluster (consider the
extreme case of every observation belonging to its own cluster). Combining compactness
and separation into a single measure is an effective way to address this issue. The Dunn
index [32] and Silhouette width [33] are scores resulting from nonlinear combination of
compactness and separation. Therefore, those scores were chosen in order to compare the
performance of different clustering techniques and to find the optimum number of clusters.
The maximum vector length allowed by R is “231-1”, therefore, the Dunn index can
only be calculated for a maximum of 46,340 observations. To address this issue, half of the
observations were randomly selected and their cluster validity values calculated by applying
PAM, CLARA, SOM and K-means techniques using 2 to 20 clusters. Then validity values for
87
the other half of observations were calculated. This procedure allowed us to evaluate the
whole set of observations considering the vector size limitation.
The AP clustering technique was initially selected as a candidate, in addition to the
aforementioned techniques. The AP technique was implemented using “APcluster” R
package [34] and was computationally expensive but ultimately unsuccessful at clustering
our huge dataset effectively. Based on our experience with this technique and the
recommendations of its developers, AP is not recommended for finding a small number of
clusters in a large dataset, but it is more appropriate for finding a large number of relatively
small clusters [35].
Non-parametric estimation of particle number size distribution and
temporal trends
In this paper, a new approach was developed to parameterise the PNSD data by
finding the local peaks and the normalised concentration at each peak. In order to find the
real local peaks, the noisy trend of PNSD data should be firstly smoothed. To achieve this,
we used the Generalised Additive Model (GAM), with a basis of penalised B-splines [36, 37],
which allows for the flexible estimation of non-linear effects without assuming, a priori, the
functional form of the non-linearity [36]. In contrast with the multi lognormal fitting
method, this approach keeps all the local peaks while smoothing the noisy data. The fitted
function was then used to find the local peaks, which were defined to be the local maxima
in a neighbourhood of five PNSD bins on each side (Figure 4. 1). The bi-variate kernel density
estimate, using the Gaussian kernel, was computed to visualise the distribution of the peaks
of the PNSD data for each cluster [38].
Understanding the temporal trend of clusters provides further information about the
nature and source of each cluster. GAMs allows for a flexible approach to investigating
these temporal trends. Therefore, the GAM, with a basis of B-spline, was employed to
calculate the temporal trend of each cluster. The resulting fitted smooth functions, and their
95% confidence intervals, indicate the temporal variation of each cluster.
88
4.3. Results and discussion
Particle number size distribution
Around 82,000 SMPS measurements with 5 min intervals were conducted during the
whole UPTECH project, which comprised about 285 days of measurements. All of the
previous long-term studies averaged the data to reduce the complexity and calculation
costs, however, averaging can conceal the transient peaks and troughs in PNSD data [9].
Therefore, we decided not to average the measured PNSD data and to use the high
performance computers for handling such a large size data instead.
Optimum clustering technique and number of clusters
Figure 4. 2 illustrates the performance of each clustering technique for two randomly
selected data. Horizontal and vertical axis show the number of clusters and validation value
(Dunn index and Silhouette value) respectively. K-means produced the highest Dunn index
and Silhouette width values for all of the number of clusters, for both sets of randomly
selected data, while the rest of techniques showed the same level of performance (Figure 4.
2). This indicates that the K-means technique was able to produce more compact clusters
that were well separated from each other. The other techniques resulted in almost the
same Dunn index value, however SOM resulted in better Silhouette width values, followed
by PAM and then CLARA. The performance of the K-means technique was significantly
higher than the rest of the techniques and was therefore selected as the preferred
technique.
Cluster schemes with 2 - 8 and 10 clusters had the highest Silhouette width values in
the first and second set of randomly selected data, respectively. Five clusters had the
highest Dunn index value for the first set of data and it resulted in a high value in the second
set as well. However, 9 - 10 clusters resulted in a higher Dunn index value in the second set
of data. As much as possible, clustering with the optimum number of clusters should have
the highest validation value, while still having an appropriate number of clusters for
distinguishing between different sources and processes [9, 11]. Therefore, five was selected
as the optimum number of clusters, as it had a high validation value, as well as scientifically
89
relevant number of clusters to relate each cluster to a source and to find the processes
related to each cluster.
Clustering PNSD data
The K-means clustering technique was applied to all PNSD data in order to find five
clusters. Figure 4. 3 shows the normalised number spectra and 95% confidence interval
associated with each cluster, together with the diurnal variation of their occurrence and
their association with solar radiation intensity, particle mass concentration, PNC and
gaseous pollutants concentrations.
Local peaks of each single PNSD spectra were found using the technique described in
section 0. Plotting the normalized concentration of the peaks versus their diameter is useful
in order to determine the frequency of peaks and their diameter for each cluster. However,
with too many data, points are over-plotted which makes it impossible to distinguish the
underlying trends and relationships. Therefore, Bi-kernel density estimation, which is a very
effective technique to address this issue, was used to visualise the distribution of peaks in
PNSD for each cluster (Figure 4. 4). Characteristics of each cluster and the associated sources
are explained in the following sections, based on Figure 4. 1 and Figure 4. 2.
Cluster1: This cluster included 4.5% of the total measured particle size spectra with a
mode at the SMPS lowest size detection limit (9 nm). The diurnal pattern of occurrence
showed nocturnal minima with a peak at midday. Cluster1 was associated with the highest
solar intensity and lowest PM2.5 among all clusters, and it was also associated with high
PNC and low CO and NOx. Local peaks predominantly occurred at diameters less than 30nm,
with the highest density and normalised concentration at 9nm. This clearly shows the
dominance of newly formed particles which have grown in size and reached the
instrument’s size detection limit. The strength of the local peaks was the highest among all
clusters, indicating the dominance of the smallest particles.
The abovementioned observations indicate that this cluster was attributed to
photochemically-induced nucleated particles in the ambient air. New particles mainly
formed during the middle of the day and grew to reach the instrument’s size detection limit.
The same type of particles were observed in Barcelona for 4% of the observations [10].
90
These types of particles were not observed in London, which could be due to long time
averaging of the data used in that study [9].
Cluster 2: This cluster included 14.1% of the total PNSD data and showed a mode for
particles with a diameter less than 20 nm. The diurnal pattern of occurrence for Cluster 2
peaked strongly at two hours after midday (14:00), with a nocturnal minimum, which was in
agreement with its association with high solar radiation intensity. Its diurnal pattern of
occurrence and association with a low concentration of traffic generated primary pollutants
(NOx and CO) indicated that it had non-traffic related sources. This cluster was also
associated with high PNC. Local peaks were mainly present at diameters less than 20 nm,
with the highest density occurring at a normalized concentration around 0.03. Local peaks
present at diameters larger than 30 nm were lower and corresponded to normalized
concentrations of less than 0.01. This cluster was attributed mainly to aged,
photochemically-induced nucleated particles in the ambient air. Minor peaks at diameters
larger than 30 nm revealed the contribution of vehicle generated particles to this cluster.
However, their contribution to the total PNC was minimal.
Cluster 1 and 2 were both attributed to the same source of particles, but they were
distinguished as relatively fresh and aged, respectively. Nucleated particles initially grew and
reached the instrument detection size limit in Cluster 1, before growing further and
gradually shifting to Cluster 2 after about two hours. Nucleated particles grew in size by
condensation and coagulation, and as they aged their number concentration decreased and
their size increased. This illustrates why Cluster 1 had higher PNC but lower PM2.5
compared to Cluster 2.
The fraction of PNSD data related to nucleated particles in this study was higher than
in another study in the same environment, which used the classic approach to find the
banana shaped new particle formation events [39]. This shows that the actual occurrence of
new particle formation events is much higher than what is generally found by classic
approaches, because in most particle formation events, the newly formed particles would
be scavenged by the pre-existing particles as a result of coagulation, before they grew by
means of condensation. However, this does not mean that the nucleation events would not
occur and that the particles resulting from them would not be present in the air.
91
Cluster 3: This cluster included 22.6% of the total data, with a mode at 60 nm. The
diurnal pattern of occurrence showed nocturnal maxima, with a minimum during the
midday and early afternoon hours, and it was associated with the highest PM2.5 among all
clusters as well as with high NOx and CO. Local peaks were present at diameters smaller
than 20 nm and between 50-70 nm. The normalised concentration corresponding to either
of these diameter ranges had a similar magnitude. These findings suggest that cluster 3 was
attributed mainly to regional background aerosols. However, a source of small particles and
gaseous emissions is also present in this cluster.
Cluster 4: This cluster included 31.6% of the total measured data, with a mode at 20
nm and a minor peak at the smallest instrument size limit (≈ 9nm), showing a strong
association with morning and afternoon rush hour traffic [23]. This cluster was associated
with low solar radiation and high vehicle generated primary pollutants, including CO and
NOx, and the particle number size modes were in the range of vehicle generated primary
particles and nucleated particles during the exhaust emissions dilution [40-42]. Local peaks
were mainly present at diameters less than 30 nm, particularly at the lowest detection size
limit of the instrument. The normalized concentration corresponding to peaks were highest
at the smallest particle size. These observations suggest that this cluster was attributed to
vehicle generated particles, including primary particles (having a diameter of around 40 nm)
and particularly secondary particles formed in the vehicle exhaust (having a diameter of
around 9nm). Similar types of particles were observed in the literature [9, 10].
Cluster 5: This cluster included 27.2% of the total PNSD data, with a mode at 40 nm.
The diurnal pattern of occurrence followed almost the same trend as Cluster 3, with a
minimum during the early afternoon and a peak during the night-time. The local peaks’
density was almost evenly distributed through the whole range of diameters, with smaller
particles having more peaks particularly for diameters less than 20 nm. However,
normalized concentrations corresponding to each local peak were highest at diameters
between 30-60 nm, showing the dominance of particles at this size range. The observations
mentioned above suggest that, like Cluster 3, this cluster was attributed to regional
background aerosols.
92
Temporal trend of clusters
A non-parametric regression model was fitted to the daily occurrence fraction of
each cluster, in order to determine the temporal trend of each cluster. The regression
model quantified the presence of each cluster as the sum of a smooth function for each
month and a trend which included the day, month and year. It should be noted that no data
were collected during the first two months of the year.
The smooth monthly function for Cluster 1 did not show any significant variation
before November, at which point it increased moderately and peaked during December.
Solar radiation intensity is also known to increase during the last two months of the year in
the southern hemisphere, which indicates that this may be the driving force of more
atmospheric nucleation events. Cluster 2 showed similar trends of increase in December in
conjunction with the expected increase in solar radiation intensity. The prevalence of
Cluster 3 increased and peaked around July-August and then decreased again. Cluster 3
showed the strongest monthly variation among all clusters. As mentioned earlier, this
cluster was associated with regional background aerosols and its monthly variation showed
a positive correlation with biomass burning within the studied region, which mostly took
place during July to September. Biomass burning associated PNSDs peak at 100-200 nm [43],
however the PNSD of Cluster 3 peaked at around 60 nm, which was 20 nm higher than the
Cluster 5, which was also associated with background aerosols. This implies that Cluster 3
included the mixture of background and biomass burning aerosols, which made the average
PNSD peak shift by 20 nm. The prevalence of Cluster 4 decreased, with a trough during
August, before increasing again, which is likely to be the result of the biomass burning which
occurred during August and consequently decreased the occurrence of particles belonging
to Cluster 4. For Cluster 5, a peak was observed during June, which decreased to show a
trough during August, before peaking again during November. This trend is the opposite to
what was observed for Cluster 3 and given that the peaks for Cluster 5 were observed when
there were less biomass burning events, it was most likely associated with regional
background aerosols, without the influence of biomass burning events. The PNSDs which
were attributed to regional background aerosols in this study had different modes
compared to the one observed in London [9], but similar ones to the observations in
Barcelona [10].
93
In summary, the K-means clustering technique was found to be the preferred
technique when compared to SOM, PAM and CLARA. The K-means clustering technique
categorised the PNSD data into five clusters and each cluster was attributed to its source
and origin. Five clusters were attributed to three major sources and origins, as follows: 1)
Regional background particles: Clusters 3 and 5, which included 49.8% of the total data,
were attributed to regional background aerosols with clear modes at 60nm and 40nm,
respectively; 2) Photochemically induced nucleated particles: Clusters 1 and 2, which
included 18.6% of the total data, were attributed to photochemically-induced nucleated
particles; and 3) Vehicle generated particles: Cluster 4, which included 31.6% of the data,
was attributed to vehicle generated particles. A new method was proposed for the
parameterisation of particle size spectra, based on the GAM, which was found to be an
effective tool and is recommended to be used for particle size data. K-means clustering
successfully attributed each particle size spectra to its source and/or origin. However, while
this technique could attribute each particle size spectra to its major contributing source,
several sources with different levels of contribution are often responsible for the pattern of
each particle size spectra. The structure of the contribution of sources can be further
investigated using other techniques, such as Bayesian infinite mixture modelling [44, 45],
Bayesian K-means, Bayesian Beta-process clustering [46] and positive matrix factorisation
[8].
4.4. Acknowledgment
This work was supported by the Australian Research Council (ARC), Department of
Transport and Main Roads (DTMR) and Department of Education, Training and Employment
(DETE) through Linkage Grant LP0990134. Our particular thanks go to R. Fletcher (DTMR)
and B. Robertson (DETE) for their vision regarding the importance of this work. We would
also like to thank all members of the UPTECH project, including Prof G. Marks, Dr P.
Robinson, Prof K. Mengersen, Prof G. Ayoko, Dr G. Johnson, Dr R. Jayaratne, Dr S. Low Choy,
Prof G. Williams, W. Ezz, M. Mokhtar, N. Mishra, L. Guo, Prof C. Duchaine, Dr H. Salonen, Dr
X. Ling, Dr J. Davies, Dr L. Leontjew Toms, F. Fuoco, Dr A. Cortes, Dr B. Toelle, A. Quinones
and P. Kidd for their contribution to this work. We thank former ILAQH, QUT members Dr M.
Falk, Dr F. Fatokun, Dr J. Mejia and Dr D. Keogh for their contributions at the early stage of
94
the project. Thanks also to Prof T. Salthammer from Fraunhofer WKI in Germany for VOC
analysis, R. Appleby and C. Labbe for their administrative assistance.
4.5. References
4.6. Figures:
Figure 4. 1: An example of a fitted smooth function (solid line) on PNSD data (circles) and the identified local peaks
(solid circles).
95
Figure 4. 2: Dunn index and Silhouette width for Kmeans, SOM, PAM, and CLARA clustering techniques for
different cluster numbers.
96
Figure 4. 3: Clustering results using SMPS data. Graphs on the left show the particle size distributions with 95%
confidence intervals, the associated graphs on the middle show the diurnal cycle of the hourly percentage of
occurrence for each cluster, and the graphs on right show the associated solar radiation, PM2.5,, PM10, NOx, CO, SO2,
total PNC.
97
Figure 4. 4: Density of peaks in particle number size data at each cluster.
98
Figure 4. 5: Temporal trend of occurrence of each cluster, with their 95% confidence intervals.
1. Stevens, B. and G. Feingold, Untangling aerosol effects on clouds and precipitation in a buffered system. Nature, 2009. 461(7264): p. 607-613.
2. Pope II, C.A. and D.W. Dockery, Health effects of fine particulate air pollution: lines that connect. Air & Waste Management Association, 2006(56): p. 709-742.
3. Lohmann, U. and J. Feichter, Global indirect aerosol effects: a review. Atmospheric Chemistry and Physics, 2005. 5(3): p. 715-737.
4. Morawska, L., et al., Submicrometer and Supermicrometer Particles from Diesel Vehicle Emissions. Environmental Science & Technology, 1998. 32(14): p. 2033-2042.
5. Harrison, R.M. and J. Yin, Particulate matter in the atmosphere: which particle properties are important for its effects on health? Science of The Total Environment, 2000. 249(1–3): p. 85-101.
6. WHO, W.H.O., Air Quality Guidelines: Global Update 2005: Particulate Matter, Ozone, Nitrogen Dioxide, and Sulfur Dioxide. 2005: World Health Organization Europe.
7. Charron, A., W. Birmili, and R.M. Harrison, Fingerprinting particle origins according to their size distribution at a UK rural site. J. Geophys. Res., 2008. 113(D7): p. D07202.
8. Harrison, R.M., D.C.S. Beddows, and M. Dall’Osto, PMF analysis of wide-range particle size spectra collected on a major highway. Environmental Science & Technology, 2011. 45(13): p. 5522-5528.
99
9. Beddows, D.C.S., M. Dall’Osto, and R.M. Harrison, Cluster Analysis of Rural, Urban, and Curbside Atmospheric Particle Size Data. Environmental Science & Technology, 2009. 43(13): p. 4694-4700.
10. Dall'Osto, M., et al., Urban aerosol size distributions over the Mediterranean city of Barcelona, NE Spain. Atmos. Chem. Phys. Discuss, 2012. 12: p. 16457-16492.
11. Wegner, T., et al., Properties of aerosol signature size distributions in the urban environment as derived by cluster analysis. Atmospheric Environment, 2012. 61(0): p. 350-360.
12. Tunved, P., J. Ström, and H.C. Hansson, An investigation of processes controlling the evolution of the boundary layer aerosol size distribution properties at the Swedish background station Aspvreten. Atmospheric Chemistry and Physics Discussions, 2004. 4(4): p. 4507-4543.
13. Charron, A., W. Birmili, and R.M. Harrison, Factors influencing new particle formation at the rural site, Harwell, United Kingdom. Journal of geophysical research, 2007. 112(D14): p. D14210.
14. Costabile, F., et al., Spatio-temporal variability and principal components of the particle number size distribution in an urban atmosphere. Atmos. Chem. Phys, 2009. 9: p. 3163-3195.
15. Hartigan, J.A. and M.A. Wong, A k-means clustering algorithm. Applied statistics, 1979: p. 100-108.
16. Kaufman, L. and P.J. Rousseeuw, Finding Groups in Data: An Introduction to Cluster Analysis. 2009, New York: John Wiley & Sons, Inc.
17. Kohonen, T., Self-organizing maps. Vol. 30. 2001, New York: Springer. 18. Frey, B.J. and D. Dueck, Clustering by Passing Messages Between Data Points. Science, 2007.
315(5814): p. 972-976. 19. Hussein, T., et al., Urban aerosol number size distributions. Atmospheric Chemistry and
Physics, 2004. 4(2): p. 391-411. 20. Hussein, T., et al., Evaluation of an automatic algorithm for fitting the particle number size
distributions. Boreal Environment Research, 2005. 10(5): p. 337-355. 21. Heintzenberg, J., et al., Aerosol particle number size distributions and particulate light
absorption at the ZOTTO tall tower (Siberia), 2006–2009. Atmos. Chem. Phys, 2011. 11: p. 8703-8719.
22. Shen, X., et al., First long-term study of particle number size distributions and new particle formation events of regional aerosol in the North China Plain. Atmos. Chem. Phys, 2011. 11(4): p. 1565-1580.
23. Salimi, F., et al., Spatial variation of particle number concentration in school microscale environments and its impact on exposure assessment. Environmental Science & Technology, 2013.
24. UPTECH. Available from: http://www.ilaqh.qut.edu.au/Misc/UPTECH%20Home.htm. 25. Wiedensohler, A., et al., Mobility particle size spectrometers: harmonization of technical
standards and data structure to facilitate high quality long-term observations of atmospheric particle number size distributions, Atmos. Meas. Tech, 2012. 5(3): p. 657-685.
26. Hinds, W.C., Aerosol technology: properties, behavior, and measurement of airborne particles. 1999, New York: Wiley.
27. Karlsson, M.N.A. and B.G. Martinsson, Methods to measure and predict the transfer function size dependence of individual DMAs. Journal of Aerosol Science, 2003. 34(5): p. 603-625.
28. Covert, D., A. Wiedensohler, and L. Russell, Particle Charging and Transmission Efficiencies of Aerosol Charge Neutralizes. Aerosol Science and Technology, 1997. 27(2): p. 206-214.
29. Brock, G., V. Pihur, and S. Datta, clValid, an R package for cluster validation. Journal of Statistical Software (Brock et al., March 2008), 2011.
30. Handl, J., J. Knowles, and D.B. Kell, Computational cluster validation in post-genomic data analysis. Bioinformatics, 2005. 21(15): p. 3201-3212.
100
31. R Development Core Team, R: A language and environment for statistical computing. 2010, R foundation for Statistical Computing: Vienna, Austria.
32. Dunn, J.C., Well-separated clusters and optimal fuzzy partitions. Journal of cybernetics, 1974. 4(1): p. 95-104.
33. Rousseeuw, P.J., Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 1987. 20: p. 53-65.
34. Bodenhofer, U., A. Kothmeier, and S. Hochreiter, APCluster: an R package for affinity propagation clustering. Bioinformatics, 2011. 27(17): p. 2463-2464.
35. FAQ, A.P.; Available from: http://www.psi.toronto.edu/affinitypropagation/faq.html. 36. Wood, S.N., Thin plate regression splines. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 2003. 65(1): p. 95-114. 37. Eilers, P.H.C. and B.D. Marx, Flexible Smoothing with $B$-splines and Penalties. Statistical
Science, 1996. 11(2): p. 89-102. 38. Wand, M., Fast computation of multivariate kernel estimators. Journal of Computational and
Graphical Statistics, 1994. 3(4): p. 433-445. 39. Cheung, H.C., Z.D. Ristovski, and L. Morawska, Observation of new particle formation in
subtropical urban environment. Atmospheric Chemistry and Physics Discussions, 2010. 10(10): p. 22623-22652.
40. Casati, R., et al., Measurement of nucleation and soot mode particle emission from a diesel passenger car in real world and laboratory in situ dilution. Atmospheric Environment, 2007. 41(10): p. 2125-2135.
41. Ntziachristos, L., et al., Particle Concentration and Characteristics near a Major Freeway with Heavy-Duty Diesel Traffic. Environmental Science & Technology, 2007. 41(7): p. 2223-2230.
42. Janhäll, S., et al., Size resolved traffic emission factors of submicrometer particles. Atmospheric Environment, 2004. 38(26): p. 4331-4340.
43. Friend, A.J., et al., Source apportionment of ultrafine and fine particle concentrations in Brisbane, Australia. Environmental Science and Pollution Research, 2011: p. 1-9.
44. Wraith, D., et al., Bayesian mixture model estimation of aerosol particle size distributions. Environmetrics, 2011. 22(1): p. 23-34.
45. Kulis, B. and M.I. Jordan, Revisiting k-means: New algorithms via Bayesian nonparametrics. arXiv preprint arXiv:1111.0352, 2011.
46. Broderick, T., B. Kulis, and M.I. Jordan, MAD-Bayes: MAP-based asymptotic derivations from Bayes. arXiv preprint arXiv:1212.2126, 2012.
101
Chapter 5: Investigation into Chemistry of New Particle Formation
and Growth in Subtropical Urban Environment
Farhad Salimi, Leigh R. Crilley, Svetlana Stevanovic, Zoran Ristovski, Mandana
Mazaheri, and Lidia Morawska*
International Laboratory for Air Quality and Health, Queensland University of
Technology, GPO Box 2434, Brisbane QLD, 4001, Australia
Farhad Salimi, Leigh R. Crilley, Svetlana Stevanovic, Zoran Ristovski, Mandana
Mazaheri, and Lidia Morawska. “Investigation into Chemistry of New Particle Formation and
Growth in Subtropical Urban Environment”, under review in Environmental Science and
Technology
102
Statement of Joint Authorship
Title: Investigation into Chemistry of New Particle Formation and Growth in
Subtropical Urban Environment
Authors: Farhad Salimi, Leigh R. Crilley, Svetlana Stevanovic, Zoran Ristovski, Mandana
Mazaheri, and Lidia Morawska
Farhad Salimi (candidate): Contributed to the measurements, performed data analysis and
wrote the manuscript.
Leigh R. Crilley: Performed the AMS measurements, calibrated AMS, processed the
collected data and reviewed the manuscript
Svetlana Stevanovic: Contributed to data analysis and interpretation of the results, wrote
part of the manuscript and reviewed it
Zoran Ristovski: Contributed to data interpretation and reviewed the manuscript
Mandana Mazaheri: Supervised the project and reviewed the manuscript
Lidia Morawska: Supervised the project, assisted with data interpretation and reviewed the
manuscript.
103
5.1. Abstract
The chemical and physical dynamics behind new particle formation (NPF) events are not
fully understood. Therefore, this study was conducted to investigate the chemistry of
aerosol particles during NPF events in an urban subtropical environment. Aerosol chemical
composition was measured along with particle number size distribution (PNSD) and several
other air quality parameters at five sites across an urban subtropical environment. An
Aerodyne compact Time-of-Flight Aerosol Mass Spectrometer (c-TOF-AMS) and a TSI
Scanning Mobility Particle Sizer (SMPS) measured aerosol chemical composition and PNSD,
respectively. Five NPF events, with growth rates in the range 3.3-4.6 nm, were detected at
two sites. The NPF events happened on relatively warmer days with lower humidity and
higher solar radiation. Temporal percent fractions of nitrate, sulphate, ammonium and
organics were modelled using the Generalised Additive Model (GAM), with a basis of
penalised spline. Percent fractions of organics increased after the NPF events, while the
mass fraction of ammonium and sulphate decreased. This uncovered the important role of
organics in the growth of newly formed particles. Three organic tracer ions (i.e. f43, f44 and
f57) were calculated and the f44 vs f43 plots were compared between nucleation and non-
nucleation days. f44 vs f43 followed a different pattern on nucleation days compared to non-
nucleation days, whereby f43 decreased for vehicle emission generated particles, while both
f44 and f43 decreased for NPF generated particles.
5.2. Introduction
Aerosol particles affect the Earth’s climate, air quality and public health, both directly and
indirectly [1-3]. Knowledge about their formation, transformation and physical/chemical
properties helps in understanding their effects on climate and human health. New particle
formation (NPF; also known as nucleation) events have been observed in different locations,
including coastal, forested, rural and urban areas [4-12], and these events are one of the
main sources of ultrafine particles (UFPs; particles smaller than 100nm), in addition to
combustion emitted particles. The significant increase in the number of UFPs after NPF
events can potentially cause adverse effects on human health, given the increasing evidence
of the toxicity of UFPs and their role in human mortality and morbidity [13-15].
NPF events include two main steps, firstly the homogenous or ion-induced nucleation of
neutral or ion clusters takes place, followed by the growth of newly formed clusters into
104
larger particle [16-18].The classical binary H2SO4-H2O nucleation theory significantly
underestimates the formation and growth rate of observed NPF events in the atmosphere
[19]. Therefore, apart from H2SO4, other species with sufficiently low volatility, such as NH3
and semi-volatile organics, have been suggested to contribute to particle formation and
growth [20-23]. However, the chemistry and dynamics of UFPs during NPF events have been
investigated in only a handful of studies that covered very limited types of climate and
environments. Zhang et al. [24] employed an Aerodyne quadrupole Aerosol Mass
Spectrometer (AMS) and two Scanning Mobility Particle Sizer (SMPS) systems in Pittsburgh,
USA, in order to investigate the chemistry of NPF and particle growth over a 16 day
measurement campaign. They observed three significant NPF events with a significant
increase in the concentration of sulphate, ammonium, nitrate and organics in the ultrafine
mode of particles. Creamean et al. [25] employed an Ultrafine Time-of-Flight Aerosol Mass
Spectrometer (UF-TOF-MS) at a remote rural mountain site for investigating individual
aerosol aerodynamic size and chemical composition during NPF events. Amines and
sulphates were found in aerosol particles during the NPF events and their concentration
increased as the particles grew.
NPF events have been observed frequently in Brisbane, an urban area in Eastern Australia,
and were associated with precursors emitted from traffic and solar radiation [10]. Cheung et
al. [10] characterised the evolution of particle number size distribution (PNSD) in Brisbane
and found 65 nucleation events over a one year period, with an average particle growth rate
of 4.6 (nmh-1). However, to date, no study has been conducted to identify the chemical
compounds responsible for NPF and growth using direct measurements in this area. More
studies need to be carried out, particularly in less studied areas, using direct measurements
techniques, in order to determine the species involved in NPF events and the nature of their
contribution. To address this gap in knowledge, the main aim of this study was to determine
the role of chemical species in NPF events in a subtropical urban environment.
105
5.3. Materials and Methods
5.3.1. Background
This study was performed as part of the ‘Ultrafine Particles from Traffic Emissions and
Children’s Health’ (UPTECH) project [26, 27]. Air quality measurements were conducted for
two consecutive weeks at each of 25 sites across the Brisbane Metropolitan Area, in
Australia, during the period October 2010 to August 2012. Chemical composition of the
aerosol particles were monitored in real-time at five sites, referred to as S1, S4, S11, S12 and
S25. The measurement sites were state primary schools as this study was conducted within
the UPTECH framework.
5.3.2. Instrumentation
A TSI SMPS measured PNSD within the size range 9-414 nm. The SMPS system consisted of a
TSI 3071 Differential Mobility Analyser (DMA) and a TSI 3782 water-based Condensation
Particle Counter (CPC). Particle size detection accuracy of the SMPS system was calibrated at
the start of measurements at each site using polystyrene latex (PSL) particles. In addition,
the system sheath and aerosol flow rate were checked every second day. The measured
PNSD data were corrected for diffusion loss inside the system and sampling tube. Further
information regarding the SMPS’s operating conditions and calibration can be found in [28].
An Aerodyne compact Time-of-Flight Aerosol Mass Spectrometer (c-ToF-AMS) was deployed
in five sites (S1, S4, S11, S12, and S25) to monitor the chemical composition of aerosol
particles in the PM1 fraction in real time. The lowest detection limit of this instrument is
50nm which limits the measured composition to particles larger than that size. A detailed
description of the sampling method and c-TOF-AMS operation can be found in Crilley et al.
[29]. Briefly, the c-TOF-AMS was housed in a vacant room within the site and sampled for 2-
3 weeks. The sampling interval was 5 minutes, alternating equally between particle time of
flight and mass spectrometer modes. The instrument was calibrated for both ion efficiency
and particle size at the beginning of measurements at each site, and ion efficiency
calibration was also performed in the middle and at the end of measurement campaigns at
each site. All calibrations were performed according to the standard procedures proposed in
the literature [30-32].
106
Solar radiation and other meteorological parameters were measured by a Monitor Sensors
µSmart Series weather station, a TSI 3781 water-based CPC was used for particle number
concentration (PNC) measurements, and a TSI DustTrak measured particle mass
concentration (PM2.5 and PM10).
5.3.3. Data analysis
Characterisation of New Particle Formation Events: NPF events were identified and
classified using the procedure developed by Dal Maso et al [33]. PNSD data were analysed
to identify a distinct new mode of particles in the nucleation mode size range (particles with
diameter less than 30 nm). If the mode grew in size and was persistent for more than one
hour, it was assigned as a NPF event. Once identified, the growth rate of the nucleation
events were calculated following the procedure described in [25].
Condensation Sink: The surface area of aerosol particles that is available for condensation
can be measured using condensation sink (CS). The CS determines the rate of condensation
of gaseous molecules on pre-existing particles and is a function of PNSD [34, 35]. CS, with
unit s-1, was calculated using the following formula [35]:
1)
i
iipMpppMp NdDdddnddDCSi ,
0
2)()(2
where D is the diffusion coefficient, dp is the particle diameter, Ni is the concentration of
particles and M can be expressed as [36]:
2)
KnKnKn
KnM
121
3
4
3
41377.0
1
where is the sticking coefficient and is assumed to be 1 [37] and Kn is the Knudsen
number which is equal to p
v
d
2. The mean free path ( v ) is a function of pressure and
temperature and can be calculated using the following formula [38]:
3) )
)110
1(
)2.296
1101(
)(2.296
)(101
(
T
T
pov
107
where P is in kPa and T in K. o = 0.039 µm, which is the mean free path of H2SO4 at standard
conditions [39]. CS’s were calculated using the above mentioned procedure on the PNSD
data.
Particle Size: Due to the difference in detection limit between the two instruments (Dva =
50nm for AMS and Dm = 10 nm for SMPS), measurement of the mass distribution for the
smallest particles detected during each nucleation was delayed for the AMS. Ratio of
mobility diameter (Dm), measured by the SMPS, to Dva is a function of size, composition,
shape and relative humidity for ambient particles. This ratio can be simplified to be equal to
particle density, assuming that particles shape to be a sphere [40]. The newly formed
particles were assumed to have the density of around 1.8 g/cm3 and therefore, Dm/Dva = 1.8
for newly formed particles [24]. Therefore, the AMS’s lowest detection limit was equivalent
to a Dm = 30 nm which will consequently limit the study of new particle formation to the
growth phase..
Particle Chemical Composition: Concentrations of sulphate, nitrate, organics and
ammonium were measured by a c-TOF-AMS, as these species have been identified as the
main contributing species to the NPF events and subsequent particulate growth in literature
[24]. These data were processed and analysed using Squirrel v1.51 in Igor Pro v6.22. The
organic aerosol (OA) can be divided into oxygenated OA (OOA) and hydrocarbon-like OA
(HOA), based upon key m/z ions that have been shown to be surrogates for the different
components. OOA can be separated further into low-volatility OOA (LV-OOA) and semi-
volatile (SV-OOA). Two main ions, m/z 44 (CO2+) and m/z 43 (mostly C2H3O+), can
characterise the evolution of OA in the atmosphere [41]. The ratios of m/z 44 and m/z 43 to
total signal in the component mass spectrum (f44 and f43, respectively) can characterise the
degree of oxidation of OA, since the SV-OOA component spectra have a lower f44 and higher
f43 compared to LV-OOA. OA has been shown to occupy a triangular region of the f44 vs f43
plot with younger OA occupying the lower part of the triangle, while more oxidised and
subsequently more aged OA are concentrated in the upper part [41, 42]. Therefore, f43 and
f44 were calculated and plotted against each other to determine the type and evolution of
OA during NPF events. m/z 57 ions have been shown to be a tracer of HOA (primary organics
from combustion sources) [43] and therefore, f57 was calculated to investigate the OA
originated from vehicle emissions.
108
Generalised Additive Model (GAM): The diurnal trends of chemical components and f57 were
further analysed using the Generalised Additive Model, with a basis of penalised B-splines
[44]. This approach allows for the flexible estimation of non-linear relationships without
assuming, a priori, the functional form of the non-linearity. The resulting fitted smooth
functions, and their 95% confidence intervals, indicate the temporal trends of each chemical
species and f57.
5.4. Results and Discussion
5.4.1. Identified New Particle Formations
After the analysis of PNSD measured at five sites, five NPF events were detected, where
three events were observed in S12 and two in S25, while no events were observed at S1, S4
and S11. Therefore, we focused only on S12 and S25, for which the NPF growth rates varied
from 3.3 to 4.6 (nmh-1) (Table 5. 1).
Table 5. 1: Average growth rates during the nucleation events.
Site start of event end of event GR(nm/hr)
s12 12/08/2011
11:00 13/08/2011
5:00 3.9
s12 13/08/2011
12:30 14/08/2011
6:20 4.6
s12 14/08/2011
15:20 15/08/2011
6:00 3.3
s25 25/08/2012
12:24 26/08/2012
2:14 4.3
s25 26/08/2012
12:00 27/08/2012
1:00 3.8
Figure 5. 1 summarises the average solar radiation, relative humidity, temperature and CS
for days with NPF events (nucleation days), as well as for the days where no NPF events
were observed (non-nucleation days) at S12 and S25. At both sites, nucleation days had
higher solar radiation intensity compared to non-nucleation days, which is aligned with
similar observations in the literature [24]. As expected, similar trends were observed for
temperature as well. Lower relative humidity before the start of NPF events was observed
on nucleation days compared to non-nucleation days at S12, whereas higher relative
humidity was observed on nucleation days after the start of NPF events. Lower relative
humidity was observed on nucleation days at S25, which is in agreement with the
109
observations made by others in similar environments [45, 46]. CS was found to be lower on
nucleation days compared to non-nucleation days, about two hours before the start of
nucleation at both sites. Similar trends have been observed in other studies [9, 47]. CS is a
measure of the available surface area for condensation of vapours, as well as for the
scavenging of particle clusters. A small CS is favourable for nucleation as it produces less
surface area for vapours and newly formed clusters.
Figure 5. 1: Average diurnal pattern of solar radiation (SR), humidity, temperature and condensation sink (CS) on nucleation and non-nucleation day at S12 and S25. Vertical dotted lines show the start of nucleation.
Figure 5.2 illustrates the average mass concentration of organics, nitrate, sulphate and
ammonium during nucleation and non-nucleation days at S12 and S25. Mass concentration
of organics and nitrate followed relatively the same trend all day until 4-5 pm (6-7 hours
after the start of the NPF event) at S12 and S25.
However, mass concentration of organics and nitrate increased dramatically between 3-10
pm on nucleation days, at higher rate compared to non-nucleation days at S12, and the
same trend was observed at S25, with an hours delay in the increase of the mass
concentration of organics and nitrate. Sulphate and ammonium mass concentration started
to increase at 12pm, reached a peak and decreased subsequently, however, a significant
increase was observed around 4pm.
110
Figure 5.2: Average diurnal pattern of organics, nitrate, sulphate and ammonium concentrations on nucleation and non-nucleation days at S12 and S25.
5.4.2. Evolution of Chemical Compositions
Figure 5.3 illustrates the evolution of PNSD and mass concentration of chemical species
(organics, nitrate, sulphate and ammonium) during the three consecutive NPF events at S12.
In general, size distribution of the aerosol mass species followed the evolution of PNSD, with
the exception of ammonium, which shows no particular trend throughout the all of the
nucleation events. Ammonium concentration was evenly distributed throughout the
nucleation events in contrast with other components as illustrated in Figure 5.3.
The mass concentration of sulphate increased significantly six hours after the start of
nucleation, when the nucleated particles grew and reached the AMS’s lowest detection size
limit. The mass concentration of nitrate and organics increased three hours after an increase
in sulphate concentration was observed. The same trends were observed in the case of all
three nucleation events. Time series of PNSD and mass concentration of particle species
during two NPF events happening on two consecutive days at S25 are illustrated in Figure
5.4. Unlike the observations at S12, no distinctive trend in sulphate mass concentration was
observed during the first event. However, organics and nitrate showed similar trends and
ammonium did not display any distinctive trend. Sulphate followed similar trend as the
events observed at S25.
111
It has been previously determined in the literature that the main contributing species to NPF
events and subsequent particulate growth are sulphate, nitrate, organics and ammonium
[24]. Therefore, the percent fraction of each of these chemical species in the size range 50-
100 nm (Dva) were calculated by dividing the mass concentration of each chemical species
by the total (sulphate + nitrate + ammonium + organics). The temporal trend of each species
mass precent fraction was modelled using the Generalised Additive Model (GAM) with a
basis of penalised B-splines. Similar trends for all chemical species were observed at both
sites. Ammonium, sulphate and nitrate mass fractions peaked around the start of nucleation
and subsequently decreased after the event. However, organics followed the opposite
trend, with a significant increase after the start of nucleation (Figure 5.5). In other words,
the fraction of organics increased and the fraction of ammonium, sulphate and nitrate
decreased with newly formed particles in the range 50-100 nm (Dva), compared to pre-
existing particles within the same range indicating that organics were the main contributor
to the growth of newly formed particles.
112
Figure 5.3: Time series of the mass concentration of particle species and particle number size distribution at S12.
113
Figure 5.4: Time series of the mass concentration of particle species and particle number size distribution at S25.
114
Figure 5.5: Mass fraction of each chemical species (sulphate, nitrate, ammonium and organics) during the nucleation events at s12 and s25. Vertical dashed lines represent the point that the newly formed particles reached the AMS’s lowest
detection limit.
5.4.3. Role of organics
As mentioned in the previous section, three significant NPF events were observed at S12 on
three consecutive days, each event lasted for rest of the day and was closely followed by the
start of another event. In addition, the involvement of organics in NPF events at S12 was
higher than at S25, therefore, further data analysis regarding the role of organics were
performed at S12. Firstly, f43 and f44 were calculated for nucleation days and non-nucleation
days. f44 was plotted against f43 at each hour of the day, as well as the triangle space
proposed by Ng et al. [41] (Figure 5. 6 & Figure 5. 7). SV-OOA components fell in the lower
half of the triangle, while LV-OOA components were concentrated in the upper half. The
base of the triangular region included a wide range of f43, which represent less oxidised SV-
OOA [41]. During the non-nucleation days, aerosol components were concentrated in the
middle and right border of the triangular region, which is the beginning of the LV-OOA
components region. As expected, the f44 decreased during morning and afternoon rush
hours, as the particles were younger and less oxidised, while, f43 remained almost
unchanged during the day. Aerosol components fell in the middle right side of the triangular
115
region before the start of morning rush hour, at which time f44 decreased while f43 remained
unchanged.
During the nucleation days, several hours after the start of nucleation, when the particles
grew enough to be detected by the AMS, f44 and f43 both decreased and the aerosol
components reached the bottom left side of the triangular region. It is expected that the
particles originating from nucleation events cluster somewhere at top of the triangle,
however, as the particles measured in our study have been coming from nucleation events
and vehicles, they clustered at the middle right hand side of the triangle.
After the start of nucleation and an increase in traffic density, OOA components moved
toward the region defined for younger particles. However, during the nucleation events,
both f44 and f43 decreased, while only f44 decreased when the aerosol particles were
generated by the traffic during non-nucleation days (Figure 5.8). Aerosol particles which
originated from traffic emissions and NPF appeared to cluster at different locations inside
the triangle, indicating that the f44 vs f43 data can be used to assign a dominant source
(traffic or NPF) to the OA components.
It has been reported in the literature that f57 is associated with combustion generated
primary organics [48, 49]. f57 was calculated and its diurnal variation was modelled using a
GAM model with a basis of panelised B-spline. f57 followed exactly the same trend on
nucleation and non-nucleation days, with an increase during the morning and afternoon
rush hours. This shows the effect of vehicle generated particles both in nucleation and non-
nucleation days and can explain the pattern observed in f44 vs f43 plot during nucleation
days.
In summary, PNSD, chemical composition and meteorological parameters were measured at
five sites across the Brisbane metropolitan area. Five NPF events, with growth rates ranging
from 3.3-4.6 nm.hr-1, were observed, and the NPF events happened on days with relatively
lower humidity, and higher solar radiation and temperature than non-event days. Higher
sulphate, nitrate, ammonium and organics were observed on nucleation days compared
with days when no nucleation was observed. Percent fractions of nitrate, sulphate,
ammonium and organic chemical species were calculated and their diurnal trends were
modelled using the GAM. Ammonium, sulphate and nitrate mass fractions increased before
116
the start of nucleation, peaked around the start of nucleation and decreased after the
event. Conversely, the organics percent fraction increased significantly after the start of
nucleation, indicating the important role of organics in the growth phase of NPF events. f44
and f43 were analysed to investigate the role of organics in further depth, as the f44 vs f43
would reveal information regarding the level of oxidation and volatility of OOA. f43 and f44
both decreased after the start of nucleation, while f44 decreased only when the particles
were vehicle emission generated. Aerosol particles generated by vehicle emissions and NPF
events clustered in different locations on the f44 vs f43 plot, indicating the application of the
f44 vs f43 plot for identification of the source/s and transformation of the OOA components.
117
Figure 5. 6: f44vs f43 at each hour of the day for all data measured at S12 during the non-nucleation days. The triangle from Ng et al. [41] is drawn as a visual aid.
118
Figure 5. 7: f44 vs f43 at each hour of the day for all data measured at S12 during the nucleation days. The triangle from Ng et al. [41] is drawn as a visual aid.
119
Figure 5.8: The pattern observed in f44 vs f43 plot after the aerosol particles were dominated by either traffic or NPF events.
Figure 5. 9: Diurnal variation of the f57 at S12 during the nucleation and non-nucleation days.
5.5. Acknowledgments
This work was supported by the Australian Research Council (ARC), QLD Department of Transport
and Main Roads (DTMR) and QLD Department of Education, Training and Employment (DETE)
through Linkage Grant LP0990134. Our particular thanks go to R. Fletcher (DTMR) and B. Robertson
(DETE) for their vision regarding the importance of this work. We would also like to thank Prof G.
Marks, Dr P. Robinson, Prof K. Mengersen, Prof G. Ayoko, Dr C. He, Dr G. Johnson, Dr R. Jayaratne, Dr
S. Low Choy, Prof G. Williams, W. Ezz, S. Clifford, M. Mokhtar, N. Mishra, R. Laiman, L. Guo, Prof C.
Duchaine, Dr H. Salonen, Dr X. Ling, Dr J. Davies, Dr L. Leontjew Toms, F. Fuoco, Dr A. Cortes, Dr B.
Toelle, A. Quinones, P. Kidd and E. Belousova, Dr M. Falk, Dr F. Fatokun, Dr J. Mejia, Dr D. Keogh,
Prof T. Salthammer, R. Appleby and C. Labbe for their contribution to the UPTECH project.
120
5.6. References
1. Stevens, B. and G. Feingold, Untangling aerosol effects on clouds and precipitation in a buffered system. Nature, 2009. 461(7264): p. 607-613.
2. Lohmann, U. and J. Feichter, Global indirect aerosol effects: a review. Atmospheric Chemistry and Physics, 2005. 5(3): p. 715-737.
3. Pope II, C.A. and D.W. Dockery, Health effects of fine particulate air pollution: lines that connect. Air & Waste Management Association, 2006(56): p. 709-742.
4. O'Dowd, C.D., et al., Marine aerosol formation from biogenic iodine emissions. Nature, 2002. 417(6889): p. 632-636.
5. Kulmala, M., et al., Toward direct measurement of atmospheric nucleation. Science (New York, N.Y.), 2007. 318(5847): p. 89-92.
6. Petäjä, T., et al., Sub-micron atmospheric aerosols in the surroundings of Marseille and Athens: physical characterization and new particle formation. Atmos. Chem. Phys, 2007. 7: p. 2705-2720.
7. Stanier, C.O., A.Y. Khlystov, and S.N. Pandis, Nucleation Events During the Pittsburgh Air Quality Study: Description and Relation to Key Meteorological, Gas Phase, and Aerosol Parameters Special Issue of Aerosol Science and Technology on Findings from the Fine Particulate Matter Supersites Program. Aerosol Science and Technology, 2004. 38(S1): p. 253-264.
8. Woo, K., et al., Measurement of Atlanta aerosol size distributions: observations of ultrafine particle events. Aerosol Science and Technology, 2001. 34(1): p. 75-87.
9. Salma, I., et al., Production, growth and properties of ultrafine atmospheric aerosol particles in an urban environment. Atmospheric Chemistry and Physics Discussions, 2010. 10(6): p. 13689-13721.
10. Cheung, H.C., Z.D. Ristovski, and L. Morawska, Observation of new particle formation in subtropical urban environment. Atmospheric Chemistry and Physics Discussions, 2010. 10(10): p. 22623-22652.
11. Mejía, J.F. and L. Morawska, An investigation of nucleation events in a coastal urban environment in the Southern Hemisphere. Atmospheric Chemistry and Physics, 2009. 9(20): p. 7877-7888.
12. Modini, R., et al., New particle formation and growth at a remote, sub-tropical coastal location. Atmos. Chem. Phys, 2009. 9(19): p. 7607-7621.
13. Donaldson, K., et al., The pulmonary toxicology of ultrafine particles. Journal of aerosol medicine, 2002. 15(2): p. 213-220.
14. WHO, W.H.O., Air Quality Guidelines: Global Update 2005: Particulate Matter, Ozone, Nitrogen Dioxide, and Sulfur Dioxide. 2005: World Health Organization Europe.
15. Ning, L., et al., Ultrafine Particulate Pollutants Induce Oxidative Stress and Mitochonrial Damage. Environmental Health Perspectives, 2003. 111(4): p. 455.
16. Kulmala, M., L. Pirjola, and J.M. Makela, Stable sulphate clusters as a source of new atmospheric particles. Nature, 2000. 404(6773): p. 66-69.
17. Seinfeld, J.H. and S.N. Pandis, Atmospheric chemistry and physics: from air pollution to climate change. 2006, Hoboken, N.J: Wiley.
18. Kulmala, M., How Particles Nucleate and Grow. Science, 2003. 302(5647): p. 1000-1001. 19. Wehner, B., et al., The contribution of sulfuric acid and non‐volatile compounds on the
growth of freshly formed atmospheric aerosols. Geophysical research letters, 2005. 32(17). 20. Ristovski, Z., et al., The role of sulphates and organic vapours in growth of newly formed
particles in a eucalypt forest. Atmos. Chem. Phys, 2010. 10(6): p. 2919-2926. 21. Wang, L., et al., Atmospheric nanoparticles formed from heterogeneous reactions of
organics. Nature Geoscience, 2010. 3(4): p. 238-242. 22. McMurry, P.H., et al., A criterion for new particle formation in the sulfur‐rich Atlanta
atmosphere. Journal of Geophysical Research: Atmospheres (1984–2012), 2005. 110(D22).
121
23. Lazaridis, M., New particle formation of ternary droplets in the atmosphere — a steady-state nucleation kinetics approach. Atmospheric Environment, 2001. 35(3): p. 599-607.
24. Zhang, Q., et al., Insights into the Chemistry of New Particle Formation and Growth Events in Pittsburgh Based on Aerosol Mass Spectrometry. Environmental Science & Technology, 2004. 38(18): p. 4797-4809.
25. Creamean, J.M., et al., Measurements of Aerosol Chemistry during New Particle Formation Events at a Remote Rural Mountain Site. Environmental Science & Technology, 2011. 45(19): p. 8208-8216.
26. UPTECH. Available from: http://www.ilaqh.qut.edu.au/Misc/UPTECH%20Home.htm. 27. Salimi, F., et al., Spatial variation of particle number concentration in school microscale
environments and its impact on exposure assessment. Environmental Science & Technology, 2013.
28. Farhad Salimi, Z.R., Mandana Mazaheri, Rusdin Laiman, Leigh R. Crilley, Sam Clifford and Lidia Morawska, Assessment and Application of Clustering Techniques to Atmospheric Particle Number Size Distribution for the Purpose of Source Apportionment. Under review in Environmental Science and Technology.
29. Crilley, L.R., et al., Aerosol mass spectrometric analysis of the chemical composition of non-refractory PM1 samples from school environments in Brisbane, Australia. Science of The Total Environment, 2013. 458–460(0): p. 81-89.
30. Drewnick, F., et al., A new time-of-flight aerosol mass spectrometer (TOF-AMS)—Instrument description and first field deployment. Aerosol Science and Technology, 2005. 39(7): p. 637-658.
31. Jayne, J.T., et al., Development of an aerosol mass spectrometer for size and composition analysis of submicron particles. Aerosol Science & Technology, 2000. 33(1-2): p. 49-70.
32. Jimenez, J.L., et al., Ambient aerosol sampling using the Aerodyne Aerosol Mass Spectrometer. Journal of Geophysical Research: Atmospheres, 2003. 108(D7): p. 8425.
33. Dal Maso, M., et al., Formation and growth of fresh atmospheric aerosols: eight years of aerosol size distribution data from SMEAR II, Hyytiälä, Finland. Boreal Environment Research, 2005. 10(5): p. 323-336.
34. Pirjola, L., et al., FORMATION OF SULPHURIC ACID AEROSOLS AND CLOUD CONDENSATION NUCLEI: AN EXPRESSION FOR SIGNIFICANT NUCLEATION AND MODEL COMPRARISON. Journal of Aerosol Science, 1999. 30(8): p. 1079-1094.
35. Lehtinen, K.E.J., et al., On the concept of condensation sink diameter. Boreal Environment Research, 2003. 8(4): p. 405-411.
36. Fuks, N., A.G. Sutugin, and S. Soo, Topics in current aerosol research. 1971: Pergamon press. 37. Clement, C.F., M. Kulmala, and T. Vesala, Theoretical consideration on sticking probabilities.
Journal of Aerosol Science, 1996. 27(6): p. 869-882. 38. Willeke, K., Temperature dependence of particle slip in a gaseous medium. Journal of Aerosol
Science, 1976. 7(5): p. 381-387. 39. Bae, M.S., et al., Characteristics of size distributions at urban and rural locations in New York.
Atmospheric Chemistry and Physics, 2010. 10(10): p. 4521-4535. 40. Jimenez, J.L., et al., New particle formation from photooxidation of diiodomethane (CH2I2).
Journal of Geophysical Research: Atmospheres (1984–2012), 2003. 108(D10). 41. Ng, N., et al., Organic aerosol components observed in Northern Hemispheric datasets from
Aerosol Mass Spectrometry. Atmospheric Chemistry and Physics, 2010. 10(10): p. 4625-4641. 42. Ng, N., et al., Changes in organic aerosol composition with aging inferred from aerosol mass
spectra. Atmospheric Chemistry and Physics, 2011. 11(13): p. 6465-6474. 43. Ng, N., et al., Real-time methods for estimating organic component mass concentrations
from aerosol mass spectrometer data. Environmental science & technology, 2010. 45(3): p. 910-916.
122
44. Wood, S.N., Thin plate regression splines. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 2003. 65(1): p. 95-114.
45. Birmili, W. and A. Wiedensohler, New particle formation in the continental boundary layer: Meteorological and gas phase parameter influence. Geophysical Research Letters, 2000. 27(20): p. 3325-3328.
46. Hamed, A., et al., The role of relative humidity in continental new particle formation. Journal of Geophysical Research: Atmospheres, 2011. 116(D3): p. D03202.
47. Hussein, T., et al., Observation of regional new particle formation in the urban atmosphere. Tellus B, 2008. 60(4): p. 509-521.
48. Canagaratna, M.R., et al., Chase studies of particulate emissions from in-use New York City vehicles. Aerosol Science and Technology, 2004. 38(6): p. 555-573.
49. Jimenez, J.L., et al., Ambient aerosol sampling using the aerodyne aerosol mass spectrometer. Journal of Geophysical Research: Atmospheres (1984–2012), 2003. 108(D7).
123
Chapter 6: Conclusions
Urban aerosols originating from natural and anthropogenic sources can have
important adverse effects on the health of people and the environment. Understanding the
physical and chemical dynamics and transformations of urban aerosols helps us to
understand and mitigate their effects on climate and human health. This research program
has contributed significantly to scientific knowledge on the dynamics and transformation of
urban aerosols.
The overall aim of this research was to investigate the physical/chemical dynamics
and transformation of UFPs in urban environments. Firstly, this study investigated the
spatial variation of PNC in microscale environments and its impact on exposure
assessments. Then, the application of cluster analysis and GAM modelling for the
categorisation and parameterisation of PNSD data was assessed and they were both
employed for source apportionment purposes. Finally, the chemistry of NPF and growth
were analysed using data collected by the AMS.
This PhD project was part of a major project called UPTECH, which sought to
determine the effect of the exposure to airborne UFPs emitted from vehicles on the health
of children in schools. Two week measurement campaigns for a comprehensive number of
air quality parameters were conducted at each of 25 primary schools within Brisbane
metropolitan area. Air quality parameters were measured with a focus on UFPs at each
school. Data collected from the CPC, SMPS and AMS comprised the main portion of data
analysed in this PhD project. All the measurements were conducted in schools, however, the
results presented in this thesis have implications on wider urban population and thus is not
limited to schools’ environments. Original and interesting insights on the dynamics and
transformation of urban UFPs were obtained during this PhD project. These findings and
their significance are explained in further details in the following section.
6.1. Principal Significance of the Findings
Spatial variation of particles is an important factor in the transformation of aerosols
in the atmosphere and plays a key role in human exposure to aerosol particles. Spatial
124
variation in macroscale environments has been extensively studied, however, spatial
variation in microscale environments has not and there is currently no data available
regarding the spatial variation of PNC in microscale environments. Therefore, the first part
of this thesis, as described in Chapter 3, set out to determine the spatial variation of PNC
within microscale environments, as well as investigate the main parameters which affect
this variation. To achieve this aim, we measured and analysed PNC at three sites within each
of 25 schools across Brisbane, Australia. PNC was measured using TSI 3781 water-based
CPCs, however based on the issues we experienced with this instrument (as described in
Section 3S.2.1), use of this particular type of CPC for environmental monitoring, especially
for long term field measurements, is not recommended. Three CPC were set up side by side
to measure the same ambient air for three consecutive days, in order to find the instrument
uncertainty. A CV = 0.3 was found to be attributable to instrument uncertainty, since CV’s
were smaller than 0.3 for 95% of the time while the CPC’s were measuring from the same
aerosol.
The spatial variation of PNC for each school was quantified using CV’s, with a
correction factor threshold of CV = 0.3 to account for instrument uncertainty. Overall,
spatial variability was observed in 12 out of 25 schools, of which 3 schools were found to
require three monitoring sites, and 9 schools were found to require two sites, in order to
account for the effect of PNC spatial variation on exposure estimation. The association of
spatial variability with traffic density and wind speed/direction was investigated using
Spearman’s partial correlation coefficient and CV’s were found to be positively correlated
with traffic density and wind direction. Higher spatial variation was observed while the wind
was blowing from the direction of the adjacent road towards the school. The effect of using
a single monitoring site on exposure estimation was determined for sites that displayed
significant spatial variation. Ratios of mean concentration at each site to the site with the
lowest mean concentration were calculated and found to have an average value of 2, which
indicates that exposure estimation results were dependent on the location of the single
monitoring site for schools which displayed spatial variation.
To further analyse the impact of traffic density and wind speed on spatial variation, a
non-parametric regression model was fitted to the data. It was found that wind speed had
no effect on the spatial variation of PNC until it reached about 1.5 m/s, while above this
125
threshold, an increase in wind speed was found to reduce the spatial variation. Traffic
density had a significant effect on spatial variation and the amount of effect increased
linearly as the density increased from 0 to 70 (5min-1). The effect of traffic plateaued after
exceeding a traffic density of 70 vehicles per five minutes. These findings enhance our
understanding of the spatial variation of PNC within microscale environments and identify
its effect on exposure estimations. Based on these findings, there is a definite need to assess
spatial variation in the microenvironment in question prior to conducting measurements for
exposure estimation. In addition, the methods used in this study, namely those used to
quantify spatial variation, determine instrument uncertainty and investigate the parameters
affecting spatial variation, can be applied to other studies dealing with similar issues. Finally,
the results of the abovementioned part of the study can be used in construction of
microscale environments for exposure considerations.
PNSD was measured in short intervals for long time periods, in order to determine
the transformations, sources and origins of aerosol particles. This generally results in a large
amount of measured PNSD data, which makes their processing and analysis a challenging
task. Therefore, advanced mathematical and statistical methods need to be employed for
the purpose of long term PNSD data analysis. Clustering and GAM can be employed for
parameterisation and categorisation purposes, and they can potentially be used for source
apportionment as well. The application of some clustering techniques on PNSD data has
been studied in literature and K-means has been identified as the optimum method,
however, there are still several methods which have not been assessed. In addition,
application of the GAM for PNSD parameterisation has never been studied.
Therefore, in this thesis, the performance of three clustering techniques on PNSD
data measured during the whole UPTECH project, which included about 84,000 SMPS scans,
were compared with K-means using Dunn and silhouette validations values and K-means
was found to be the best method, with an optimum of five clusters. A new method based on
GAM was introduced for the first time in this thesis. This method was able to parameterise
the PNSD data with advantages over traditional multi-lognormal parameterisation. All of the
measured particle size data were clustered into five clusters using the K-means method and
they were all parameterised by the new approach which employed GAM. Each cluster was
attributed to its source and origin using the properties of the particle size data, diurnal
126
variation patterns and other air quality parameters. It was found that the five clusters could
be attributed to three major source/origins: 1) Regional background particles: Clusters 3 &
5, which included 49.8% of the total data, were attributed to regional background, with
clear modes at 60 nm and 40 nm, respectively. PNSD attributed to regional background in
this study had different modes compared to the one observed in London, but they were
similar to the observations in Barcelona. 2) Photochemically induced nucleated particles:
Clusters 1 & 2, which included 18.6% of the total data, were attributed to photochemically
induced nucleated particles. The same types of particles were observed in Barcelona for 4%
of the observations, but excluding measurements July, when photochemical nucleation is
expected to be more frequent. These types of particles were not observed in London, which
could be due to the long-time averaging of the data in that study and/or much lower
insolation. The classic approach (i.e. to find banana shaped new particle formation events),
which had previously been performed in the same environment, resulted in a lesser fraction
compared to the fraction of PNSD data related to nucleated particles in this thesis. This
indicates a much higher occurrence of new particle formation than what can be found using
classic approaches, because in most formation events, the newly formed particles would be
scavenged by pre-existing particles. 3) Vehicle generated particles: Cluster 4, which included
31.6% of the data, was attributed to vehicle generated particles. Similar types of particles
were observed in the literature. The cluster distribution at each of the 25 sites was
determined and the mixture of clusters at each site was in agreement with the location of
the site.
NPF events are one of the main sources of particles in the atmosphere which affect
the climate and human health. Despite significant progress in this area during recent years,
due to the development of new instruments, the physical and chemical mechanisms
involved in NPF events are not fully understood. The chemical composition of the species
involved in NPF events relates to the toxicity of newly formed particles and can determine
the mechanisms behind this phenomenon. Online measurements of the chemical
composition of particles with a high time resolution are possible by employing the AMS
instrument. In this thesis, the AMS was used to measure the chemical composition of
aerosol particles at five schools. The PNSD for these five schools were analysed based on a
standard criteria to look for NPF events and five NPF events were observed at S12 and S25.
127
The observed NPF events had growth rates of 3.3-4.6 (nm/hr) and they occurred on days
with lower humidity, and higher solar radiation and temperature. CS was calculated as a
measure of the surface area of the pre-existing particles and event days displayed a lower
CS compared with days when no events were observed. The mass concentration of
sulphates, ammonium, nitrates and organics increased significantly after the start of
nucleation (with the exception of ammonium). Trends in the mass fraction of each chemical
species were modelled using GAM. Ammonium, sulphate and nitrate mass precent fractions
increased before the start of nucleation, peaked around the start of nucleation and
decreased after the start of event. However, the organics percent fraction increased
significantly after the start of nucleation, showing the important role of organics in NPF
growth. In other words, aerosol particles consisted of more organics species as a result of
the NPF event.
f44 and f43 were analysed to investigate the role of organic aerosol, as the f44 vs f43
reveals oxidation level of the OOA. f43 and f44 both decrease after the start of nucleation,
while during non-nucleation days, only f44 decreased when the particles were vehicle
generated. Vehicle emission and NPF generated particles clustered in different locations on
the f44 vs f43 plot, demonstrating the application of this plot for source identification and
transformation of the OOA components.
Overall outline of this thesis is illustrated in Figure 6.1. In summary, the research
conducted in this thesis led to the following conclusions:
Significant spatial variation of PNC may be present in microscale
environments, depending on local traffic density and meteorological
parameters.
Wind speeds of higher than 1.5 m/s will result in less spatial variation in PNC,
while slower wind has no effect.
Traffic density of adjacent roads to a microscale environment correlates
positively with spatial variation of PNC until the density reaches about 70
(5min-1). Above this threshold correlation decreases.
128
Spatial variation of PNC can potentially affect the results of exposure
estimation by as much as an order of magnitude, depending on the location
of monitoring site, if a single monitoring site is in use.
K-means was found to be the optimum clustering technique for PNSD
compared with PAM, SOM and CLARA.
The GAM method was able to successfully parameterise the PNSD with
advantages over multi lognormal methods.
Measured PNSD were successfully attributed to three main sources/origins
including: 1) traffic generated; 2) urban background; and 3) newly formed
particles.
Organics played a key role in newly formed particle growth.
Both f44 and f43 decreased after the NPF events, while only f44 decreased
when the particles were generated by vehicle emissions.
Vehicle generated and newly formed particles clustered in different locations
in the f44 vs f43 plot.
f44 vs f43 was found to be an appropriate tool for identifying the source and
transformation of the OOA components.
129
Physical/chemical dynamics and transformations of
ultrafine particles
Spatial variation of particle
number concentration
Quantification of
spatial variation
Modelling variation
trend and its affecting
parameters
Significant spatial variations
in microscale environments
were identified
Important contribution of
sources other than traffic
Sources Chemistry of New Particle
Formation (NPF)
Cluster analysis of
PNSD data
Development of a
parameterisation
method
Three main sources/origins
were identified (traffic, NPF
and background)
Generalised additive model
was employed successfully
for the purpose of particle
size parameterisation
Chemical analysis of
NPF events
Modelling of chemical
composition trends
Important role of organics
in NPF events
Newly formed particles
clustered at different
locations in f44 vs f43 graph
Figure 6.1: Outline of the overall aims, methods and results of this thesis
130
6.2. Directions for future research
As stated in Chapter 3, the work presented in the 1st paper identified significant
spatial variation of PNC within microscale environments, such as schools. It was the first
study of its kind that confirmed the significant impact of spatial variation on exposure
assessments in such environments. This study was able to find a relationship between
spatial variation and its affecting parameters (e.g. meteorological conditions and traffic
density), however it was not able to draw comprehensive conclusions regarding the
conditions under which spatial variation can be expected at a microscale environment. This
information can be helpful when designing studies to investigate exposure to UFPs in
microscale environments. This is important and original research which needs to be pursued
in future. In addition, three outdoor sites were used for quantification of spatial variations in
this thesis, however, having more outdoor sites can result in a more comprehensive
assessment of spatial variation, which should be considered in future studies. In addition,
the results presented herein cannot be generalised to environments with different
meteorological and/or traffic conditions, therefore, studies in other types of environments
need to be carried out.
As described in Chapter 4, the work presented in the 2nd paper was able to
successfully categorise all SMPS scans measured during the UPTECH project into similar
groups using a clustering technique. However, the pattern of each particle size spectra is the
result of contributions of different sources which cannot be quantified by the method used.
Therefore, investigation of other techniques, such as Bayesian infinite mixture modelling,
Bayesian K-means, Bayesian Beta-process clustering and positive matrix factorisation can
enhance our understating further. In addition, clustering techniques can also be applied in a
temporal sense to separate days based on their diurnal pattern. This has never been studied
before and should be addressed in future research.
As described in Chapter 5, the 3rd paper identified the importance of ammonium
and sulphates in the early stage of nucleation events, while organics contributed
significantly to particle growth. The contribution of organics and their nature was
investigated further by calculating f44 and f43, and comparing their diurnal pattern for
nucleation and non-nucleation days. Since there were only five nucleation events observed,
131
we were not able to draw a comprehensive conclusion regarding the role of organics. In
addition, due to its limitation, AMS did not cover the first steps of nucleation. Therefore,
longer online measurements of chemical composition in the same environment using
instruments with lower detection limit are needed. Moreover, this study only focused on
urban areas and it is recommended that similar studies be conducted in other
environments, such as rural and coastal, to compare the chemistry of NPF events in those
areas with the results presented in this study.