Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Land use regression modelling of NO₂, NO, PM₂.₅ and black carbon in Hong Kong Lee, Martha 2016

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2016_november_lee_martha.pdf [ 8.01MB ]
Metadata
JSON: 24-1.0314324.json
JSON-LD: 24-1.0314324-ld.json
RDF/XML (Pretty): 24-1.0314324-rdf.xml
RDF/JSON: 24-1.0314324-rdf.json
Turtle: 24-1.0314324-turtle.txt
N-Triples: 24-1.0314324-rdf-ntriples.txt
Original Record: 24-1.0314324-source.json
Full Text
24-1.0314324-fulltext.txt
Citation
24-1.0314324.ris

Full Text

LAND USE REGRESSION MODELLING OF NO2, NO, PM2.5 AND BLACK CARBON IN HONG KONG by  Martha Lee  B.A., The University of British Columbia, 2013  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Occupational and Environmental Hygiene)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)   September 2016  © Martha Lee, 2016 ii  Abstract Land use regression (LUR) modelling is a common method for estimating pollutant concentrations. This project created two-dimensional LUR models for nitrogen dioxide (NO2), nitric oxide (NO), fine particulate matter (PM2.5), and black carbon (BC) for Hong Kong, a prototypical high-density high-rise city. Two sampling campaigns (April-May and November-January) were carried out in Hong Kong. Measurements of NO2 and NO (2-3-week averaged) and PM2.5 and BC (24-hour averaged) were adjusted for instrument bias and temporal variation, and offered to multiple linear regression models along with 365 potential geospatial predictor variables.  Variables were created from a number of geospatial metrics including land use and traffic variables (road length, average annual daily traffic [AADT], traffic loading [AADT * road length]).  Measurement averaged across both campaigns were: a) NO2 (M = 106 μg/m3, SD = 38.5, N = 95), b) NO (M = 147 μg/m3, SD = 88.9, N = 40), c) PM2.5 (M = 35 μg/m3, SD = 6.3, N = 64), and BC (M = 10.6 μg/m3, SD = 5.3, N = 76).  Thirty-six LUR models were created (4 pollutants * 3 combined and separate sampling campaigns * 3 traffic variable type).  The annual (combined values from both campaigns) road length models were selected as preferred models based on data reliability and overall model fit.  Road length, car park density, and land use types were commonly selected predictors in the final preferred models.  The preferred models had the following parameters: a) NO2 (R2 = 0.46, RMSE = 28 μg/m3) b) NO (R2 = 0.50, RMSE = 62 μg/m3), c) PM2.5 (R2 = 0.59; RMSE = 4 μg/m3), and d) BC (R2 = 0.50, RMSE = 4 μg/m3).  NO2 predictions were strongly influenced by traffic and higher around Kowloon and northern Hong Kong Island.  PM2.5 predictions had a strong northwest (high) to southeast (low) gradient.  BC had a similar gradient and high predictions around the port.   This matched with existing literature of spatial variation and sources in Hong Kong.  Spatial patterns varied by pollutant. iii  The success of this modelling suggests LUR modelling is appropriate in high-density high-rise cities.   iv  Preface The creation of two-dimensional LUR models for NO2, NO, PM2.5, and black carbon, as a means of exploring spatial variation in exposure to poor air quality in Hong Kong, was originally posed in the context of the development of the Hong Kong dynamic, three-dimensional (HKD3D) LUR model.  In this project, the two dimensional models described here are to be combined with information on pollution vertical structure, indoor infiltration and population mobility to form a more comprehensive model of population exposure to ambient air pollution.   Roadside sampling in Hong Kong was planned and carried out by myself, members of the Hong Kong University’s School of Population and Public Health (notably Crystal Choi, Anthony Tsui, and Dr. T.Q. Thach), members of the Department of Geography, Hong Kong University (notably Dr. Paulina Wong, Jenny Cheng, and Dr. P.C. Lai), Dr. Benjamin Barratt (the lead investigator for the HKD3D project), and Dr. Michael Brauer (my supervisor).  AECOM (Asia Company Ltd.) was hired as a contactor to physically mount the samplers on the lampposts and to drive to all sampling sites.  The sampling campaigns were carried out in conjunction with Hong Kong’s Environmental Protection Department.  I prepared Ogawa samplers for ion chromatography analysis, which was executed by the Environmental and Occupational Hygiene laboratory staff at the University of British Columbia. Laboratory analysis of the diffusion tubes samplers was conducted at Gradko Environmental in the United Kingdom.  PM2.5 and black carbon raw data reduction were performed by Dr. Benjamin Barratt.    v  Air pollution measurement adjustments for NO2 instrument bias and temporal variation, geographical information extraction, statistical analysis, model building and model validation, sensitivity tests, results interpretation, and thesis writing were conducted by myself with input from Dr. Michael Brauer, Dr. Benjamin Barratt, and others involved in the HKD3D project.     vi  Table of Contents  Abstract .......................................................................................................................................... ii Preface ........................................................................................................................................... iv Table of Contents ......................................................................................................................... vi List of Tables ..................................................................................................................................x List of Figures .............................................................................................................................. xii List of Symbols ........................................................................................................................... xiv List of Abbreviations ...................................................................................................................xv Acknowledgements .................................................................................................................... xvi Chapter 1: Introduction ................................................................................................................... 1 1.1 Study Area .................................................................................................................. 2 1.2 Outdoor Air Quality .................................................................................................... 4 1.2.1 General Characteristics of Hong Kong’s Outdoor Air Pollution ............................ 4 1.2.2 Health Impacts ........................................................................................................ 5 1.2.3 Determinants of Air Pollution Levels and Spatial Variation .................................. 6 1.2.3.1 Local Emission Sources .................................................................................. 7 1.2.3.2 Factors Affecting Transport and Dispersion ............................................... 13 1.3 Spatial Variation of Air Quality ................................................................................ 20 1.4 Land Use Regression Analysis ................................................................................. 23 1.4.1 Literature Review of Land Use Regression .......................................................... 23 1.4.2 Land Use Regression Methodology Overview ..................................................... 25 1.4.3 Comparison to Other Models ................................................................................ 26 vii  1.5 Research Goals and Rationale .................................................................................. 29 Chapter 2: Methods ....................................................................................................................... 31 2.1 Field Sampling .......................................................................................................... 31 2.1.1 Timeline of Campaigns ......................................................................................... 31 2.1.2 Sampling Equipment ............................................................................................. 33 2.1.3 Site Selection ........................................................................................................ 34 2.1.3.1 NO2 and NO Sites ......................................................................................... 35 2.1.3.2 PM2.5 and Black Carbon Sites ....................................................................... 37 2.1.3.3 Co-location and Duplicate Sites.................................................................... 38 2.1.4 Ogawa Badge Preparation..................................................................................... 38 2.1.5 Deployment Arrangement ..................................................................................... 39 2.1.6 Procedures for Data Quality Control .................................................................... 42 2.1.7 Post Campaign ...................................................................................................... 42 2.2 Data Calculations and Corrections............................................................................ 43 2.2.1 NO2 and NO .......................................................................................................... 43 2.2.1.1 Laboratory Analyses and Calculation ........................................................... 43 2.2.1.2 Corrections .................................................................................................... 44 2.2.2 PM2.5 and Black Carbon........................................................................................ 45 2.2.2.1 Corrections .................................................................................................... 45 2.3 Prediction Variable Extraction and Aggregation ...................................................... 47 2.4 Model Building ......................................................................................................... 53 2.4.1 Pre-Selection of Variables .................................................................................... 53 2.4.2 Selection and Review of Predictors ...................................................................... 54 viii  2.5 Evaluation ................................................................................................................. 55 2.5.1 Concentration Evaluation...................................................................................... 55 2.5.2 Model Prediction Evaluation ................................................................................ 55 2.5.3 Model Diagnostic .................................................................................................. 56 2.5.3.1 Diagnostic Plots ............................................................................................ 57 2.5.3.2 Variance Inflation Factor (VIF) .................................................................... 57 2.5.3.3 Moran’s I ....................................................................................................... 58 2.5.4 Sensitivity Tests .................................................................................................... 58 Chapter 3: Results ......................................................................................................................... 59 3.1 Correction Factors ..................................................................................................... 59 3.1.1 Blank Correction ................................................................................................... 59 3.1.2 Bias Correction ..................................................................................................... 60 3.1.3 Temporal Correction ............................................................................................. 61 3.2 Data Quality Checks ................................................................................................. 62 3.2.1 Data Loss and Outliers .......................................................................................... 62 3.2.2 Comparison with Chemiluminescence Monitors .................................................. 64 3.2.3 Duplicates ............................................................................................................. 67 3.2.4 Descriptive Statistics ............................................................................................. 67 3.2.5 Seasonality ............................................................................................................ 74 3.3 Model Results and Evaluation .................................................................................. 75 3.4 Sensitivity Tests ........................................................................................................ 85 Chapter 4: Discussion ................................................................................................................... 88 4.1 Basic Assessment of Exposure ................................................................................. 89 ix  4.1.1 NO2 ....................................................................................................................... 89 4.1.2 NO ......................................................................................................................... 90 4.1.3 PM2.5 ..................................................................................................................... 91 4.1.4 Black Carbon ........................................................................................................ 94 4.2 Performance of HK2D LUR Models ........................................................................ 97 4.3 Viability of LUR Modelling in Other High-Density High-Rise Cities................... 101 4.4 Limitations of the Study.......................................................................................... 102 4.5 Future Work ............................................................................................................ 103 4.6 Conclusion .............................................................................................................. 104 Bibliography ...............................................................................................................................105 Appendices ..................................................................................................................................125 Appendix A All roadside sampling sites...................................................................................... 125 Appendix B Site selection flowchart ........................................................................................... 128 Appendix C PM2.5 and black carbon data cleaning and scaling ................................................... 129 Appendix D Predictor variables for modelling ............................................................................ 132 Appendix E Pre-selection ranking script ..................................................................................... 133 Appendix F Table of a priori hypotheses ..................................................................................... 135 Appendix G Temporal correction multiplication factors for PM2.5 and black carbon ................. 137 Appendix H Temporal correction multiplication factors for NO2 and NO .................................. 140 Appendix I All 36 model results .................................................................................................. 141  x  List of Tables  Table 2.1 Possible prediction variables offered to variables selection procedure ........................ 48 Table 3.1 Blank correction factors ................................................................................................ 59 Table 3.2 Sample size breakdown ................................................................................................ 62 Table 3.3 Comparison of passive samplers and chemiluminescence monitors (without outliers) 65 Table 3.4 Descriptive statistics of pollutant concentrations ......................................................... 68 Table 3.5 Seasonality of pollutant statistics (in μg/m3) ................................................................ 74 Table 3.6 Preferred models’ results for gaseous pollutants .......................................................... 76 Table 3.7 Preferred models' results for particulate pollutants....................................................... 76 Table 3.8 Models' results for gaseous pollutants without New Territories sites .......................... 86 Table 3.9 Models' results for particulate pollutants without New Territories sites ...................... 86 Table 3.10 Natural log models' results for gaseous pollutants...................................................... 87 Table 3.11 Natural log models’ results for particulate pollutants ................................................. 87 Table 4.1 LUR modelling for Other Asian Cities ......................................................................... 99 Table C.1 Virkkula results (period = 20 minutes) ...................................................................... 129 Table I.1 NO2 annual models ...................................................................................................... 141 Table I.2 NO2 SC1 models.......................................................................................................... 142 Table I.3 NO2 SC2 models.......................................................................................................... 143 Table I.4 NO annual models ....................................................................................................... 144 Table I.5 NO SC1 models ........................................................................................................... 144 Table I.6 NO SC2 models ........................................................................................................... 145 Table I.7 PM2.5 annual models .................................................................................................... 146 xi  Table I.8 PM2.5 SC1 models........................................................................................................ 147 Table I.9 PM2.5 SC2 models........................................................................................................ 148 Table I.10 Black carbon annual models ...................................................................................... 149 Table I.11 Black carbon SC1 models ......................................................................................... 150 Table I.12 Black carbon SC2 models ......................................................................................... 151         xii  List of Figures  Figure 1.1 Regions and districts in Hong Kong .............................................................................. 3 Figure 1.2 Developed areas in Hong Kong ..................................................................................... 3 Figure 1.3 Major local emission sources in Hong Kong ................................................................ 7 Figure 1.4 Air pollutants emitted in Hong Kong in 2013 (EPD, 2015c) ........................................ 8 Figure 1.5 Location of Hong Kong’s air quality monitoring stations ........................................... 21 Figure 2.1 Sampling timeline ........................................................................................................ 32 Figure 2.2: HK2D and EPD campaign sites ................................................................................. 35 Figure 2.3 Ogawa badge mounting setup ..................................................................................... 39 Figure 2.4 SidePak and microAeth arrangement in monitoring boxes ......................................... 40 Figure 2.5 Sampler's lamppost deployment .................................................................................. 41 Figure 2.6 Monitoring boxes' lamppost setup ............................................................................... 41 Figure 3.1 Ogawa badge and diffusion tube comparison ............................................................. 60 Figure 3.2 Time series plot of temporal correction effectiveness at site WCA1 in SC2 .............. 61 Figure 3.3 Ogawa and chemiluminescence comparison for NO2 ................................................. 65 Figure 3.4 Diffusion tube and chemiluminescence comparison for NO2 ..................................... 66 Figure 3.5 Ogawa and chemiluminescence comparison for NO .................................................. 67 Figure 3.6 Boxplot of gaseous pollutant concentrations ............................................................... 68 Figure 3.7 Boxplot of particulate pollutant concentrations .......................................................... 69 Figure 3.8 Measured annual NO2 concentrations ......................................................................... 70 Figure 3.9 Measured annual NO concentrations........................................................................... 71 Figure 3.10 Measured annual PM2.5 concentrations ..................................................................... 72 xiii  Figure 3.11 Measured annual black carbon concentrations .......................................................... 73 Figure 3.12 NO2 prediction surface (preferred model) ................................................................. 78 Figure 3.13 NO prediction surface (preferred model) .................................................................. 79 Figure 3.14 PM2.5 prediction surface (preferred model) ............................................................... 80 Figure 3.15 Black carbon prediction surface (preferred model) ................................................... 81 Figure 3.16 NO2 residual spatial plot ............................................................................................ 83 Figure 3.17 NO residual spatial plot ............................................................................................. 83 Figure 3.18 PM2.5 residual spatial plot .......................................................................................... 84 Figure 3.19 Black carbon residual spatial plot .............................................................................. 84 Figure C.1 Comparison of co-location SidePak units prior to application of the ratification process......................................................................................................................................... 131 Figure C.2 Comparison of co-location SidePak units following ratification.............................. 131  xiv  List of Symbols  R2 – coefficient of determination μg – micrograms xv  List of Abbreviations  AQMS – air quality monitoring stations CV – coefficients of variation EPD – Environmental Protection Department HK – Hong Kong  HK2D – Hong Kong 2 Dimensional Land Use Regression HKD3D – Hong Kong Dynamic 3 Dimensional Land Use Regression  HKU – Hong Kong University LUR – land use regression MBE – mean bias error μg/m3 – micrograms per cubic meter NO – nitric oxide NO2 – nitrogen dioxide NOX – nitrogen oxides OGV – ocean going vehicles PM2.5 – particulate matter with an aerodynamic diameter < 2.5 μm RMSE – root mean square error SC1 – first sampling campaign SC2 – second sampling campaign SD – standard deviation TEA - triethanolamine UBC – University of British Columbia xvi  Acknowledgements  I wish to express my sincere gratitude to, first and foremost, my supervisor Dr. Michael Brauer.  His support, encouragement, patience, and knowledge paved the way for the completion of this thesis.  I could not have asked for a better supervisor.  I would like to thank my thesis committee, Dr. Benjamin Barratt and Dr. Ian McKendry for their time, comments, and encouragement.  I would also like to thank Dr. Barratt for all his work as PI of the HKD3D project and allowing me to carry out an internship in his lab.    I also thank everyone else involved with the HK2D project: Crystal Choi, Jenny Chung, Prof. Lai, Dr. Thach, Anthony Tsui and Paulina Wong.  I would like to thank CREATE-AAP for their funding as well as the Health Effects Institute, under program RFA13-1, which funded the fieldwork for this thesis.  Lastly, I would like to thank my family and friends for being so supportive.  In particularly I would like to thank my father for taking the time to read over this thesis.  1  Chapter 1: Introduction  The goal of this thesis was to produce two-dimensional ground level land use regression models (LUR) of Hong Kong for the following air pollutants: NO, NO2, PM2.5, and black carbon.  These pollutants are commonly used as air quality and health indicators. The models provide spatially continuous estimates of long-term (seasonal or annual average) concentrations of these pollutants to describe spatial variation throughout the territory.  From a human health prospective these models provide a proxy for human exposure to poor air quality and highlight areas of potentially high exposure.  This information is important when looking at the public health outcomes associated with poor air quality and has been used in a large number of epidemiologic studies in diverse locations (Abernethy et al., 2013; Allen et al., 2013; Amini et al., 2014; Brauer et al., 2006; Briggs et al., 2000; Chen et al., 2010; Dionisio et al. 2010; Eeften et al. 2013; Habermann & Gouveia, 2012; Hoek et al, 2008; Kashima et al., 2009; Ryan & LeMasters, 2007; Saraswat et al., 2013; Su, Brauer, Ainslie, Steyn, Larson, & Buzzelli, 2008; Tang et al., 2013; Wang, 2011).  These models were developed in the context of a larger project in which the spatial models are to be combined with measurements of vertical decay of pollutants, indoor infiltration and population mobility to develop a dynamic three-dimensional exposure model (HKD3D).  This overall model aims to provide a more complete view of human exposure in Hong Kong by considering the substantial proportion of the population residing in high-rise buildings and their movement throughout the territory.    This section provides overviews of the following:  the study area (Hong Kong); 2   outdoor air quality in Hong Kong;  the spatial variation in air quality in Hong Kong;   LUR analysis and modelling; and  the rationale for this study.  1.1 Study Area Hong Kong is a Special Administrative Region (SAR) of the People’s Republic of China and is located in the southeast corner of the Pearl River Delta (PRD). The SAR (~ 41,700 km2) includes the lowlands around the Pearl River estuary and a number of major industrial Chinese cities (Lo, et al., 2006).  Mountains encircle this region on all but the southern border where the Pearl River drains into the South China Sea (Ye et al, 2013). Hong Kong is located in the south of an estuary, near the mouth of the Pearl River.  Hong Kong is divided into three geographic areas: Hong Kong Island, Kowloon, and the New Territories (Figure 1.1).  Within these lie Hong Kong’s 18 districts.  Hong Kong had a population of 7,240,000 as of mid-2014 (Information Services Department, 2015a).  One of the most densely populated regions in the world, Hong Kong has an average population density of 6,690 people/km2 as of mid-2014 (Information Services Department, 2015a).  Due to the clustering of developments, less than 25% of the total territory of 1,104 km2 is developed (Figure 1.2), population density can be significantly higher, e.g. 57,250 people/km2 in the district of Kwun Tong (GovHK, 2015; Information Services Department, 2015a).   3   Figure 1.1 Regions and districts in Hong Kong  Figure 1.2 Developed areas in Hong Kong CENTRAL& WESTERN EASTERNSOUTHERNWANCHAIKOWLOONCITYKWUNTONGSHAMSHUI POWONGTAI SINYAUTSIMMONGISLANDSKWAITSINGNORTHSAI KUNGSHA TINTAI POTSUENWANTUEN MUNYUENLONGDistricts of Hong KongNew TerritoriesKowloonHong Kong IslandSitesBuildingsUndeveloped landHong Kong4   The clustering effect is further enhanced by the prevalence of high-rise buildings in Hong Kong.  Hong Kong is a high-density high-rise city. It has the second highest number of high-rise buildings for a city, after Moscow, and has one of the highest ratios of high-rises to population (Emporis, 2016a).  The resultant densification can increase population exposure to poor air quality since people and emission sources (e.g. traffic) are concentrated in the same areas and associated urban build up may limit dispersion of air pollutants.    1.2 Outdoor Air Quality The poor air quality in Hong Kong is characterized by high concentrations in a number of major pollutants. The overlap of a large population with this poor air quality gives rise to significant public health risks. Air quality varies spatially in Hong Kong largely due to spatial variation in local emission sources, the inflow of regional pollution, and interactions with the temporal effects of meteorology.  These factors lead to differences in exposure throughout the population.   1.2.1 General Characteristics of Hong Kong’s Outdoor Air Pollution Poor air quality is a concern in Hong Kong.  Pollutants regularly monitored throughout the territory to assess air quality include:  nitrogen dioxide (NO2);  ozone (O3);  sulphur dioxide (SO2);  carbon monoxide (CO); and  particulate matter (PM10 and PM2.5). 5  Of these pollutants nitrogen dioxide (NO2) and particulate matter (PM)1 are the largest contributors to overall air pollution levels in Hong Kong (Lu et al., 2011).   According to the World Health Organization (WHO) the annual average PM10 concentration in Hong Kong was 49 µg/m3, in 2013, and the PM2.5 was 29 µg/m3, in 2014 (World Health Organization [WHO], 2016).  These numbers are in excess of the WHO’s annual mean air quality guidelines for PM10, 20 µg/m3, and PM2.5, 10 µg/m3 (WHO, 2014).  They are close to the WHO’s short-term 24-hour exposure guideline of 50 µg/m3 and 25 µg/m3 respectively (WHO, 2014). In comparison, Vancouver, BC had an annual average PM10 concentration of 12 µg/m3 and a PM2.5 concentration of 7 µg/m3 in 2013 (WHO, 2016).  Average annual NO2 concentrations over the last 10 years measured at both rooftop (50 - 60 µg/m3) and roadside (90 – 130 µg/m3) air quality monitoring stations in Hong Kong are consistently greater than the WHO’s annual mean air quality guideline of 40 µg/m3 (Environment Bureau [ENB], 2013; School of Population and Public Health, 2016; WHO, 2014).    1.2.2 Health Impacts At current levels, air pollution concentrations pose a significant public health risk.  Different studies have found that the ambient concentrations of NO2, O3, SO2, PM10, and PM2.5 in Hong Kong are linked to increases in hospital admissions, mortalities, and symptoms of respiratory (Gao et al., 2014; Ko et al., 2007; Lai et al., 2010; C. Wong et al., 2002; Wong et al., 2001;                                                  1 PM refers to solid particles or liquid drops in the atmosphere.  Particle size is generally inversely linked to the location of deposition within the respiratory system, with smaller particles penetrating into the lower airways with greater impact on health risk.  PM with a diameter of 10 µm or less is referred to as PM10 or coarse particulate matter.  Less than 2.5 µm, is referred to as PM2.5 or fine particulate matter. 6  Wong et al., 1999; T. Wong et al., 2001; Yu et al., 2001) and cardiovascular (Lin et al., 2012; C. Wong et al., 2002; Wong et al., 2001; Wong et al., 1999; T. Wong et al., 2002) diseases.  Based on these local studies there appears to be a significant relationship between NO2, O3, SO2, PM10, and PM2.5 and respiratory and cardiovascular diseases.  The magnitude of the measures of association do not strongly indicate any of the pollutants being more influential on health impacts.  Authors of these studies and additional reviews found the relationship between air pollution levels and health outcomes in the local Hong Kong studies match those seen in other studies both in terms of the direction and magnitude (Brajer et al., 2006; Ko et al., 2007; Lin et al., 2013; C. Wong et al., 2002; T. Wong et al., 2002).  C. Wong et al. (2002) suggested that traffic may be a particularly influential emission source in reference to health risks due to the strong association between hospital admissions for respiratory conditions and NO2 and PM10 found in their study. This fits with other literature however the study was not able to provide strong evidence for this connection (C. Wong, 2002).    1.2.3 Determinants of Air Pollution Levels and Spatial Variation In the long term, concentrations of air pollution in an area are driven largely by a combination of emissions, transport, and dispersion.  Local emissions refer to the pollutants being produced in the region.  Emission inventories reflect regulations and local economies as they report amount of pollutants released from given sources.  Transport and dispersion are connected to atmospheric science.  Transport of pollutants in and out of a region is linked to synoptic scale meteorology (e.g. prevailing winds), while dispersion is largely dependent on synoptic, meso- and microscale meteorological phenomenon.  These factors also drive the spatial variation in air quality, as the effects of these factors are not felt evenly over the territory. 7   1.2.3.1 Local Emission Sources Local and regional emissions affect air quality in Hong Kong (Lau et al, 2007; Louie et al., 2005; So & Wang, 2003; Yuan et al, 2013; Zheng et al, 2011).  A joint study sponsored by the Hong Kong government using 2006 data, suggested that 36% of poor air quality events (132 days in 2006) were primarily due to regional air pollution while 59% (192 days) were primary due to local emissions (Lau et al., 2007). Any investigation of air quality in Hong Kong needs to consider both regional and local sources and the spatial locations of the latter (Figure 1.3).  Figure 1.3 Major local emission sources in Hong Kong  8  According to the Hong Kong emission inventory (ENB, 2013; Environmental Protection Department [EPD], 2015c), the major local air pollution source categories are:  marine transport;  vehicle traffic; and   power stations.   These sources account for over half of the SO2, NOX, PM10, PM2.5, and CO emitted locally (EPD, 2015c; Figure 1.4).  Civil aviation, other fuel combustion activities, and non-combustion activities are also sources (EPD, 2015c).  Since 1997, local emissions have declined for all pollutants the inventory covers (EPD, 2015c).   Figure 1.4 Air pollutants emitted in Hong Kong in 2013 (EPD, 2015c)  Marine Transport Marine transport is the largest local emitter of SO2, PM2.5, PM10, and NOX (EPD, 2015c).  This is in part due its proportional increase rather than total emission increase; marine emissions from 1997 to 2013 have been relatively stable in the face of decreasing power station and traffic 020000400006000080000100000120000SO2 Nox PM10 PM2.5 VOC COTonnes2013 Emission InventoryNon CombustionOther Fuel CombustionCivil AviationNavigationRoad TransportPublic Electricity Generation9  emissions (EPD, 2015c).  Within marine transport, the largest emissions contributors are ocean-going vessels (OGV) (EPD, 2015c).  In 2011, OGV contributed 79% of the SO2, 43% of the NOX, and 69% of the PM10 of shipping emissions (ENB, 2013).  The prevalence of OGV emissions is due to Hong Kong being a major port city and the historically having only limited regulations for air emissions on OGVs (ENB, 2013).  The lack of regulations allows and inadvertently promotes the use of low-grade fuel, use of less efficient engines, and poor maintenance of engines.  Marine fuel tends to be low-grade with a relatively high sulphur content (estimate 2.8%) compared to other more refined fuels (Lack et al., 2011; ENB, 2013). In comparison, fuel for motor vehicles has sulphur content capped at 0.001% in Hong Kong (EPD, 2009; Information Services Department, 2015b).  The use of this low-quality fuel leads to higher emissions of SO2 and PM (Winnes, 2010).  PM from marine emissions is composed mainly of organic carbon, elemental/black carbon, sulphates, and inorganic metal containing compounds (Kasper et al., 2007; Moldanová et al. 2013; Winnes, 2010).    The concentration of OGVs at the container port has the potential to be a major point source of emissions. According to Lau et al. (2004, as cited in Lau et al., 2005) Hong Kong’s container port is located in the same airshed as 3.8 million inhabitants. Shipping lanes are likely another area of significant emissions.  Since July, 2015 (within the period of this thesis) the Hong Kong government mandated that OGVs must use low-sulphur fuel (<0.5% by weight), liquefied natural gas, or an approved alternative fuel that releases at the SO2 at the same reduced levels as low 10  sulphur marine fuel while berthing (EPD, 2015c; EPD, 2015f).  This suggests emissions from the port but not necessary from shipping lanes may have decreased.  Vehicle Traffic Hong Kong had 2,095 km of roads with 699,540 licensed vehicles as of the end of 2014 (Transport Department, 2015).  The average vehicle density was therefore 334 vehicles per km of road.  The number of motor vehicles in Hong Kong is increasing. There was a 2.7% increase from 2013 to 2014 (Transport Department, 2015).  It is predicted that vehicle registration will continue to rise until 2017 (ENB, 2013).  Motor vehicle emissions were considered, until recently, to be the main local source of air pollution in Hong Kong.  Over the last couple of decades, vehicle emissions have been decreasing, despite increases in the number of vehicles and vehicle-kilometer-travelled (EPD, 2015c; Transport Department, 2015).  Between 1997 and 2013, the traffic contribution to the emissions inventory for all six pollutants included (NOX, PM10, CO, SO2, VOC, and PM2.5) decreased (EPD, 2015c).  The largest decrease (98%) was SO2 while the smallest decrease (36%) was NOX (EPD, 2015c).  These decreases were likely linked to more stringent regulations regarding fuel and vehicle emission standards (EPD, 2009; Information Services Department, 2015b; Tian et al., 2011).  The government has also encouraged industries to move from diesel to vehicles powered by liquefied petroleum gas or electricity, to replace older commercial vehicles, and to replace catalytic convertors and oxygen sensors (END, 2013; EPD, 2015c; EPD, 2009; Information Services Department, 2015b; Tian et al., 2011).    11  While roadside NOX concentrations and ambient concentrations of NO2 have decreased, roadside NO2 concentrations still continue to rise, ~20% increase since 1999 (END, 2013; Tian et al., 2011).  A study by Tian et al. (2011) suggested the increase in the total fraction of primary NO2 in NOX emissions from vehicles may have, at least in part, given rise to this phenomenon.  Primary NO2 is emitted directly from vehicles rather than formed in the atmosphere.  The majority of NO2 is formed by secondary production in the atmosphere (Tian et al., 2011). The study’s results indicated an increase in the ratio of primary NO2 emissions from 2% (1998) to 13% (2008) over the three roadside sites sampled.  The study suggested that current strategies for reducing vehicles, though successful for reducing NOX, should be tailored to reducing NO2 as well.  While marine transport is now considered a larger source of air pollution, motor vehicles remain the significant contributor (END, 2013).  In addition, motor vehicles emissions are an important source when looking at human exposure since human populations live in close proximity to roadways.  Elevated concentrations along roadsides can impact indoor air quality of buildings within the immediate area (Morawska et al., 1999).  Exposure to these pollutants is also high for drivers, pedestrians, or other individuals who spend time near or on roads, e.g. roadside merchants.    Power Plants Currently, there are four power plants functioning in the Hong Kong territory operated by the CLP and HK Electric (END, 2013).  These plants are all located in the New Territories, the outlying, relatively sparsely populated area of the Hong Kong territory. Power plants are located 12  on Lamma Island, Lautau Island, and in the district of Tuen Mun (CLP, 2010; HK Electric Investments Limited, 2014; Wang et al., 2001).  The station on Lautau Island is a backup facility (CLP, 2015; CLP, 2014).    The plants generate electricity by burning coal and natural gas.  Coal is the dominant fuel; it accounted for 71% of the fuel used in 2011 (END, 2013). The major emissions from coal-powered plants are SO2, NOX, and PM10 (EPD, 2015c).  The power plants are located outside the core of Hong Kong but still impact air quality in this region and throughout the wider territory (END, 2013; Guo et al., 2004; Wang et al., 2001).    Hong Kong has been working to reduce emissions by banning the construction of new coal burning power plants since 1997, imposing emissions caps beginning in 2005, and forcing companies to use cleaner burning fuels and emissions control equipment (EPD, 2015d; EPD, 2015c).  Between 1997 and 2013, these efforts reduced levels of SO2 (73%), NOX (38%), and PM10 (64%), despite increasing electricity demands (END, 2013; EPD, 2015c).  There was a slight increase in NOX, CO, and VOC emissions from 2010 – 2013 due to increased coal burning in this period (EPD, 2015c).   Civil Aviation Aviation accounted for less than 6% of total emissions locally in 2013 (EPD, 2015c).  There had been an increase in SO2 and NOX emissions from 1997 to 2013 of about 120% from this sector, which was about the same as growth in air traffic in the territory during this period (EPD, 2015c).  Studies by Hudda, et al. (2014), Unal et al. (2005), K. Yu et al. (2004) have suggested airports and flight routes may impact regional air quality even if they make up only a small proportion of the local emissions.   13   Other Fuel Combustion Sources Other fuel combustion sources accounts for a significant percentage of PM10 and PM2.5 emissions, 14% and 16% respectively in 2013 (EPD, 2015c).    Emissions from this sector have decreased in the last two decades, most notably in SO2 production where there was a 95% drop from 1997 to 2013 (EPD, 2015c).  CO, NOX, PM, and VOC have also dropped significantly during this period ranging between ~20 to ~30% (EPD, 2015c).  Non-mobile machinery, i.e. off-road machinery using internal combustion engines, is the largest contributor within this category, particularly construction machinery (END, 2013; EPD, 2015c).  The Air Pollution Control (Fuel Restriction) Regulation, introduced in 2008, capped commercially and industrially used diesel sulphur content at 0.005% and has significantly helped decrease SO2 emissions from this sector (EPD, 2015c).    Non-combustion Sources This category refers to sources that release air pollution through non-combustion processes such as mechanical (e.g. quarrying and explosives) or off-gassing (e.g. from paints and consumer products).  This category is noteworthy as it accounts for substantial contributions to emissions of VOCs, PM10, and PM2.5, in 2013 58%, 16% and 10% respectively (EPD, 2015c).  There have also been cuts to emissions from this category in a large part due to regulations (EPD, 2015c).    1.2.3.2 Factors Affecting Transport and Dispersion Meteorology drives regional transport of and dispersion of pollutants into the greater atmosphere.  Synoptic (global), meso (regional), and micro (local) meteorological scales all 14  affect air quality in Hong Kong.  The interactions amongst these atmospheric scales and air pollution concentrations are complex and not always well understood.  This section presents a basic overview of meteorology and atmospheric phenomena that potentially affect air pollution levels in Hong Kong and relate to later selection of variables and modelling decisions.   Synoptic Scale Meteorology Synoptic scale meteorology covers large-scale weather systems.  Synoptic meteorological conditions tie Hong Kong’s air quality to the greater regional air quality from the PRD and beyond.  The effects of regional air pollution sources on Hong Kong is worse in winter and less in the summer (Lau et al., 2007; Louie, et al., 2005; Wang & Lu, 2006; J. Yu et al., 2004).  The direction of the prevailing winds is the major causal mechanism creating this seasonality as it changes the source region of air parcels transported into Hong Kong, and with that, imported air pollution.  Relative humidity and atmospheric stability also play significant roles.  The spring (March and April/May) and fall (September and October) act as transition seasons between the dominant summer and winter seasons (Louie et al., 2005; J. Yu et al. 2004).  During this time both marine and continental air masses affect synoptic conditions and prevailing winds are from the east (J. Yu et al. 2004; Wang et al, 2001).  Asian Monsoons dominate climate in this part of the world. During winter, the radiative cooling of the Asian landmass creates a high-pressure system that persists over the continent leading to a dominant anticyclonic outflow from the continent of dry, cool air.  In Hong Kong, this produces strong winds from the north and weak/moderate winds from the northeast during the winter (Louie, et al., 2005; Murakami, 1979). In turn, this produces mass air transport into the territory 15  from the higher latitudes of the Asian landmass, typically northern China (J. Yu et al., 2004).  These air masses are transported along the Chinese coastline or through central China, into the PRD, and then funnelled by meteorological and topographic conditions into Hong Kong (Louie, et al., 2005; J. Yu et al., 2004).  These transit regions are known for high air pollution concentrations.  Transport of the aged pollution into Hong Kong adds to local emissions and increases pollution concentrations.  Because the northern prevailing winds occur mainly in the winter and due to the presence of large emissions sources in the Pearl River Delta and beyond, regional pollution has the largest impact on Hong Kong air quality during this time (Lau et al., 2007), compared to summer when prevailing winds bring air masses from over the South China Sea.    In addition to regional input, winter air quality in Hong Kong is affected by reduced removal and dispersion rates of pollutants compared to summer.  Lower relative humidity means that particulate matter and soluble gases are less likely to be washed out by precipitation (wet deposition).  More stable atmospheric conditions and a shallower boundary layer are found in the winter (Chiu & Lok, 2011; EPD, 2003) This has the net effect of decreasing the volume in which concentrations of pollutants can be mixed and potentially diluted.   The meteorology reverses in the summer.  A low-pressure system is established over the continent due to the radiative heating of the landmass and produces cyclonic winds that draw warm, humid ocean air inland.  From May to August, the prevailing winds in Hong Kong are from the south and southwest (Louie et al., 2005; J. Yu et al., 2004).  These winds bring in cleaner marine air masses from lower latitudes of the South China Sea (Louie et al., 2005; J. Yu 16  et al., 2004). The transport of relatively clean air masses dilutes local emissions and transport pollutants out of the territory. This means local sources contribute a greater mean proportion of total pollutant concentrations during the summer (Lau et al., 2007). The high relatively humid air mass generates rainfall in the territory, allowing wet disposition (Chiu & Lok, 2011; So & Wang, 2003).  Sixty-five percent of Hong Kong’s precipitation occurs in the summer (J. Yu et al., 2004). Greater atmospheric instability during the summer means a higher boundary layer with more volume to disperse pollutants (Chiu & Lok, 2011).    In general, air quality in summer is better than in winter.  J. Yu et al. (2004) found that organic carbon aerosol concentrations were at least two times higher during the winter months than summer months. Using 24-hour samples taken every six days at three sampling sites (a roadside, an urban, and a rural), Louie et al. (2005) found similar results for PM2.5 with concentrations ranging from 33-69 μg/m3 in the winter compared with 15-51 μg/m3 in the summer.  Fall had the next highest concentrations (27-56 μg/m3).  Wang and Lu (2006) also found generally better air quality in the summer as the average mean Air Quality Index (AQI) value was ~60 while in winter the values were between 75 and 80.  These findings matched health studies that found increases in hospitalization associated with air pollution during winter months, particularly on days with cool and dry weather, though other factors, such as use of air conditioning and time spent outdoors, may confound this relationship (Qiu et al., 2012; Wong et al., 2001).  Seasonal variations in concentrations and sources are significant and must be considered when sampling and modelling air quality.   17  Seasonal patterns have not been detected for all pollutants. Locally emitted pollutants appear to be less affected by seasonal variation (Louie et al., 2005; J. Yu et al., 2004).  An example is black carbon2.  J. Yu et al. (2004) and Louie et al. (2005) found little seasonal variation in elemental carbon (EC) concentrations throughout the territory as a whole.  In a study by J. Yu et al. (2004), winter EC values ranged from 4-8 μg/m3 while summer values ranged from 2-9 μg/m3.   They did, however, find seasonal difference in prevailing winds affected where EC emitted from the port would be found in the territory.  These effects were limited to areas close to the port.  Louie et al. (2005) found EC concentrations ranged from 2-19 μg/m3 in winter and from 1-22 μg/m3 in summer, at their three monitoring sites.  This lack of seasonal variation suggests local emissions of EC are relatively constant seasonally and, excluding the area around the port, seasonal meteorology does not significantly impact spatial variation in long-term EC concentrations (Louie et al., 2005; J. Yu et al. 2004).  Mesoscale Meteorology Underlying the synoptic conditions are mesoscale meteorological phenomena, such as land-sea breezes and mountain-valley breeze circulations (Guo, et al., 2013; Lo, et al., 2006).  These processes primarily affect ambient air quality by controlling dispersion of pollutants.  When synoptic winds are weak and the sky is clear, thermal mesoscale phenomena become more apparent as differential heating, driven by regional topography, is more apparent (END, 2013; Liu & Chan, 2002; Steyn, 2003).                                                    2 Black carbon is a term defining carbon that is measured optically through light attenuation.  Elemental carbon refers to the same pollutant but measurements are done thermal-optically. 18  The land-sea breeze circulation is the result of the difference in the specific heating capacities of water and land, and the ensuing pressure gradients attempt to equilibrate. During the day, the land’s surface is warmer than the ocean’s surface causing air over the land to warm and rise. The cooler air over the ocean is pulled inland to fill the void.  This is the sea breeze.  At night and in the early morning, this cycle reverses as the land cools more quickly than water.  This creates a land breeze that flows from the land surface out over the ocean.  When the land-sea breeze is well expressed in Hong Kong (during spells of low synoptic winds on days with clear skies), its circulations may not overlap significantly with the greater atmosphere and exchange with the greater atmosphere is limited (ENB, 2013). Pollutants become trapped on land and become concentrated.   The mountain-valley breeze circulation forms due to the temperature difference between mountains slopes and valley floors. During the day, air above the mountain slope warms more quickly causing air to flow upslope (valley breeze). At night, the slopes lose heat faster.  The cooler air descends downslope into the valley (mountain breeze). Limited research has been done on defining the mountain-valley circulation in Hong Kong. A study by Guo et al. (2013) suggested this circulation does, at a minimum, have an impact on the daily profile of gaseous pollutants in Hong Kong.  As 70 -75% of the Hong Kong territory is hilly terrain, this suggests elevation could be correlated with air quality with other factors such as difference in wind speed with elevation effecting this relationship (Guo et al., 2013; So & Wang, 2004).    Microscale Meteorology 19  Microscale meteorology is defined as meteorological phenomena with a horizontal extent of 1 km or less.  Within Hong Kong as with other large cities, urban morphology creates microscale phenomena by altering larger scale wind flow and increasing turbulence (Eeftens et al., 2013).  Street canyons are a particularly dominant urban feature in Hong Kong (END, 2013).   Street canyons are streets enclosed on both sides by tall buildings.  When wind flow is perpendicular to the orientation of the street canyon, a vortex (or vortices) form in the street canyon and circulations within the canyon are self-contained relative to atmospheric flow above the canyon (Eeftens et al., 2013; Gu et al., 2011; Vardoulakis et al., 2003).   Aspect ratio (building height divided by road width) is used to characterize street canyons, with ratios greater than 0.7 commonly used to indicate a street canyon (Allard & Ghiaus, 2012; Georgakis & Santamouris, 2008; Murena et al., 2009; Nakamura & Oke, 1988; Vardoulakis et al., 2003).  Though a vortex may form in situations below 0.7 (Murena et al., 2009).  Street canyons pose an air quality concern.  In cases of high of motor vehicle emissions and limited dispersion into the atmosphere, concentrations of air pollutants in street canyons can be significantly higher than the surrounding area (Zhou & Levy, 2008).  These areas are potential areas of high exposure for people on or along the roads and for people living in the buildings making up the street canyon.  A number of characteristic street canyons are found in Hong Kong (Tian et al., 2011).  A study by Rakowska et al. (2014) in central Hong Kong found that the presence of a street canyon significantly elevated NOX and black carbon concentrations.  Within the range of a few blocks 20  they found significantly higher NOX and black carbon levels on roadways lined with tall buildings than a nearby open roadway with about ten times more traffic.  Street canyons should be a consideration when looking at roadside air pollution levels.  1.3 Spatial Variation of Air Quality  Spatial variation in air pollution exposure within Hong Kong, particularly on a neighbourhood level, is not well described.  The few studies that examined spatial variation used only a few sampling sites, mainly the government’s air quality monitoring network (AQMS) that consist of 16 continuous monitors, 13 general and 3 roadside (Figure 1.5).  This dataset provides good temporal data as the monitoring record are continuous but has relatively poor spatial resolution because of the small number of monitors.   Monitoring networks tend not to provide an accurate representation of neighbourhood level spatial variation, a level important when looking at human exposure (Brauer et al., 2006). In part this is because they are generally set up to characterize typical urban background levels for regulation, have a limited number of samplers, and are, in the case of Hong Kong, at rooftop level.  Reliance on the monitoring network for data means only large-scale spatial variation are likely to be observed.  21   Figure 1.5 Location of Hong Kong’s air quality monitoring stations  Despite limited research, some general spatial trends have been observed, mostly at the territory level.  Broadly these high level trends reflect the effects of regional transport and proximity to local sources. Multiple studies reported poorer air quality in the northwest of the territory.  Compared to the eastern regions of Hong Kong, the western regions are influenced more by north and northeasterly synoptic winds and this had been suggested as a major contributor to this pattern (Liu & Chan, 2002).  Kok et al., (1997) sampled air quality around the territory from Oct. to Nov. in 1994.  They found the west and north regions of Hong Kong had poorer air quality during this time due to the mixing of regionally transported air pollution and locally emitted 22  pollutants. These results were supported by a study by Chiu & Lok (2011) that examined PM10 concentrations at the Hong Kong rooftop air quality monitoring stations between 2000 and 2008.  They observed high PM10 concentrations in the northwest and lower concentrations in the southeast and along the coast.  Based on the synoptic meteorology of the territory and the results from Chiu & Lok (2011), it is likely that this gradient is seasonal. Liu & Chan (2002), looking at a poor air quality event from Dec. 29 to 30, 1999, acknowledged poorer air quality in the west but also pointed to the land-sea breeze creating a convergence zone in this part of the territory. This latter effect allows concentrations to build.  Air quality spatially varies between urban and rural areas. A comparison of three rooftop sites representing different levels of urban build-up indicated higher O3 concentrations in the study’s northeast rural site but the lowest NOX, CO, SO2, and PM10 concentrations (So & Wang, 2003).  The reverse was true for the urban site despite total potential O3 levels (defined as O3 plus NO2 in the paper) being similar for all sites (So & Wang, 2003).  For some pollutants, local emissions appeared to contribute a larger proportion to total pollutant concentrations at urban sites whereas regional emissions affect rural sites (So & Wang, 2003); regionally transported pollutants have a more constant spatial distribution hence their dominant influence on rural air quality. Also, there are fewer local emission sources in rural areas (Qin et al., 1997).  Historically, the New Territories have had better air quality due to less urban development, however, as development within this area increases to match the rest of Hong Kong, air quality has declined (Chao & Wong, 2002).  Roads also affect the spatial distribution of air pollution.  Not surprisingly, roadside pollution concentrations (e.g. PM10, C3-C12 non-methane hydrocarbons, and NO2) are 23  notably higher than ambient concentrations in Hong Kong due to proximity to vehicle traffic (EPD, 2003, EPD, 2012; So & Wang, 2004; Tian et al., 2011).   In summary, air quality is worse in the northwest of Hong Kong and in urban locations, particularly by roadsides.  Air quality is better in the southeast and in coastal areas as well as in rural areas where vehicle densities are low.  Local emissions give rise to greater spatial variation since they have not had a chance to mix into the atmosphere.    In a territory with degraded air quality such as Hong Kong, knowledge of the spatial variation in exposure is important because differences in concentrations of pollutants can have significant public health implications.  Not knowing who is at risk, and what the magnitude of this risk is, is a problem for the development of public health policies and air pollution regulations.    1.4 Land Use Regression Analysis  1.4.1 Literature Review of Land Use Regression LUR analysis is a 'niche' modelling technique mainly associated with environmental exposure science and epidemiology. Originally termed regression mapping, it was first used in a SAVIAH (Small-Area Variations in Air Quality and Health) study by Briggs et al. (2000) to estimate long-term air pollution exposure for members of a cohort study.  Since then, the popularity of this method has grown. This procedure is mainly used to analyze spatial variation in air pollution concentrations as a means of predicting spatial variation in human exposure to poor air quality. When applying LUR analysis to air quality, the most commonly modelled pollutants are traffic-24  related including NOX, PM, and black carbon (Hoek et al, 2008).  Recently, LUR models are increasingly used to map other environmental risks such as temperature variation (Kestens et al., 2011; Zhou, et al., 2014), noise (Aguilera et al., 2015; Goudreau et al., 2014; Xie et al., 2011) and soil pollution (Deschenes et al., 2013).   As LUR modelling for air quality has become more popular, the predictor variables and range of areas modelled has expanded. Variables commonly used as predictors include traffic density, elevation, land use, emission sources, and population density (Amini et al., 2014; Brauer et al., 2006; Hoek et al., 2008; Ryan & LeMasters, 2007; Wang, 2011).  Recent models include more complex explanatory variables including meteorological variables, e.g. wind speed and direction (Abernethy et al., 2013; Su, Brauer, Ainslie, Steyn, Larson, & Buzzelli, 2008), and vertical dispersion predictors such as sky-view factor and building angle (Eeften et al. 2013), as well as building volume (Tang et al., 2013).    LUR development has traditionally been mainly focused on European and North American cities. However, a number of models have been developed for air pollution concentrations in cities from other parts of the world including: Africa (Dionisio et al. 2010), Asia (Allen et al., 2013; Amini et al., 2014; Chen et al., 2010; Kashima et al., 2009; Lee et al., 2012; Li et al., 2010; Li et al., 2015; Meng et al., 2015; Meng et al., 2016; Saraswat et al., 2013), and South America (Habermann & Gouveia, 2012).  Geographically, broadening the model’s application highlights the differences in the spatial metrics and pollution sources affecting exposure variations.  For example, European and North American cities generally have lower pollution and building densities, and fewer small-scale dispersed pollution sources (Saraswat et al., 2013). In some 25  cases, this may make the application of the prototypical LUR regression model problematic. In these cases, the addition of more explanatory variables may solve these issues.  1.4.2 Land Use Regression Methodology Overview The underlying principle for LUR is that the value of the response variable is empirically associated with quantitative spatial characteristics.  Therefore, the value of the response variable at a given location can be predicted by spatial predictors. In short, LUR is a form of multiple linear regression where variables represent a spatial aspect. Assuming that spatial predictors are available for the entire area of interest the regression equation allows for the estimation of concentrations at unmeasured sites (at individual points or as a surface) and provides insights into the effects of spatial predictors on a response variable and is represented by   𝐶𝑖 = 𝑎 + (𝑏𝑋1)𝑗𝑘 +  (𝑏𝑋2)𝑗𝑘 + ⋯ + (𝑏𝑋𝑛)𝑗𝑘 + 𝜀                          (Equation 1)  where:  Ci is the modelled levels at point i;  a  is the y intercept;  bjk is a weight (coefficients) for predictor j for buffer k;  Xijk is a value for variables j (a function of spatial metric for buffer k around point i); and  𝜀 is the error term.  LUR models can either be predictive or interpretive (Amini et al., 2014), although these are not mutually exclusive.  Predictive models aim to maximize the predictive power for the response variable, Ci.  Interpretive models maximize the accuracy of variable inclusion (Xi selected) and 26  regression coefficients (values of b), so the effects of the selected spatial metrics on the response variable are easier to interpret.  These models mainly differ in variable selection. Interpretive models rely on the existing literature to determine the variables for the final model and may be more conservative in predictor selection, potentially sacrificing R2 if needed.  In practice, LUR models are generally a mix of the two.  1.4.3 Comparison to Other Models LUR is not the only method for spatial estimating air pollution exposure at unmeasured locations.  Other methods include proximity models, geo-spatial interpolation, and dispersion models.  Proximity models estimate concentrations based on the distance from emission sources.  The underlying assumption is the receptor’s pollution exposure is a function of their distance from an emission source.  This is generally considered a primitive method of modelling exposure with a number of limitations and opportunities for error (Abernethy, 2012; Jerret et al., 2005; Zou et al, 2009a).  Proximity models only consider emission sources, and generally only the closest source.  This excludes emission sources further afield that may impact exposure as well as meteorological factors (such as transport and dispersion) that are known to affect air quality.  Emissions inventories are known to have inaccuracies as they are labour intensive and hard to verify (Brauer et al., 2006). This can lead to misclassification of exposure sources when formulating proximity models (Zou et al., 2009b).  Due to these issues and difficulties in assessing exposure risk, the results produced by proximity models are generally used as an 27  exploratory step or for relatively simple exposure and health effect assessment (Zou et al., 2009a).  Geo-spatial interpolation combines stochastic and deterministic statistical methods with monitoring data to estimate concentrations.  This method depends on spatial autocorrelation and assumes that values are spatially related.  Approaches include kriging, inverse distance weighting, and Theissen triangles.  A major limitation of this modelling method is the requirement for a significant number of sampling sites to capture spatial variation over a landscape as this is the only variable on which values are estimated.  The range of concentrations also tends to be amplified (Jerret et al., 2005).  To improve the performance of modelling, geo-spatial interpolation is sometimes combined with regression analysis. An example is regression-kriging in which the residuals of a regression model are interpolated over the study surface using kriging and then added back into the regression as a predictor variable (Araki et al., 2015).   Dispersion models apply known atmospheric processes to emissions inputs and simulate their concentrations over space and time.  These models are able to provide high resolution of concentration profiles without the use of large monitoring networks and calculate concentrations for different time frames.  This type of modelling has large data and computational requirements.  These are the main limitations.  Data are required on source emissions and location, background concentrations, meteorological conditions (e.g. wind direction, temperature, and atmospheric stability).  These data may not be available, and if they are there may be comparability issues. As more emissions data and meteorological processes are included in the model, the prediction accuracy should be improved. However, this also increases the computational power and 28  expertise requirements.  A potential source of error is the application of idealized atmospheric processes to define dispersion within the model. These idealized conditions may lead to an oversimplification of the model and not reflect real world conditions.    Compared to these other models, the main advantages of the LUR models are they provide a relatively robust method for spatial prediction while having a lower sampling effort/data requirement.  According to Hoek et al. (2008), LUR models performed better than or on par with geo-statistical methods and conventional dispersion models when predicting air pollution concentrations at unmeasured locations in a city, though this was based on a relatively small sample of studies.  Fewer sampling sites are required compared with geo-statistical modelling, as the resolution of spatial variation is a function of the spatial resolution of parameters and not the number of sites within the sampling network (Hoek et al., 2008; Wang, 2011).  This requires high-resolution spatial predictor data and takes advantage of its general availability in many cities.  LUR requires less data input and computational power than dispersion models, as the model does not directly include atmospheric processes.    There are limitations of LUR modelling (Hoek et al, 2008; Isakov et al., 2012; Jerrett et al., 2005).  They tend to only model long-term concentrations and only one pollutant per model.  Efforts can be made, through the selection of the pollutant data use or by modelling multiple pollutants, to reduce this limitation.  The created models tend to be city specific and not easy to transfer directly to other cities (Isakov et al., 2012; Jerrett et al., 2005).  Extreme local variations produced at close proximity to emission source may not be able to be captured by LUR modelling (Hoek et al., 2008).  This is primary a limitation of available input data (Hoek et al., 29  2008).  LUR predictions may also not be directly comparable to regulatory standards, because spatial campaigns tend not to have the temporal coverage to confidently calculate a concentration suitable for direct comparison with air quality standards for regulatory purposes (Hoek et al., 2008).  While these limitations should be considered, they are not at odds with the objective to this thesis.  1.5 Research Goals and Rationale Hong Kong has poor air quality, which poses a public health risk.  Modelling spatial variation in air quality and estimating human exposure allows for a better understanding the health risk and to inform the necessary policies and research to address air pollution levels.  The value of modelling exposure is not just because air quality is poor, but because Hong Kong is a high density, high-rise city featuring a predominantly vertical population growth. The combination of densification, vertically stratified growth and complex topography in the context of high levels of regional pollution places unique challenges on conventional estimates of exposure and applied air quality models, suggesting a more complex variation in pollution concentrations than seen in North American and European cities.  Findings from these cities may not be applicable to the Hong Kong content, or even to other Asian cities.  As cities with a high density of high-rise buildings are increasingly common in Asia, it is important to understand how spatial variation and exposure may differ in this environment.  Hong Kong serves as a prototype for the viability of LUR modelling of these cities as it is not only a high-density high-rise city, it is one of the most advanced examples of this city type.     30  The objective of this thesis is to create two-dimensional (ground level) LUR models for NO2, NO, PM2.5, and black carbon in Hong Kong.  These models would provide an understanding of the current spatial distribution of air pollutants and point to spatial features that are important determinants for pollutants. NO2, NO, PM2.5, and black carbon where selected for modelling as they are by-products of anthropogenic activities, and are commonly used as indicators of air quality and adverse health impacts. These pollutants should have different seasonal and spatial profiles due to emission sources, atmospheric chemistry, and atmospheric residence times. The model outputs would provide spatially relevant data on the distribution of pollutants and potential human exposure.   This study feeds into a larger study designed to create a dynamic three-dimensional model of Hong Kong (HKD3D).  This larger 3-D model will characterize vertical gradients through supplementary canyon campaigns while also looking to predict exposure based on movement of individuals throughout the territory. Inclusion of these factors will lead to an increase in the accurate accounting of human air quality exposure. This information can then be used to identify at-risk groups, and allow city planners and developers to use zoning laws to plan, retro-develop and develop the city in a manner that minimizes the health risks to its population.    An interpretive LUR model was deemed to be the best method for assessing air quality exposure based on the goals of this study, the available data, and the integration with the larger 3-D model.  It provided the best insights into the relationship between spatial variables and concentrations without requiring a large number of monitoring sites, unavailable datasets, and an a priori understanding of the underlying processes.   31  Chapter 2: Methods  For this thesis, the methodology involved five primary phases:    Field sampling;  Outcome variable preparation;  Prediction variable extraction and aggregation;  Model building and executing; and   Evaluation.  2.1 Field Sampling Two sampling campaigns, corresponding to warm and cool seasons, were run in Hong Kong to collect roadside NO2, NO, PM2.5, and black carbon concentrations for the two dimensional aspect of the HKD3D project (hereafter referred to as HK2D).  NO2 and NO were collected together, with NO2 results supplemented with data from the sampling campaign executed by the Hong Kong Environmental Protection Department (EPD).  PM2.5 and black carbon were sampled together, on a separate cycle from the NO2 and NO sampling, though the exposure period did partially overlap.   2.1.1 Timeline of Campaigns Air pollution concentrations in Hong Kong differ between the warm and cool season.  Sampling both seasons was necessary to capture the range of long term concentrations found in the territory.  To effectively sample this underlying trend, the first campaign (SC1) ran in the late spring/early summer, April 24, 2014 to May 30, 2014 (37 days), and the second in late fall/early winter, Figure 2.1. The second campaign (SC2) was split.  PM2.5 and black carbon sampling ran 32  from Nov. 18, 2014 to Jan. 6, 2015 (50 days).  NO2 and NO sampling ran from Jan. 3, 2015 to Jan. 26, 2015 (24 days).    Figure 2.1 Sampling timeline34  The campaign dates were chosen to cover seasonal fluctuation and to coincide with the EPD sampling campaigns. The HK2D campaigns were originally planned without knowledge of the EPD campaigns.  When it became known that there was a large-scale ground level sampling effort occurring around the same time with similar parameters as this study, an agreement with the EPD was reached to join the campaigns.  The produced a number of benefits in terms of reducing costs, synchronizing logistics, data sharing, and producing a more robust NO2 model because of the additional data for both parties.  This study was able to use the NO2 results from the EPD campaign and the HK2D NO2/NO samplers were deployed with the EPD samplers.  Using the EPD sampling sites made permitting for the sampling locations for this study easier.  However when the EPD delayed sampling from Nov. 2014 to Jan. 2015 due to public protests, consequent street closures, and changes in traffic patterns in Hong Kong, it meant SC2 sampling                                                  3 PM2.5 and Black Carbon are grouped as PM and NO falls under NOX and NO2. 4 Only shows the main EPD sampling campaign.  A few sites were resampled from May 13, 2014 to June 3, 2014 in the first campaign.  In the winter campaign a few sites were sampled from Dec. 29, 2014 to Jan. 19, 2015. 33  was split.  Deployment of the HK2D NO2/NO samplers was delayed to coincide with the EPD sampling campaign.  The PM2.5 and black carbon sampling period remained as initially planned, as modelling efforts for these pollutants are independent of NO2 and NO.  2.1.2 Sampling Equipment Four types of samplers were used for data collection in this study:   Ogawa badges (NO2 and NOX);   palmes-type diffusion tube (NO2);   SidePak (PM2.5); and   microAeth (black carbon).   Diffusion tubes were used by the EPD.  The HK2D campaign deployed Ogawa badges, SidePaks, and microAeth, plus of limited number of diffusion tubes which were deployed at a few rooftop AQMS during SC2 for specific use in model development.    Both Ogawa badges and diffusion tubes are passive samplers.  Passive samplers do not actively intake air and pollutants are collected using diffusion.  Ogawa badges are plastic cylinders with a glass fibre filter impregnated with an absorbent placed at the open ends of the cylinder.  In this study, the filter at one end captured NO2, while the other captured NOX.  NO concentrations were estimated by subtracting NO2 concentrations from NOX concentrations.  Diffusion tubes are acrylic tubes covered with a rubber cap containing an absorbent (Gradko, 2012). For NO2 capture (Ogawa badges and diffusion tubes), triethanolamine (TEA) is the absorbent (Durant et al., 2014; Gradko, 2012; Hagenbjörk-Gustafsson et al., 2010).  For NOX collection, Ogawa badge filters are impregnated with TEA and an oxidizing agent, 2-phenyl-4,4,5,5-tetramethylimidazoline-1-oxyl-3-oxide (Hagenbjörk-Gustafsson et al., 2010).  34   SidePaks and microAeths are active samplers.  Air is drawn into the sampler and data is continuously recorded. These samplers require a power source.  The SidePak is equipped with a laser photometer and PM2.5 mass concentrations are calculated based on light scattering (TSI, 2012).  The microAeth is equipped with a stabilized 880 nm LED light and photo diode detector (AethLabs, 2014).  It calculates real time black carbon concentrations by measuring the changes in light absorption of the LED light as it passes through a filter onto which captured particulate matter is deposited (AethLabs, 2014).    2.1.3 Site Selection Overall 100 sampling locations were utilized in this study (Appendix A   Not all pollutants were sampled at all these sites. For the EPD campaign, samples were collected at 173 roadside sites placed throughout the territory (Figure 2.2). These sites were predominantly located in urbanized areas and all were located at roadside.  The majority of the 100 HK2D sites were selected from these sites.  EPD sites close to bridges and elevated roads were removed, as they into introduced a vertical traffic component, leaving 90 remaining EPD sites.  The HKU Geography Department then selected another 10 sites (P-sites) to cover extremes in population density and annual average daily traffic (AADT).  From these 100 HK2D sites, the subsets for the Ogawa badge sampling locations and SidePak and microAeth sampling locations were chosen (Appendix B  as described in more detail below.   35   Figure 2.2: HK2D and EPD campaign sites  2.1.3.1 NO2 and NO Sites In the first sampling campaign, 40 Ogawa sites were selected from the 100 sampling locations, along with three roadside AQMS sites for co-location.  Ogawa site selection was based on geographic location, AADT, land use, and population density, aiming to capture a full range of values for these factors.  To systematically capture these factors, a five-step process was applied to the selection of monitoring at each site. Step 1: Cluster sites To ensure geographic coverage, thirteen spatial clusters were identified amongst the 100 sites.  The 14th cluster contained remaining non-clustered sites. One sampling station from each of the 14 clusters was selected.  Priority was given to sites in residential, commercial, or mixed land 36  use areas.  If multiple sites in the cluster were in these land use classes, selection from among these sites was random.  If none of the sites were located in the above land use areas, selection was random. Step 2: Paired sites Each of the ‘cluster sites’ were matched with one of a remaining 83 possible sites.  ‘Cluster sites’ with AADT values of 1,2,4, or 5 were paired with a site with the opposite AADT (e.g. AADT category 2 paired with 4) and the same land use classification.  ‘Cluster sites’ with AADT values of 3 were paired with another AADT level 3 site with a different land use.  If there was more than one potential pair for a ‘cluster site’, the population density categories had to be different. Random selection was used if multiple sites fit these further selection criteria or if criteria were not met.   Once a ‘paired site’ had been selected it was dropped from the pool of candidates, and could not be selected again. These criteria helped fill out the range of AADT, land use classes, and population density.   Step 3: Remaining P sites After the above process, all remaining “P” sites were selected. These sites were chosen to extend coverage to cover the full range of population densities and AADT values. In total seven P-sites were included.   Step 4: Remaining sites The remaining five sites were selected to balance representation of AADT, land use classes, and population density after accounting for all of the other sites previously identified.  The lowest representation of levels within these categories was identified.  A site matching that variable level was then selected (randomly if multiple sites were available).   37  Step 5: Spatial check Lastly, selected sites within 500 m of another site were reassigned to the closest unselected site. If it was not clear which site should be removed to remedy this proximity situation, one site was randomly selected.    In SC2, an additional 20 sites were selected from the 100 HK2D sites to add to the 40 Ogawa sites used in SC1.  Selection was based on the EPD’s diffusion tubes results from the first sampling campaign.   Within each district, the diffusion tube NO2 concentration range was calculated using all HK2D sites as well as only sites where Ogawa badges were also deployed. Districts were ranked based the difference between the two (i.e. ranked based on which districts the Ogawa badges did not capture the NO2 range measured by the diffusion tubes).  One site was randomly selected from each of the top ten ranked districts (10 sites selected).  This was then repeated with the top 7 and then top 3.  The goal of selection was to increase the range of measured NO concentrations.    2.1.3.2 PM2.5 and Black Carbon Sites The 100 HK2D sites were utilized for both the 24-hour and two-week SidePak and microAeth monitor deployments.  The twenty two-week sites were selected first.  The three roadside AQMS sites were selected for the remaining 17 x two-week sites dispersed amongst the 40 SC1 Ogawa sites.  To ensure these sites were spatially distributed, fourteen sites were selected from the previously described cluster sites. In cases where the cluster site had been replaced in Ogawa sampling sites because it was too close to another site, another site in the same cluster was randomly chosen.  The last three sites were selected to round out representation of land use: 38  specifically, government, recreational park, and industrial.  The eighty 24-hour sites were composed of the remaining 100 HK2D sites.   2.1.3.3 Co-location and Duplicate Sites In addition to the roadside sampling, samplers were placed at rooftop and roadside AQMS sites for comparison to reference measurements.  Ogawa badges, SidePaks, and microAeths were placed at the three roadside AQMS (Central, Causeway Bay, and Mong Kok).  EPD also deployed diffusion tubes at these sites.  Ogawa badges were placed at ten of the rooftop AQMS in SC1 and SC2 (Central/Western, Eastern, Kwai Chung, Kwun Tong, Sham Shui Po, Shatin, Tai Po, Tsuen Wan, Tung Chung, Yuen Long).  Diffusion tubes were deployed at these same rooftop sites in SC2.  In SC2, PM2.5 and black carbon monitors were co-located with the Tsuen Wan AQMS site for two weeks.  All rooftop sampling was used only for comparison of monitoring method and not used directly in modelling.  Duplicate Ogawa badges were deployed in both campaigns to assess measurement precision.  In SC1, eight of the Ogawa sites were randomly selected to house duplicates.  In SC2, there were three additional duplicate sites were added to match the additional Ogawa badge sites in this measurement period.    2.1.4 Ogawa Badge Preparation Ogawa badges were prepared at UBC prior to each sampling campaign.  NO2 and NOX filters were ordered from Ogawa & Co., USA.  All Ogawa badge parts were washed using de-ionized water and air dried prior to assembly.  Parts, including the coated collection filters, were 39  assembled as per the protocol of Ogawa & Co., USA (2006) and tweezers were used to avoid contamination.  Loaded samplers were directly transferred to sealable plastic bags, sealed in vials, and stored to a refrigerator.  As per the manufacturer’s protocol, samplers were used within 60 days (Ogawa & Co., USA, 2006).    2.1.5 Deployment Arrangement In SC1 77 Ogawa badges were used. This included Ogawa badges HK2D sampling sites (n=40), co-located with AQMS (n=13), and samplers used for quality control: field blanks (n=8), lab blanks (n=8), and site duplicates (n=8).  In SC2 106 Ogawa badges were used.  This included Ogawa badges HK2D sampling sites (n=60), co-located with AQMS (n=13), and samplers used for quality control: field blanks (n=11), lab blanks (n=11), and site duplicates (n=11).  Ogawa badges were deployed for between 15 to 21 days with white shelters to protect from sunlight and rain (Figure 2.3).    Figure 2.3 Ogawa badge mounting setup  40  A total of 13 SidePaks and 11 microAeths were used in this study, though not all were available at all times.  As there were more sites than monitors, they were rotated through both the 24-hour and two-week sites.  SidePaks and microAeths were deployed in pairs in waterproof monitoring boxes.  As both the SidePak and microAeth are active samplers, they were deployed with external battery packs (Figure 2.4).   Figure 2.4 SidePak and microAeth arrangement in monitoring boxes  The majority of the samplers were attached to lampposts (Figure 2.5 & Figure 2.6).  At sites where lampposts were not available, traffic signs were used.  Trees and portable posts were also used as a last resort. Roadside samplers were fixed approximately 2.5 m off the ground.  Samplers co-located with the rooftop AQMS were placed 0.5 to 1 m off the surface of the building rooftop.  Site Ogawa badges were oriented facing the road, the same direction as the EPD diffusion tubes.  Duplicate Ogawa badges were exposed at a 90° angle from the site sampler. The SidePak and microAeth were attached on the opposite side facing away from the road.  This was done for safety considerations, as the monitoring boxed required the individual to 41  be on the same side of the post as it was being mounted on.  Lastly due to distances that had to be covered and the amount of sampling equipment, a contractor was hired to drive staff between locations and to deploy the samplers.  The contractor was the same as used by the EPD for their campaign.    Figure 2.5 Sampler's lamppost deployment   Figure 2.6 Monitoring boxes' lamppost setup 42   2.1.6 Procedures for Data Quality Control Calibration, mass instrument co-location, and blanks were built into sampling procedures to ensure data quality.  SidePaks were zero calibrated using a HEPA filter prior to site deployment.  SidePaks and microAeth had their flow checked at least once every week. Mass co-location of all SidePaks and microAeths of at least 48 hours were carried out at the beginning and end of the sampling periods.  This allowed for inter-instrument correction.  Lab blanks were assembled with the other Ogawa badges but left at University of British Columbia (UBC) in a refrigerator.  Field blanks were transport around with the Ogawa badge samplers while they were deployed and then stored in the office, in the dark, during the sampling period.   The target number for lab and field blanks was ~20% of the number of Ogawa badges deployed at HK2D sites.  In SC1 there were eight field and lab blanks.  This target (12 lab and field blanks) could not be reached for SC2 due to a shortage of Ogawa badges so only 11 lab and 11 field blanks were used.  2.1.7 Post Campaign After completion of campaigns Ogawa badges were transported back to UBC.  As per the protocol of the manufacturer, samples were extracted within 14-21 days after exposure and refrigerated prior to extraction (Ogawa & Co., USA, 2006; Ogawa & Co., USA, 2014).  All samples were analyzed with 90 days of extraction (Ogawa & Co., USA, 2006).  Diffusion tubes were shipped to Gradko Environmental in the UK for processing and analysis.     43  2.2 Data Calculations and Corrections  2.2.1 NO2 and NO  2.2.1.1 Laboratory Analyses and Calculation Ogawa badge samplers were analyzed at UBC in the School of Occupational and Environmental Hygiene lab. after each campaign using ion chromatography.  Filters were extracted into 6 ml of distilled and deionized water.  The resulting solution was then filtered (Pall IC Acrodisc syringe filters with a diameter of 13mm and a 0.45 um pore size) strained and injected into a high-pressure ion chromatograph (Dionex ICS-900) equipped with a conductivity detector.  Anion concentrations are converted into NO2 and NO concentrations with standardized equations (Equations 2 and 3, respectively):  𝑁𝑂2[𝑝𝑝𝑏] =𝑀𝑎𝑠𝑠 𝑜𝑛 𝑁𝑂2𝑓𝑖𝑙𝑡𝑒𝑟[𝑛𝑔]∗ 𝛼𝑁𝑂2𝑠𝑎𝑚𝑝𝑙𝑖𝑛𝑔 𝑡𝑖𝑚𝑒 [𝑚𝑖𝑛]                     (Equation 2)  𝑁𝑂[𝑝𝑝𝑏] =(𝑚𝑎𝑠𝑠 𝑜𝑛 𝑓𝑖𝑙𝑡𝑒𝑟−𝑚𝑎𝑠𝑠 𝑜𝑛 𝑁𝑂𝑥𝑓𝑖𝑙𝑡𝑒𝑟) [𝑛𝑔]∗𝛼𝑁𝑂 𝑠𝑎𝑚𝑝𝑙𝑖𝑛𝑔 𝑡𝑖𝑚𝑒 [𝑚𝑖𝑛]                     (Equation 3)      where: mass on filter = nitrite *6ml*1000; αNO2 = 10000/((0.667 * P5 *relative humidity) + (2.009 * temperature) + 89.9); and αNO = 10000/((-0.78 * P *relative humidity) + 220). Temperature and relative humidity data was retrieved from 47 Hong Kong Observatory (2015b) weather stations and values at the sites were interpolated using inverse distance weighting.                                                  5 Dimensionless vapour pressure coefficient dependent on temperature (Ogawa & Co., USA, 2006) 44   The EPD and HK2D diffusion tubes were prepared and analysed by Gradko Environmental, sampling, by U.V. spectrophotometry (UV 04 Camspec M550) (Gradko, 2012).   2.2.1.2 Corrections Measured NO2 concentrations were corrected for concentrations captured outside of site exposure (blank correction), sampler differences (bias correction), and temporal variation (temporal correction).  Blank and temporal corrections were applied to the NO concentrations.   The average field blank concentrations for the sampling period were subtracted from the calculated concentration of each sampler.  The sampler correction accounts for the systematic sampling bias of the diffusion tubes relative to Ogawa badges.  Specifically, a zero-intercept linear regression analysis with Ogawa concentrations as the outcome variable was applied to the diffusion tube concentrations to normalize the two measurement datasets.  Temporal corrections remove the effects of temporal variation on captured concentrations, due to fluctuations in the baseline air quality in the territory.  It is necessary since NO2 and NO sampler exposure was staggered during an overlapping 3-4 week period. The temporal correction factor was calculated by dividing the baseline concentration for the sampling period by the baseline concentration for the period the sampler was deployed (Equation 4): 𝐶𝐹𝑇 =∑ 𝑀𝑒𝑎𝑛𝑆𝑃,𝑘11𝑘=111∑ 𝑀𝑒𝑎𝑛𝑅,𝑘11𝑘=111                     (Equation 4) where: 45  k is the air quality monitoring stations; meanSP,k is the averages for the sampling period for each station; meanR,k is the averages for the sampler’s exposure for each station; and CFT is the temporal correction factor. Baseline concentration were calculated from the average daily NO2 and NO concentrations from the urban rooftop AQMS.  2.2.2 PM2.5 and Black Carbon  2.2.2.1 Corrections PM2.5 concentrations were corrected for instrument bias (scaled bias correction) and for temporal variation.  Black carbon concentrations were corrected for filter loading and temporal variation. No instrument bias correction was applied to the black carbon.  During the co-location period the Hong Kong Government’s reference black carbon monitor was not functioning so no comparison dataset was available.     PM2.5 and black carbon raw data were cleaned, corrected for instrument bias and filter loading by the Environmental Research Group at King’s College London, as described in the Appendix C  Briefly, for PM2.5 and black carbon the first and last 5 minutes of site sampling were dropped to account to set-up and collection.  Spurious periods were manually identified, by examining the datasets in conjunction with the log sheets, and dropped from the dataset.  Spurious conditions were: a) values below -0.5 μg/m3, b) instability following filter changes (indicated by large deviation in values) c) knocks to equipment (indicated by a saw tooth pattern), or d) significant 46  noise that was not smoothed over by the minimum averaging period (15 minutes).  Linear regression was used to correct for inter instrument variation between SidePaks and variation between the SidePaks and the reference Filter Dynamic Measurement System monitors in the AQMS.  Using the data from the 48-hour co-location runs of all SidePaks and microAeths at the beginning and end of the sampling campaigns, instrument bias corrections were applied to the bias correction for the individual SidePaks.  The filter loading correction adjusted for the under sampling of black carbon that occurred as the microAeth filter collected more black carbon mass;  as an optical sampler, the microAeth uses light attenuation to calculate black carbon mass, however this relationship is not linear.  The more heavily loaded a filter becomes, the lower the decrease in light attenuation per additional unit mass of black carbon collected.  A k factor was calculated using the Virkkula method (Virkkula et al., 2007).    The temporal correction applied to PM2.5 was similar to that one applied to the NO2 and NO data, although in this case hourly averages from the 11 rooftop AQMS were used instead of a daily average.  Temporal corrections were specific to each sampling site.  For black carbon only one reference AQMS measurement location (Hok Tsui) was operating in Hong Kong throughout the full study period.  A second site (Tung Chung) was only operational for only a portion of the period.  Available measurements from both black carbon sampling stations were merged with the Hok Tsui site measurements adjusted to the higher concentration Tung Chung station values using a correction factor given by  𝐶𝐹𝑆𝑡𝑛 =  𝐵𝐶𝑇𝑢𝑛𝑔 ̅̅ ̅̅ ̅̅ ̅̅ ̅̅𝐵𝐶𝐻𝑜𝑘 ̅̅ ̅̅ ̅̅ ̅̅ ̅                     (Equation 5) where: 47   𝐵𝐶𝑇𝑢𝑛𝑔̅̅ ̅̅ ̅̅ ̅̅ ̅̅  is the averaged black carbon concentrations at Tung Chung;  𝐵𝐶𝐻𝑜𝑘 ̅̅ ̅̅ ̅̅ ̅̅ ̅ is the averaged black carbon concentrations at Hok Tsui; and 𝐶𝐹𝑆𝑡𝑛 is the correction factor adjusting Hok Tsui concentrations to Tung Chung concentrations. A centred 24-hour rolling average was applied to the merged reference monitoring station dataset.    For PM2.5 and black carbon, sites with less than 18 hours (75% capture rate) of data were removed.  Datasets from sites with over 35 hours of data were subset evenly into periods of 18 to 35 hours.  Hourly concentrations were then averaged into a single value for that sampling period.   At sites with more than 36 hours of sampling, one subset was randomly chosen.  2.3 Prediction Variable Extraction and Aggregation Candidate spatial metrics of interest were selected based on background studies, information from Hong Kong’s public policies on growth, development, and public health, and regulations on air quality.  Potential predictor variables (spatial metrics) were divided into two groups: a) variables representing a point value and b) variables representing the cumulative values of an area (buffer variables).    Point variables include: distance from features, coordinates, aspect ratio, flight path, and predicted NO2 concentrations.  Buffer variables include: all traffic, urban build-up, feature point and land use variables.  The same buffer radii were used for all buffered variables (25 m, 50 m, 100 m, 200 m, 300 m, 500 m, 1000 m, 1500 m, 2000 m, 3000 m, 4000 m, and 5000 m).   48   Spatial data layers populated with the candidate metrics of interest were examined using ArcGIS, v.10.1 and 10.2 (ESRI, 2012; ESRI, 2013).  Vector layers were preferred to raster layer, where possible, to increase accuracy in extractions. R (R Core Team, 2015) was used in some cases to aggregate the raw extracted data and calculate the values of each spatial matric at the sampling sites.  In total 48 categories of potential predictor variables were identified (Table 2.1).  In total 373 spatial predictor variables were calculated – 364 from spatial data layers and 9 from the NO2 LUR models solely for NO modelling (Appendix D  ). Table 2.1 Possible prediction variables offered to variables selection procedure Type Potential Variables  Source Annual Average Traffic Density (AADT) (sum in buffer) Expressway Transport department (Intelligent Road Network Package Annual Traffic Census) Main roads Transport department (Intelligent Road Network Package Annual Traffic Census) Secondary roads Transport department (Intelligent Road Network Package Annual Traffic Census) Road Length (sum in buffer) Expressway Lands department (RG1000 & B10000) Main roads Secondary roads Elevated roads Traffic Loading (AADT * road length in buffer) Expressway Lands department (B10000) & Transport department (Intelligent Road Network Package Annual Traffic Census) Main roads Secondary roads Urban Build-up  (area or volume/buffer area) Building volume density Lands department (B5000) Building area density Lands department (B10000) Population density Census and statistics department (SB_2011) On street parking density Transport department (Intelligent Road Network Package Road Network Data) Point feature Street intersections Transport department 49  Type Potential Variables  Source (count/buffer area) (Intelligent Road Network Package Road Network Data) Bus terminus density Lands department (B10000) Car park density Mini bus terminus density Temple density Food stall density Land Use (total area in buffer) Commercial HKU Geography Government Park Mixed Residential Open area Industrial Undeveloped lands Lands department (B10000) Value Extracted at Point Elevation Aster GDEM In aircraft flight path Civil Aviation department Aspect ratio Manually created Longitude Latitude Predicted NO2 (NO model only) HKD2D NO2 LUR Model Distance (in meters) - from nearest sampling sites (Euclidean and natural log)  Ferry Lands department (B10000) Cross boundary vehicle terminus Toll gates Incinerator Crematorium MTR line MTR stations Coastline Shipping lanes ERM  High volume unrestricted shipping lanes  Airport Manually created Port Shenzhen Power stations 50   General Buffer Variable Extraction Most of the buffered variables were created in the same manner, as little or no modification to the GIS layers was necessary prior to extraction of road length, urban build up (except population density), point features, and land use categories (Table 2.1).  Features of interest overlapping with sampling site buffers were extracted and attributed to the appropriate sampling site.  Extractions were in the form of:   area for land use and building area;   volume for building volume;  count for point features; and   length for road length.   A variable category was represented with either a density value (standardized value) or a total value (unstandardized values) based on what appeared to be commonplace in the literature.  To calculate density for building area, building volume, and point features, the summed values were then divided by the extraction buffer’s area.  AADT Extraction AADT values required a significant amount of modification.  Measured AADT values were limited to counting stations (n = 1643) and values were interpolated for the road network.  Count stations were spatially joined to the road network centreline.  This allowed for road type (an attribute of the road centreline layer) to be assigned to the stations and used to allocate AADT counts to the road type groups: expressways, main roads, and secondary roads.  Elevated roads are divided into these groups as there were not enough count stations on elevated roads for interpolation.  Empirical Bayesian kriging (in ArcGIS) was used to interpolate AADT for each of 51  the three road types (Association of Central Oklahoma Governments, 2013; Shamo et al., 2014; Zou et al., 2012).  This removed the spatial transition between smaller and larger road groups.  Interpolated AADT values were then joined back to the appropriate road type segment using the midpoint of the road segment as the extraction point.  For each site buffer, intersecting road segments were extracted and AADT values were summed by road type.   Traffic Loading Traffic loading was calculated from road length and AADT.  AADT, in a buffer, was multiplied by the matching road length type in the same buffer to determine the extent of traffic in the area.  Elevated roads were divided into the three AADT road groups.   Population Density Population counts from the 2011 census were available at the street block level (n = 4993).  Street block density was calculated by dividing the block’s population count by its area.  Final site population density values were calculated by the weighted average extracted values (weighted by area as a portion of total buffer area).   Coordinates and Elevation Extraction Latitude and longitude data was extracted using the ‘calculate geometry’ function in ArcGIS.  Elevation was extracted for the sampling site from a digital elevation model (DEM).      52  Flight Path Aircraft flight path was a binary variable indicating if the sampling site was within the flight path area or not.  A study by Hudda et al. (2014) mapped out elevated PM2.5 around LAX and found a trapezoidal plume extended 16 km with parallel edges of 6.5 km and 1.5 km.  These plume dimensions were used for this study to define the flight path variable.  While this plume was defined outside of Hong Kong, it was applied to this study as it is the best information available on the possible maximum extent of the flight path impact.  The airport is located on a man-made island and the surrounding terrain is relatively flat.  Aspect Ratio Extraction The aspect ratios of building height to road width were calculated manually in ArcGIS using the ruler tool.  Road width directly in front of sampling sites was measured (90° angle from the street).  Building heights at the site were noted from the building layers and were averaged.  Average building heights were then divided by the road width.    Distance Extraction For single or few point or area emission sources, distance seemed a better predictor compared to point density.  Euclidean distance variables were calculated for each sampling site to the nearest point or edge (for polygons) of the feature layer of interest.  Airport, port, shipping lanes, high volume shipping lanes, Shenzhen, and power stations layers were drawn in ArcGIS, since pre-existing layers where not found.  The power station at Penny’s Bay was not included in the power stations variable as it is a backup facility and has a production capacity an order of magnitude lower than the other power stations (CLP, 2014; CLP, 2015).  All other distance 53  variables were calculated using pre-existing point feature layers. Distance values were also natural log transformed to account for the decay of pollutants from their emission sources (Amini et al., 2014; Weichenthal et al., 2015).   2.4 Model Building For each of the pollutants, models were built using the SC1 data, the SC2 data, and the averaged SC1 and SC2 data (hereinafter referred to as annual) sites with measurements for both campaigns. Within each of these subsets, a model was built for each of the three traffic predictor types (i.e. road length, AADT, and traffic loading), resulting in a total of 36 (4 pollutants x 3 datasets x 3 traffic predictors) models.  Variable selection consisted of a pre-selection and a selection and review phase.  All models were built in R (R Core Team, 2015).  2.4.1 Pre-Selection of Variables The pre-selection phase narrowed the number of variables moved to the automated selection (Appendix E  ). All variables had to have at least two non-zero values. Within each of the buffered variables group (e.g. car park density or industrial land use) buffer radii were ranked by correlation (Pearson’s) with the pollutant subset being used in the model (SC1, SC2, or both).  The top ranked buffer radius was identified for each.  Any variables in that group highly correlated with the top-ranked buffer (r > 0.6) were dropped.  The second and third ranked (based on correction with pollutant) buffer radii for the group of the remaining variables were then identified.  The remaining buffer radii were dropped. For distance variables, either the Euclidean or natural log variable made it to the automated selection. The variable with the greater correlation with the pollutant was selected. 54   2.4.2 Selection and Review of Predictors The automated phase took the selected variables from the pre-selection phase and entered them through an automated variable selection, using the regsubset function in the leaps package in R (Lumley, 2009).  The regsubset function relies on a branch-and-bound algorithm to highlight major variable combination pathways, which are then tested (Goodenough et al., 2012; Lumley, 2009; McLeod & Xu, 2010).  This allows the search algorithm to simulate an exhaustive search but decreases the computational power required, though this process is still more computationally intensive relative than a stepwise regression selection.  The maximum adjusted R2 was used to select an optimal model from the regsubset function’s results.  The 1 to 10 rule (1 predictor variable for every 10 observations) was used to set the maximum number of predictor variables that could remain in the model.  As the goal was to create interpretable models, selected variables had to be consistent with priori hypotheses and be statistically significant.  Consistency with priori hypotheses meant the direction of the variable fit with the current literature (Appendix F  .    An example of inconstancy with priori hypotheses would be road length having a negative relationship with concentrations. After variables were checked, variables were also removed if p ≥ 0.10.  The remaining variables in the dataset were rerun through the regsubset function. Direction and statistical significant dropping was repeated till no variables were removed.  After, if any model had two buffers with radii within 1500 m with coefficients of opposite signs the one with the lower p-value was dropped.  Automated variable selection was run on the remaining variables.  This only affected variables without a predetermined direction (e.g. residential landuse). 55   2.5 Evaluation  2.5.1 Concentration Evaluation  NO2 and NO concentrations were compared to the concentrations captured by the AQMS chemiluminescene monitors linear regression and by mean ratio (Durant et al., 2014; Van Reeuwijk et al., 1998).  Ogawa precision was established by comparing duplicates at sites by calculating the coefficient of variation (CV) as given by  𝐶𝑉 =√∑ (𝑂𝑖 − 𝐷𝑖)2𝑛𝑖=12 ∗ 𝑛∑ (𝑂𝑖 − 𝐷𝑖)𝑛𝑖=12 ∗ 𝑛∗ 100                     (Equation 6)  where: n is the number of samplers; O is the site Ogawa sampler; and  D is the site duplicate sampler (Eeftens et al., 2012).  2.5.2 Model Prediction Evaluation Two methods of model evaluation were used: leave-one-out cross validation (LOOCV) and held-out evaluation (HEV).  LOOCV removes one sampling site (N-1) while holding the same predictor variables in the model.  The model beta coefficients, residuals, and R2 values are recalculated for all sites each time a site is dropped (N times). The LOOCV R2 (comparison between the observations and the predictions of the LOOCV runs) is used as the evaluation criteria. This is an internal validation method. Bias is a potential issue as the dropped 56  observations were already used to create the model.  Studies (Basagaña et al., 2012; Wang et al, 2012) have found that LOOCV leads to overly optimistic predictions of the model’s ability by inflating R2 values.    HEV holds out a set number of sites from model building. The model is used to predict concentrations at held-out sites.  Site observations and predictions are compared and the R2 is used as an indicator of the effectiveness of the model.  This is considered a stronger indicator of a model predictive power the evaluation dataset was not used to build the model.  LOOCV was applied to all models.  The HEV was applied only applied to the NO2 dataset as it included the largest number of measurement sites.  A random selection 20 sites were used as the held-out dataset.  2.5.3 Model Diagnostic Diagnostic tests were run on the final models to ensure all the necessary statistical assumptions for linear regression had been meet.  Linear regression models implicitly make assumptions for:  linearity;   homoscedasticity;  normality; and   independence of predictor variables (multicollinearity). In the case of these LUR model it was also assumed that there should be no spatial autocorrelation with residual.  These assumptions were tested using diagnostic plots, variance inflation factor (VIF), and Moran’s I.  57  2.5.3.1 Diagnostic Plots Residuals vs. fitted plots were used to examine model linearity and homoscedasticity.  For linearity and homoscedasticity to be meet the residuals should have an even distribution around the 0 line and have no apparent slopes or trends.  Q-Q plots were used to inspect the normality of residuals.  From these plots, skewness and kurtosis can be assessed. Treatment methods for non-normality include removing outliers (if appropriate) or transforming variables.  2.5.3.2  Variance Inflation Factor (VIF) VIF was used to assess multicollinearity.  VIF values range from 1 to infinite and are given by  𝑉𝐼𝐹 =11 − 𝑅𝑖2                     (Equation 7) where: 1 − 𝑅𝑖2 is the tolerance and  𝑅𝑖2 is ith predictor variable’s squared multiple correlation with the other predictor variables in the model (Amini et al., 2014; O’Brien, 2007). Within statistics and the field of LUR modelling (Amini et al., 2014), there is no consensus on the VIF cut-off for unacceptably high multicollinearity.  VIF values ranging from 4-10 are used in statistics as rules of thumb for signalling excess multicollinearity (O’Brien, 2007).  Within LUR modelling VIF cut-offs tend to be lower, with studies applying cut-offs of 2 (Clougherty et al., 2013), 3 (Eeftens et al., 2012; Gulliver et al., 2013), and 10 (Aguilera et al, 2008; Amini et al., 2014).  Based on this information a VIF of 3 or greater was chosen to trigger remedial action in the form of dropping the highest value VIF predictor.  58  2.5.3.3  Moran’s I Moran’s I was used to test for spatial autocorrelation in residuals.   2.5.4 Sensitivity Tests Two sensitivity tests were performed on all models.  The first removed New Territory sites from modelling to see if differences in the urban development patterns in New Territories affected model fits.  The second test was to use naturally logged concentrations and examine if this improved model fit.  59  Chapter 3: Results  The LUR models developed in this thesis demonstrate the relationships spatial features have on air pollutant concentrations in Hong Kong.  The chapter will present the results of each step, specifically, the Results section will address:  corrections for concentrations;  data quality checks;  concentration description;   model results and evaluation; and   sensitivity tests.  3.1 Correction Factors  3.1.1 Blank Correction Blank corrections factors were relatively small (Table 3.1).  Both the field and lab blanks had nitrite levels close to or below the limit of detection.  This indicated capture by the passive samplers during loading, transport, and lab preparation was minimal and concentrations were representative of capture at sites even without the blank correction.  The blank corrections represented air concentrations to be subtracted raw concentrations, not nitrite levels to be subtracted from extracted nitrite concentrations. Table 3.1 Blank correction factors  SC1 (μg/m3) SC2 (μg/m3) NO - Ogawa 0.9 0.0 NO2 - Ogawa 0.6 1.3 NO2 – EPD diffusion tube 1.06  1.06  NO2 – Rooftop diffusion tube - 0.05   60  3.1.2 Bias Correction The bias correction for adjusting diffusion tubes to Ogawa badges was 0.88.  The correction factor was based on the slope of the regression (forced through zero) for Ogawa and diffusion tube concentrations of NO2.  The large sample size (n = 91), the wide range of concentrations, and relatively robust fit of the linear regression line (Figure 3.1) suggested this correction was a reasonable approximation of the relationship between the samplers.  Figure 3.1 Ogawa badge and diffusion tube comparison  The bias correction factors for the Sidepak’s PM2.5 results (and the adjustment for filter loading in black carbon concentrations) were performed outside of this thesis (see Section 2.2.2.1). The corrected versions of these concentrations were used in the modelling.     0 50 100 150 200 250 300050150250Ogawa versus Tube NO2 ResultsTube - NO2 (mg/m3)Ogawa - NO2 (mg/m3) y = 0.88xR2 = 0.7961  3.1.3 Temporal Correction  Temporal correction factors varied for each pollutant.  PM2.5 and black carbon factors were substantial due to the relatively short sampling period compared to the overall sampling campaign (Appendix G   Since there was more overlap with the NO and NO2 sampling periods, the correction factors were not as significant (Appendix H  ).  Evidence for the effectiveness of temporal corrections using the 11 rooftop sites was only based on comparison of resampled PM2.5 sites due to the lack of resampling by passive samplers within campaigns.   On average, the temporal corrections reduced standard deviation (SD) by 6.6 μg/m3 (to an average SD of 4.4 μg/m3) at the two-week PM2.5 sites (Figure 3.2). These results suggested that the temporal correction based on the rooftop AQMS sites was effective in removing temporal variation, though it does not remove all variation. The black carbon corrections were less successful.  This correction factor was developed from data from only two monitoring sites.  On average, the temporal correction increases the SD of the black carbon concentrations by 0.76 μg/m3 when looking at the two-week sites.  Figure 3.2 Time series plot of temporal correction effectiveness at site WCA1 in SC2 Nov 21 Nov 26 Dec 01020406080Temporal Correction - WCA1DatePM2.5 (mg/m3)Temporally correctedNon corrected62   3.2 Data Quality Checks  3.2.1 Data Loss and Outliers Annual models had the following sample sizes: NO2 – 75 (plus n = 20 in held out validation dataset), NO – 40, PM2.5 – 64, and black carbon – 76.  These sample size were lower than intended due to data lose caused by samplers not being deployed, samplers going missing or issues with data capture, and outliers (Table 3.2). Table 3.2 Sample size breakdown  All pollutants were affected by samplers not deployed at selected sites.  One hundred NO2 sites (a combination of diffusion tube and Ogawa badge) were to be sampled.  However the Ogawa selection process knocked out P12 and P7 for being to close to other Ogawa sites and the EPD did not deploy diffusion tubes at site NTA1.    Planned Deployed Collected Modelled SC1 NO2 - Ogawa 43 43 43 43 NO2 - Tube 90 89 89 89 NO2 100 97 97 97  NO 43 43 43 43 PM2.5 100 84 70 70 Black carbon 100 84 76 76 SC2 NO2 - Ogawa 63 62 60 60 NO2 - Tube 90 89 85 85 NO2 100 97 95 95  NO 63 62 60 58 PM2.5 100 84 77 77 Black carbon 100 84 82 82 63  In the original proposal, 20 sites were to have two-week PM2.5 and black carbon sampling period.  This proved to be too difficult logically because of the limited number of samplers (particularly with water damage to samplers and resampling) and the start of another sampling campaign for the larger HKD3D immediately after the HK2D campaigns.  In the end only four two-week sites were sampled.    In SC2 there was a loss of two of the Ogawa samplers in SC2 (sites P11 and TWB7).  Three EPD samplers (TWB6, NTA3, and TPA3) in SC2 went missing or were listed as void samplers.   Two of these sites (NTA3 and TPA3) were covered by Ogawa badges.  A number of issues arose during sampling that affected data capture.  These occurred more during the first sampling campaign.  During SC1, weather, battery life, and battery connection were major issues.  In Hong Kong, heavy rains are common during the warm season and a number of the SidePaks and microAeth’s suffered water damage. This resulted in data loss from monitors breaking during deployment and the decrease in sensors to deploy.  Batteries were also an issue for the active samplers.  Limited storage of the built-in battery and poor connections to external battery packs lead to data loss, particularly in the beginning of SC1.  By SC2 these issues were mostly eliminated.  Data loss did not appear to be systematic and did not have bias results.   To make up for data loss, sites were resampled when possible.    Only two outliers were removed from the model input. These were the NO observations from MKA1 and CWA1 in SC2.  These observations were less than half of the values from the 64  roadside monitor and the SC1 Ogawa badge concentrations.  It appeared these two samples had been contaminated and were removed.     3.2.2 Comparison with Chemiluminescence Monitors  In each campaign, 13 Ogawa badges were co-located with the AQMS chemiluminescence monitors (three roadside AQMS and ten rooftop).  Diffusion tubes were also placed at the roadside AQMS in both campaigns and at the rooftop AQMS in SC2.  This provided an understanding of how the results collected during the campaign compare to results collected from the equipment used for regulatory monitoring.  In general, the results of the passive samplers and the chemiluminescence monitors had similar relative rankings of sites, however, the presence of a weak relationship suggested other factors were at play and samplers were not directly adjustable to each other.  The mean ratios of Ogawa over chemiluminescence were 0.92 (NO) and 0.89 (NO2), Table 3.3. One NO2 site (KTAQMS) from SC2 and two NO sites (CWA1 and MKA1) from SC2 were removed as outliers.  The ratios from these outliers suggest that the Ogawa badges under sampled the chemiluminescence monitors concentrations for both NO and NO2 by ~10% on average (Table 3.3).    There were large spreads and variations in the ratios from site to site.  This can be seen in the large standard deviation of the Ogawa/Chemiluminescence ratios, particularly for NO (Table 3.3).  On average, diffusion tube concentrations agreed with the chemiluminescence concentrations as their ratio was 1.01 (Table 3.3).  The standard deviation was large with 34.4% of ratios excepted to fall within a 0.78 to 1.24 ratio of chemiluminescence concentrations (Table 3.3).  65   Table 3.3 Comparison of passive samplers and chemiluminescence monitors (without outliers) Comparison Ratio R2 Regression equation n NO2 Ogawa/NO2 Chem 0.89 (SD = 0.25) 0.78 Chem = 0.60±0.14*Ogawa + 34.01±11.49 25 NO2 Tube/NO2 Chem 1.01 (SD = 0.23) 0.76 Chem = 0.41±0.13*Tube + 45.94±16.14 16 NO Ogawa/NO Chem 0.92 (SD = 0.36) 0.92 Ogawa = 0.71±0.10*Chem + 16.17±8.21 21  For NO2, the R2 was 0.78 for Ogawa badges and 0.76 for diffusion tubes.  Agreement between values was not perfect but still showed a linear trend (Table 3.3; Figure 3.3; Figure 3.4). The y-intercept value should theoretically be zero if the relationship between the instruments was linear as all the instruments measured zero when there were no pollutants. Forcing the regression line through zero fits conceptually but did not necessarily fit the data.    Figure 3.3 Ogawa and chemiluminescence comparison for NO2 0 50 100 150050100150AQMS versus Ogawa NO2 ResultsOgawa - NO2 (mg/m3)AQMS - NO2 (mg/m3)y = 0.60x + 34.0R2 = 0.7866   Figure 3.4 Diffusion tube and chemiluminescence comparison for NO2  This issue was less pronounced with the NO results, where the y-intercept was lower and the linear fit was stronger (Table 3.3; Figure 3.5).   The lack of correlation and large y-intercepts indicated that while the Ogawa results were within the correct range, they were not the directly representative of what the regulatory network captured.  While having NO2 concentrations in the model adjusted to the regulatory network would have been ideal, in this context adjusting either the Ogawa or diffusion tube data to the chemiluminescence results would have been misleading.  Given the current data any bias correction factor would have been weak and would have introduced significant error to the dataset.  The resulting concentrations would not have represented chemiluminescence concentrations and would have been moved away from Ogawa concentrations, a known measurement method.  The strong correlation between the Ogawa badges and diffusion tubes (Figure 3.1) and lack of agreement with the chemiluminescence results suggested the difference in pollutant capture methods may have played a large part in the poor agreement. 0 50 100 150 200 250050100150AQMS versus Tube NO2 ResultsTube - NO2 (mg/m3)AQMS - NO2 (mg/m3)y = 0.41x + 45.9R2 = 0.7667   Figure 3.5 Ogawa and chemiluminescence comparison for NO  3.2.3 Duplicates Duplicates Ogawa badges and diffusion tubes were deployed at randomly selected sampling sites in both campaigns.  Overall the within site agreement at these sites was acceptable.  For the Ogawa badges, the coefficient of variation (CV) for NO2 was 10.8% (n = 18) and 13.6% for NO (n = 15).  Three sites were dropped as outliers in SC2 for NO (SSPB7, NTB6, and P9). Diffusion tubes from the EPD campaigns should have had CV of 10% or less as sites with CV greater than 10% had the outlier dropped.  3.2.4 Descriptive Statistics The averaged SC1 and SC2 concentrations captured for NO2, NO, PM2.5 and black carbon across both sampling campaigns were 106, 147, 35, and 11 μg/m3, respectively (Table 3.4).  Concentration distributions for all the pollutants were skewed slightly right (Figure 3.6; Figure 3.7).  Results from the Anderson-Darling normality test suggested the NO (p = 0.067) and PM2.5 (p = 0.30) distributions did not significantly depart from a normal distribution as the null hypothesis that the data was normal could not be rejected.  For the NO2 (p = 0.00065) and black 0 50 100 150 200 250050100200AQMS versus Ogawa NO ResultsOgawa - NO (mg/m3)AQMS - NO (mg/m3)y = 0.71x + 16.2R2 = 0.9268  carbon (p = 0.00031) distributions the null hypothesis of normality was rejected.  Comparison with the boxplots agrees, as the right skew to the data more notable then in the other pollutants.   The highest concentrations of gaseous pollutants were found in Kowloon and Hong Kong Island (Figure 3.8; Figure 3.9).  This pattern was not reflected for PM2.5 and black carbon concentrations where a number of the highest concentrations are found in the northern regions of the New Territories (Figure 3.10; Figure 3.11).   Table 3.4 Descriptive statistics of pollutant concentrations  NO2 (μg/m3) NO (μg/m3) PM2.5 (μg/m3) BC (μg/m3) Mean 105.8 146.7 35.1 10.6 Median 98.4 130.5 34.5 9.5 Minimum 42.9 20.8 24.8 3.5 Maximum  212.7 376.1 51.4 27.7 SD 38.5 88.9 6.3 5.3 n 95 40 64 76     Figure 3.6 Boxplot of gaseous pollutant concentrations NO2 NO50100150200250300350Gas PollutantsConcentrations in mg/m369   Figure 3.7 Boxplot of particulate pollutant concentrations PM2.5 BC1020304050Particulate PollutantsConcentrations in mg/m370   Figure 3.8 Measured annual NO2 concentrations 71   Figure 3.9 Measured annual NO concentrations 72   Figure 3.10 Measured annual PM2.5 concentrations 73   Figure 3.11 Measured annual black carbon concentrations  74  3.2.5 Seasonality While PM2.5 concentrations were statistically and substantively higher during the cold season sampling campaign, this trend was not as apparent for the other pollutants sampled (Table 3.5).    Table 3.5 Seasonality of pollutant statistics (in μg/m3)  NO2 (SC1) NO2 (SC2) Mean 108.5 103.2 Median 101.1 97.1 Minimum 37.2 40.9 Maximum  223.8 212.0 SD 43.5 38.2 n 95 95 Paired t-test results t(94) = 1.98, p = 0.051  NO (SC1) NO (SC2) Mean 156.0 137.4 Median 139.9 118.8 Minimum 22.5 11.4 Maximum  362.8 450.0 SD 95.1 94.5 n 40 40 Paired t-test results t(39) = 1.79, p = 0.081  PM2.5 (SC1) PM2.5 (SC2) Mean 23.4 46.8 Median 22.1 45.7 Minimum 11.0 30.3 Maximum  49.4 73.1 SD 6.8 9.4 n 64 64 Paired t-test results t(63) = -17.80, p < 0.001  BC (SC1) BC (SC2) Mean 10.4 10.8 Median 8.2 10.3 Minimum 1.2 2.2 Maximum  36.6 25.9 SD 7.6 5.3 n 76 76 Paired t-test results t(75) = -0.47, p = 0.638  75  3.3 Model Results and Evaluation In total 36 models were developed, with nine models for each of the pollutants (Appendix I  ).   These models were evaluated for model fit, meeting linear regression assumptions, and predictive power.  Model fit was evaluated with R2, adjusted R2, RMSE, MBE, and MAE.   R2 values across all the models ranged from 0.27 to 0.63.  In general, higher R2 values were noted in the SC1 PM2.5 models while the lower R2 values were noted in the SC1 black carbon models.  The mean R2 values (based on all nine models for each pollutant) were 0.43, 0.49, 0.56, and 0.43 for NO2, NO, PM2.5, and black carbon, respectively.  The RMSE measures the absolute fits of the models. The mean RMSE values for the same models of these pollutants were 29 μg/m3, 64 μg/m3, 5.1 μg/m3, and 4.6 μg/m3, respectively.  The RSME values were particularly high for NO and black carbon and were about 43% of their respective mean concentrations.    Based on data reliability, intended model application, and overall model fit, the road length annual models were picked from the nine pollutant models as the preferred model for explaining the spatial distribution and concentration of pollutants in Hong Kong (Table 3.6; Table 3.7).  As the goal of the models was to predict long-term exposure, using of the annual models was most appropriate as they were better indicators of long-term concentrations. Of all the traffic variables (road length, AADT, and traffic loading), there was greatest confidence in road length as AADT and traffic loading were based on interpolation.   In generally road length models fit slightly better than the other two traffic variables.    76  Table 3.6 Preferred models’ results for gaseous pollutants  NO2 Road length NO Road Length R2 0.46 0.50 Adjusted R2 0.43 0.48 LOOCV R2 0.39 0.28 HEV R2 0.56 - n 75 40 MBE 5.89e-16 3.42e-15 RMSE 27.65 62.10 MAE 21.31 51.69 Moran’s I of residuals -0.26 -0.0068 Variables6 (β; VIF; Partitioned R2 - Dominance)  Intercept ExpRL.1000 MainRL.50 ElvRL.5000 OpArT.300  7.84e+01 1.61e-03 9.67e-02 3.02e-04 -1.27e-04   1.02 1.29 1.29 1.08   0.060 0.134 0.221 0.047  Intercept ElvRL.500 BldVolD.25 IndT.25 WPopDen.100  7.07e+01 1.06e-02 3.73e+00 1.98e-01 5.29e+02   1.01 1.10 1.02 1.08    0.192 0.148 0.078 0.082   Table 3.7 Preferred models' results for particulate pollutants  PM2.5 Road length BC Road length R2 0.59 0.50 Adjusted R2 0.54 0.44 LOOCV R2 0.43 0.31 n 64 76 MBE -5.38e-17 1.76e-17 RMSE 3.99 3.70 MAE 3.22 3.02 Moran’s I of residuals -0.23 -0.13 Variables7 (β; VIF; Partitioned R2 - Dominance)  Intercept ExpRL.25 Dist_ShenzhenP CarPD.1000 CarPD.25 GovT.100 IndT.25   3.67e+01 8.91e-02 -3.09e-04 4.17e+05 1.68e+04 -3.81e-04 1.38e-02   1.35 1.16 1.26 1.05 1.05 1.06   0.056 0.168 0.151 0.057 0.066 0.089  Intercept ExpRL.3000 ExpRL.50 Long CarPD.50 ComT.500 ResT.50 MixT.500 Lands.500  2.51e+03 9.48e-05 1.76e-02 -2.19e+01 3.23e+04 -2.74e-05 -8.74e-04 -2.32e-05 -1.50e-04   1.27 1.42 1.23 1.20 1.22 1.39 1.27 1.23   0.091 0.075 0.089 0.044 0.037 0.050 0.053 0.065                                                   6 Appendix C provides full form of variable abbreviations  7 Appendix C provides full form of variable abbreviations 77   The final model surfaces were created for the preferred models (Figure 3.12; Figure 3.13; Figure 3.14; Figure 3.15).  These prediction surfaces had truncated values, except for the NO2 surface.  Values were truncated to ±20% of the range of the corrected annual concentrations entered into the models.     78   Figure 3.12 NO2 prediction surface (preferred model) NO2 (ug/m3)High : 205Low : 430 6 123KilometersNO2 Prediction Surface79   Figure 3.13 NO prediction surface (preferred model) NO (ug/m3)High : 451Low : 70.690 6 123KilometersNO Prediction Surface80   Figure 3.14 PM2.5 prediction surface (preferred model) PM2.5 (ug/m3)High : 62Low : 200 6 123KilometersPM2.5 Prediction Surface81   Figure 3.15 Black carbon prediction surface (preferred model) Black Carbon (ug/m3)High : 33Low : 2.80 6 123KilometersBlack Carbon Prediction Surface82  Generally, the model assumptions of linearity, homoscedasticity, normality, spatial independence, and multicollinearity were met for all 36 models (Table 3.6; Table 3.7; Appendix I  ).  Diagnostic plots showed that the models meet the conditions of linearity, homoscedasticity, and normality.  Moran’s I values of the residual s ranged from -0.36 to 0.15, meaning spatial correlation ranged from slightly dispersive to slightly clustered.   In the preferred model, the range was narrower, -0.26 to -0.0068.  All values were close to zero or slightly negative.  This indicates that the residuals were randomly to slightly dispersive in spatial distribution.  This can also be seen in the spatial plots of residuals (Figure 3.16; Figure 3.17; Figure 3.18; Figure 3.19).  Hence, spatial autocorrelation of residuals was not a concern. This also meant interpolated residuals could not be used as an additional predictor variable.   Multicollinearity was a concern during the modelling process and a few of the initial 36 models were rerun due to high VIF values.  A major source of multicollinearity was large buffered predictor variables, particularly 4000 m and 5000 m, as correlation was high for variables of this buffer size between most spatial metrics.  All final models had VIF values under 3.   83   Figure 3.16 NO2 residual spatial plot  Figure 3.17 NO residual spatial plot 113.8°E 113.9°E 114°E 114.1°E 114.2°E 114.3°E 114.4°E22.1°N22.2°N22.3°N22.4°N22.5°N22.6°NPlot of Residuals from NO2 Model-+ -++----++-+--+++-+++--+ ++-+-- -----+-++--++--+++-++-----+-+++++-+++--Residuals mg/m3-50-252550113.8°E 113.9°E 114°E 114.1°E 114.2°E 114.3°E 114.4°E22.1°N22.2°N22.3°N22.4°N22.5°N22.6°NPlot of Residuals from NO Model- ++-+++++++ -- ----+----+--+-+-++----+++++--Residuals mg/m3-100-505010084   Figure 3.18 PM2.5 residual spatial plot  Figure 3.19 Black carbon residual spatial plot 113.8°E 113.9°E 114°E 114.1°E 114.2°E 114.3°E 114.4°E22.1°N22.2°N22.3°N22.4°N22.5°N22.6°NPlot of Residuals from PM2.5 Model-+ -- ++---+-++--+-+--+-+-+ - -+++-+-++-++-++------+++---+++--Residuals mg/m3-10-5510113.8°E 113.9°E 114°E 114.1°E 114.2°E 114.3°E 114.4°E22.1°N22.2°N22.3°N22.4°N22.5°N22.6°NPlot of Residuals from Black Carbon Model+- + ++----++----+-+--+-+--++ + -+-+-++-+-++-++-+--+++-++----+++---+++-++--Residuals mg/m3-6-33685  The predictive power of the models was evaluated using LOOCV and HEV.  The LOOCV value for each model was lower than the model’s R2 value.  On average, LOOCV R2 values were 0.09, 0.18, 0.15, and 0.16 lower for each of their respective NO2, NO, PM2.5, and black carbon models across all 36 models.  The final preferred models had the following LOOCV R2 values: a) NO2 (LOOCV R2 = 0.39), b) NO (LOOCV R2 = 0.28), c) PM2.5 (LOOCV R2 = 0.43), and d) black carbon (LOOCV R2 = 0.31) (Table 3.6; Table 3.7).    The HEV R2 values for the NO2 models were greater than most of the LOOCV R2 with the exception of two of the traffic loading models.  This was unexpected as it is generally assumed that HEV values should be lower than the LOOCV R2.  In three models (annual road length, annual AADT, and SC2 road length), the HEV R2 values were higher than the R2.  While is this not impossible, as a model can fit the held out dataset better, it was unexpected, as the model was not built based on that dataset.    3.4 Sensitivity Tests To evaluate the effect of changing input parameters and the robustness of model relationships, two sensitivity tests were applied to the preferred models.  The first sensitivity test limited modelling to Hong Kong Island and Kowloon.  This was done to determine if the different development patterns of the New Territories, dispersed but density built-up clusters of development, decreased model performance.  The annual road length modelling was run with truncated datasets with New Territories sites removed (Table 3.8; Table 3.9). The second test used the full dataset with the pollutant concentrations naturally logged (Table 3.10; Table 3.11). Neither of these test produced significantly better model performances.   86  Table 3.8 Models' results for gaseous pollutants without New Territories sites  NO2 Road length without New Territories NO Road length without New Territories R2 0.38 0.56 Adjusted R2 0.34 0.52 LOOCV R2 0.22 0.46 n 50 25 MBE -1.13e-15 2.49e-15 RMSE 30.75 60.31 MAE 24.46 48.53 Moran’s I of residuals -0.24 -0.25 Variables8 (β; VIF; Partitioned R2 - Dominance)  Intercept ExpRL.100 MainRL.25 BldArD.500   76.70 1.47 0.47 83.86   1.01 1.01 1.01   0.108 0.229 0.043  Intercept PrkArD.200 MixT.25  1.12e+02 9.52e+03 2.87e-01   1.03 1.03   0.092 0.470  Table 3.9 Models' results for particulate pollutants without New Territories sites  PM2.5 Road length without New Territories BC Road length without New Territories R2 0.55 0.51 Adjusted R2 0.49 0.45 LOOCV R2 0.45 0.22 n 36 40 MBE -1.32e-17 -4.06e-18 RMSE 3.81 2.94 MAE 2.89 2.38 Moran’s I of residuals -0.31 -0.27 Variables9 (β; VIF; Partitioned R2 - Dominance)  Intercept LnDist_Coast CarPD.25 CarPD.500 GovT.100  1.29e+01 3.00 2.22e+04 1.65e+05 -7.49e-04   1.23 1.09 1.03 1.15   0.131 0.176 0.088 0.154   Intercept MainRL.25 PrkArD.100 IndT.5000 WPopDen.25  4.25 2.60e-02 2.23e+02 9.50e-06 3.13e+01   1.01 1.01 1.02 1.01   0.077 0.059 0.277 0.098                                                     8 Appendix C provides full form of variable abbreviations  9 Appendix C provides full form of variable abbreviations  87  Table 3.10 Natural log models' results for gaseous pollutants   LN NO2 Road length LN NO Road Length R2 0.46 0.51 Adjusted R2 0.43 0.46 LOOCV R2 0.40 0.35 n 95 40 MBE -1.20e-17 -1.98e-17 RMSE 0.27 0.48 MAE 0.21 0.39 Moran’s I of residuals 0.034 -0.096 Variables10 (β; VIF; Partitioned R2 - Dominance)  Intercept ExpRL.1000 MainRL.50 BldArD.4000 OpArT.200 ResT.25   4.23 1.51e-05 1.09e-03 3.08 -2.22e-06 -8.57e-05   1.04 1.33 1.19 1.14 1.12   0.048 0.163 0.168 0.043 0.034  Intercept ElvRL.500 PrkArD.1000 ParkT.25 NO2.Len  3.82 5.33e-05 -1.62e+02 -3.10e-03 1.15e-02   1.43 1.17 1.06 1.38   0.110 0.080 0.132 0.189  Table 3.11 Natural log models’ results for particulate pollutants  LN PM2.5 Road length LN BC Road Length R2 0.58 0.37 Adjusted R2 0.54 0.32 LOOCV R2 0.50 0.25 n 64 76 MBE -2.75e-18 -9.19e-20 RMSE 0.11 0.38 MAE 0.093 0.31 Moran’s I of residuals -0.26 0.029 Variables11 (β; VIF; Partitioned R2 - Dominance)  Intercept ExpRL.25 Lat LnDist_Coast CarPD.1000 CarPD.25 GovT.100   -1.78e+01 3.23e-03 9.38e-01 2.97e-02 1.18e+04 4.99e+02 -1.54e-05   1.33 1.14 1.16 1.32 1.07 1.06   0.076 0.154 0.052 0.151 0.058 0.093  Intercept Dist_Coast CarPD.50 ParkT.2000 ComT.25 ResT.50 IndT.1500  2.47 1.97e-04 2.75e+03 -1.13e-07 -6.52e-04 -8.20e-05 9.64e-07   1.09 1.18 1.11 1.05 1.20 1.10   0.072 0.029 0.086 0.061 0.057 0.067                                                    10 Appendix C provides full form of variable abbreviations  11 Appendix C provides full form of variable abbreviations  88  Chapter 4: Discussion  Poor air quality, and the public health risk it poses, is a concern in Hong Kong.  The objective of this study was the creation of two-dimensional (ground level) LUR models for NO2, NO, PM2.5, and black carbon in Hong Kong.  These models provide an understanding of the spatial distribution in the territory of air pollutants commonly emitted from anthropogenic activities and linked to adverse health effects.  These pollutants were assumed to have different seasonal and spatial profiles due to emission sources, atmospheric chemistry, and atmospheric residence times.  The resulting models and their associated prediction surfaces provide a starting point for estimating exposure of individuals in Hong Kong while highlighting spatial factors that potentially influence concentrations. Lastly, these models test the viability of LUR modelling in high-density high-rise Asian cities similar to Hong Kong.    Warm and cool season sampling campaigns were carried out in Hong Kong for NO2, NO, PM2.5, and black carbon. The campaigns used the same 100 site sampling framing and each pollutant was measured at between 43 to 97 of the sites.  Concentrations of all four pollutants were high relative to other cities and to air quality standards.  Significant seasonality in concentrations was only detected in PM2.5, M = 23.4 μg/m3 (SC1) and M = 46.8 μg/m3 (SC2).  A large number of potential predictors (373), covering a wide array of geospatial metrics, were extracted for each sampling site.  These were combined with the concentration measurement data to produce 36 different LUR models (including the four pollutants, seasonal and annual models, and separate models for road length and traffic density predictors).  Traffic variables, land use, and car park density were the most commonly selected model predictors.  The final preferred models (annual 89  road length) provided a modest explanation of variation in the four pollutants with explained variance in PM2.5 the highest (59%) and NO2 the lowest (46%).  Spatial patterns varied by pollutant.  Both PM2.5 and black carbon predictions exhibited a northwest-southeast gradient, with higher concentrations in the north.  For black carbon, the port was also an area of elevated predicted values. The NO2 prediction surface values were highest in Kowloon and the northern region of Hong Kong Island.  The NO2 predictions were heavily influenced by traffic predictors.    4.1 Basic Assessment of Exposure  4.1.1  NO2 Measured two-week average concentrations of NO2 ranged from 43 – 213 μg/m3 with a mean of 106 μg/m3.  Predicted annual concentrations ranged from 43 – 205 μg/m3, with a mean of 80 μg/m3.  All measured and predicted concentrations were in exceedance of the WHO 2005 air quality guidelines (WHO, 2014) and the Hong Kong objectives (EPD, 2016) for annual NO2 concentrations, i.e. 40 μg/m3.  The 200 μg/m3 1-hour WHO and Hong Kong threshold (EPD, 2016; WHO, 2014) was exceed at two of the 95 measured sites12. While uncommon, such high concentrations represent potential for high human exposure.  Spatial variation in annual NO2 prediction surfaces was heavily influenced by traffic. In the preferred model, three of the four significant predictors were road length variables.  This fits with the literature that suggests that traffic is an important source of NO2 (ENB, 2013; Tian et                                                  12 Concentrations were rounded to the nearest whole number. 90  al., 2011).   NO2 concentrations were highest in the densely populated northern region of Hong Kong Island and Kowloon.  The lack of seasonality in NO2 levels suggests the ground level NO2 is not as affected by regional sources. Hence, without the regional influence the NO2 concentrations were highest in the more density-populated regions of Hong Kong where local emission sources density (e.g. traffic) is highest.    In conclusion, even though there was substantial spatial variation in the concentrations of NO2, concentrations exceeded health guidelines throughout the territory with the highest concentrations in the most densely populated regions.  4.1.2 NO NO had the widest range of the four pollutants that were assessed, with annual measured and annual predicted levels ranging from 21 – 376 μg/m3 and 71 – 451 μg/m3, respectively.  NO is a primary pollutant emitted during high-temperature combustion (e.g. traffic).  Under most ambient atmospheric conditions, NO rapidly oxidizes into NO2 (Altshuller, 1956; WHO, 2006).  Its transient temporal state in the ambient atmosphere and the large range of values, suggest the high likelihood of fine scale spatial variation.  Few LUR models have focused on NO, however they tend to find the highest concentrations along roads with concentrations dropping off quickly and increasing distances (Su et al., 2009; Wang et al., 2013).   The preferred NO model explained a similar amount of variance as the models of other pollutants.  However, NO modelling in this study produced worse R2 results than the other pollutants at the 10m resolution prediction surface.  The prediction calculation was dominated by 91  the building volume density coefficient whereas the prediction surface was expected to be similar to other NO LUR models with high concentrations around roads and quick roadside drop off (Su et al., 2009; Wang et al., 2013).   This does not suggest that the model does not capture features unique to the NO distribution, relative to the other pollutants, or that is not representative to broader scale NO distribution.  The predictors selected did fit with the transient temporal state, and therefore assumed more heterogeneous spatial distribution, of NO, relative to NO2 or PM2.5 that should have smoother spatial distributions.  The NO predictors for the preferred NO model were all buffered at 500 m or smaller.  Smaller buffer predictors are consistent with the spatial distribution of a mainly short-term primary pollutant.   At a coarser resolution or when smoothed, the NO prediction surface described a reasonable estimate of the regional spatial distribution.  Higher concentrations were most frequent in Kowloon and northern Hong Kong Island.  This pattern was similar to the NO2 model.  As NO and NO2 are chemically linked, the overlap was expected given that background concentrations of both pollutants should be spatially correlated.  This is the pattern expected on the regional scale.  The model and the prediction surface are useful as long as one accounts for smoothing of values.  4.1.3 PM2.5 Measured annual concentrations of PM2.5 ranged from 25 – 51 μg/m3 with a mean of 35 μg/m3.  There was a strong seasonal component to concentrations with a ~ 20 μg/m3 increase in the 92  mean, minimum, and maximum from SC1 to SC2.  The direction and magnitude of seasonality was not surprising. Chiu and Lok (2011) and J. Yu et al. (2004) both observed winter concentrations roughly twice high as summer levels at each individual AQMS for PM10, and organic carbon aerosol, respectively.  Predicted annual concentrations, with truncation, ranged from 20 – 62 μg/m3 with a mean of 30 μg/m3.   All measured and predicted annual PM2.5 concentrations surpassed the WHO annual guideline of 10 μg/m3 (WHO, 2014).  Even pre-truncated annual predictions, which had a minimum value of 12 μg/m3, and SC1 measurements, minimum 11 μg/m3, exceeded this guideline.  The WHO 24-hour guideline of 25 μg/m3 (WHO, 2014) was also exceeded by all annual measurements and 94.4% of the annual prediction surface. The majority of areas not in exceedance were in the less populated south and outlying islands.  In SC1, only 30% of the measurements exceeded this guideline, but all did in SC2. The Hong Kong PM2.5 annual objective, 35 μg/m3, is not as stringent (EPD, 2016) and only 50% of the annual measurements and 8.0% of the prediction surface were over this level.   The PM2.5 annual prediction surface had a strong northwest to southeast gradient with higher concentrations closer to the border with Mainland China. The ‘distance to Shenzhen’ predictor was responsible for this pattern. This variable was also the strongest explanatory variable of unique variance in the model, according to dominance analysis results.  This suggested that regional sources from Shenzhen and further upwind were a significant factor in explaining the spatial variation of PM2.5 in the territory.  A regional northeast/southwest gradient in PM10 concentrations had be noted previously (Chiu & Lok, 2011) in an analysis of the rooftop 93  monitoring stations, however, our study expanded the pattern beyond the few rooftop regulatory network monitors to the larger territory and more clearly separated the impacts of regional and local emissions.    The spatial gradient also appears to vary by season.  ‘Distance to Shenzhen’ and latitude, which represent regional effects in the model, were present in all the annual and SC2 models, except the SC2 AADT model, but were not present in the SC1 models.  This was expected having been implied in the literature, which identified regional sources as significant contributors to local particulate matter levels mainly during the winter (Chiu & Lok, 2011; Kok et al, 1997; Lau et al., 2007; J. Yu et al, 2004).  The selection of regional predictors in the annual models despite only being relevant in one of the sampling campaigns again highlights the impact of these regional sources.    In addition to the regional gradient, predictors representing local sources indicated localized gradients related to vehicles and landuse. The highest PM2.5 levels occurred in developed regions in the New Territories and appeared to results from the combined effects of regional and local sources.  This was opposite to the observed spatial pattern of NO2. Population density is lower in these areas but they still present a large potential population for exposure.  All significantly populated areas had predicted annual values above 25 μg/m3.  Such levels are associated with a 9% increase in risk of premature mortality compared to the 10 μg/m3 WHO annual guideline (WHO, 2006).  The annual guideline does not necessarily represent a threshold below which there are no adverse health risks, rather the threshold represents the minimum value 94  for which the current scientific literature has found a statistically significant increase in premature mortality. At 35 μg/m3 (Hong Kong PM2.5 annual objective), the long-term risk of premature mortality rises to 15% relative to the 10 μg/m3 guideline (WHO, 2006), however, a smaller proportion of the Hong Kong population was exposed to these levels.  Assuming an individuals’ exposure was only a function of mean ground level concentrations in their residential street block, a crude exposure approximation, ~1,500,000 individuals would be exposed to levels greater than 35 μg/m3, i.e. ~20% of the population.  This subpopulation was centred in the north of the territory.  While only a small percentage of the overall population, the notable increase in relative risk indicates the exposure maybe a significant public health risk and a greater risk in the winter.  4.1.4 Black Carbon Black carbon annual concentrations ranged from 3.5 – 28 μg/m3 with a mean of 11 μg/m3.  Truncated annual predictions ranged from 2.8 – 33 μg/m3 with a mean of 6.7 μg/m3.  Although there are no WHO or Hong Kong standards for black carbon, the measured black carbon concentrations were greater than observed in many other urban centers.  For example, a report by the Environmental Protection Agency indicated global ambient urban annual black carbon observations from available monitoring networks ranged from 0.3 - ~15 μg/m3 (Environmental Protection Agency [EPA], 2012).  Urban observations from North American and continental European ranged from 0.3 – 3.0 μg/m3 (n = 262), while Chinese observations (n = 5) ranged from 5 – 14.2 μg/m3 (EPA, 2012).  An urban roadside site in London (Marylebone Road monitoring station) had an annual average black carbon concentration of 7 μg/m3 (Butterfield et al., 2015).  A study in Beijing, which collected black carbon at an urban site next to a major 95  expressway, found average summer concentrations of 12.3 μg/m3 and average winter concentrations of 17.9 μg/m3 (Song et al., 2013).  While there is limited data upon which to compare Hong Kong to other Chinese cities, it does appear that black carbon levels are significantly greater than many North American and European cities.    Similar to PM2.5, regional sources notably impacted ground level black carbon concentrations. A west-east gradient existed in the preferred black carbon prediction surface.  Kok et al. (1997) also reported elevated black carbon concentrations in the western regions of the territory and attributed these higher levels to regional sources.  Using data collected in 1994 by airplane, they found that black carbon was greater in the western areas of Hong Kong with maximum concentrations of 7 μg/m3 (60 m above ground), while concentrations to the south and east ranged from 1-2 μg/m3.  They noted in the west, black carbon made-up 10% of aerosol mass relative to 5% in other regions.  The authors believed a major source of the elevated black carbon in west Hong Kong was coal burning, both industrial and domestic, in the PRD.  This based on the black carbon to carbon monoxide (CO) ratio, which indicates the black carbon was transported, and comparison CO/NOX emission ratios between PRD and Hong Kong (Kok et al., 1997).   Unlike PM2.5, local emissions appeared more influential in predicted black carbon spatial patterns.  Undeveloped regional areas had significantly lower levels of black carbon despite regional influences, whereas, areas directly adjacent to expressways had significantly higher levels, and the port had the greatest levels. These spatial patterns dominated the regional gradient. The highest annual predictions were along expressways notability in the vicinity of the 96  port and in the north.  The port and motorway had previously been identified as important local sources.  J. Yu et al. (2004) noted concentrations recorded at network monitoring stations around the port changed seasonally due to the direction of the prevailing wind.  In the winter, when the prevailing wind is from the north, EC concentrations were higher at the two monitoring stations south of the port.  In the summer, when the prevailing wind is from the south, the five monitoring stations immediately to the north of the port had higher EC concentrations.  To have this impact on background concentrations indicates the port is a significant source.  This was further supported by the seasonal and spatial patterns of SO2 and the trace elements V and Ni, which are pollutants associated with heavy marine fuel. The dominances of vehicle emissions as a source were seen in the highest EC concentrations at the roadside monitoring station (J. Yu et al., 2004).  Black carbon is a contributor to overall PM2.5 concentrations and health outcomes associated with PM2.5 are generally the same as black carbon making it difficult to separate these effects (EPA, 2012).  However, black carbon is a valuable additional exposure metric as it is an indicator of primary combustion related particles, which are believed to a more toxic component of general PM2.5 (EPA, 2012; Janssen et al., 2011).  In the case of Hong Kong high levels in the vicinity of the port are found for the black carbon prediction surface, but not the PM2.5 surface.  The PM2.5 surface does have higher concentrations to the north of the port. This highlights the spatial variation of primary combustion particles and potentially elevated health risks posed to the population in this area. Specifically, it reinforces the importance of regulating marine pollutants in Hong Kong.   97  4.2 Performance of HK2D LUR Models Overall, the Hong Kong roadside LUR modelling was successful.  As discussed in Section 4.1, the spatial pattern depicted in the prediction surfaces for NO2, PM2.5, and black carbon aligned with existing literature on spatial distribution of air pollution concentrations within Hong Kong. This suggests that the models captured the important spatial parameters. There is also confidence that the models represent the best results given the available resources due to the robustness of the modelling methodology and comparison with other LUR models’ performances.   The HK2D modelling followed many of the best practices outlined for LUR modelling.   Multiple pollutants were sampled to characterize air quality.  Two extensive roadside sampling campaigns were carried out to account to seasonality in pollution levels and give a better indication of long-term concentrations.  While there is no set number of sites required for LUR modelling, at least 40-50 training sites is suggested (Hoek et al., 2008; Wang et al., 2012).  All models in this study had at least 40 training sites.  Sampling sites were selected that maximized the range of values in pollution concentrations and spatial variables.    LUR models can use monitoring network data; it is easier to obtain and less costly than running a dedicated sampling campaign (Hoek et al, 2008).  Monitoring network sites are generally located to capture ambient concentrations for regulatory purposes (e.g. on rooftop or at locations without nearby emission sources).  Sites set up to capture ambient concentrations tend to capture a muted range of concentrations and predictor variables. Modelling can appear more successful as there is less variation to be explained due to less sites or a muted range.   98  During model building a substantial number of plausible spatial predictors were offered, to ensure no relevant spatial metric was overlooked, and predictor selection was limited to one for every ten observations.  This rule of thumb was applied to avoid overfitting (Harrell, 2015; Harrell et al, 1996).  Sensitivity tests indicated that truncating the dataset and transforming pollution concentrations to their natural logarithms did not affect the performance of the models. The selected predictors were similar, again suggesting robustness in the modelling and that the final models were the best possible models based on the available data. The approach to the modelling may have led to more conservative results.  However, this was weighed against other procedures that could have had the potential to artificially inflated R2 values.  The final preferred models explained 46 – 59% of the variance: a modest performance.  This was lower than seen in many European and North American models, where for example NO2 models typically explain 60-70% of the variance in measured concentrations, though still within the range seen overall (Hoek et al., 2008). Hong Kong has a more complex urban morphology compared to many European and North American cities.  This complex urban morphology and geography may play a role in the lower R2 values in the HK2D models. When compared to LUR models for other Asian cities, which may have more similarities in urban morphology, vehicle use, and building designs, R2 values are closer (Table 4.1).  Hong Kong has more high-rises (7,837) and a higher ratio of high-rises to population than any of the other Asian cities modelled using LUR (Table 4.1; Emporis, 2016b).  This suggests Hong Kong has a more extreme and complex urban morphology in comparison to these other cities (Table 4.1). When reviewing these studies it is also important to note that many used the regulatory monitoring network as their source information.  As indicated previously, this approach can artificially create models 99  with higher R2 values.  Using only monitoring network sites tends to result in a lower number of sampling sites and sampling sites located to capture ambient concentrations for regulatory purposes.  Overfitting can also be an issue when there are few sampling sites.  Table 4.1 LUR modelling for Other Asian Cities Location Reference No. of high-rise buildings Population Sampling Sites (n) Predictors (R2; no. of predictors) Changsha, China Li et al., 2015 159 3,617,469 in city1  Field campaign  (n = 80) NO2 (n = 40 PM10) NO2 (0.41-0.51) PM10 (0.32-0.39)13 Jinan, China Li et al., 2010 153 2,540,000 in city2  Regulatory air quality monitoring sites  (n = 14) SO2 (0.62; 3) NO2 (0.64; 3) PM10 (0.60; 3) Shanghai, China Meng et al., 2015 1,500 17,836,133 in city 13,053,754 in metro3 Ambient air quality monitoring sites  (n = 38) NO2 (0.82; 4) Shanghai, China Meng et al., 2016 1,500 17,836,133 in city 13,053,754 in metro3 Ground level air pollution monitoring sites  (n = 28) PM10 (0.80; 4) Tianjin, China Chen et al., 2010 356 6,825,105 in city 9,612,220 in metro4 Ambient air quality monitoring sites  (n = 20; another 10 used to validate) NO2 (0.74; 5, heating season) (0.61; 4, non-heating season)  PM10  (0.72; 4, heating season)  (0.49; 3, non-heating season) Korea Lee et al., 2012 [Conference paper]  Not provided in abstract NO2 (0.51) PM2.5 (0.52) Incheon, Korea Lee et al., 2012 [Conference paper] 2,141 2,710,579 in city 20,982,273 in metro5 Field campaign  (n = 97) NO2 (0.60; 4) Metropolitan area of Seoul, Kyunggi, and Incheon, Korea Lee et al., 2012 [Conference paper]  Regulatory air quality monitoring sites  (n = 102)  NO2 (0.59; 5)                                                   13 For Li et al. (2015) only the circular buffer models were included as they are directly comparable.  100  Location Reference No. of high-rise buildings Population Sampling Sites (n) Predictors (R2; no. of predictors) Seoul, Korea Choi, Giehae et al., 2014, [Conference paper] 4,456 10,581,728 in city 20,982,273 in metro6 Distract ambient air quality monitoring sites (n= 25) NO2 (0.53; 4) New Delhi, India Saraswat et al., 2013 73 301,000 in city 12,791,458 in metro7 Field campaign  (n = 18, morning; n = 37, afternoon) Ultrafine particle number concentrations (0.28; 1 morning) (0.23; 2, afternoon)14 Shizuoka prefecture, Japan Kashima et al., 2009  Regulatory air quality monitoring sites  (n = 83) NO2 (0.54; 5) SPM (0.11; 1) Ulaanbaatar, Mongolia Allen et al., 2013 523 989,900 in city8 Field campaign deployment  (n =37) Wintertime  NO2 (0.74; 5) SO2 (0.78;2) 1=Emporis 2016c; 2=Emporis 2016d; 3=Emporis 2016e; 4=Emporis 2016f; 5=Emporis 2016g; 6=Emporis 2016h; 7=Emporis 2016i; 8=Emporis, 2016j;   A recent study also produced PM2.5 and PM10 LUR models of Hong Kong (Shi et al., 2016).  In contrast to the HK2D modelling described in this thesis, this study only modelled Kowloon and northern Hong Kong Island, used mobile monitoring, sampled only during the summer months, and aggregated data at a 300 m spatial resolution.  The PM2.5 spatial pattern of the Shi et al. prediction surface is similar to the HK2D PM2.5 surface.   While the Shi et al. (2016) 300 m resolution PM2.5 model has a higher adjusted R2 (0.63) compared to the HK2D PM2.5 (0.54), this difference appears to be due to the spatial aggregation of data.   Models from the Shi et al. study at the aggregation scales of 200 m, 100 m, and 50 m all have adjusted R2 of 0.51 or lower.  Unlike the HK2D study, Shi et al. found urban /building morphology variables were important in PM2.5 modelling.                                                    14 Only models without rooftop concentrations as a predictor are shown. 101  4.3 Viability of LUR Modelling in Other High-Density High-Rise Cities As previously discussed, complex urban morphology found in high-density high-rise Asian cities may affect the feasibility of LUR modelling.  The viability of LUR modelling in such cities was tested by applying LUR modelling to Hong Kong. To replicate the success of the HK2D models in these other cities, the elements required for the successful modelling of Hong Kong must be logistically feasible to repeat in these other cities.    The inputs required for the HK2D models are feasible to achieve in studies in other high-density high-rise Asian cities.  A large number of spatial predictors were offered to HK2D modelling.  More complex predictors, such as aspect ratio, that one would assume would be important in a high-rise city, were not selected in any of the final models.  The final preferred NO2, PM2.5, and black carbon models contained fairly traditional LUR predictors (traffic, land use, coordinates, and distance for a large regional emission source).  Road length, the least labour-intensive traffic variable to calculate, was the best preforming traffic variable.   Car park density was a common predictor in the HK2D models.  This predictor was likely a proxy for traffic flow and urban development and other predictors maybe able to fill this role.  In the application of LUR modelling to high-rise high-density Asian cities, accessibility to spatial data likely varies depending on the city.  Most of these predictors are relatively basic, hence more likely to be easy to acquire, and in many cases the data may be found from open sources.     Despite the added complexity which high-density high-rise cities bring to the spatial distribution of air quality and therefore to LUR modelling, this study suggests that this complexity is not reflected in the selection of predictors but rather in the model performance. This effect may be 102  reflected in the more modest performances of LUR models from Asian cities, though not all are high-density high-rise cities (Table 4.1). Despite the large number of spatial predictor variables offered to the HK2D models, the unexplained variance was not reduced. This suggests a complex relationship between spatial metrics and pollution concentrations and not that potential spatial predictors were missed. This limitation may also extend to other methods of air quality modelling.  Modelling air quality and exposure is important in these cities, as they tend to have large populations and higher air population levels meaning air quality is a large potential public health risk.  Though there are limitations, LUR modelling can be successful and be a useful public health tool for identifying key predictors for air quality and making policy makers aware of the complexity of relationships and spatial variation.  This can help avoid simple, one size fits all policies or policies transferred and incorrectly applied from other jurisdictions.  4.4 Limitations of the Study The sampling focused on developed lands and used only roadside sites, indicating that the models are more suited to predicting concentrations in developed regions and may be biased towards selecting traffic variables. This makes it difficult to validate predictions made in undeveloped park areas with the current dataset. However, as the modelling was focused on human exposure it is reasonable to only consider developed areas.     The NO corrected and combined (SC1 and SC2) models had the greatest range in concentrations but the least number of sampling sites (n = 40). This was the result of a reduced sampling campaign in the SC1, the loss of data from three sites, and inability to use NO2 to predict NO levels due to variation in the NO:NO2 ratios at different sampling sites. Forty sites is at the low 103  end of the suggested number of sites for LUR modelling (Hoek et al. 2008). A larger sample may have improved modelling by capturing a fuller range of the NO concentration, though model fit was not stronger for the SC2 models where the sample size was 58.  The lack of bias correction for NO2, NO, and black carbon meant these concentrations may not be directly comparable to the concentrations from the reference monitoring network.  Many LUR studies, such as the ESCAPE studies, do not correct measured concentrations to the reference monitors in regulatory networks (Beelen et al., 2013; Cyrys et al., 2012).  As presented in Chapter 3, the NO2 and NO co-location comparison with the reference chemiluminescence monitor had ratio values of 0.89 and 0.92 and R2 values of 0.78 and 0.92, respectively.  These ratios and R2 values were within the range (R2 values for NO2 ranged from 0.60 to 0.98) reported in the ESCAPE study (Beelen et al., 2013; Cyrys et al., 2012).  This range was deemed acceptable given the difference in the sampler’s capture methods. As with PM2.5, the original intention was to correct the black carbon concentrations for potential bias, however, the regulatory monitors were offline during the co-location period, precluding co-location comparisons.  Due to the lack of bias correction, the NO2, NO and black carbon concentrations included potential sampling biases. When averaged, Ogawa badges under sample by about 10% compared to the monitoring network. This should be considered when using results.  4.5 Future Work The HK2D models described here will be used as baseline models and comparison for the larger HKD3D study’s exposure models.  Integrating the baseline HK2D models into the larger HKD3D models involves using a vertical correction factor to adjust predicted concentrations 104  based on the building floor.  Air pollutant concentrations may vary vertically due to factors such as dispersion.  Therefore street level concentration and levels at floor of residency in a high-rise may vary substantially.  This should be considered when estimating human exposure, particularly in cities with numerous high-rises like Hong Kong.  While the HK2D LUR models cannot be directly applied to other high-density high-rise cities, it is possible the vertical correction factor may be directly applicable.    A dynamic component will also be added. This dynamic component will incorporate the probable daily travel patterns of individuals in a given area and age group.  These components allow exposure prediction at a finer population resolution.    4.6 Conclusion This study met its research objectives. Ground level LUR models were created for NO2, NO, PM2.5, and black carbon.   These models provided information on the spatial distribution of air pollution in Hong Kong, air quality exposure, and significant sources, as well served as a test of LUR modelling in a high-density high-rise city.  The concentrations measured and predicted for all pollutants in this study were high throughout the study area and suggest exposure to poor air quality poses a significant public health risk throughout Hong Kong. All the pollutants had different spatial distributions.  NO2 exhibited the highest values in the central region of Hong Kong while PM2.5 and black carbon had higher concentrations in the northwest of the territory.  Black carbon also had high concentrations around the port. The methodology and results of this study suggests LUR modelling is a viable method for modelling air quality in other high-density high-rise Asian cities.   105  Bibliography  Abernethy, R. (2012). A Land Use Regression Model for Ultrafine Particles in Vancouver, Canada. [Masters thesis]. Retrieved from cIRcle. (http://hdl.handle.net/2429/41895).  Abernethy, R., Allen, R., McKendry, I., & Brauer, M. (2013). A land use regression model for ultrafine particles in Vancouver, Canada. Environmental Science & Technology, 47(10), 5217-5225. doi:10.1021/es304495s  AethLabs. (2014). microAeth Model AE51 Operating Manual. Retrieved Feb. 11, 2014 from http://aethlabs.com/sites/all/content/microaeth/microAeth%20Model%20AE51% 20Operating%20Manual%20rev02.pdf  Aguilera, I., Foraster, M., Basagaa, X., Corradi, E., Deltell, A., Morelli, X.. . Kunzli, N. (2015). Application of land use regression modelling to assess the spatial distribution of road traffic noise in three european cities. Journal of Exposure Science and Environmental Epidemiology, 25(1), 97-105. doi:10.1038/jes.2014.61  Aguilera, I., Sunyer, J., Fernández-Patier, R., Hoek, G., Aguirre-Alfaro, A., Meliefste, K., ... & Brunekreef, B. (2008). Estimation of outdoor NOx, NO2, and BTEX exposure in a cohort of pregnant women using land use regression modeling. Environmental science & technology, 42(3), 815-821. doi: 10.1021/es0715492  Allard, F., & Ghiaus, C. (Eds.). (2012). Natural ventilation in the urban environment: assessment and design. London: Earthscan.  Allen, R. W., Gombojav, E., Barkhasragchaa, B., Byambaa, T., Lkhasuren, O., Amram, O., ... & Janes, C. R. (2013). An assessment of air pollution and its attributable mortality in Ulaanbaatar, Mongolia. Air Quality, Atmosphere & Health, 6(1), 137-150. doi: 10.1007/s11869-011-0154-3  Altshuller, A. (1956). Thermodynamic Considerations in the Interactions of Nitrogen Oxides and Oxy-Acids in the Atmosphere, Journal of the Air Pollution Control Association, 6(2), 97-100, doi: 10.1080/00966665.1956.10467740  Amini, H., Taghavi-Shahri, S. M., Henderson, S. B., Naddafi, K., Nabizadeh, R., & Yunesian, M. (2014). Land use regression models to estimate the annual and seasonal spatial variability of sulfur dioxide and particulate matter in Tehran, Iran. Science of the Total Environment, 488, 343-353. doi: 10.1016/j.scitotenv.2014.04.106  106  Araki, S., Yamamoto, K., & Kondo, A. (2015). Application of Regression Kriging to Air Pollutant Concentrations in Japan with High Spatial Resolution. Aerosol and Air Quality Research, 15(1), 234-241. doi: 10.4209/aaqr.2014.01.0011  Association of Central Oklahoma Governments. (2013). Traffic Count Analysis/ Vehicle Miles Traveled (VMT) Report. Retrieved Jun. 8, 2015 from http://www.acogok.org/websites/acogok/images/downloads2013/20131021b.pdf  Bank of Canada. (2015). Daily currency converter. Retrieved Nov. 13, 2015 from http://www.bankofcanada.ca/rates/exchange/daily-converter/  Basagaña, X., Rivera, M., Aguilera, I., Agis, D., Bouso, L., Elosua, R., ... & Künzli, N. (2012). Effect of the number of measurement sites on land use regression models in estimating local air pollution. Atmospheric Environment,54, 634-642. doi: 10.1016/j.atmosenv.2012.01.064  Beelen, R., Hoek, G., Vienneau, D., Eeftens, M., Dimakopoulou, K., Pedeli, X., ... & Eriksen, K. T. (2013). Development of NO 2 and NO x land use regression models for estimating air pollution exposure in 36 study areas in Europe–the ESCAPE project. Atmospheric Environment, 72, 10-23. doi: 10.1016/j.atmosenv.2013.02.037  Bernard, S. M., Samet, J. M., Grambsch, A., Ebi, K. L., & Romieu, I. (2001). The potential impacts of climate variability and change on air pollution-related health effects in the united states. Environmental Health Perspectives, 109(Suppl 2), 199-209.  Brajer, V., Mead, R. W., & Xiao, F. (2006). Valuing the health impacts of air pollution in Hong Kong. Journal of Asian Economics, 17(1), 85-102. doi:10.1016/j.asieco.2005.12.002  Brauer, M., Henderson, S., & Marshall, J. (2006).  A land use regression road map for the Burrard Inlet area local air quality study. Retrieved Dec. 17, 2013 from http://www.cher.ubc.ca/PDFs/GVRD_BI_Report.pdf   Brauer, M., Hystad, P., & Reynolds, C. (2012). Development With Care 2012:  Environmental Guidelines for Urban and Rural Land Development in British Columbia,  Supporting Information – Air Quality.  Retrieved Jun. 13, 2014 from http://www.env.gov.bc.ca/wld/documents/bmp/devwithcare2012/DWC-Air-Quality.pdf  Briggs, D. J., de Hoogh, C., Gulliver, J., Wills, J., Elliott, P., Kingham, S., & Smallbone, K. (2000). A regression-based method for mapping traffic-related air pollution: application and testing in four contrasting urban environments. Science of the Total Environment, 253(1), 151-167. doi: 10.1016/S0048-9697(00)00429-0  107  Butterfield, D., Beccaceci, S., Quincey, P., Sweeney, B., Lilley, A., Bradshaw, C., Fuller, G., Green, D., & Font Font, A. (2015). 2014 Annual Report for the UK Black Carbon Network. Retrieved Apr. 30, 2016 from https://uk- air.defra.gov.uk/assets/documents/ reports/ cat05/1511251301_ AQ0644_2014_Black_carbon_Network_Annual_Report_ Final.pdf  Chao, C. Y., & Wong, K. K. (2002). Residential indoor PM10 and PM2.5 in Hong Kong and the elemental composition. Atmospheric Environment, 36(2), 265-277. doi: 10.1016/S1352-2310(01)00411-3  Chen, L., Bai, Z., Kong, S., Han, B., You, Y., Ding, X., ... & Liu, A. (2010). A land use regression for predicting NO 2 and PM 10 concentrations in different seasons in Tianjin region, China. Journal of Environmental Sciences, 22(9), 1364-1373. DOI: 10.1016/S1001-0742(09)60263-1  Chiu, S. L., & Lok, R. (2011). Variations of respirable suspended particulates in Hong Kong: cross-boundary and local effects. Annals of GIS, 17(2), 113-124. doi: 10.1080/19475683.2011.576269   Clougherty, J. E., Kheirbek, I., Eisl, H. M., Ross, Z., Pezeshki, G., Gorczynski, J. E., ... & Matte, T. (2013). Intra-urban spatial variability in wintertime street-level concentrations of multiple combustion-related air pollutants: the New York City Community Air Survey (NYCCAS). Journal of Exposure Science and Environmental Epidemiology, 23(3), 232-240. doi: 10.1038/jes.2012.125  CLP. (2015). Infrastructure and Fuel Mix. Retrieved Dec. 11, 2015 from https://www.clp.com.hk/en/about-clp/power-generation/infrastructure-and-fuel-mix  CLP. (2014). Facility Performance Statistics Penny’s Bay. Retrieved Dec. 11, 2015 from https://www.clpgroup.com/en/Sustainability-site/Facility%20Statistics/sr_fps_HK_ pennybay.pdf  CLP. (2010). Power Generation. Retrieved Jun. 7, 2014 from https://www.clp.com.hk/ouroperations/power/Pages/powergeneration.aspx  Cyrys, J., Eeftens, M., Heinrich, J., Ampe, C., Armengaud, A., Beelen, R., ... & Wang, M. (2012). Variation of NO 2 and NO x concentrations between and within 36 European study areas: results from the ESCAPE study. Atmospheric Environment, 62, 374-390. doi: 10.1016/j.atmosenv.2012.07.080  Delucchi, M. A., Murphy, J. J., & McCubbin, D. R. (2002). The health and visibility cost of air pollution: a comparison of estimation methods. Journal of Environmental Management, 64(2), 139-152. doi: 10.1006/jema.2001.0515  108  Deschenes S, Setton E, Demers P A, Keller P C. (2013). Modelling Arsenic and Lead Surface Soil Concentrations using Land Use Regression. In E3S Web of Conferences: proceedings of the 16th International Conference on Heavy Metals in the Environment. Retrieved April 10, 2015 from http://www.e3s-conferences.org/articles/e3sconf/abs/ 2013/01/e3sconf_ichm13_08004/e3sconf_ichm13_08004.html  Ding, A., Wang, T., Zhao, M., Wang, T., & Li, Z. (2004). Simulation of sea-land breezes and a discussion of their implications on the transport of air pollution during a multi-day ozone episode in the Pearl River Delta of China. Atmospheric environment, 38(39), 6737-6750. doi:10.1016/j.atmosenv.2004.09.017  Dionisio, K. L., Rooney, M. S., Arku, R. E., Friedman, A. B., Hughes, A. F., Vallarino, J., ... & Ezzati, M. (2010). Within-neighborhood patterns and sources of particle pollution: mobile monitoring and geographic information system analysis in four communities in Accra, Ghana. Environmental health perspectives (Online), 118(5), 607. doi: 10.1289/ehp.0901365   Dunlea, E. J., Herndon, S. C., Nelson, D. D., Volkamer, R. M., San Martini, F., Sheehy, P. M., ... & Molina, M. J. (2007). Evaluation of nitrogen dioxide chemiluminescence monitors in a polluted urban environment. Atmospheric Chemistry and Physics, 7(10), 2691-2704. doi: 10.5194/acp-7-2691-2007  Durant, J. L., Beelen, R., Eeftens, M., Meliefste, K., Cyrys, J., Heinrich, J., ... & Hoek, G. (2014). Comparison of ambient airborne PM 2.5, PM 2.5 absorbance and nitrogen dioxide ratios measured in 1999 and 2009 in three areas in Europe. Science of the Total Environment, 487, 290-298. doi: 10.1016/j.scitotenv.2014.04.019  Edussuriya, P., Chan, A., & Ye, A. (2011). Urban morphology and air quality in dense residential environments in Hong Kong. Part I: district-level analysis.Atmospheric Environment, 45(27), 4789-4803. doi: 10.1016/j.atmosenv.2009.07.061  Eeftens, M., Beekhuizen, J., Beelen, R., Wang, M., Vermeulen, R., Brunekreef, B., ... & Hoek, G. (2013). Quantifying urban street configuration for improvements in air pollution models. Atmospheric Environment, 72, 1-9. doi: 10.1016/j.atmosenv.2013.02.007  Eeftens, M., Tsai, M. Y., Ampe, C., Anwander, B., Beelen, R., Bellander, T., ... & Hoek, G. (2012). Spatial variation of PM 2.5, PM 10, PM 2.5 absorbance and PM coarse concentrations between and within 20 European study areas and the relationship with NO 2–Results of the ESCAPE project. Atmospheric Environment, 62, 303-317. doi: 10.1016/j.atmosenv.2012.08.038  Emporis. (2016a). Skyline Ranking. Retrieved July 10, 2016 from http://www.emporis.com/statistics/skyline-ranking  109  Emporis. (2016b). Hong Kong. Retrieved July 10, 2016 from http://www.emporis.com/city/101300/hong-kong-china  Emporis. (2016c). Changsha. Retrieved July 10, 2016 from http://www.emporis.com/city/100232/changsha-china  Emporis. (2016d). Jinan. Retrieved July 10, 2016 from http://www.emporis.com/city/100227/jinan-china  Emporis. (2016e). Shanghai. Retrieved July 10, 2016 from http://www.emporis.com/city/100213/shanghai-china  Emporis. (2016f). Tianjin. Retrieved July 10, 2016 from http://www.emporis.com/city/100215/tianjin-china  Emporis. (2016g). Incheon. Retrieved July 10, 2016 from http://www.emporis.com/city/100432/incheon-south-ko  Emporis. (2016h). Seoul. Retrieved July 10, 2016 from http://www.emporis.com/city/100429/seoul-south-korea  Emporis. (2016i). New Delhi. Retrieved July 10, 2016 from http://www.emporis.com/city/100258/new-delhi-india  Emporis. (2016j). Ulaanbaatar. Retrieved July 10, 2016 from http://www.emporis.com/city/100365/ulan-bator-mongolia  Environment Bureau. (2013).  Air Clean Plan for Hong Kong. Retrieved Nov. 28, 2013 from http://www.enb.gov.hk/en/files/New_Air_Plan_en.pdf  Environmental Protection Agency, Office of Radiation and Indoor Air, Office of Transportation and Air Quality, Office of Air Quality Planning and Standards, Office of Atmospheric Programs, United States. . Congress. (2012). Report to congress on black carbon: Department of the interior, environment, and related agencies, appropriations act, 2010. Research Triangle Park, NC: U.S. Environmental Protection Agency, Office of Air and Radiation, Office of Air Quality Planning and Standards.  Environmental Protection Department. (2016). Air Quality Objectives: Hong Kong’s Air Quality Objectives. Retrieved April 30, 2016 from http://www.epd.gov.hk/epd/english /environmentinhk/air/air_quality_objectives/air_quality_objectives.html  Environmental Protection Department. (2015a). Air Quality in Hong Kong 2014.  Retrieved Nov. 8, 2015 from http://www.aqhi.gov.hk/api_history/english/report/files/ AQR2014e_final.pdf  110  Environmental Protection Department. (2015b). Air: Data & statistics – Hong Kong Air Pollutant Emission Inventory. Environmental Protection Department. Retrieved Nov. 8, 2015 from http://www.epd.gov.hk/epd/english/environmentinhk/air/data/emission_ inve.html  Environmental Protection Department. (2015c). 2013 Hong Kong Emission Inventory Report [Report No. EPD/TR 2/14].  Retrieved Nov. 12, 2015 from http://www.epd.gov.hk/epd/sites/default/files/epd/2013EIReport_eng_1b.pdf  Environmental Protection Department. (2015d). Air. Retrieved Nov. 12, 2015 from http://www.epd.gov.hk/epd/english/environmentinhk/air/air_maincontent.html  Environmental Protection Department. (2015f). A Guide to the Air Pollution Control (Ocean Going Vessels) (Fuel at Berth) Regulation. Retrieved from http://www.epd.gov.hk/epd/english/environmentinhk/air/prob_solutions/guide-air-pollution-control-ocean-going-vesselsfuel-berth-regulation.html#require  Environmental Protection Department. (2013). Air Quality in Hong Kong 2012.  Retrieved Jun. 7, 2014 from http://www.aqhi.gov.hk/api_history/english/report/files/ AQR2012e_final.pdfhttp://www.epd.gov.hk/epd/english/environmentinhk/air/studyrpts/files/prd_tech_ann8.pdf  Environmental Protection Department. (2012). Air: Data & statistics – Air Quality Statistics. Environmental Protection Department. Retrieved Jan 22, 2014 from http://www.epd.gov.hk/epd/english/environmentinhk/air/data/aq_stat.html  Environmental Protection Department. (2009). Introduction of Euro V Standard for Motor Vehicle Fuel. Retrieved Nov. 13, 2015 from http://www.epd.gov.hk/epd/sites/ default/files/epd/english/boards/advisory_council/files/ACE_Paper_17_2009.pdf  Environmental Protection Department. (2006a). Air Quality Monitoring Equipment. Retrieved Jun. 7, 2014 from http://www.aqhi.gov.hk/en/monitoring-network/air-quality-monitoring-equipment.html  Environmental Protection Department. (2006b). EPD’s Air Quality Monitoring Network. Retrieved Jun. 7, 2014 from http://www.epd.gov.hk/epd/english/environmentinhk/air/air_quality/backgdf_3_1.html  Environmental Protection Department. (2003). Air Quality in Hong Kong 2003. Epd-asg.gov.hk. Retrieved Jan. 19, 2013 from http://www.epd.gov.hk/epd/english/environmentinhk/air/air_quality/files/aqr03e.pdf  ESRI. (2012). ArcGIS Desktop (Version 10.1) [Computer software]. Redlands, CA: Environmental Systems Research Institute 111   ESRI. (2013). ArcGIS Desktop (Version 10.2) [Computer software]. Redlands, CA: Environmental Systems Research Institute  Fan, X., Lam, K. C., & Yu, Q. (2012). Differential exposure of the urban population to vehicular air pollution in Hong Kong. Science of the Total Environment, 426, 211-219. doi: 10.1016/j.sciotenv.2012.03.057  Fenn, M. E., Sickman, J. O., Bytnerowicz, A., Clow, D. W., Molotch, N. P., Pleim, J. E., ... & Campbell, D. H. (2009). Methods for measuring atmospheric nitrogen deposition inputs in arid and montane ecosystems of western North America. Developments in environmental science, 9, 179-228. doi: 10.1016/S1474-8177(08)00208-8  Fung, W.Y. & Renguang, W. (2014). Relationship between intraseasonal variations of air pollution and meteorological variables in Hong Kong, Annals of GIS, 20(3), 217-226, doi: 10.1080/19475683.2014.945480   Gao, Y., Chan, E., Li, L., Lau, P., & Wong, T. (2014). Chronic effects of ambient air pollution on respiratory morbidities among Chinese children: A cross-sectional study in Hong Kong. Bmc Public Health, 14(1), 105-105. doi:10.1186/1471-2458-14-105  Georgakis, C., & Santamouris, M. (2008). On the estimation of wind speed in urban canyons for ventilation purposes—Part 1: Coupling between the undisturbed wind speed and the canyon wind.  Building and Environment, 43(8), 1404-1410. doi:10.1016/j.buildenv.2007.01.041  Giridharan, R., Ganesan, S., & Lau, S. S. Y. (2004). Daytime urban heat island effect in high-rise and high-density residential developments in Hong Kong.Energy and Buildings, 36(6), 525-534. doi: 10.1016/j.enbuild.2003.12.016  Goodenough, A. E., Hart, A. G., & Stafford, R. (2012). Regression with empirical variable selection: description of a new method and application to ecological datasets. PLoS One, 7(3), e34338-e34338. doi: 10.1371/journal.pone.0034338  Goudreau S, Plante C, Fournier M, Brand A, Roche Y, Smargiassi A. (2014). Estimation of Spatial Variations in Urban Noise Levels with a Land Use Regression Model. Environment and Pollution 3(4): 48-58. DOI: 10.5539/ep.v3n4p48  GovHK. (2015). Hong Kong – the Facts. Retrieved Nov. 8, 2015 from http://www.gov.hk/en/about/abouthk/facts.htm  Gradko. (2012). Technical Data Sheet: TDS 1. Retrieved Nov. 16, 2015 from http://www.gradko.com/environmental/products/no2-and-nitrogen-oxides-diffusion-tubes.shtml  112  Gu, Z. L., Zhang, Y. W., Cheng, Y., & Lee, S. C. (2011). Effect of uneven building layout on air flow and pollutant dispersion in non-uniform street canyons. Building and Environment, 46(12), 2657-2665. doi: 10.1016/j.buildenv.2011.06.028  Gulliver, J., de Hoogh, K., Hansell, A., & Vienneau, D. (2013). Development and back-extrapolation of NO2 land use regression models for historic exposure assessment in Great Britain. Environmental science & technology,47(14), 7804-7811. doi: 10.1021/es4008849  Guo, H., Ling, Z. H., Cheung, K., Jiang, F., Wang, D. W., Simpson, I. J., ... & Blake, D. R. (2013). Characterization of photochemical pollution at different elevations in mountainous areas in Hong Kong. Atmospheric Chemistry and Physics, 13(8), 3881-3898. doi:10.5194/acp-13-3881-2013  Guo, H., Wang, T., & Louie, P. K. K. (2004). Source apportionment of ambient non-methane hydrocarbons in Hong Kong: application of a principal component analysis/absolute principal component scores (PCA/APCS) receptor model. Environmental Pollution, 129(3), 489-498. doi: 10.1016/j.envpol.2003.11.006  Habermann, M., & Gouveia, N. (2012). Application of land use regression to predict the concentration of inhalable particular matter in São Paulo city, Brazil. Engenharia Sanitaria e Ambiental, 17(2), 155-162. doi: 10.1590/S1413-41522012000200004   Hagenbj€ork-Gustafsson, A., Tornevi, A., Forsberg, B., & Eriksson, K. (2010). Field validation of the Ogawa diffusive sampler for NO2 and NOx in a cols climate. Journal of Environmental Monitoring, 12, 1315-1324. doi: 10.1039/b924615k  Harrell, F. (2015). Regression modelling strategies: with applications to linear models, logistic and ordinal regression, and survival analysis. Springer.  Harrell, F. E., Lee, K. L., & Mark, D. B. (1996). Tutorial in biostatistics multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Statistics in medicine, 15, 361-387.   Harman, I. N., Barlow, J. F., & Belcher, S. E. (2004). Scalar fluxes from urban street canyons part II: model. Boundary-Layer Meteorology, 113(3), 387-410. doi: 10.1007/s10546-004-6205-7  Hedley, A. J., McGhee, S. M., Barron, B., Chau, P., Chau, J., Thach, T. Q.. . Wong, C. (2008). Air pollution: Costs and paths to a solution in Hong Kong-understanding the connections among visibility, air pollution, and health costs in pursuit of accountability, environmental justice, and health protection. Journal of Toxicology and Environmental Health, Part A, 71(9), 544-554. doi:10.1080/15287390801997476 113   HK Electric Investments Limited. (2014). Electricity Generation. Retrieved Jun. 12, 2014 from http://www.hkelectric.com/web/AboutUs/Generation/Index_en.htm  Ho, K. F., Lee, S. C., Chan, C. K., Yu, J. C., Chow, J. C., & Yao, X. H. (2003). Characterization of chemical species in PM2.5 and PM10 aerosols in Hong Kong. Atmospheric Environment, 37(1), 31-39. doi: 10.1016/S1352-2301(02)00804-X  Ho, S. S. H., Ho, K. F., Lee, S. C., Cheng, Y., Yu, J. Z., Lam, K. M., ... & Huang, Y. (2012). Carbonyl emissions from vehicular exhausts sources in Hong Kong. Journal of the Air & Waste Management Association, 62(2), 221-234. doi: 10.1080/10473289.2011.642952  Hoek, G., Beelen, R., de Hoogh, K., Vienneau, D., Gulliver, J., Fischer, P., & Briggs, D. (2008). A review of land-use regression models to assess spatial variation of outdoor air pollution. Atmospheric Environment, 42(33), 7561-7578. doi: 10.1016/j.stmosenv.2008.05.057  Hoek, G., Brunekreef, B., Goldbohm, S., Fischer, P., & van den Brandt, P. A. (2002). Association between mortality and indicators of traffic-related air pollution in the Netherlands: a cohort study. The lancet, 360(9341), 1203-1209. doi: 10.1016/S0140-6736(02)11280-3  Hoffmann, B., Moebus, S., Stang, A., Beck, E. M., Dragano, N., Möhlenkamp, S., ... & Jöckel, K. H. (2006). Residence close to high traffic and prevalence of coronary heart disease. European Heart Journal, 27(22), 2696-2702. doi: 10.1093/eurheartj/ehl278  Hong Kong Observatory (2015a). Numbers of hours of reduced visibility observed at the Hong Kong Observatory since 1968. Hong Kong Observatory. Retrieved Nov. 14, 2015 from http://www.hko.gov.hk/cis/statistic/hko_redvis_statistic_e.htm  Hong Kong Observatory. (2015b). Information of Weather Station. Retrieved Nov. 21, 2015 from http://www.weather.gov.hk/cis/annex/hkwxstn_e.htm  Hong Kong Observatory (2014). Observed Climate Change in Hong Kong: Wind speed. Retrieved July 2nd from http://www.hko.gov.hk/climate_change/obs_hk_wind_e.htm  Hsu, H. H., Adamkiewicz, G., Houseman, E. A., Zarubiak, D., Spengler, J. D., & Levy, J. I. (2013). Contributions of aircraft arrivals and departures to ultrafine particle counts near Los Angeles International Airport. Science of the Total Environment, 444, 347-355. doi: 10.1016/j.scitotenv.2012.12.010  114  Hudda, N., Gould, T., Hartin, K., Larson, T. V., & Fruin, S. A. (2014). Emissions from an international airport increase particle number concentrations 4-fold at 10 km downwind. Environmental science & technology, 48(12), 6628-6635. doi: 10.1021/es5001566  Information Services Department. (2015a). Population factsheet. Retrieved Nov. 8, 2015 from http://www.gov.hk/en/about/abouthk/factsheets/docs/population.pdf  Information Services Department. (2015b). Environmental Protection. Retrieved Nov. 13, 2015 from http://www.gov.hk/en/about/abouthk/factsheets/docs/environmental_ protection.pdf  Isakov, V., Johnson, M., Touma, J., & Özkaynak, H. (2012). Development and evaluation of land-use regression models using modeled air quality concentrations. (pp. 717-722). Dordrecht: Springer Netherlands. doi:10.1007/978-94-007-1359-8_117  Janssen, N., Hoek, G., Simic-Lawson, M., Fischer, P., van Bree, L., ten Brink, H.. . Cassee, F. (2011). Black carbon as an additional indicator of the adverse health effects of airborne particles compared with PM10 and PM2.5. Environmental Health Perspectives,119(12), 1691-1699. doi:10.1289/ehp.1003369  Jerrett, M., Arain, A., Kanaroglou, P., Beckerman, B., Potoglou, D., Sahsuvaroglu, T., ... & Giovis, C. (2005). A review and evaluation of intraurban air pollution exposure models. Journal of Exposure Science and Environmental Epidemiology, 15(2), 185-204. doi: 10.1038/sj.jea.7500388   Johnson, M., Isakov, V., Touma, J. S., Mukerjee, S., & Özkaynak, H. (2010). Evaluation of land-use regression models used to predict air quality concentrations in an urban area. Atmospheric Environment, 44(30), 3660-3668. doi: 10.1016/j.stmosenv.2010.06.041  Kashima, S., Yorifuji, T., Tsuda, T., & Doi, H. (2009). Application of land use regression to regulatory air quality data in Japan. Science of the Total Environment, 407(8), 3055-3062. doi: 10.1016/j.scitotenv.2008.12.038  Kasper, A., Aufdenblatten, S., Forss, A., Mohr, M., & Burtscher, H. (2007). Particulate emissions from a low-speed marine diesel engine. Aerosol Science and Technology, 41(1), 24-32. doi: 10.1080/02786820601055392  Kestens Y, Brand A, Fournier M, Goudreau S, Kosatsky T, Maloley M, Smargiassi A. 2011. Modelling the variation of land surface temperature as determinant of risk of heat-related health events. Int J Health Geogr 10(7): 1-9. DOI: 10.1186/1476-072X-10-7  115  Keuken, M. P., Jonkers, S., Zandveld, P., & Voogt, M. (2012). Elemental carbon as an indicator for evaluating the impact of traffic measures on air quality and health. Atmospheric Environment, 61, 1-8. doi: 10.1016/j.atmosenv.2012.07.009  Ko, F. W. S., Tam, W., Wong, T. W., Lai, C. K. W., Wong, G. W. K., Leung, T. F., ... & Hui, D. S. C. (2007). Effects of air pollution on asthma hospitalization rates in different age groups in Hong Kong. Clinical & Experimental Allergy, 37(9), 1312-1319. doi: 10.1111/j.1365-2222.2007.02791.x  Kok, G. L., Lind, J. A., & Fang, M. (1997). An airborne study of air quality around the Hong Kong territory. Journal of Geophysical Research, 102(D15), 19043-19. doi: 10.1029/97JD01306  Künzli, N., Kaiser, R., Medina, S., Studnicka, M., Chanel, O., Filliger, P., ... & Sommer, H. (2000). Public-health impact of outdoor and traffic-related air pollution: a European assessment. The Lancet, 356(9232), 795-801. doi: 10.1016/S0140-6736(00)02653-2  Lack, D. A., Cappa, C. D., Langridge, J., Bahreini, R., Buffaloe, G., Brock, C., ... & Williams, E. (2011). Impact of fuel quality regulation and speed reductions on shipping emissions: Implications for climate and air quality. Environmental science & technology, 45(20), 9052-9060. doi: 10.1021/es2013424  Lai, H. K., Ho, S. Y., Wong, C. M., Mak, K. K., Lo, W. S., & Lam, T. H. (2010). Exposure to particulate air pollution at different living locations and respiratory symptoms in Hong Kong–an application of satellite information. International journal of environmental health research, 20(3), 219-230. doi: 10.1080/09603120903511119  Lai, H. K., Tsang, H., & Wong, C. M. (2013). Meta-analysis of adverse health effects due to air pollution in Chinese populations. BMC public health, 13(1), 360. doi: 10.1186/1471-2458-13-360  Lau, A., Lo, A., Grey, J., Yuan, Z., & Loh, C. (2007). Relative Significance of Local vs. Regional Sources: Hong Kong’s Air Pollution. Retrieved Dec. 7, 2013 from http://www.civic-exchange.org/eng/upload/files/200703_HKAirPollution.pdf  Lau, J., Hung, W. T., & Cheung, C. S. (2012). Observation of increases in emission from modern vehicles over time in Hong Kong using remote sensing. Environnemental pollution, 163, 14-23. doi: 10.1016/j.envpol.2011.12.021  Lau, K, Wu, W.M., Fung, J.C.H., Henry, R.C, and Barron, B. (2005). Significant Marine Source for SO2 Levels In Hong Kong. Civic-exchange.org. Retrieved Dec. 8, 2013, http://www.civic-exchange.org/eng/upload/files/200506_MarineSourceSO2.pdf 116   Lee, J. Y., Leem, J. H., Kim, H. C., Hwang, S. S., Jung, D. Y., Park, M. S., ... & Kang, S. C. (2012). Land Use Regression Model for Assessing Exposure and Impacts of Air Pollutants in School Children. Journal of Korean Society for Atmospheric Environment, 28(5), 571-580. DOI: 10.5572/KOSAE.2012.28.5.571  Lee, S., Tinsley, H., Chau, J., Lai, H., Thach, T., Ip, P., ... Hedley, A. (2013). Review of outdoor air pollution and child health in Hong Kong. Hong Kong Journal of Paediatrics, 18(2), 69-76.  Li, C., Du, S. Y., Bai, Z. P., Shao-fei, K., Yan, Y., Bin, H., ... & Li, Z. Y. (2010). Application of land use regression for estimating concentrations of major outdoor air pollutants in Jinan, China. Journal of Zhejiang University SCIENCE A, 11(11), 857-867. doi: 10.1631/jzus.A1000092  Li, X., Liu, W., Chen, Z., Zeng, G., Hu, C., León, T., ... & Yan, W. (2015). The application of semicircular-buffer-based land use regression models incorporating wind direction in predicting quarterly NO 2 and PM 10 concentrations. Atmospheric Environment, 103, 18-24. DOI: 10.1016/j.atmosenv.2014.12.004  Lin, H., An, Q., Luo, C., Pun, V., Chan, C., & Tian, L. (2013). Gaseous air pollution and acute myocardial infarction mortality in Hong Kong: A time-stratified case-crossover study. Atmospheric Environment, 76, 68-73. doi:10.1016/j.atmosenv.2012.08.043  Liu, H., & Chan, J. C. (2002). An investigation of air-pollutant patterns under sea–land breezes during a severe air-pollution episode in Hong Kong. Atmospheric Environment, 36(4), 591-601. doi: 10.1016/S1352-2310(01)00504  Lo, J. C., Lau, A. K., Fung, J. C., & Chen, F. (2006). Investigation of enhanced cross‐ city transport and trapping of air pollutants by coastal and urban land‐sea breeze circulations. Journal of Geophysical Research: Atmospheres (1984–2012), 111(D14). doi: 10.1029/2005JD006837  Louie, P. K., Watson, J. G., Chow, J. C., Chen, A., Sin, D. W., & Lau, A. K. (2005). Seasonal characteristics and regional transport of PM2.5 in Hong Kong. Atmospheric Environment, 39(9), 1695-1710. doi: 10.1016/j.atmosenv.2004.11.017  Lu, W. Z., He, H. D., & Leung, A. Y. (2011). Assessing air quality in Hong Kong: A proposed, revised air pollution index (API). Building and Environment,46(12), 2562-2569. doi: 10.1016/j.buildenv.2011.06.011  Lumley, T. (2009). Leaps: regression subset selection. R package version 2.9. Retrieved Nov. 21, 2015 from http://CRAN.R-project.org/package=leaps 117   McLeod, A. I., & Xu, C. (2010). Bestglm: Best subset GLM. URL http://CRAN. R-project. org/package= bestglm. Retrieved Nov. 21, 2015 from https://cran.r-project.org/web/packages/bestglm/vignettes/bestglm.pdf  Meng, X., Chen, L., Cai, J., Zou, B., Wu, C. F., Fu, Q., ... & Kan, H. (2015). A land use regression model for estimating the NO 2 concentration in shanghai, China. Environmental research, 137, 308-315. doi: 10.1016/j.envres.2015.01.003  Meng, X., Fu, Q., Ma, Z., Chen, L., Zou, B., Zhang, Y., ... & Liu, Y. (2016). Estimating ground-level PM 10 in a Chinese city by combining satellite data, meteorological information and a land use regression model. Environmental Pollution, 208, 177-184. doi: 10.1016/j.envpol.2015.09.042  Moldanová, J., Fridell, E., Winnes, H., Holmin-Fridell, S., Boman, J., Jedynska, A., ... & Niessner, R. (2013). Physical and chemical characterisation of PM emissions from two ships operating in European Emission Control Areas. Atmospheric Measurement Techniques Discussions,6(2), 3931-3982. doi: 10.5194/amtd-6-3931-2013  Morawska, L., Thomas, S., Gilbert, D., Greenaway, C., & Rijnders, E. (1999). A study of the horizontal and vertical profile of submicrometer particles in relation to a busy road. Atmospheric Environment, 33(8), 1261-1274. doi: 10.1016/S1352-2310(98)00266-0  Murakami, T. (1979). Winter monsoonal surges over East and Southeast Asia. Journal of the Meteorological Society of Japan, 57. 133-158  Murena, F., Favale, G., Vardoulakis, S., & Solazzo, E. (2009). Modelling dispersion of traffic pollution in a deep street canyon: Application of CFD and operational models. Atmospheric Environment, 43(14), 2303-2311. doi:10.1016/j.atmosenv.2009.01.038  Nakamura, Y., & Oke, T. (1988). Wind, temperature and stability conditions in an east west oriented urban canyon. Atmospheric Environment, 22(12), 2691-2700. doi: 10.1016/0004-6981(88)90437-4  Ng, E., Yuan, C., Chen, L., Ren, C., & Fung, J. C. (2011). Improving the wind environment in high-density cities by understanding urban morphology and surface roughness: a study in Hong Kong. Landscape and Urban Planning,101(1), 59-74. doi: 10.1016/j.landurbplan.2011.01.004  O’Brien, R. M. (2007). A caution regarding rules of thumb for variance inflation factors. Quality & Quantity, 41(5), 673-690. doi: 10.1007/s11135-006-9018-6  Ogawa & Co., USA. (2014). FAQs. Retrieved Jun. 7, 2014 from http://ogawausa.com/new-faqs/  118  Ogawa & Co., USA. (2006). NO, NO2, NOX and SO2 Sampling Protocol Using The Ogawa Sampler. Retrieved Feb. 11, 2014 from http://www.ogawausa.com/pdfs/prono-noxno2so206.pdf  Pope, C. A., Dockery, D. W., & Schwartz, J. (1995). Review of epidemiological evidence of health effects of particulate air pollution. Inhalation toxicology, 7(1), 1-18. doi:10.3109/08958379509014267  Pun, V. C., Yu, I. T. S., Qiu, H., Ho, K. F., Sun, Z., Louie, P. K., ... & Tian, L. (2014). Short-term associations of cause-specific emergency hospitalizations and particulate matter chemical components in Hong Kong. American journal of epidemiology, 179(9), 1086-1095. doi: 10.1093/aje/kwu026   Qin, Y., Chan, C. K., & Chan, L. Y. (1997). Characteristics of chemical compositions of atmospheric aerosols in Hong Kong: spatial and seasonal distributions. Science of the total environment, 206(1), 25-37. doi: 10.1016/S0048-9697(97)00214-3  Qiu, H., Yu, I. T. S., Wang, X., Tian, L., Tse, L. A., & Wong, T. W. (2012). Season and humidity dependence of the effects of air pollution on COPD hospitalizations in Hong Kong. Atmospheric Environment. doi: 10.1016/j.atmosenv.2012.07.026  R Core Team. (2015). R: A language and environment for statistical computing [Computer software]. Vienna, Austria: R Foundation for Statistical Computing.  Retrieved Nov. 21, 2015 from http://www.r-project.org/  Rakowska, A., Wong, K. C., Townsend, T., Chan, K. L., Westerdahl, D., Ng, S., ... & Ning, Z. (2014). Impact of traffic volume and composition on the air quality and pedestrian exposure in urban street canyon. Atmospheric Environment, 98, 260-270. doi: 10.1016/j.atmosenv.2014.08.073  Ryan, P. H., & LeMasters, G. K. (2007). A review of land-use regression models for characterizing intraurban air pollution exposure. Inhalation Toxicology,19(S1), 127-133. doi: 10.1080/08958370701495998  Saraswat, A., Apte, J. S., Kandlikar, M., Brauer, M., Henderson, S. B., & Marshall, J. D. (2013). Spatiotemporal land use regression models of fine, ultrafine, and black carbon particulate matter in New Delhi, India. Environmental science & technology, 47(22), 12903-12911. doi: 10.1021/es401489h  School of Population and Public Health. (2016). Hedley Environmental Index. Retrieved Nov.14, 2015 from http://hedleyindex.sph.hku.hk/html/en/  119  Shamo, B., Asa, E., & Membah, J. (2014). Linear Spatial Interpolation and Analysis of Annual Average Daily Traffic Data. Journal of Computing in Civil Engineering, 29(1), 04014022. doi: 10.1061/(ASCE)CP.1943-5487.0000281  Shi, Y., Lau, K. K., & Ng, E. (2016). Developing street-level PM2.5 and PM10 land use regression models in high-density Hong Kong with urban morphological factors. Environmental Science & Technology, 50(15), 8178. doi: 10.1021/acs.est.6b01807  Shumake-Guillemot, J., Jalkanen, L., Adhair-Rohani, H. (2014). Air Quality and Human Health, a Priority for Joint Action. Weather and Climate – Understanding risks and preparing for variability and extremes, 63(2). Retrieved Nov. 8, 2015 from http://www.wmo.int/bulletin/en/content/air-quality-and-human-health-priority-joint-action  So, K. L., & Wang, T. (2004). C3-C12 non-methane hydrocarbons in subtropical Hong Kong: spatial–temporal variations, source–receptor relationships and photochemical reactivity. Science of the Total Environment, 328(1), 161-174. doi: 10.1016/j.scitotenv.2004.01.029  So, K. L., & Wang, T. (2003). On the local and regional influence on ground-level ozone concentrations in Hong Kong. Environmental Pollution, 123(2), 307-317. doi: 10.1016/S0269-7491(02)00370-6  Song, S., Wu, Y., Xu, J., Ohara, T., Hasegawa, S., Li, J.. & Hao, J. (2013). Black carbon at a roadside site in beijing: Temporal variations and relationships with carbon monoxide and particle number size distribution. Atmospheric Environment, 77, 213-221. doi:10.1016/j.atmosenv.2013.04.055  Styen, D. G. (2003). Mesoscale Circulations and Regional Air Pollution. In D. Melas & D. Syrakov (Eds.). Air Pollution Processes in Regional Scale. (pp. 281–294). doi:  10.1007/978-94-007-1071-9_31  Su, J. G., Brauer, M., Ainslie, B., Steyn, D., Larson, T., & Buzzelli, M. (2008). An innovative land use regression model incorporating meteorology for exposure analysis. Science of the total environment, 390(2), 520-529. doi: 10.1016/j.scitotenv.2007.10.032   Su, J. G., Brauer, M., & Buzzelli, M. (2008). Estimating urban morphometry at the neighborhood scale for improvement in modeling long-term average air pollution concentrations. Atmospheric Environment, 42(34), 7884-7893. doi: 10.1016/j.atmosenv.2008.07.023  120  Su, J. G., Jerrett, M., Beckerman, B., Wilhelm, M., Ghosh, J. K., & Ritz, B. (2009). Predicting traffic-related air pollution in Los Angeles using a distance decay regression selection strategy. Environmental Research,109(6), 657-670. doi:10.1016/j.envres.2009.06.001  Tang, R., Blangiardo, M., & Gulliver, J. (2013). Using building heights and street configuration to enhance intraurban PM10, NOx, and NO2 Land use regression models. Environmental science & technology, 47(20), 11643-11650. doi: 10.1021/es402156g  Tian, L., Hossain, S. R., Lin, H., Ho, K. F., Lee, S. C., & Ignatius, T. S. (2011). Increasing trend of primary NO2 exhaust emission fraction in Hong Kong. Environmental geochemistry and health, 33(6), 623-630. doi: 10.1007/s10653-011-9375-5  Transport Department. (2015). The Annual Traffic Census 2014.[no. 15CAB1] Retrieved Nov. 13, 2015 from http://www.td.gov.hk/en/publications_and_press_releases/ publications/free_publications/the_annual_traffic_census_2014/index.html  Tsai, M. Y., & Chen, K. S. (2004). Measurements and three-dimensional modeling of air pollutant dispersion in an urban street canyon. Atmospheric Environment, 38(35), 5911-5924. doi: 10.1016/j.atmosenv.2004.07.008  TSI. (2012). Sidepak Personal Aerosol Monitor Model AM510. Retrieved Feb. 11, 2014 from http://www.tsi.com/uploadedFiles/_Site_Root/Products/Literature/Manuals/ SidePak_AIM510_US_1980456-web.pdf  Unal, A., Hu, Y., Chang, M. E., Odman, M. T., & Russell, A. G. (2005). Airport related emissions and impacts on air quality: Application to the Atlanta International Airport. Atmospheric Environment, 39(32), 5787-5798. doi:10.1016/j.atmosenv.2005.05.051   Unger, N. (2012). Global Climate Forcing by Criteria Air Pollutants. Annual Review of Environment and Resources, 37, 1-24.  Van Reeuwijk, H., Fischer, P. H., Harssema, H., Briggs, D. J., Smallbone, K., & Lebret, E. (1998). Field comparison of two NO2 passive samplers to assess spatial variation. Environmental Monitoring and Assessment, 50(1), 37-51. doi: 10.1023/A:1005703722232  Vardoulakis, S., Fisher, B. E., Pericleous, K., & Gonzalez-Flesca, N. (2003). Modelling air quality in street canyons: a review. Atmospheric environment,37(2), 155-182. doi: 10.1016/S1352-2310(02)00857-9  Virkkula, A., Mäkelä, T., Hillamo, R., Yli-Tuomi, T., Hirsikko, A., Hämeri, K., & Koponen, I. K. (2007). A simple procedure for correcting loading effects of aethalometer data. Journal of the Air & Waste Management Association, 57(10), 1214-1222. doi:10.3155/1047-3289.57.10.1214  121  Wang, M., Beelen, R., Basagana, X., Becker, T., Cesaroni, G., de Hoogh, K., ... & Brunekreef, B. (2013). Evaluation of land use regression models for NO2 and particulate matter in 20 european study areas: the ESCAPE project.  Environmental science & technology, 47(9), 4357-4364. doi: 10.1021/es305129t  Wang, M., Beelen, R., Eeftens, M., Meliefste, K., Hoek, G., & Brunekreef, B. (2012). Systematic evaluation of land use regression models for NO2.Environmental science & technology, 46(8), 4481-4489. doi: 10.1021/es204183v  Wang, R. (2011). Assessment of the Temporal Stability of Land Use Regression Models for Traffic-Related Air Pollution. [Masters thesis]. Retrieved from cIRcle. http://hdl.handle.net/2429/39558.  Wang, R., Henderson, S., Sbihi, H., Allen, R., & Brauer, M. (2013). Temporal stability of land use regression models for traffic-related air pollution. Atmospheric Environment, 64, 312-319. doi:10.1016/j.atmosenv.2012.09.056  Wang, T., Poon, C. N., Kwok, Y. H., & Li, Y. S. (2003). Characterizing the temporal variability and emission patterns of pollution plumes in the Pearl River Delta of China. Atmospheric Environment, 37(25), 3539-3550. doi:10.1016/S1352-2310(03)00363-7  Wang, T., Wu, Y. Y., Cheung, T. F., & Lam, K. S. (2001). A study of surface ozone and the relation to complex wind flow in Hong Kong. Atmospheric Environment, 35(18), 3203-3215. doi: 10.1016/S1352-2310(00)00558-6  Wang, X. K., & Lu, W. Z. (2006). Seasonal variation of air pollution index: Hong Kong case study. Chemosphere, 63(8), 1261-1272. doi: 10.1016/j.chemosphere.2005.10.031  Weichenthal, S., Van Ryswyk, K., Goldstein, A., Shekarrizfard, M., & Hatzopoulou, M. (2015). Characterizing the spatial distribution of ambient ultrafine particles in Toronto, Canada: A land use regression model.Environmental Pollution. doi: 10.1016/j.envpol.2015.04.011  Winnes, H. (2010). Air pollution from ships: Emission measurements and impact assessments. [Ph.D. Dissertation]. Chalmers University of Technology, Gothenburg, Sweden,   Wong, C. M., Atkinson, R. W., Anderson, H. R., Hedley, A. J., Ma, S., Chau, P. Y. K., & Lam, T. H. (2002). A tale of two cities: effects of air pollution on hospital admissions in Hong Kong and London compared. Environmental health perspectives, 110(1), 67-77. doi: 10.1289/ehp.0211067  122  Wong, C. M., Ma, S., Hedley, A. J., & Lam, T. H. (2001). Effect of air pollution on daily mortality in Hong Kong. Environmental health perspectives, 109(4), 335. doi: 10.1289/ehp.01109335  Wong, T. W., Lau, T. S., Yu, T. S., Neller, A., Wong, S. L., Tam, W., & Pang, S. W. (1999). Air pollution and hospital admissions for respiratory and cardiovascular diseases in Hong Kong. Occupational and environmental medicine, 56(10), 679-683. doi: 10.1136/oem.56.10.679  Wong, T. W., Tam, W. S., Yu, T. S., & Wong, A. H. S. (2002). Associations between daily mortalities from respiratory and cardiovascular diseases and air pollution in Hong Kong, China. Occupational and environmental medicine,59(1), 30-35. doi: 10.1136/oem.59.1.30  World Health Organization. (2016). Ambient (outdoor) air pollution database, by country and city [Data file]. Retrieved May 12, 2016 from http://www.who.int/phe/health_topics/outdoorair/databases/cities/en/  World Health Organization. (2014). Ambient (outdoor) air quality and health - Fact sheet N°313. Retrieved Jul. 4, 2014 from http://www.who.int/mediacentre/factsheets/fs313/en/  World Health Organization. (2013). Review of Evidence on Health Aspects of Air Pollution—REVIHAAP Project: Technical Report. WHO, Copenhagen. Retrieved Nov. 8, 2015 from http://www.euro.who.int/__data/assets/pdf_file/0004/193108/REVIHAAP-Final-technical-report-final-version.pdf?ua=1  World Health Organization. (2006). WHO air quality guidelines for particulate matter, ozone, nitrogen dioxide and sulfur dioxide: Global update 2005: summary of rick assessment. Retrieved June 14, 2016 from http://apps.who.int/iris/bitstream/10665/69477/1/WHO_SDE_PHE_OEH_06.02_eng.pdf  World Health Organization. (2002). The world health report 2002: Reducing risks, promoting healthy life. Retrieved Jan. 23, 2014 from http://www.who.int/whr/2002/en/whr02_en.pdf  Xiao, F., Brajer, V., & Mead, R. W. (2006). Blowing in the wind: The impact of China's Pearl River Delta on Hong Kong's air quality. Science of the total environment, 367(1), 96-111. doi: 10/1016/j.scitotenv.2006.01.010  Xie D, Liu Y, Chen J. (2011). Mapping urban environmental noise: a land use regression method. Environmental science & technology 45(17): 7358-7364. doi: 10.1021/es200785x  Ye, Y., Zhang, H., Liu, K., & Wu, Q. (2013). Research on the influence of site factors on the expansion of construction land in the Pearl River Delta, China: By using GIS and remote 123  sensing. International Journal of Applied Earth Observation and Geoinformation, 21, 366-373. doi: 10.1016/j.jag.2011.10.012  Yu, K. N., Cheung, Y. P., Cheung, T., & Henry, R. C. (2004). Identifying the impact of large urban airports on local air quality by nonparametric regression. Atmospheric Environment, 38(27), 4501-4507. doi: 10.1016/j.atmosenv.2004.05.034  Yu, J. Z., Tung, J. W. T., Wu, A. W. M., Lau, A. K. H., Louie, P. K. K., & Fung, J. C. H. (2004). Abundance and seasonal characteristics of elemental and organic carbon in Hong Kong PM10. Atmospheric Environment, 38(10), 1511-1521. doi: 10.1016/j.atmosenv.2003.11.035  Yu, T. I., Wong, T. W., Wang, X. R., Song, H., Wong, S. L., & Tang, J. L. (2001). Adverse effects of low-level air pollution on the respiratory health of schoolchildren in Hong Kong. Journal of Occupational and Environmental Medicine, 43(4), 310-316. doi:10.1097/00043764-200104000-00004  Yuan, Z., Lau, A. K. H., Zhang, H., Yu, J. Z., Louie, P. K., & Fung, J. C. (2006). Identification and spatiotemporal variations of dominant PM10 sources over Hong Kong. Atmospheric Environment, 40(10), 1803-1815. doi: 10.1016/j.atmosenv.2005.11.030  Yuan, Z., Yadav, V., Turner, J. R., Louie, P. K., & Hon Lau, A. K. (2013). Long-term trends of ambient particulate matter emission source contributions and the accountability of control strategies in Hong Kong over 1998-2008. Atmospheric Environment, 76, 21-31. doi: 10.1016/l.atmosenv.2012.09.026  Zheng, M., Cheng, Y., Zeng, L., & Zhang, Y. (2011). Developing chemical signatures of particulate air pollution in the Pearl River Delta region, China. Journal of Environmental Sciences, 23(7), 1143-1149. doi: 10.1016/S1001-0742(10)60526-8  Zhong, L., Louie, P. K., Zheng, J., Yuan, Z., Yue, D., Ho, J. W., & Lau, A. K. (2013). Science-policy interplay: air quality management in the Pearl River Delta region and Hong Kong. Atmospheric Environment, 76, 3-10. doi: 10.1016/j.atmosenv.2013.03.012  Zhou W, Ji S, Chen T H, Hou Y, Zhang K. (2014). The 2011 heat wave in Greater Houston: effects of land use on temperature. Environmental research 135: 81-87. DOI: 10.1016/j.envres.2014.08.025  Zhou, Y., & Levy, J. I. (2008). The impact of urban street canyons on population exposure to traffic-related primary pollutants. Atmospheric environment, 42(13), 3087-3098. doi: 10.1016/j.atmosenv.2007.12.037  124  Zhu, Y., Fanning, E., Yu, R. C., Zhang, Q., & Froines, J. R. (2011). Aircraft emissions and local air quality impacts from takeoff activities at a large International Airport. Atmospheric Environment, 45(36), 6526-6533. doi: 0.1016/j.atmosenv.2011.08.062  Zou, B., Wilson, J. G., Zhan, F. B., & Zeng, Y. (2009a). Air pollution exposure assessment methods utilized in epidemiological studies. Journal of Environmental Monitoring, 11(3), 475-490. doi: 10.1039/b813889c   Zou, B., Wilson, J. G., Zhan, F. B., & Zeng, Y. (2009b). An emission-weighted proximity model for air pollution exposure assessment. Science of the total environment, 407(17), 4939-4945. doi: 10.1016/j.scitotenv.2009.05.014  Zou, H., Yue, Y., Li, Q., & Yeh, A. G. (2012). An improved distance metric for the interpolation of link-based traffic data using kriging: a case study of a large-scale urban road network. International Journal of Geographical Information Science, 26(4), 667-689. doi: 10.1080/13658816.2011.609488  125  Appendices  Appendix A  All roadside sampling sites   Site ID Latitude Longitude Region District Planned Collection1516 CW-A-1 22.28178 114.15812 Hong Kong Island Central and Western NO2, NO, PM (2w) CW-B-2 22.28103 114.15556 Hong Kong Island Central and Western NO2, PM CW-C-3 22.28135 114.15610 Hong Kong Island Central and Western NO2, NO, PM (2w) EN-A-1 22.28226 114.18340 Hong Kong Island Wan Chai NO2, PM EN-A-2 22.28207 114.18372 Hong Kong Island Wan Chai NO2, NO, PM EN-B-5 22.29105 114.20040 Hong Kong Island Eastern NO2, PM EN-B-7 22.28804 114.19340 Hong Kong Island Eastern NO2, NO (SC2 only), PM EN-B-8 22.28931 114.19460 Hong Kong Island Eastern NO2, NO (SC2 only), PM EN-C-10 22.29088 114.19957 Hong Kong Island Eastern NO2, NO, PM EN-C-9 22.29034 114.19627 Hong Kong Island Eastern NO2, NO, PM (2w) IS-A-2 22.30768 113.93743 New Territories Islands NO2, NO (SC2 only), PM IS-A-3 22.32826 114.02649 New Territories Islands NO2, PM IS-A-4 22.29365 113.95429 New Territories Islands NO2, NO, PM (2w) IS-B-6 22.28851 113.94170 New Territories Islands NO2, PM KC-A-1 22.32926 114.19281 Kowloon Peninsula Kowloon City NO2, NO, PM KC-A-3 22.30871 114.18347 Kowloon Peninsula Kowloon City NO2, NO (SC2 only), PM KC-A-4 22.30822 114.18261 Kowloon Peninsula Kowloon City NO2, NO (SC2 only), PM KC-B-5 22.30531 114.18902 Kowloon Peninsula Kowloon City NO2, NO, PM KC-B-6 22.30431 114.18978 Kowloon Peninsula Kowloon City NO2, NO (SC2 only), PM KC-B-7 22.32227 114.18778 Kowloon Peninsula Kowloon City NO2, NO (SC2 only), PM KC-B-8 22.32252 114.18849 Kowloon Peninsula Kowloon City NO2, PM KT-A-1 22.33100 114.20925 Kowloon Peninsula Kwun Tong NO2, PM KT-A-2 22.32981 114.21057 Kowloon Peninsula Kwun Tong NO2, NO, PM KT-B-6 22.31347 114.22508 Kowloon Peninsula Kwun Tong NO2, NO, PM (2w) KW-A-3 22.35916 114.12008 New Territories Kwai Tsing NO2, NO, PM (2w) MK-A-1 22.32244 114.16849 Kowloon Peninsula Yau Tsim Mong NO2, NO, PM (2w) NT-A-1 22.47931 114.15367 New Territories North NO2, PM NT-A-2 22.48393 114.15294 New Territories North NO2, NO (SC2 only), PM NT-A-3 22.48523 114.15163 New Territories North NO2, NO (SC2 only), PM                                                  15 NO2 refers to Ogawa badge, diffusion tube, or both. 16 PM refers to 24-hour PM2.5 and black carbon collection, while PM (2w) refers to 2 week collection. 126  NT-A-4 22.48642 114.14893 New Territories North NO2, PM NT-B-5 22.50270 114.12674 New Territories North NO2, PM NT-B-6 22.50184 114.12775 New Territories North NO2, NO, PM NT-B-7 22.50315 114.12671 New Territories North NO2, NO (SC2 only), PM NT-B-8 22.50291 114.12920 New Territories North NO2, NO (SC2 only), PM NT-B-9 22.50276 114.12901 New Territories North NO2, PM P1 22.28321 114.12749 Hong Kong Island Central and Western NO2, NO, PM P10 22.29624 114.16919 Kowloon Peninsula Yau Tsim Mong NO2, NO, PM P11 22.37122 114.11923 New Territories Tsuen Wan NO2, NO, PM P12 22.28729 114.21942 Hong Kong Island Eastern PM P2 22.28697 114.14045 Hong Kong Island Central and Western NO2, NO, PM P3 22.27956 114.17536 Hong Kong Island Wan Chai NO2, NO, PM P4 22.28289 114.22140 Hong Kong Island  Eastern NO2, NO, PM P5 22.26472 114.24122 Hong Kong Island  Eastern NO2, NO, PM P6 22.31605 114.16383 Kowloon Peninsula Yau Tsim Mong  NO2, NO, PM P7 22.31536 114.17038 Kowloon Peninsula Yau Tsim Mong  PM P8 22.31582 114.17333 Kowloon Peninsula Yau Tsim Mong  NO2, NO, PM (2w) P9 22.30532 114.17164 Kowloon Peninsula Yau Tsim Mong  NO2, NO, PM (2w) SK-A-2 22.31920 114.26438 New Territories Sai Kung NO2, NO, PM (2w) SK-A-3 22.30481 114.25361 New Territories Sai Kung NO2, NO, PM SK-A-6 22.29540 114.27331 New Territories Sai Kung NO2, PM SK-A-7 22.31810 114.25892 New Territories Sai Kung NO2, PM SK-A-9 22.38220 114.27097 New Territories Sai Kung NO2, PM SK-B-11 22.38203 114.27342 New Territories Sai Kung NO2, NO, PM SO-A-1 22.25036 114.17056 Hong Kong Island Southern NO2, NO, PM SO-A-2 22.25024 114.16969 Hong Kong Island Southern NO2, PM SO-A-3 22.24850 114.16159 Hong Kong Island Southern NO2, PM SO-A-4 22.24844 114.16287 Hong Kong Island Southern NO2, NO, PM (2w) SO-B-5 22.24896 114.15454 Hong Kong Island Southern NO2, NO, PM SO-B-6 22.24886 114.15366 Hong Kong Island Southern NO2, PM SSP-A-4 22.33551 114.15913 Kowloon Peninsula Sham Shui Po NO2, NO (SC2 only), PM SSP-B-5 22.32982 114.16348 Kowloon Peninsula Sham Shui Po NO2, PM SSP-B-6 22.33171 114.16102 Kowloon Peninsula Sham Shui Po NO2, NO (SC2 only), PM SSP-B-7 22.33087 114.16315 Kowloon Peninsula Sham Shui Po NO2, NO, PM (2w) SSP-C-10 22.33699 114.14700 Kowloon Peninsula Sham Shui Po NO2, NO, PM SSP-C-9 22.33701 114.14764 Kowloon Peninsula Sham Shui Po NO2, PM ST-A-1 22.38753 114.19191 New Territories Sha Tin NO2, NO (SC2 only), PM ST-B-5 22.37579 114.17734 New Territories Sha Tin NO2, NO, PM (2w) ST-B-6 22.37391 114.17856 New Territories Sha Tin NO2, PM ST-B-7 22.37374 114.17873 New Territories Sha Tin NO2, PM 127  ST-B-8 22.37349 114.17843 New Territories Sha Tin NO2, NO (SC2 only), PM TM-A-1 22.39960 113.97668 New Territories Tuen Mun NO2, NO, PM TM-A-4 22.40735 113.97831 New Territories Tuen Mun NO2, PM TM-B-6 22.39261 113.97841 New Territories Tuen Mun NO2, NO, PM (2w) TP-A-1 22.42411 114.21151 New Territories Tai Po NO2, NO, PM TP-A-2 22.43240 114.20111 New Territories Tai Po NO2, NO (SC2 only), PM TP-A-3 22.46646 114.15106 New Territories Tai Po NO2, NO, PM (2w) TP-A-4 22.46747 114.15176 New Territories Tai Po NO2, PM TP-B-5 22.44859 114.16702 New Territories Tai Po NO2, PM TP-B-6 22.44465 114.16666 New Territories Tai Po NO2, PM TP-C-10 22.44841 114.16740 New Territories Tai Po NO2, NO, PM (2w) TP-C-9 22.44903 114.16604 New Territories Tai Po NO2, PM TW-A-3 22.36251 114.11794 New Territories Tsuen Wan NO2, NO (SC2 only), PM TW-A-4 22.36397 114.11977 New Territories Tsuen Wan NO2, NO, PM TW-B-5 22.37062 114.11487 New Territories Tsuen Wan NO2, NO (SC2 only), PM TW-B-6 22.37004 114.11598 New Territories Tsuen Wan NO2, PM TW-B-7 22.36908 114.11636 New Territories Tsuen Wan NO2, NO (SC2 only), PM WC-A-1 22.28006 114.18436 Hong Kong Island Wan Chai NO2, NO, PM (2w) WC-B-2 22.28014 114.18578 Hong Kong Island Wan Chai NO2, PM WC-C-3 22.28080 114.18551 Hong Kong Island Wan Chai NO2, NO (SC2 only), PM WTS-A-1 22.33540 114.20355 Kowloon Peninsula Wong Tai Sin NO2, PM WTS-A-3 22.33620 114.20721 Kowloon Peninsula Wong Tai Sin NO2, NO, PM (2w) WTS-A-4 22.33705 114.20657 Kowloon Peninsula Wong Tai Sin NO2, PM WTS-B-5 22.34175 114.19530 Kowloon Peninsula Wong Tai Sin NO2, PM WTS-B-6 22.34175 114.19315 Kowloon Peninsula Wong Tai Sin NO2, NO, PM (2w) YL-A-1 22.46584 114.05380 New Territories Yuen Long NO2, NO, PM YL-A-2 22.45906 114.05216 New Territories Yuen Long NO2, PM YL-A-3 22.43547 114.02089 New Territories Yuen Long NO2, NO, PM (2w) YL-A-4 22.43460 114.02527 New Territories Yuen Long NO2, PM YL-B-5 22.44463 114.02935 New Territories Yuen Long NO2, PM YL-B-6 22.44432 114.02934 New Territories Yuen Long NO2, PM 128  Appendix B  Site selection flowchart    129  Appendix C  PM2.5 and black carbon data cleaning and scaling   This document was taken from documents written by Dr. Benjamin Barratt from King’s College London.  Data cleaning and scaling process (BC) 1. Carry out co-location of all instruments prior to campaign for at least 24 hours in order to identify any units not operating to within 10% of the mean of all units. Do not use any outlying units unless a consistent scaling factor can be derived using RMA regression. 2. Check site code assignments against field logs. 3. Import campaign data, retaining instrument status codes. 4. Exclude first and last 5 minute data points to allow for set up and take down period after instrument is switched on. 5. Manually examine data to exclude spurious periods not flagged by the unit and correct any time drift issues. 6. Canyon campaigns: Calculate mean k factors for each canyon using the Virkkula method (separate procedure). 7. Spatial campaigns: Calculate mean k factors for the campaign using the Virkkula method applied to co-location exercises prior to and following campaigns. Note that filter loading during the spatial campaigns was typically <70%, therefore filter correction had less impact on results. 8. Apply Virkkula filter correction to all data. 9. Export the dataset as hourly means.  Table C.1 Virkkula results (period = 20 minutes) Site Summer Winter Spatial 0.008 0.007 CHO 0.006 0.009 JDC 0.011 0.007 MKC 0.019 0.004 SWO 0.009 0.006 HHC  0.009 NPC  0.005 130    Data cleaning and scaling process (PM2.5)  1. Carry out co-location of all instruments prior to campaign for at least 24 hours. 2. Derive a Sidepak to reference ambient correction factor using co-location of a ‘reference’ unit with HK EPD PM2.5 reference monitor during the campaign using linear regression. If more than one co-location during campaign (spatial) use site with best R2. 3. Use ‘reference’ instrument to derive an offset and scaling factor for each other instrument based on linear regression with the reference. 4. Check site code assignments against field logs. 5. Import campaign data, retaining instrument status codes. 6. Apply unit campaign scaling factor if > +/- 10%. 7. Apply ambient correction factor. 8. Manually examine data to exclude spurious periods not flagged by the unit and correct any time drift issues.  9. Assign campaign site codes 10. Exclude first and last 5 minute data points to allow for set up and take down period after instrument is switched on. 11. Export the dataset as hourly means.  Figures C.1 and C.2 show the impact of scaling factors on PM2.5 SidePak data from the post SC2 instrument co-location.  131   Figure C.1 Comparison of co-location SidePak units prior to application of the ratification process  Figure C.2 Comparison of co-location SidePak units following ratification  132  Appendix D  Predictor variables for modelling  This appendix is located in the file “FullPredictorDataset.csv”.  133  Appendix E  Pre-selection ranking script  Based on R script found in Wang (2011) and Abernethy (2012).  # remove predictors (columns) that have less than 2 sites with non-zero value for (i in dim(org.data)[2]:4) {     if (length(which(org.data[,i]>0))<2) {ourdata <- ourdata[,-i]} } attach (ourdata)  ######################################## #   functions needed ########################################  # r table function # Rank table function get_ranks_table = function(y, modeltype){     if (modeltype == "length"){     covariates = names(cbind(ourdata[2], ourdata[14:61], ourdata[134:371]))   }   if (modeltype == "density"){     covariates = names(cbind(ourdata[2], ourdata[62:97],ourdata[134:371]))   }   if (modeltype == "loading"){     covariates = names(cbind(ourdata[2], ourdata[98:371]))   }   r2 = numeric()   for (variable in covariates){     model = lm(get(y)~get(variable), na.rm = T)     r2 = c(r2, as.double(summary(model)$r.squared))}   ranks = as.data.frame(covariates)   ranks$r2 = r2     ranks$abs.r = ranks$r2^(1/2)   ranks$covariates = as.character(ranks$covariates)   ranks$vartype = strsplit(ranks$covariates,"\\.")   ranks$vartype = sapply(ranks$vartype, '[[', 1)   ranks$vartype = sapply(ranks$vartype, paste, ".", sep='')   return(ranks) }  ############################################################################### # Function to return the names of variables within the same group correlated  # by less than 0.6 # Function variable group should be of form "RD1." # Function output is a list of character class, names of variables ############################################################################### include_in_stepwise = function(ranks, x){   vargroups = as.character(unique(ranks$vartype))   stepwiselist = character()   for (group in vargroups){     maxr2 = max(ranks$r2[ranks$vartype == group])     maxvar = ranks$covariates[ranks$r2 == maxr2]     subdata = as.data.frame(ourdata[,grep(group, names(ourdata))])     if (dim(subdata)[2] > 1){       varcor = cor(subdata)[maxvar,]       valid = c(maxvar, names(varcor)[varcor < 0.6])       if (length(valid) >1) {          remain <- subdata[,valid]         remain <- cbind(x, subdata[,valid])         test = cor(remain)[1,]         test <- test^2         y <- sort(test, decreasing = T)         left <- head(y, 4)         left <- left[-1] 134          ext = names(left)        }else {ext = c(maxvar)}     }     else {ext = c(maxvar)}     stepwiselist = c(stepwiselist, ext)   }   return(stepwiselist) }       135  Appendix F  Table of a priori hypotheses  Variable Group Code Direction of effects Expressway road length ExpRL + Main road length MainRL + Secondary road length SecRL + Elevated ElvRL + Expressway AADT AADTExp + Main AADT AADTMain + Secondary AADT AADTSec + Expressway traffic loading ExpTrL + Main Traffic Loading MainTrL + Secondary Traffic Loading SecTrrL + Longitude Long NA Latitude Lat NA Aspect Ratio AspRatio + Elevation Elevation NA Distance to Airport (Ln)Dist_Airport  - Distance from Coast (Ln)Dist_Coast N/A Distance from Port (Ln)Dist_Port - Distance from Shipping Lane (Ln)Dist_ShippingLanes - Distance from High Volume  (Ln)DistUnRHighVol - Distance from MTR Line (Ln)Dist_MTRLines NA Distance from MTR Station (Ln)Dist_MTRstnts NA Distance from Ferry Terminal (Ln)Dist_FerryTerm - Distance from Incinerator (Ln)Dist_Incin - Distance from Crematorium (Ln)Dist_Crematorium - Distance from Cross Border Vehicle Terminal (Ln)Dist_CrossBrdVehTm - Distance from Toll Gates (Ln)Dist_TollGate - Distance from Shenzhen (Ln)Dist_ShenzhenP - Distance from Power Stations (Ln)Dist_PSntswoPenny - Building Area  BldArD + Building Volume  BldVolD + Bus Terminus  BusTD + Car Park  CarPD + Food Stall  FoodStD + Temple  TmplD + Parking Area PrkArD NA Open Area OpArT - Undeveloped Land Lands - Park ParkT - 136  Government GovT N/A Industrial IndT + Commercial ComT N/A Residuals ResT N/A Mixed MixT N/A Traffic Intersections InterD + Mini Bus Terminus MiniBusD + Flight Path Buffer (B) FlightRouteBuf - Population Density WPopDen +    137  Appendix G  Temporal correction multiplication factors for PM2.5 and black carbon  Site Black carbon (SC1) Black carbon (SC2) PM2.5 (SC1) PM2.5 (SC2) CWA1 0.74 1.77 0.95 2.09 CWB2 0.91 0.73 0.78 0.75 ENA1 1.14 0.63 1.03 0.64 ENA2  -  - 0.86 1.24 ENB5 1.80 0.85 0.94 0.85 ENB7 1.15 0.63 1.65 0.64 ENB8 0.90 0.76 0.78 0.82 ENC10 1.87 0.79 0.88 0.82 ISA2 1.79 1.39 1.31 0.80 ISA3 1.81 1.40 1.31 1.27 ISB6 1.79 1.39 1.31 0.76 KCA1 0.88 0.70 1.03  - KCA3 1.82 0.73 0.79 0.83 KCA4 0.50 1.59 0.49 2.15 KCB5 0.90 0.67 0.90 0.78 KCB6 0.38 1.42 1.58 0.69 KCB7 0.37 0.68  - 0.80 KCB8 0.47 1.57  - 2.17 KTA1 0.44 1.53 0.45 0.98 KTA2 1.09 1.54 1.00 1.17 MKA1 0.85 0.62 0.49 0.57 NTA1 2.29 1.49 1.53  - NTA2 0.90 1.23 1.43 1.20 NTA3 1.22 1.03 1.15 0.59 NTA4 2.17 1.25 1.53 1.25 NTB5  - 1.63 1.61 1.71 NTB6 0.90 1.48 1.49 1.21 NTB7 1.20 1.61 1.13 1.66 NTB8 0.78 1.48 1.41 1.21 NTB9 0.89 1.23 1.42 1.20 P1 1.13 0.61 1.69 0.85 P10 0.38 0.70 0.39 0.87 P11 1.79 1.68 1.14 1.26 P12  - 0.83 0.87 0.85 P2 1.14 0.61 1.66 0.65 P3 1.23 0.95 1.66 0.84 138  P4 0.78 0.82 0.82 0.83 P5 1.73 0.61 0.80 0.65 P6 0.78 1.59 1.38 0.59 P7 1.27 1.20 1.19 1.18 P8 0.42 0.67 0.53 0.66 SKA3 1.70 1.08  - 0.95 SKA6 1.71 1.08  - 0.95 SKA7 1.69 1.08  - 0.95 SKA9 1.70 1.07 1.11 0.95 SKB11  - 1.07 1.13 0.95 SOA1 0.75 1.11 1.00  - SOA2 0.55 1.69 0.49 1.60 SOA3  - 2.26 0.86  - SOB5 0.75 1.43  - 2.17 SOB6 0.55 1.14 0.49 1.26 SSPA4 1.30 2.05 1.26 1.18 SSPB5  - 1.11 0.46 1.78 SSPB6 0.77 1.62  - 1.44 SSPC10 1.32 1.50 0.80 2.17 SSPC9 1.83 1.62 1.48 1.44 STA1  - 1.55 1.12 1.17 STB6 1.31 1.29 0.70 1.33 STB7 2.45 1.02  - 1.39 STB8 1.03 1.54 1.04 1.18 TMA1 2.32 1.68 1.64 1.26 TMA3 0.92 2.04 1.52 2.14 TMA4 1.23 1.74  - 1.43 TPA1 2.26 1.49 1.60 1.20 TPA2 0.90 1.60 1.36 0.59 TPA4 1.22 1.59 1.15 1.64 TPB5 1.28 1.29 1.70  - TPB6 2.26 1.28 1.38 1.30 TPC9 1.27 0.95  - 0.78 TWA4 0.98 1.24 1.52 1.26 TWB5 1.06 2.14 0.85 2.32 TWB6 1.15 1.68  - 1.26 TWB7 0.98 1.41 0.96 1.24 WCA1 0.92 1.08 0.68 0.87 WCB2 1.09 2.52 0.80 1.26 WCC3 1.92 1.71 1.38 1.41 139  WTSA1 1.19 1.00 0.96 0.84 WTSA4 0.44 1.53 0.44 0.98 WTSB5 0.48 1.58 0.44  - YLA1 0.90 1.97 1.36 1.23 YLA2 2.31 1.97  - 1.99 YLA4 2.30 1.85 1.56 1.39 YLB5 1.26 1.74 1.28 1.32 YLB6 1.35 2.02  - 2.15  140  Appendix H  Temporal correction multiplication factors for NO2 and NO Group Date NO NO2 Notes SC1 1 Apr. 25 – May 12 1.14 0.84 Roadside AQMS sites 2 Apr. 25 – May 13 1.05 0.84 Rooftop AQMS sites 3 Apr. 26 – May 12 1.13 0.83 EPD sites 4 Apr. 26 – May 13 1.03 0.84 Rooftop AQMS sites 5 Apr. 26 – May 14 0.98 0.86 HK2D exclusive sites 6 Apr. 27 – May 12 1.18 0.84 EPD sites 7 Apr. 27 – May 14 1.01 0.87 HK2D exclusive sites 8 Apr. 21 – May 12 - 0.85 EPD diffusion tubes 9 Apr. 28 – May 12 - 0.87 Replaced EPD diffusion tubes 10 May 13 – Jun. 3 - 1.22 Resampled diffusion tubes SC2 1 Jan. 3 – Jan. 24 0.98 1.01 EPD sites 2 Jan. 5 – Jan. 26 0.97 1.01 Rooftop AQMS and exclusive HK2D sites 3 Jan. 6 – Jan. 23 0.89 0.97 Rooftop AQMS and exclusive HK2D sites 4 Dec. 29 – Jan. 19 - 1.05 Diffusion tubes at SK sites        141  Appendix I  All 36 model results  Table I.1 NO2 annual models  Road length AADT Traffic loading R2 0.462 0.460 0.410 Adjusted R2 0.431 0.421 0.377 LOOCV R2 0.391 0.344 0.327 HEV R2 0.564 0.603 0.363 n 75 75 75 MBE 5.89e-16 1.71e-16 -4.60e-16 RMSE 27.7 27.7 28.9 MAE 21.3 20.9 21.8 Moran’s I of residuals -0.256 -0.218 -0.230 Variables (β; VIF; Partitioned R2 - Dominance) Intercept ExpRL.1000 MainRL.50 ElvRL.5000 OpArT.300 7.84e+01 1.61e-03 9.67e-02 3.02e-04 -1.27e-04  1.02 1.29 1.29 1.08  0.060 0.134 0.221 0.047 Intercept AADTMain.25 AADTSec.300 Elevation ParkT.2000 WPopDen.500 7.35e+01 2.53e-04 1.67e-05 6.02e-01 -5.95e-06 2.78e+02  1.10 1.17 1.30 1.18 1.40  0.133 0.072 0.087 0.061 0.107 Intercept ExpTrL.3000 MainTrL.5000 OpArT.300 ResT.25 9.75e+01 8.44e-11 8.64e-13 -1.97e-04 -1.35e-02  1.01 1.11 1.06 1.08  0.068 0.221 0.071 0.051      142  Table I.2 NO2 SC1 models  Road length AADT Traffic loading R2 0.400 0.398 0.376 Adjusted R2 0.367 0.355 0.341 LOOCV R2 0.324 0.288 0.293 HEV R2 0.341 0.382 0.233 n 77 77 77 MBE 9.20e-16 6.12e-16 -1.27e-15 RMSE 32.1 32.2 32.7 MAE 25.1 25.9 25.5 Moran’s I of residuals -0.223 -0.233 -0.207 Variables (β; VIF; Partitioned R2 - Dominance) Intercept ExpRL.1000 MainRL.50 ElvRL.5000 OpArT.300 8.20e+01 1.93e-03 8.48e-02 3.06e-04 -1.70e-04  1.02 1.29 1.30 1.08  0.070 0.096 0.178 0.056 Intercept AADTExp.1000 AADTMain.25 BusTD.5000 OpArT.300 ParkT.2000 1.03e+02 1.26e-05 1.80e-04 2.50e+07 -2.17e-04 -6.44e-06  1.07 1.34 1.39 1.07 1.25  0.065 0.072 0.131 0.070 0.060 Intercept ExpTrL.3000 MainTrL.5000 OpArT.300 ResT.25 1.02e+02 9.97e-11 8.19e-13 -2.39e-04 -1.60e-02  1.01 1.11 1.06 1.09  0.074 0.171 0.079 0.053         143  Table I.3 NO2 SC2 models  Road length AADT Traffic loading R2 0.446 0.479 0.423 Adjusted R2 0.422 0.441 0.390 LOOCV R2 0.385 0.373 0.321 HEV R2 0.583 0.471 0.271 n 75 75 75 MBE -2.71e-16 4.91e-16 3.18e-16 RMSE 27.8 27.0 28.4 MAE 21.3 20.9 22.2 Moran’s I of residuals -0.286 -0.223 -0.246 Variables (β; VIF; Partitioned R2 - Dominance) Intercept ExpRL.1000 MainRL.50 ElvRL.5000 7.01e+01 1.36e-03 1.13e-01 3.14e-04  1.02 1.25 1.27  0.042 0.161 0.243 Intercept AADTMain.25 AADTSec.300 Elevation LnDist_ ShippingLanes ParkT.2000 1.51e+02 2.11e-04 2.36e-05 4.39e-01 -9.01e+00  -5.65e-06  1.32 1.10 1.41 1.65  1.20  0.113 0.114 0.059 0.130  0.062 Intercept LnDist_ ShippingLanes ParkT.2000 GovT.1000 InterD.25 2.22e+02 -1.56e+01  -6.30e-06 6.65e-05 1.12e+04  1.05  1.05 1.02 1.02  0.203  0.071 0.104 0.045       144  Table I.4 NO annual models  Road length AADT Traffic loading R2 0.500 0.515 0.486 Adjusted R2 0.475 0.460 0.427 LOOCV R2 0.276 0.33 0.325 n 40 40 40 MBE 3.42e-15 4.27e-15 -2.26e-15 RMSE 62.1 61.1 63.0 MAE 51.7 50.3 49.0 Moran’s I of residuals -0.00675 -0.00652 -0.0427 Variables (β; VIF; Partitioned R2 - Dominance) Intercept ElvRL.500 BldVolD.25 IndT.25 WPopDen.100 7.07e+01 1.06e-02 3.73e+00 1.98e-01 5.29e+02  1.01 1.10 1.02 1.08  0.192 0.148 0.078 0.082 Intercept AADTMain.200 BldVolD.25 TmplD.5000 IndT.25 2.18e+01 7.43e-05 2.56e+00 2.19e+08 2.23e-01  1.14 1.16 1.25 1.03  0.139 0.106 0.175 0.094 Intercept SecTrL.5000 ComT.300 MixT.4000 IndT.25 5.32e+01 1.15e-12 -9.09e-04 1.25e-04 2.40e-01  1.23 1.32 1.41 1.06  0.150 0.022 0.214 0.100  Table I.5 NO SC1 models  Road length AADT Traffic loading R2 0.530 0.510 0.477 Adjusted R2 0.481 0.458 0.437 LOOCV R2 0.408 0.319 0.375 n 43 43 43 MBE 1.29e-15 9.85e-15 -1.22e-15 RMSE 62.5 67.2 66.0 MAE 48.0 52.8 53.6 Moran’s I of residuals -0.0451 0.153 -0.0362 Variables (β; VIF; Partitioned R2 - Dominance) Intercept ElvRL.500 BldVolD.25 TmplD.5000 IndT.25 4.57e+01 1.03e-02 3.74e+00 1.68e+08 2.26e-01  1.14 1.15 1.29 1.02  0.178 0.135 0.133 0.085 Intercept AADTMain.200 BldVolD.25 PrkArD.25 IndT.25 8.82e+01 1.06e-04 4.27e+00 -2.67e+03 2.14e-01  1.02 1.05 1.04 1.02  0.192 0.145 0.045 0.075 Intercept FoodStD.500 MixT.4000 IndT.25 6.54e+01 2.42e+07 1.40e-04 2.60e-01  1.01 1.05 1.05  0.084 0.297 0.096 145  Table I.6 NO SC2 models  Road length AADT Traffic loading R2 0.460 0.469 0.436 Adjusted R2 0.396 0.406 0.369 LOOCV R2 0.221 0.285 0.226 n 58 58 58 MBE -9.34e-16 2.24e-15 1.48e-16 RMSE 65.8 65.2 67.2 MAE 53.8 52.3 53.9 Moran’s I of residuals -0.134 -0.126 -0.150 Variables (β; VIF; Partitioned R2 - Dominance) Intercept LnDist_ ShenzhenP BldVolD.25 PrkArD.200 ComT.25 IndT.25 NO2.SC2.Len 492 -60.1  3.68 3.70e+3 -0.106 0.173 1.94  1.49  1.51 1.03 1.33 1.03 1.44  0.025  0.077 0.056 0.030 0.046 0.227 Intercept AADTExp.300 AADTMain.100 AADTSec.200 BldVolD.25 ComT.25 Lands.1500 3.47e+01 1.02e-04 1.65e-04 9.56e-05 3.80e+00 -1.01e-01 -2.94e-05   1.02 1.16 1.17 1.33 1.33 1.03  0.059 0.107 0.145 0.089 0.027 0.042 Intercept MainTrL.100 SecTrL.200 TmplD.5000 ParkT.25 MixT.50 IndT.25 2.02e+01 2.60e-07 1.70e-08 1.72e+08 -2.33e-01 9.39e-03 1.75e-01  1.38 1.12 1.25 1.23 1.06 1.04  0.094 0.075 0.123 0.050 0.047 0.046        146  Table I.7 PM2.5 annual models  Road length AADT Traffic loading R2 0.587 0.564 0.589 Adjusted R2 0.544 0.518 0.546 LOOCV R2 0.431 0.442 0.482 n 64 64 64 MBE -5.38e-17 2.78e-17 7.11e-17 RMSE 3.99 4.10 3.98 MAE 3.22 3.38 3.23 Moran’s I of residuals -0.228 -0.294 -0.290 Variables (β; VIF; Partitioned R2 - Dominance) Intercept ExpRL.25 Dist_ShenzhenP CarPD.1000 CarPD.25 GovT.100 IndT.25 3.67e+01 8.91e-02 -3.09e-04 4.17e+05 1.68e+04 -3.81e-04 1.38e-02  1.35 1.16 1.26 1.05 1.05 1.06  0.056 0.168 0.151 0.057 0.066 0.089 Intercept AADTExp.25 Lat LnDist_Coast CarPD.1000 CarPD.25 GovT.100 -6.49e+02 8.20e-05 3.01e+01 9.38e-01 3.83e+05 1.80e+04 -5.23e-04  1.36 1.24 1.15 1.27 1.07 1.07  0.093 0.138 0.046 0.140 0.062 0.085 Intercept ExpTrL.25 Lat LnDist_Coast CarPD.1000 CarPD.25 GovT.100 -6.59e+02 2.02e-06 3.06e+01 9.63e-01 4.20e+05 1.80e+04 -5.07e-04  1.39 1.19 1.15 1.33 1.07 1.07  0.102 0.142 0.048 0.152 0.062 0.084        147  Table I.8 PM2.5 SC1 models  Road length AADT Traffic loading R2 0.495 0.617 0.634 Adjusted R2 0.455 0.581 0.600 LOOCV R2 0.244 0.506 0.524 n 70 70 70 MBE 1.58e-16 7.05e-17 -1.21e-16 RMSE 4.68 4.08 3.98 MAE 3.71 3.31 3.19  Moran’s I of residuals -0.267 -0.361 -0.348 Variables (β; VIF; Partitioned R2 - Dominance) Intercept BusTD.100 CarPD.300 GovT.25 IndT.2000 IndT.25  1.74e+01 1.81e+05 1.97e+05 -5.97e-03 5.74e-06 1.54e-02  1.03 1.02 1.03 1.15 1.13  0.103 0.236 0.039 0.062 0.055 Intercept AADTExp.25 BusTD.100 CarPD.300 OpArT.100 GovT.25 IndT.2000 1.65e+01 1.25e-04 2.06e+05 2.53e+05 -2.23e-04 -5.61e-03 5.12e-06  1.32 1.04 1.19 1.07 1.05 1.10  0.092 0.113 0.284 0.027 0.038 0.064 Intercept ExpTrL.25 BusTD.100 CarPD.300 OpArT.100   GovT.25 IndT.2000  1.62e+01 2.78e-06 2.07e+05 2.62e+05 -2.12e-04 -5.37e-03 5.21e-06  1.32 1.04 1.22 1.06 1.05 1.09  0.102 0.113 0.291 0.026 0.038 0.064         148  Table I.9 PM2.5 SC2 models  Road length AADT Traffic loading R2 0.446 0.541 0.565 Adjusted R2 0.415 0.487 0.521 LOOCV R2 0.341 0.360 0.388 n 77 77 77 MBE 1.30e-16 3.60e-16 -2.48e-18 RMSE 7.63 6.95 6.77 MAE 5.93 5.44 5.58 Moran’s I of residuals -0.0925 -0.200 -0.153 Variables (β; VIF; Partitioned R2 - Dominance) Intercept SecRL.1000 Dist_ShenzhenP FoodStD.100 GovT.300 5.23e+01 2.85e-04 -5.01e-04 4.27e+05 -9.49e-05  1.06 1.08 1.00 1.03  0.132 0.191 0.060 0.063 Intercept AADTSec.50 Dist_Coast FoodStD.100 GovT.100 MixT.5000 IndT.25 Lands.1500 InterD.300 4.31e+01 3.90e-05 3.87e-03 4.39e+05 -9.34e-04 -5.80e-06 2.32e-02 -4.91e-06 2.55e+04  1.08 1.34 1.29 1.10 1.31 1.06 1.42 1.33  0.056 0.100 0.062 0.101 0.067 0.050 0.068 0.037 Intercept SecTrL.50 Lat Dist_Coast CarPD.1500 FoodStD.100 GovT.100 IndT.25  -6.80e+02 1.78e-07 3.21e+01 4.90e-03 6.08e+05 3.99e+05 -8.70e-04 1.83e-02  1.27 1.55 1.24 1.10 1.08 1.09 1.09  0.083 0.103 0.116 0.074 0.057 0.093 0.039        149  Table I.10 Black carbon annual models  Road length AADT Traffic loading R2 0.505 0.441 0.444 Adjusted R2 0.445 0.383 0.377 LOOCV R2 0.311 0.269 0.246 n 76 76 76 MBE 1.76e-17 -2.91e-16 1.20e-17 RMSE 3.70 3.94 3.93 MAE 3.02 3.08 3.34 Moran’s I of residuals -0.129 -0.0389 -0.0872 Variables (β; VIF; Partitioned R2 - Dominance) Intercept ExpRL.3000 ExpRL.50 Long CarPD.50 ComT.500 ResT.50 MixT.500 Lands.500 2.51e+03 9.48e-05 1.76e-02 -2.19e+01 3.23e+04 -2.74e-05 -8.74e-04 -2.32e-05 -1.50e-04  1.27 1.42 1.23 1.20 1.22 1.39 1.27 1.23  0.091 0.075 0.089 0.044 0.037 0.050 0.053 0.065 Intercept AADTExp.3000 AADTSec.25 Long CarPD.50 ComT.500 ResT.50 MixT.500 3.12e+03 1.10e-06 3.51e-05 -2.73e+01 3.10e+04 -3.19e-05 -9.57e-04 -1.83e-05   1.17 1.05 1.08 1.18 1.06 1.25 1.14  0.108 0.021 0.117 0.038 0.048 0.062 0.047 Intercept ExpTrL.3000 Long  Lat Elevation CarPD.50 ResT.50 MixT.500 MixT.4000 1.38e+03 2.40e-11 -1.65e+01 2.29e+01 -7.44e-02 2.88e+04 -7.39e-04 -2.61e-05 4.12e-06  1.60 1.17 1.54 1.94 1.16 1.32 1.60 2.72  0.120 0.074 0.074 0.021 0.038 0.049 0.050 0.018        150  Table I.11 Black carbon SC1 models  Road length AADT Traffic loading R2 0.444 0.271 0.332 Adjusted R2 0.377 0.219 0.263 LOOCV R2 0.246 0.118 0.148 n 76 76 76 MBE 1.66e-16 1.67e-16 -1.37e-16 RMSE 5.64 6.46 6.18 MAE 4.28 4.89 4.76 Moran’s I of residuals -0.192 -0.0205 -0.0392 Variables (β; VIF; Partitioned R2 - Dominance) Intercept ExpRL.3000 ExpRL.50 MainRL.25 CarPD.50 PrkArD.50 ComT.3000 ResT.1000 Lands.500  1.13e-04 4.77e-02 5.67e-02 3.69e+04 3.30e+02 -4.98e-06 -4.68e-06 -2.57e-04  1.16 1.59 1.21 1.29 1.27 1.67 1.64 1.43  0.067 0.084 0.052 0.047 0.060 0.025 0.034 0.074 Intercept CarPD.50 PrkArD.50 ParkT.1500 ComT.3000 ResT.50 2.05e+01 4.20e+04 3.61e+02 -4.16e-06 -9.78e-06 -1.18-03  1.37 1.30 1.85 1.82 1.20  0.046 0.062 0.061 0.055 0.047  Intercept CarPD.50 PrkArD.50 ParkT.1500 ComT.500 GovT.3000 ResT.50 IndT.1500  1.89e+01 3.13e+04 3.76e+02 -3.07e-06 -6.08e-05 -1.61e-06 -1.11e-03 1.26e-05  1.34 1.33 1.31 1.16 1.14 1.20 1.15  0.037 0.061 0.052 0.060 0.025 0.047 0.049        151  Table I.12 Black carbon SC2 models  Road length AADT Traffic loading R2 0.416 0.522 0.492 Adjusted R2 0.378 0.469 0.444 LOOCV R2 0.337 0.389 0.371 n 82 82 82 MBE -5.49e-16 2.40-16 1.10e-16 RMSE 3.99 3.62 3.72 MAE 3.12 2.96 2.95 Moran’s I of residuals -0.0194 -0.0252 -0.0243 Variables (β; VIF; Partitioned R2 - Dominance) Intercept ExpRL.3000 Long Dist_MTRstn ResT.4000 MixT.50 3.63e+03 7.98e-05 -3.17e+01 1.15e-03 7.05e-07 -5.02e-04  1.26 1.09 1.22 1.13 1.20  0.059 0.171 0.034 0.093 0.058 Intercept AADTExp.2000 AADTSec.25 Long Dist_Coast Dist_MTRstn LnDist_MTRline GovT.200 GovT.1500 3.38e+03 7.36e-07 3.48e-05 -2.96e+01 2.69e-03 2.16e-03 -8.55e-01 -8.55e-05 4.08e-06  1.18 1.11 1.26 1.18 2.19 1.83 1.50 1.33  0.043 0.035 0.153 0.158 0.055 0.020 0.028 0.030 Intercept Long Dist_Coast Dist_MTRstn LnDist_ MTRline GovT.200 GovT.1500 IndT.2000 2.63e+03 -2.29e+01 2.89e-03 2.27e-03 -9.73e-01  -1.15e-04 4.45e-06 4.99e-06  1.42 1.15 2.12 1.94  1.39 1.32 1.25  0.126 0.162 0.057 0.024  0.037 0.034 0.051    

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0314324/manifest

Comment

Related Items