Microscale Air Temperature Mapping in Greater Vancouver, British Columbia

MICROSCALE AIR TEMPERATURE MAPPING IN GREATER VANCOUVER, BRITISH COLUMBIA

by Pak Keung Tsin

B.Sc., 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)

August 2018

Microscale air temperature mapping in greater Vancouver, British Columbia

submitted in partial fulfilment of the requirements by Pak Keung Tsin for the degree of Master of Science in Occupational and Environmental Hygiene

Examining Committee:

Sarah Henderson Co-supervisor

Michael Brauer Co-supervisor

Matilda van den Bosch Additional Examiner

Additional Supervisory Committee Members: Anders Knudby Supervisory Committee Member

Scott Krayenhoff Supervisory Committee Member

Abstract

Background: Mobile air temperature monitoring is a promising method to better understand temperature distributions at fine spatial resolutions across urban areas and to minimize extreme hot weather health impacts. The first study objective was to collect pedestrian microscale air temperature data to evaluate different methods for assessing spatial variation in urban heat exposure in greater Vancouver, Canada. The second objective was to develop microscale land use regression (LUR) air temperature models using the data collected.

Methods: Mobile air temperature monitoring was conducted on foot at least twice for 20 routes chosen to represent potential heat exposures. The mobile data were compared with 1-minute measurements from the nearest fixed site, with satellite-derived land surface temperature (LST) for runs corresponding with Landsat overpass days, and with estimates from a previously- developed heat map for the region based on satellite generated geographic data. Six independent variables were considered for use in constructing a 30 x 30m LUR model for each run and within all routes in greater Vancouver. All models were evaluated using a spatial leave-ten-out cross- validation (LTOCV) approach.

Results: Mobile measurements were typically higher and more variable than simultaneous fixed site measurements. The relationship between mobile measurements and LST were weak and highly variable. The mobile measurement and heat map z-score differentials suggested that spatial temperature variability was well-captured by the previously-developed heat map. The

Distance to Large Water Body, Distance to Major Road, Normalized Difference Water Index, and Sky-View Factor were selected as the most predictive independent variables. On average, the iii

best individual route models explained 38.6% of the variation in microscale air temperatures at

20 routes. The overall model explained only 10.0% of the variation in the route areas of the greater Vancouver region.

Conclusion: The microscale measurements confirmed that fixed sites did not characterize the thermal variability within nearby streetscapes. They could also be used to generate LUR models for some locations. The strength of daytime mesoscale atmospheric processes may weaken the predictive power of land use variables. Future studies intending to use microscale modelling should collect data within a restricted time range and across fewer routes.

Lay Summary

Mobile air temperature data were collected by walking 20 routes in greater Vancouver using a specialized thermometer, a global positioning system (GPS), and other equipment. These routes were chosen to reflect heat vulnerability, then compared with data from weather stations, satellite images, and a heat map. The data were also used to develop microscale air temperature models for each route and for all of greater Vancouver. Mobile air temperature measurements were higher than at fixed weather stations, had weak relationships with satellite measurements, and suggested that previously-developed heat map was generally representative of spatial temperature variability. On average, the individual route models explained 39% of the air temperature differences along each route. The overall greater Vancouver model only explained

10% of the air temperature differences in the greater Vancouver route areas. The strength of large scale daytime meteorological processes may have a larger influence than potentially predictive land use variables.

Preface

The research idea for this study was originally identified by Dr. Sarah Henderson and Dr.

Anders Knudby. Design of the research program was conducted collaboratively by Dr. Sarah

Henderson and myself, with input from Dr. Michael Brauer, Dr. Anders Knudby, and Dr. Scott

Krayenhoff. I conducted all field measurements of air temperature, with occasional assistance and accompaniment from volunteers Annie Wang and William Leung. Analysis of the research data was conducted by the author with guidance from all committee members.

Chapter 2 is based on research published in Tsin, P. K. et al. Microscale mobile monitoring of urban air temperature. Urban Clim. 18, 58–72 (2016). Chapter 3 is currently in the process of publication in a to-be selected journal. For both chapters, I conducted all field measurements and wrote the manuscript with guidance from Dr. Sarah Henderson and input from Drs Michael Brauer, Anders Knudby, Scott Krayenhoff, and Derrick Ho.

Table of Contents

Abstract ...... iii

Lay Summary ...... v

Preface ...... vi

Table of Contents ...... vii

List of Tables ...... x

List of Figures ...... xi

List of Abbreviations ...... xiii

Acknowledgements ...... xiv

Chapter 1: Introduction ...... 1

1.1 High Air Temperatures and Extreme Heat Events ...... 1

1.1.1 Extreme Heat Events...... 1

1.1.2 Extreme Heat Event Susceptibility in Temperate Regions ...... 2

1.1.3 Responses to Extreme Heat Events...... 3

1.1.4 Future Trends in Air Temperature and Extreme Heat Events ...... 4

1.2 Background Climatology ...... 5

1.3 Overview of Variables Affecting Temperature Determinants ...... 6

1.4 In-Situ Air Temperature Measurement Studies ...... 7

1.4.1 Studies Examining the Urban Heat Island ...... 8

1.4.2 Studies Examining Other Factors Related to Urban Temperatures ...... 10

1.5 Remote Sensing Studies of Air Temperature ...... 12

1.6 Land Use Regression Models of Air Temperature ...... 13

1.7 Evidence Gaps for Greater Vancouver ...... 15 vii

Chapter 2: Microscale Mobile Monitoring of Urban Air Temperature ...... 17

2.1 Introduction ...... 17

2.2 Methods...... 17

2.2.1 Study Area ...... 17

2.2.2 Route Selection ...... 18

2.2.3 Mobile Air Temperature Data Collection ...... 19

2.2.4 Temperature Data for Comparison with Microscale Measurements ...... 21

2.2.5 Data Analysis ...... 23

2.3 Results ...... 25

2.3.1 Mobile Data Compared with Fixed Site Data ...... 25

2.3.2 Mobile Data Compared with Land Surface Temperature Data ...... 27

2.3.3 Mobile Data Compared with the Greater Vancouver Heat Map (GVHM) ...... 28

2.4 Discussion ...... 31

Chapter 3: Land Use Regression Modelling of Microscale Urban Air Temperatures...... 35

3.1 Introduction ...... 35

3.2 Methods...... 35

3.2.1 Air Temperature Data ...... 35

3.2.2 Independent Variables ...... 36

3.2.3 Land Use Regression Modelling ...... 38

3.2.4 Model Evaluation ...... 39

3.3 Results ...... 40

3.3.1 Independent Variable Selection ...... 40

3.3.2 Individual Route Models...... 40 viii

3.3.3 Overall Route Model...... 44

3.4 Discussion ...... 45

Chapter 4: Conclusion ...... 49

4.1 Summary of Key Findings ...... 49

4.2 Contribution to Body of Knowledge...... 50

4.3 Strengths ...... 51

4.4 Limitations ...... 51

4.5 Future Studies ...... 53

4.6 Conclusions ...... 54

Bibliography ...... 56

Appendices ...... 67

Appendix A - Supplemental Tables ...... 67

Appendix B - Supplemental Figures ...... 70

Appendix C - Wind Speed and Direction Analysis, YVR Station ...... 77

List of Tables

Table 2-1: Station air temperature sensor heights and Local Climate Zone classifications ...... 22

Table 2-2: Correlation between mobile run air temperature data and land surface temperature

(LST) ...... 27

Table 3-1: Variables that were evaluated as potential predictors of microscale air temperature . 42

Table 3-2: Overall air temperature model variable coefficients and percentage contribution to r2 value...... 45

Table A-1: Summary statistics for each of the 42 mobile air temperature monitoring runs and their corresponding nearest fixed site ...... 67

Table A-2: Z-score differential (GVHM z-score minus mobile air temperature data z-score) distribution for raster cell values from overall sampling campaign and all runs ...... 68

List of Figures

Figure 2-1: Map of all 20 greater Vancouver mobile air temperature sampling routes and all

Metro Vancouver fixed stations used in this study ...... 19

Figure 2-2: Mobile monitoring setup. Left image shows the white radiation shield containing the

Met One 064-2 sensor and the yellow Kestrel 4500 Portable Weather Station...... 21

Figure 2-3: Mobile and fixed air temperature mesasures for the 42 runs ...... 26

Figure 2-4: Scatter plots of mobile air temperature versus land surface temperature (LST) as well as the LST histograms ...... 28

Figure 2-5: Differences in the z-scores between the Greater Vancouver Heat Map (GVHM) and mobile air temperatures for all 42 runs ...... 30

Figure 3-1: Map showing r2 and relative importance (variable percentage contribution to model r2) for the best individual route models ...... 44

Figure 3-2: Plot showing raw air temperature metadata of individual run models used to create overall air temperature...... 45

Figure A-1: Scatterplots and correlation coefficients (r) of z-score differentials (GVHM z-score minus mobile air temperature data z-score) between all replicate runs of each route ...... 70

Figure A-2: Map of the 50m buffer Normalized Difference Water Index (NDWI) data used to build the individual route models and the overall model...... 71

Figure A-3: Map of the 50m buffer Normalized Difference Vegetation Index (NDVI) data considered for inclusion into the individual route models and the overall model...... 72

Figure A-4: Map of the 200m buffer Sky-View Factor (SVF) data used to build the individual route models and the overall model...... 73

Figure A-5: Map of the unbuffered Distance to Major Roads data used to build the individual route models and the overall model ...... 74

Figure A-6: Map of the 200m buffer Distance to Large Water Body data used to build the individual route models and the overall model...... 75

Figure A-7: Map of the 200m buffer Elevation data considered for inclusion into the individual route models and the overall model...... 76

xii

List of Abbreviations

ANOVA – Analysis of Variance

CIMS5 – Coupled Model Intercomparison Project, Phase 5

DEM – Digital Elevation Model

GVHM – Greater Vancouver Heat Map

HCTI- Human Comfort Thermal Index

LCZ – Land Classification Zone

LST – Land Surface Temperature

LTOCV – Leave-Ten-Out-Cross-Validation

LUR – Land Use Regression

NDVI – Normalized Difference Vegetation Index

NDWI – Normalized Difference Water Index

RCP4.5 – Representative Concentration Pathway, corresponding to an increased radiative forcing of 4.5W/m2 worldwide

RCP8.5 – Representative Concentration Pathway, corresponding to an increased radiative forcing of 8.5W/m2 worldwide

SVF – Sky-View Factor

UHI – Urban Heat Island

YVR – Vancouver International Airport

xiii

Acknowledgements

There are many people who have helped made this study possible. Firstly, I would first like to thank Dr. Sarah Henderson for her expertise, guidance, patience, support, time, and for giving me just the right amount of freedom to conduct the study while knowing when I needed a push in the right direction. I would also like to thank Drs. Michael Brauer, Anders Knudby, and

Scott Krayenhoff for providing their feedback, expertise, and time. As well, I wish to thank the

UBC OEH program, staff, and students for their support during this occasionally challenging but worthwhile and rewarding experience.

I would also like to thank the Pacific Institute for Climate Solutions and the University of

British Columbia for their support and funding of this research. I would also like to acknowledge

Metro Vancouver for providing us with 1-minute frequency fixed weather station data, as well as

Derrick Ho and Anders Knudby for providing data from their heat map studies.

Lastly, I would like to thank my friends and family. I would like to thank my friends for their support, and for always checking on the progress of my thesis and keeping me motivated.

Special thanks to my parents for their care, love, and support, both morally and financially.

xiv

Chapter 1: Introduction

1.1 High Air Temperatures and Extreme Heat Events

High air temperatures increase the risk of extreme heat exposure and heat stress, which can result in a wide range of heat-related illnesses, including heat cramps, heat exhaustion, heat syncope, heat stroke, and even death (1,2). The human body dissipates generated metabolic and environmental heat through conduction, convection through blood circulation to the skin, sweat evaporation, and radiation (1,2). The human body overheats when it cannot dissipate heat quickly enough to keep the core temperature in optimal range, such as during extreme heat events. Overheating can cause direct cell damage and increased inflammation, which can lead to heat-related illnesses (1,2). The elderly, young children, those with pre-existing medical conditions, those with mental illness, and those living alone are at greater risk from extreme heat events (1–4).

1.1.1 Extreme Heat Events

Extreme heat events are defined as consecutive days that have maximum and minimum air temperatures above the normal climatic range and/or exceeding a certain location-specific threshold (5,6). There are many examples of extreme heat events being associated with large increases in population mortality, including at least 30,000 excess deaths during the 2003 heatwave in Europe (a 1-35% increase in mortality depending on the country) (7–9). In the US, extreme heat is associated with more weather-related deaths annually than deaths from earthquakes, floods, hurricanes, and tornados combined (1,10). During the summer of 2009 there was an extreme heat event in greater Vancouver, British Columbia, associated with an estimated

110 excess deaths, or a 40% increase in weekly mortality compared with mortality during previous summer weeks (3,11).

1.1.2 Extreme Heat Event Susceptibility in Temperate Regions

Within North America, temperate regions, such as greater Vancouver, likely have higher heat public health risks associated with extreme heat events when compared with non-temperate regions (2,5,12,13). For example, relatively cooler temperate regions in the United States have higher mortality during summer weather than the warmer, non-temperate southern regions

(5,12,13).

One reason that extreme heat has greater impacts in temperate areas is their lack of acclimatization to hot weather (2,12). Acclimatization is the process of physiological and biochemical adjustment by the body to heat, which includes improved circulation and salt and water retention (1,2). Acclimatization requires several weeks to occur, and allows individuals who experienced prolonged exposures to tolerate conditions that may result in morbidity or mortality for non-acclimatized individuals, although there are physiological limits to how much acclimatization humans can experience (1,2). Acclimatization to consistently warmer summer temperatures in the non-temperate southern US regions was suggested as one reason for lower summer weather mortality in these relative to temperate US regions, which comparatively have lower mean maximum temperatures, fewer warmer temperature days, and therefore less opportunities for acclimatization (12).

Public health risks may also be higher in temperate regions because they tend to lack the protective measures and adaptations found in hotter areas (1,15). For example, air conditioning is less common in temperate regions (1). In addition, many municipalities lack response plans, such

as warning systems, cooling centers, and direct care facilities for extreme hot weather events, which may increase the risk of morbidity and mortality (15).

1.1.3 Responses to Extreme Heat Events

Many cities have taken measures to reduce the potential public health impacts from extreme hot weather events (1,11,16). Current measures in greater Vancouver include a heat health warning system, which is used to notify local health authorities and emergency responders of potential danger within the coming days (11). This system was developed based on historical relationships between temperature and mortality in the region, and it is triggered for the following day when the average of the observed 14:00 air temperature and the Environment

Canada forecasted high exceed 29oC on the coast and/or 34oC inland (11). If triggered, the system launches a comprehensive emergency response from the two local health authorities and local municipalities. This includes such activities as active outreach to ensure the safety of the homeless, as well as temporary cooling stations in the form of air conditioned libraries and community centers (17,18).

Understanding how air temperature is related to geographic and human factors can help focus heat mitigation/adaptation resources where they are needed, to better inform the spatial targeting of emergency interventions. Long-term heat mitigation/adaptation plans can also benefit from this understanding, which can inform mitigation/adaptation strategies in urban design. There are multiple examples of how urban design can be used to lower air temperature and increase thermal comfort: awnings for shading; lighter building colours to increase albedo; vegetation that provides shade and facilitates evapotranspiration; and streets oriented with prevailing summer winds for increased airflow (19,20).

1.1.4 Future Trends in Air Temperature and Extreme Heat Events

Focusing resources to reduce, mitigate, and adapt to heat exposure is increasingly important in the context of global climate change. Mean global land surface air temperature has increased by approximately 0.7oC between 1951 and 2012, and is projected to increase on average by 0.3-0.7 oC worldwide between 2016-2035 relative to the period of 1986-2005 (21).

Hot days, hot nights, and extreme hot weather events have become more frequent since 1950 and are also projected to increase in frequency (1,6,21). In addition, the duration and severity of these events will also increase in the future, as will their related mortality (6,21,22).

The RCP4.5 scenario (Representative Concentration Pathway, which corresponds to an increased radiative forcing of 4.5W/m2 worldwide) using 42 CMIP5 (Coupled Model

Intercomparison Project, Phase 5) models predicts that the global and Western North American median temperatures will increase by 0.7oC and 1.0oC, respectively, during the 2016-2035 period relative to the 1986-2005 period (21). Another study using the RCP4.5 scenario that modelled air temperature changes in Washington state predicted that greater Vancouver would experience

0.67-2.67 extra heat waves annually (defined as 3-day events where the heat index exceeds 32oC) between 2030-2059 relative to 1970-1999 (23). The same study predicted that greater Vancouver would experience 11-15 more warm nights annually from 2030-2059, exceeding the baseline 90th percentile minimum daily temperature calculated using all days from 1970-1999 (23). Another study which modelled the RCP8.5 scenario (Representative Concentration Pathway, which corresponds to an increased radiative forcing of 8.5W/m2 worldwide) throughout Canada found that relative to the period of 1986-2005, the number of days in BC exceeding 30oC was projected to increase by 2.5 days by 2031-2050 and by 16 days by 2081-2100 (24).

1.2 Background Climatology

Air temperature, technically known as “near surface air temperature”, is measured at 1.5 m to 2 m above the ground (25–28). For urban areas, air temperature is defined as the temperature within the urban canopy layer, which is the height below the highest local surface features, which can be trees or buildings (28). Within non-urban areas, the recommended screen height of a thermometer to measure air temperature is 1.25-2 m (28). A thermometer height of 1 m has shown very little difference compared with higher heights within the urban canopy layer, although higher heights reduce the influence of vehicle exhaust and dust contamination (28,29).

Many variables, such as land cover, elevation, distance from bodies of water, and surface temperature can influence air temperature (30–40). Many studies of spatial heat variability and heat health effects use air temperature to quantify heat, and air temperature is an important environmental variable due to its common use in multiple scientific disciplines including climatology, disease vector mapping, meteorology, terrestrial hydrology, and public health (25–

28,30,32,34,41–43).

Climatic phenomena, including air temperature, are usually classified using scales based on horizontal extent, including: microscale (less than 100m); toposcale or local scale (100m-

3000m); mesoscale (3,000m-100,000m); large scale or macroscale (100,000m-3,000,000m); and planetary scale or synoptic scale (larger than 3,000,000m). There are slight differences in these extent definitions depending on the source (28,44,45). Buildings, trees, and roads have microscale extents, neighbourhoods with similar levels of development and geography have local scale extents, while entire cities have mesoscale extents (28).

Exposures to extreme temperatures and their population health impacts are commonly evaluated at municipal or regional scales using time series data from a single fixed monitoring 5

station (3,46–48). However, previous work in greater Vancouver involving microscale vulnerability and exposure map data has shown that risk of mortality is greater in areas with higher exposure (46). Given that thermal exposure is not uniform across the urban environment, there is value in modelling areas with relatively higher and lower temperatures for epidemiologic studies, risk communication, and heat mitigation/adaptation. This is especially important given that socially vulnerable groups, such as those who have lower socioeconomic status, may be more likely to live in areas with higher heat exposure (3,5,7,46). Indeed, previous work in greater

Vancouver found that both increased exposure and decreased participation in the labour force were associated with increased risk of mortality on extremely hot days (46,49). While it is relatively simple to assess the intra-urban variability in social vulnerabilities using routine census data, assessing the intra-urban variability in temperature exposures is more challenging. Human thermal exposure relates to multiple factors, including time spent indoors (>90% on average)

(50), ability to regulate the indoor climate, the radiative, humidity, and wind environments during time spent outdoors, the clothing worn, and physical activity levels. Even so, outdoor air temperature has been most consistently linked to mortality at city-wide, regional, and continental scales, but its microscale variation has not been assessed in this capacity.

1.3 Overview of Variables Affecting Temperature Determinants

There are many different variables that affect the air temperature in any given place at any given time. Generally, these factors can be considered as part of the built environment or the natural environment. In the built environment, the albedo of construction materials indicates the amount of solar radiation they reflect. Materials with higher albedo absorb less sunlight, have lower surface temperature, and emit less radiative heat that materials with lower albedo (19). In

addition to having low albedo, buildings and roads indicate urban development, human activity, and vehicular traffic, and they experience relatively high storage heat flux, whereby heat is stored throughout the day and then released at night (51,52). Sky-view factor (SVF) indicates how much of the sky above is visible and unobstructed by overhead objects. Lower SVF generally correlates with higher nocturnal air temperature due to decreased cooling rates, but previous studies have reported mixed results for the relationship between SVF and daytime temperatures (53–60).

Although the distribution of vegetation is it often part of the built environment, it is the natural properties of vegetation that affect air temperature. Most importantly, vegetation reduces air temperature through both increased evapotranspiration and increased shading (58,61). Large bodies of water also have a cooling effect on nearby areas and, in the case of coastal cities such as greater Vancouver, they are related to the land-sea breeze process, which causes cool air to be blown inland during the day (43,62). Higher elevation has been correlated with lower air temperature in the atmosphere and in previous models (57,58,63). In addition, all of these variables are associated with land surface temperature (LST), which has shown a positive linear correlation with air temperature in multiple studies (64–67).

1.4 In-Situ Air Temperature Measurement Studies

Most in-situ air temperature measurement studies conducted to date have been for the purpose of assessing the urban heat island (UHI) effect on air temperatures, both at night and during the day. Other in-situ studies have focused on how human and geographic factors co-vary with air temperature.

1.4.1 Studies Examining the Urban Heat Island

The term UHI refers to the phenomenon that urban (or more developed and populated) areas are warmer than corresponding rural (or less developed and populated) areas nearby

(26,38,39,43,68,69). There are many proposed reasons for the differences in temperature between urban and rural areas, most of which focus on the modified geography in urban areas.

One explanation of the UHI effect is increased solar radiation absorption due to building walls and vertical surfaces increasing reflection and radiation trapping, as well as the lower albedo of urban building materials (20,39,69). Infrared light retention in urban areas is also higher due to their restricted SVF, which quantifies the percent visibility of the sky when looking upward at any given location (39,69). Furthermore, urban heat uptake is higher and heat release is delayed in urban areas because of the higher heat capacity of concrete building materials, the trapping of solar and infrared radiation, and the reduced convective heat loss from reduced urban airflow, with these effects having the most influence at nighttime (39,69). The introduction of impermeable paved surfaces to replace soil and plants is another consequence of urbanization, which reduces the potential for evapotranspiration and shading (39,69). As a result, a large portion of heat in urban areas is in the form of sensible heat and not latent heat, such as water vapor (39,69). One last explanation is the contribution of sensible and latent heat from anthropogenic fluxes, such as from automobiles, heating/cooling of buildings, and industry in urban areas (39,69). Generally, as the urban area becomes larger, more developed, increases in building density, or increases in population density, the larger the UHI effect (27,38,39,70).

Several mobile monitoring studies using vehicle traverses have been completed along routes that pass between urban and rural areas to observe the difference in air temperature attributed to the UHI effect (26,27,35,71). Many studies conducted vehicle temperature traverses 8

at night, because the UHI effect is most prominent 3-5 hours after sunset. This is due to the higher cooling rates in rural areas relative to urban areas and the stronger lower atmosphere stability during these times. However, different studies have used different times of day for measurement, different route lengths, and differing numbers of replicate runs. Within greater

Vancouver, urban and rural areas had air temperature differences of 5-12oC after sunset, depending on the exact time of day and weather conditions (26).

Until recently, the majority of UHI studies classified different parts of study areas and routes as urban or rural. This leads to inadequate characterizations, because the terms urban and rural are too general, and their definitions change depending on the study areas (38,39). In addition, the spatial variation within urban areas and the temporal changes across study areas were not well captured by the standard urban and rural classifications (38,39). Recent UHI studies have incorporated the Land Climate Zones (LCZ) classification scheme, which defines different zones as “regions of uniform surface cover, structure, material, and human activity that span hundreds of meters to several kilometers in horizontal scale” based on ten climatic factors

(27,35,39,68).

Accounting for these climatic factors helps to describe spatial and temporal urban variability in geographic features, including: SVF; building or tree aspect ratio; mean building/tree height; terrain roughness; building surface fraction; impervious surface fraction; pervious surface faction; surface admittance; albedo; and anthropogenic heat flux (27,35,39).

Several studies have shown that the LCZ classification scheme can be used to quantify interurban UHI effect differences at city-wide scales (local or mesoscale), particularly at night.

Air temperature differences as large as 2oC were present between different urban areas at night, while daytime differences were smaller when present (27,35,39,68). 9

1.4.2 Studies Examining Other Factors Related to Urban Temperatures

In-situ temperature measurements have also been conducted in other studies for different purposes, such as identifying and characterizing geographic or socio-economic factors that are correlated with warmer temperatures. One study in Phoenix, Arizona, examined the influence of socioeconomic factors on heat exposure by placing a data logging thermometer/humidity meter in the backyards of residences in eight neighbourhoods with differing socioeconomic status (5).

Data were logged from June 1 through August 31, 2003 at 5-minute intervals, and air temperature at 5pm from all days as well as a Human Comfort Thermal Index (HCTI) were used for the analysis. Researchers found that having a lower socioeconomic status and/or belonging to an ethnic minority group were correlated with higher probability of living in warmer neighbourhoods and a higher probability of heat stress exposure. Compounding this, residents in the warmer neighbourhoods had fewer material and social resources to cope with warmer temperatures due to the lower socioeconomic status. These warmer neighbourhoods tended to have higher population density, less open space, and less vegetation than cooler areas, all factors correlated with a higher UHI effect (5).

The effect of less vegetation is not surprising, given that other studies have shown that vegetation can significantly mitigate heat, and is even linked to decreased heat-related ambulance calls during extreme heat events. A study in downtown Toronto, Canada examined the effect of shading on temperature by installing thirteen pairs of thermometers at 5m heights directly on building surfaces, with one thermometer shaded by trees or vines and one thermometer unshaded (20). The average built surface temperature difference between shaded and non-shaded thermometers during high solar intensity was 11.7oC, with shaded built surfaces being cooler than non-shaded surfaces by 1oC for as long as 10-12 hours within the day. As well, 10

shading was the most effective method for reducing temperatures on the west sides of buildings, and perennial vines were found to be as effective in reducing built surface temperatures as trees

(20). As well, a study in downtown Toronto, Canada also found that heat-related ambulance calls in census tracts with less than 5% canopy cover had five and fifteen times more heat-related ambulance calls than census tracts with over 5% canopy cover and over 70% canopy cover respectively (72).

Another study analyzed tree effects on microclimate by measuring air temperature, relative humidity, and radiation at a height of 1.5m under tree shade and under non-shaded areas in the urban parks of Thessaloniki, Greece (33). Air temperature and relative humidity were reduced by as much 24% and 41%, respectively, under tree shade when compared with unshaded areas. The tree species had a significant effect on the percentage reduction of both variables, and the reduction in relative humidity had an exponential relationship with the amount of solar radiation that passed through the tree foliage. The reduction in air temperature had an exponential relationship with both relative humidity and the amount of solar radiation that passed through the tree foliage (33).

Besides vegetation, many other geographic factors influence air temperature. A study in

Göteborg, Sweden collected temperature data from thirty radiation-shielded thermometers placed at 2m heights throughout the urban district in different representative land-use areas (32). The researchers conducted a step-wise multiple regression on 1-hour air temperature data during a 1- month period using geographic predictors, and found that “distance from the sea, sky-view factor, and surface cover (percent built up or impervious)” explained most of the spatial air temperature variability during clear and calm conditions. The best model, which described the

2pm measurements during clear and calm conditions, had a coefficient of determination (r2) 11

value of 0.70. An analysis of variance (ANOVA) test was also conducted on air temperature and land use classes, which included “urban dense, multi-family, single houses, other built-up, and green”. These analyses found that significant temperature variations were associated with land use classes both during the day (represented by noon) and during the night (represented by three hours after sunset), with night variations being stronger due to the UHI effect (32).

1.5 Remote Sensing Studies of Air Temperature

Remote sensing uses satellites and other tools to image and measure objects without making physical contact with the objects or being physically present in their area (73). Remote sensing air temperature is challenging, and values are usually estimated from land surface temperature (LST) data as well as other remote sensing measurements, such as humidity indices, elevation, and solar insolation (30,34,36,40). Surface temperatures and air temperatures do not have constant relationships, but surface temperatures do influence air temperatures. Remote sensing data may be better for estimating and capturing spatial heterogeneity in air temperatures than conventional meteorological measurements because a single image from a single instrument can cover a very large area (30,74). However, remote sensing has some drawbacks, such as cloud interference, which directly shields sensors from the measured area (30,40). In addition, the spatial resolution of the data may be low, and the instruments may have long return periods for repeat overpasses. For example, the Landsat TM/Landsat ETM+ overpasses every 16 days with an approximate resolution of 120m x120m (TM) or 60m x 60m (ETM) for thermal mapping, whereas the MODIS instruments overpass every day at 1030 and 1330, with an approximate resolution of 1000m x 1000m for thermal mapping (75–77).

One study used a Landsat 5 TM satellite image from September 3, 2008 at 10am to analyze LST in the City of Toronto, Canada (37). The researchers conducted an ANOVA on 12

surface temperature and land use categories, which included “commercial, government and institutional, open area, parks and recreation, residential, resource and industrial, and water body” and found significant relationships between all categories except “commercial relative to resource/industrial land uses, and residential land use in relation to government/institutional areas.” The study also reported that land use polygon area was correlated with LST for commercial (r = 0.405), resource/industrial (r = 0.259), water bodies (r = −0.323) and parks and recreational (r = −0.264). In comparison, the residential (r = 0.067), government/institutional (r

=-0.083), and open area (r = -0.098) correlations were weak (37). These findings agree with observations in other studies and the LCZ classification scheme that more “built up” urban areas tend to have higher temperatures and a stronger UHI effect (27,38,39,70).

1.6 Land Use Regression Models of Air Temperature

Another common use of remote sensing data is to model spatial variability in environmental conditions, such as air temperature or air pollution. Multiple studies have used different methods to model spatial variability in temperature exposures including geostatistical kriging and land use regression (LUR) (57,58,78). In general, LUR can be used to predict values of a continuous dependent variable at any given location using statistical regression models of its relationship with independent variables that describe the surrounding environment (58,79,80). In air pollution applications, LUR is most often used to model the spatial distribution of pollutants such as nitrogen dioxide oxide using information about the surrounding road network, land use, and population density (79–84). Most LUR models are trained with data measured at a local scale density (with 100-10,000 m between measurements) and used to make predictions at the microscale (<100m). However, not many models have been built or evaluated with data measured at a microscale density (80–88). 13

Similar methods have been used to model and map urban air temperatures, typically using remotely measured LST, normalized difference vegetation index (NDVI), and other geographic features such as distance from water bodies and SVF, as independent variables

(30,32,37,57,58,64,65). As with LUR models of air pollution, most models of urban temperatures have used measurements from a limited number of fixed sites spread across a large geographic area to make estimates at the microscale. According to Ho et al (2014), most remote sensing studies that estimated air temperature have done so for areas of larger than 100,000 km2

(30,40,67,74,89). These studies used satellite data with a spatial resolution of 1000m x 1000m or larger, which does not adequately represent smaller scale air temperature heterogeneity, such as the heterogeneity present within a city.

There have been recent studies that use microscale remote sensing data to quantify the heterogeneity of temperature at a city-wide scale (local scale or mesoscale). One such study used

ASTER satellite image data at a 90m x 90m resolution and emissivity data at a 10m x 10m resolution to create two daytime and two nighttime urban heat stress maps in Hong Kong, which were verified with in-situ air temperature measurements (36). The study also aimed to quantify

UHIs in Hong Kong and used two models between LST and air temperature to estimate daytime and nighttime values. Despite the fact that remote sensing images are instantaneous snapshots, based on correlations with air temperature station data, the daytime and nighttime heat stress maps could significantly represent air temperature for 7.5 daytime hours at 12 stations and 13 nighttime hours at 11 stations on the day the image was taken. In addition, the daytime and nighttime heat stress maps could significantly represent 23 and 64 days, respectively, of the seasons in which the images were taken based on correlations with air temperature station data.

The heat maps were also validated with time-matched air temperature data from the Hong Kong 14

fixed automatic weather station network, and the daytime and nighttime heat maps explained

75% and 84% of the air temperature variability from the weather station network (36).

Recent studies in greater Vancouver have generated high resolution air temperature maps based on land use regression techniques applied to fixed site monitoring data (57,58). These 60m x 60m resolution maps were developed to facilitate epidemiologic research and risk communication. They estimated temperatures using a random forest regression model calibrated with six remote sensing images taken on cloudless summer days that exceeded 25oC at the reference weather station. The most recent greater Vancouver heat map (GVHM) included independent variables such as LST, normalized difference water index (NDWI), solar radiation,

SVF, elevation, distance from ocean, and atmospheric water vapor. Both maximum apparent temperature and maximum air temperature were estimated relative to a reference station (57,58).

1.7 Evidence Gaps for Greater Vancouver

Although the GVHM provides useful air temperature information for greater Vancouver, there was no microscale evaluation of the spatial validity of the GVHM, which was constructed using both fixed site and LST data (57). As well, there is currently little published literature on the relationship between highly resolved remote sensing measurements of LST and air temperature measurements taken at similar resolution aside from studies using fixed site air temperature data. Finally, prior to this study, there was no evidence about how well the greater

Vancouver fixed weather monitoring network captures spatial variability in air temperatures across the city. The first set of study objectives aim to address these gaps through the collection and analysis of microscale air temperature data, and to demonstrate the general utility of microscale measurements for evaluating such questions.

First, twenty urban routes with varying land use, land cover, and other microenvironmental and human features were identified to monitor and collect air temperature data at an ambulatory pace on hot summer days. The collected mobile air temperature data was then compared with measurements from (1) the fixed air temperature monitoring network, (2) available Landsat measurements of LST, and (3) the GVHM.

The subsequent availability of microscale air temperature data raises the question of whether they can be used to develop LUR models of air temperature for each route. Further, it would be useful if such models could be applied beyond the route to reliably predict microscale variability in unmeasured areas. Our second set of study objectives aims to address the above questions through building and evaluating microscale air temperature models.

First, 42 microscale LUR models were developed such that there was one model for each run replicate. The best model (based on model r2) for each of the 20 routes was then selected and used to evaluate between-route similarity, as well as evaluate differences between areas where air temperature is and is not well predicted. Finally, data from the 20 best models was used to develop a model for all the routes within the greater Vancouver area to determine if the microscale data could be used to model air temperatures for the region as a whole.

Chapter 2: Microscale Mobile Monitoring of Urban Air Temperature

2.1 Introduction

Mobile air temperature monitoring is a promising method to better understand temperature distributions at fine spatial resolutions across urban areas but has not been well studied in recent scientific literature. One of the goals of this study was to identify twenty urban routes with varying land use, land cover, and other microenvironmental and human features for the purpose of monitoring and collecting air temperature data at an ambulatory pace on hot summer days. Another goal was to compare the collected mobile air temperature data with measurements from (1) the fixed air temperature monitoring network, (2) available Landsat measurements of LST, and (3) the GVHM.

2.2 Methods

2.2.1 Study Area

The study was conducted in greater Vancouver, British Columbia (BC), located on the southwestern coast of Canada. The 2016 population was 2.46 million residents, with an average year-over-year increase of 1.3% (90). Approximately 15.7% of the population is over the age of

65 years (90). Typical summer weather in Vancouver is characterized by low pressure, anti- cyclonic systems associated with warm, sunny weather, and light winds (34,43,62). High pressure, cyclonic systems associated with cloudy, wet, and rainy weather originating from the

Pacific are usually deflected north of Vancouver, but occasional cloudy and rainy periods occur during the summer when a cyclonic system manages to intrude southward (62).

Greater Vancouver is bounded by the Salish Sea, which is a straight of water connected to the Pacific Ocean in the west, by mountain ridges as well as the Burrard Inlet, a body of water

connected to the Salish Sea, in the north, and by the Fraser Valley in the east (34,43,62).

Stronger winds are more common near coast (62). Both land-sea breeze and mountain-valley

(mesoscale) systems are commonly present during the summer and complement each other

(43,62). During the day, the two systems cause cooler surface marine air to travel inland, up- valley, and upslope (towards the east), while at night, the two systems causes warmer air from the interior to travel downslope, down-valley, and shoreward (43,62).

2.2.2 Route Selection

We selected 20 walking routes with an approximate length of 8-10 km per route, which would allow each to be walked in two to three hours (Figure 2-1). Route selection was based on factors that could increase exposure as well as individual vulnerability to heat stress.

Specifically, walking routes were chosen in areas that fulfilled either of the following criteria: (1) two of the conditions listed below were met for at least 30% of the area, or (2) one of the conditions listed below was met for at least 75% of the area. The conditions were: (1) a population density >100 persons/ha as determined from the 2006 census (91,92); (2) an estimated air temperature greater than 3.5 °C warmer than the YVR according to the GVHM

(57); and (3) an average household income below $30,000 from the 2006 census (91,92). Census tracts with zero income occurred when census results had been suppressed, so population density and GVHM air temperature were used as the only selection criteria in these areas. After initial identification, the final walking routes were spatially distributed across the study area to maximize coverage, especially in areas with sparse fixed monitoring data. The routes were also selected to provide a range of distances to the nearest major bodies of water. Routes were created manually by retaining streets that optimized logistical efficiency (Figure 2-1).

Figure 2-1: Map of all 20 greater Vancouver mobile air temperature sampling routes and all Metro Vancouver fixed stations used in this study. The Burnaby North and Richmond South stations were in grass fields located in residential neighbourhoods with detached dwellings. The Burnaby South station was on a school rooftop in a residential neighbourhood with detached dwellings, and the Vancouver International Airport (YVR) station was in a grass field hundreds of meters from an airport runway (93).

2.2.3 Mobile Air Temperature Data Collection

Mobile temperature data were collected on foot using a Met One 064-2 temperature sensor inside a radiation shield, a Kestrel 4500 Portable Weather Station, a GoPro Hero 3 video camera, and a Garmin GPSMAP 78s GPS (Figure 2-2). The Met One 064-2 and Kestrel 4500 were mounted to a PVC pipe frame at a height of 1.5m and a distance of 50cm away from the body to prevent direct heat transfer. The Kestrel and GoPro video data were not used in the analyses presented here. The Met One 064-2 temperature sensor was chosen for this study because its technical specifications indicate an accuracy of ±0.1°C and a response time of 10

seconds in still air (94). Our walking speed calculations and the Kestrel 4500 data indicate that there was approximately 1.2m/s of ventilation for the air temperature sensor. Our preliminary testing indicated that, at walking speed, the sensor responded to a 2°C fluctuation within 10 seconds and returned to 70% of baseline within 30 seconds. The return to baseline required five minutes. All data were logged at 10-second intervals. Study data were collected on foot for two reasons. First, travelling by foot means smaller distances were covered within set time periods compared with travelling by vehicle or bike, which allows higher spatial resolution. Second, many areas in greater Vancouver were only accessible by pedestrians or bicycles to transportation network design.

We collected 42 sets of data for the 20 sampling routes, with each route monitored at least twice. All data collection was conducted from May to September 2014, between the hours of 15:00 and 18:00 to capture the hottest hours of summer days in the region. Data collection occurred mainly on days when the maximum air temperature exceeded 22°C at YVR to ensure high temperatures were present in the dataset. Overcast days were not sampled, and any periods of cloudiness that occurred during a sampling run were manually recorded. Instrument times were synchronized prior to each run to maximize temporal matching. Replicate runs of each route were designed such that they were walked on the opposite side of the street and in opposite directions from each other.

Figure 2-2: Mobile monitoring setup. Left image shows the white radiation shield containing the Met One 064-2 sensor and the yellow Kestrel 4500 Portable Weather Station. The right image shows the typical sampling setup with the thermometers worn and the GoPro video camera mounted on the left shoulder. 2.2.4 Temperature Data for Comparison with Microscale Measurements

We compared mobile measurements with three different sources of temperature data: (1) fixed site air temperature measurements from the local monitoring network; (2) Landsat measurements of LST; and (3) hot day air temperature estimates from the GVHM. Fixed air temperature measurements at 1-minute intervals were received for four sites (Figure 2-1 and

Table A-1, Appendix A) from Metro Vancouver, which is a partnership providing joint services to all the municipalities in the region (95). The fixed site air temperature sensors had heights ranging from 5.5m at YVR to 19.3m at Burnaby South, and all instruments were within the urban canopy layer for their respective areas (Table 2-1) (93).

Table 2-1: Station air temperature sensor heights and Local Climate Zone classifications

Site Name Air Temperature Sensor Height (m) Local Climate Zone Fixed Sites Richmond South 8.2 9- Sparsely Built Burnaby South 19.3 8- Large low-rise Burnaby North 5.7 4- Open high-rise YVR 5.5 E- Bare Rock or paved Mobile Routes Route 1 1.5 5- Open mid-rise Route 2 1.5 2- Compact mid-rise Route 3 1.5 3- Compact low-rise Route 4 1.5 4- Open high-rise Route 5 1.5 6- Open low-rise Route 6 1.5 6- Open low-rise Route 7 1.5 4- Open high-rise Route 8 1.5 6- Open low-rise Route 9 1.5 4- Open high-rise Route 10 1.5 6- Open low-rise Route 11 1.5 4- Open high-rise Route 12 1.5 4- Open high-rise Route 13 1.5 5- Open mid-rise Route 14 1.5 4- Open high-rise Route 15 1.5 5- Open mid-rise Route 16 1.5 4- Open high-rise Route 17 1.5 6- Open low-rise Route 18 1.5 6- Open low-rise Route 19 1.5 6- Open low-rise Route 20 1.5 5- Open mid-rise

Landsat satellite images are publicly available and can be processed into multiple products, including LST (77,96). Level 1 data from June 3 (Landsat 7), July 13 (Landsat 8), and

July 29 (Landsat 8) 2014, all acquired at approximately 11:00 pacific standard time, were obtained from the US Geological Service (97). These data were collected 4-7 hours prior to the spatially matched mobile data, which were measured during the hottest hours of the day. We chose to conduct the comparison this way because it was consistent with the use of LST data in the GVHM, which also modelled air temperature during the hottest hours of a typical hot summer day. While Landsat thermal data are originally sampled at horizontal resolutions of 60m 22

x 60m (Landsat 7) or 100m x 100m (Landsat 8) at nadir, they are resampled and provided to users at a 30m x 30m spatial resolution, which was used for our analyses (98). All three images were processed into LST values using established methods (57,99,100) in the R statistical computing environment (101). These methods included a correction for thermal emissivity using

NDVI values (102), and a conversion from top of atmosphere temperature values to LST values

(103,104). The Landsat 7 data gaps due to the scan line corrector error were left as is (105), while the Landsat 8 stray light effect was minimized by using band 10 as the thermal band (106).

The GVHM (57) was constructed using a random forest algorithm, which modelled the relationship between maximum air temperature at 59 weather stations and several predictor variables. Random forest is a non-parametric machine learning method which uses a large number of regression trees to model such relationships (57,107). The predictor variables in the

GVHM we used were LST averaged within a 1000m buffer, distance from the ocean, elevation,

NDWI, SVF, solar radiation, and atmospheric water vapor (57). The Landsat LST data used in the GVHM were restricted to six hot days with air temperatures exceeding 25 °C at the YVR fixed site (57). This restriction was made because the GVHM was designed to reflect spatial variation in air temperature on relatively hot days in the region. The map performed well at the local and mesoscale when compared with similar products for other cities (57). The published map expresses modelled temperatures relative to YVR because it is the station most frequently reported by the media, but we used the map of absolute values for this study.

2.2.5 Data Analysis

All data analysis and mapping were conducted with R version 3.1.0 and ESRI ArcGIS

10.2 (101,108). To compare the mobile air temperatures with the nearest fixed site air temperatures we generated 1-minute time series plots for each run and calculated the following 23

variables: mean and standard deviation (SD) mobile temperature; mean and SD fixed temperature; mean and SD difference between each 10-second value and the closest 1-minute value.

Because mobile temperatures were measured over the course of 2-3 hours during the hottest period of the day, there was typically an increase in ambient temperature over the measurement period. Given our interest in the spatial rather than temporal variability for comparison with the LST and GVHM data, we adjusted for these short-term temporal trends

(Equation 2-1). The reference site used was the Metro Vancouver fixed weather station data at

YVR (Figure 2-1) (109).

푇푠푎푚푝푙푖푛푔.푟푎푤 Equation 2-1: 푇푟푢푛 푎푑푗 = 푇 푟푒푓 푠푖푡푒,푟표푙푙푖푛푔⁄ 푇푟푒푓 푠푖푡푒,푟푢푛 푚푒푎푛

Where: Trun adj is the adjusted mobile air temperature in 10-second intervals; Tsampling,raw is the raw mobile sampling air temperature in 10-second intervals; Tref site, rolling is the reference site air temperature centered rolling mean (over 1 min intervals), time matched with Tsampling,raw; Tref site,run mean is the overall mean for the reference site air temperature during the same period when

Tsampling,raw was collected.

After the mobile air temperature data were adjusted, they were rasterized to a 30m x 30m resolution by taking the average of all point values within each cell. To compare the rasterized mobile data with the rasterized LST data, we generated scatter plots of the spatially matched values for the three sampling dates covered by Landsat, and histograms showing the LST distributions on each day were also generated. The r2, its p-value, the line of best fit, and the

LST/mobile air temperature difference were calculated for each plot.

The mobile data were collected on 42 summer days, whereas the GVHM was developed to reflect a typical hot summer day in greater Vancouver (57). To compare the rasterized mobile data with the rasterized GVHM data we converted both to z-scores, which indicate how many standard deviations each value is from the mean of its distribution. The use of z-scores allowed us to compare relative air temperatures between the mobile and GVHM datasets. Each individual mobile air temperature value was converted to z-score based on the mean and standard deviations for its specific run. Each individual GVHM value was converted to a z-score by cropping the entire GVHM raster to the extent of each route and using the mean and standard deviation of the cropped area. The difference between the GVHM z-scores and the mobile z- scores was calculated to create z-score differential maps for each run. Finally, correlation values for each run were calculated by taking spatially-matched mobile air temperature z-score data from two replicates of each run, creating a scatterplot for that run, and then calculating the correlation,

2.3 Results

2.3.1 Mobile Data Compared with Fixed Site Data

Complete 10-second interval air temperature datasets were collected for 41 out of 42 runs, while run 4B had data that were collected at 1-minute intervals due to an instrument error.

The routes ranged in length from 6.4km (route 6) to 9.9km (route 17), and in area from 0.5km2

(route 6) to 2.5km2 (route 16) (Table A-1, Appendix A). As expected, mobile air temperature measurements were typically higher and more variable than simultaneous air temperature measurements at the nearest fixed site (Figure 2-3, Table A-1). There was little consistency in the air temperature magnitude or variability between replicate measurements on the same route,

possibly due to differences in meteorology and regional thermal state on the different measurement days, and possibly due to the fact that replicates were measured in opposite directions and on opposite sides of the street. The fixed stations did generally characterize short- term temporal trends of nearby routes, even when the closest station was almost 10 km away

(Figure 2-1). The trends were most similar when the distance between the fixed site and the route was short. However, even for routes that were less than 1 km from the nearest fixed site (7A and

B), there was considerable variability in the mobile measurements that was not evident in the fixed station measurements.

Figure 2-3: Mobile and fixed air temperature mesasures for the 42 runs. Raw mobile (red) and fixed measurements from the closest Metro Vancouver weather station (blue). Y-axis shows temperature in oC while the x-axis shows time . Note that both x-axis and y-axis scales vary from run to run.

2.3.2 Mobile Data Compared with Land Surface Temperature Data

Landsat satellites overpassed the study area on June 3, July 13, and July 29, which corresponded with runs 15A, 18B, and 12B, respectively. Run 15A was monitored between 4:19-

6:39 PM, run 18B was monitored between 4:01-5:53 PM, and run 12B was monitored between

3:36-5:22 PM. The mean temperatures at YVR during those runs were 18.5 °C, 27.0 °C, and 26.4

°C, and the mean mobile air temperatures for those runs were 20.1 °C, 31.9 °C, and 26.8 °C respectively. In comparison, the mean LST values were 35.1 °C for run 15A, 38.1 °C for run 18B, and 34.8 °C for run 12B. Runs 15A and 12B were both sunny throughout, while route 18B was sunny with intermittent cloudiness. The linear regression relationship between the mobile air temperatures and LSTs were weak (Figure 2-4), and there was no clear similarity between the three runs other than a relatively narrow range for air temperatures and a relative wide range for

LSTs. The relationship with LST was likely driven by route location, as the correlations on the matched overpass days were similar to those on the unmatched days, with the exception of route

12B (Table 2-2).

Table 2-2: Correlation between mobile run air temperature data and land surface temperature (LST). Data include three overpass dates, and bold values indicate the correlation on the matched date.

LST from three overpass dates Mobile Run June 3 (Run 15A) July 13 (Run 18B) July 29 (Run 12B) 15A 0.37 0.31 0.27 18B 0.22 0.20 0.16 12B 0.00 0.64 0.61

Figure 2-4: Scatter plots of mobile air temperature versus land surface temperature (LST) as well as the LST histograms. Mobile monitoring runs coincided with Landsat on three overpass days. The r2 value between mobile air temperature and LST, the p-value for the calculated r2 value, the fit line equation, and the mean temperature difference between LST and mobile air temperature are provided on each plot. The LST data were measured at approximately 11:00, while the mobile air temperature data were mostly measured between 15:00 and 18:00.

2.3.3 Mobile Data Compared with the Greater Vancouver Heat Map (GVHM)

When the z-scores of mobile measurements were compared with the z-scores of the

GVHM estimates, the correlation between run replicates varied from route to route, but most run replicates were similar in visual comparison (Figure 2-5). The mean correlation (r) between route replicates z-scores was 0.46, the maximum correlation was 0.83, and the minimum correlation

was -0.01 (see Figure A-1 in Appendix B for correlations and scatterplots of all runs). The

GVHM described spatial variability in air temperatures for parts of many routes, as indicated by the large areas that were in agreement (Figure 2-5).

The total number of raster cells between all 42 runs was 11,379. Of these, 57.4% had a z- score between -1 and 1, 23.3% had a z-score > 1, and 19.3% had a z-score < -1 (Table A-2,

Appendix). Route 16A had the largest percentage of cells with z-scores > 1 (32.3%, 41.9% between -1 and 1, and 25.8% < -1). Route 18B had the largest percentage of cells with z-scores between -1 to 1 (76.5%, 17.9% > 1, and 5.6% < -1). Route 8A had the largest percentage of cells with z-scores < -1 (32.9%, 39.2% between -1 and 1, and 27.9% > 1). Of the 42 runs, 33 had over

50% of raster cells with z-scores between -1 and 1 (Table A-2, Appendix A), which suggests the

GVHM generally captured the spatial variability in air temperatures.

Figure 2-5: Differences in the z-scores between the Greater Vancouver Heat Map (GVHM) and mobile air temperatures for all 42 runs. When interpreting, a positive z-score differential means that GVHM z-scores were higher than the mobile air temperature z-scores, suggesting that the GVHM overestimated air temperature. A negative differential suggests the GVHM underestimated air temperature.

2.4 Discussion

There is limited evidence on the relationships between microscale urban air temperature measurements and other existing methods for measuring or modelling air temperature.

Comparisons between mobile and fixed site air temperatures indicated that mobile measurements were generally higher than time-synchronized fixed site measurements. Comparison between mobile air temperatures and LST showed that the two variables were positively correlated, although the relationships were weak. Finally, comparison between mobile air temperatures and the GVHM suggested that the GVHM captured spatial variability in temperatures within many routes, but that it underestimated and overestimated variation in some areas when compared with measured data.

Some previous studies have measured both fixed and mobile air temperatures (110,111), but we found none that had compared these two variables directly. Although the differences we observed were expected, we report them here to provide others with some information about the range of variability that might be found in other contexts. The differences are likely driven by the built environment (19,26,42,59,68,69), because most mobile monitoring routes were in urban areas, while the fixed weather stations were in less developed areas. In addition, temperatures at the fixed stations were potentially cooler due to vertical temperature gradients. The mobile air sensor was mounted at a height of 1.5 m, while the fixed station temperature sensors were at heights of 5.5-19.3 m (Table 2-1). The differences between fixed site and mobile measurements highlight the reality that heat exposure and related health impacts are likely to vary with the location, and that the limitations of fixed site data should be acknowledged when planning public health interventions. Many epidemiologic studies use these data to reflect population exposures

(3,7,8,11,112–114), but estimates that are more representative at the microscale, such as heat 31

maps, may result in better models and more targeted interventions. For example, the GVHM provided air temperature data that were more representative of relative air temperature spatial distributions, as measured by microscale monitoring, than the fixed site data (46).

The r2 values we report for the relationship between mobile air temperature and LST

(0.04-0.38) were lower than those reported in previous studies. These ranged from 0.64-0.87 when LST was compared with time-matched air temperature (64–66) or 0.83 when LST was compared with daily maximum air temperature (67). One possible explanation for our lower correlations was the 4-7 hour time gap between LST and mobile air temperature measurements, which would weaken any existing relationship. Another possible explanation is that we used mobile measurements from a single day, while most previous studies used years of data from fixed weather stations or observations from sites throughout all seasons (64,65). In addition, most of the other studies compared LST with time-matched or daily maximum air temperature at specific fixed pixels over time, which removes potential confounding by geographic variables such as land use/land cover and temporal variation of thermal state. One study in Hong Kong did compare LST to 148 km of vehicular air temperature traverse data at a 10-meter resolution and found a r2 of 0.80 (66). However, when the results were split based on urban or rural classification of data, the r2 value dropped to 0.42 for urban areas (66).

Another possible factor in our lower r2 values is the way that temperatures are measured by the different instruments. Air temperature sensors have an ellipsoid source area with both horizontal and vertical influences upwind of the sensor (28,39,115), while LST sensors have a circular source area below the sensor (28,116). While this mismatch affects all measurements of the two variables, it may be exaggerated for microscale variables because of smaller overlapping areas. Another consideration is that all satellite-based LST measurements have some blurring 32

between adjacent pixels, primarily due to the scanning process of the moving sensor. As such, these LST measurements may not truly represent variability at the microscale. Furthermore,

Landsat 7 and 8 thermal data were natively recorded on 60m x 60m and 100m x 100m grids respectively, and are then resampled to a 30m x 30m grid, which results in blurring between neighbouring 30m x 30m pixels (98). Calibration of LST measurements from Landsat 7 ETM+ and its predecessor instruments have been ongoing for decades (103), but spatial blurring is not relevant during these calibrations as they focus on homogeneous areas where absolute in-situ

LST values can be measured. Without specific information on the influence of this blurring, we followed numerous other studies and compared in-situ measurements to the LST values from the pixels in which they were taken (30,36,64,66,67,117).

Heat maps generated by regression models provide useful estimates and visualizations of air temperature differences between areas. However, like any model estimates, they must be evaluated with real data to assess their validity, preferably at the same resolution as the maps themselves. Microscale mobile air temperature data fill a gap that other data sources cannot by offering accurate measurements collected at very high spatial resolutions with temporal coverage limited only by the study design and resources. Spatial coverage is limited by the distance a person can cover on foot, restricting the utility of this approach to smaller areas. However, it could have a wide range of applications such as heat map development, studying the land use/land cover drivers of microscale air temperature variability, or evaluation of more common and less resource-intensive methods.

Most heat maps are evaluated using cross-validation. Leave-one-out cross-validation for the GVHM involved (1) leaving out all observations from one weather station, (2) running the model, (3) calculating the difference between the modelled temperature value and actual station 33

temperature values, (4) repeating the procedure until the model has been run with each station missing once, and (5) calculating the average mean absolute error and root means square error for all runs to quantify accuracy (57). One study in Hong Kong constructed 10m x 10m resolution daytime and nighttime air temperature heat maps by regressing a ASTER image from approximately 11am and one from approximately 11pm with air temperature measurements collected within 1.5 hours of the corresponding overpass (36). These heat maps were validated with time-matched air temperature data from the Hong Kong fixed automatic weather station

(AWS) network. The daytime heat map and AWS air temperature data had a correlation of 0.75, while the nighttime heat map and the AWS air temperature data had a correlation of 0.84 (36). In comparison, we found a wide range of spatial correlation between patterns observed in the

GVHM and measured in our data. One possible explanation is that the air temperature data used to construct the Hong Kong model had higher temperatures and generally had a larger range

(approximately 5oC) compared with our individual route models (minimum range = 1.4oC, maximum range = 5.8oC, and mean range = 2.8oC).

Chapter 3: Land Use Regression Modelling of Microscale Urban Air

Temperatures

3.1 Introduction

Extreme heat events have been associated with excess morbidity and mortality worldwide, especially in cities where temperatures are magnified by the urban heat island effect.

Most previous research has evaluated extreme heat exposures at the municipal and local scales, but individuals are exposed at much smaller areas, and being able to model air temperature exposures accurately at the microscale would be useful for future urban planning and resource allocation. This study’s goal was to determine if microscale air temperature data could be used to develop LUR models, and if such models could be applied beyond the route to reliably predict microscale variability in unmeasured areas.

Firstly, 42 microscale LUR models were developed such that there was one model for each run replicate. The best model (based on model r2) for each of the 20 routes was then selected and used to evaluate between-route similarity, as well as evaluate differences between areas where air temperature is and is not well predicted. Finally, data from the 20 best models were used to develop a model for all the routes within the greater Vancouver area to determine if the microscale data could be used to model air temperatures for the region as a whole.

3.2 Methods

3.2.1 Air Temperature Data

The microscale temperature data and the methods used for collection have been described in detail in Chapter 2 (118). In brief, all data were collected during May to September 2014.

Each of 20 monitoring routes was sampled at least twice between the hours of 15:00 – 18:00 for 35

a total 42 runs. The routes were 8-10 km in length, were sampled on foot, and most took approximately two hours to complete. Air temperatures were measured every 10 seconds using a

Met One 064-2 air temperature sensor, and global position was measured every 10 seconds with a Garmin 78s GPS. Data from the Metro Vancouver fixed station at YVR (Figure 2-1 and

Equation 2-1) were used to remove the short-term temporal trends from the time-matched mobile data. Once adjusted, the mobile air temperature data were converted to raster format by averaging all data points within each 30m x 30m cell. To remove the long-term temporal trend for the combined overall route model, the mobile temperatures were divided by the average temperature measured at YVR during each of the mobile runs, similar to Equation 2-1.

3.2.2 Independent Variables

A set of potentially predictive independent variables was assembled based on our a priori understanding of factors driving microscale urban air temperatures. For each variable, we assigned an expected direction of effect, and for most variables we calculated the average values within 50, 100, and 200m buffers of the central point of each 30m x 30m raster cell (Table 3-1).

All variable construction and regression modeling was done in the R statistical computing environment (101). Variables were retained for further consideration if the average univariate correlation with air temperatures across 42 runs had a magnitude of 0.05 or larger in the expected direction, and if the relationship for at least 50% of the runs was consistent with the expected direction.

The first variables were the NDWI and the NDVI. Both the NDWI and the NDVI reflect the amount of vegetation within a given area (58,61,119), but the NDWI is more sensitive to the liquid water content within a canopy and is less sensitive to atmospheric interference than the

NDVI (61). We expected both NDWI and NDVI to have a negative relationship with air temperature because higher values correspond to increased evapotranspiration and increased shading (58,61). Landsat 8 surface reflectance data from July 13, 2014 and July 29, 2014 were processed into 30m x 30m rasters using established methods and then combined to create composite NDWI and NDVI rasters (61,97,119).

The next variable was SVF, which indicates how much of the sky above is visible and unobstructed by overhead objects. The SVF affects the ratio of radiation received by a surface to the total radiation from the environment, as well as the emission of infrared radiation

(53,54,57,58). While lower SVF is generally correlated with higher nocturnal air temperature due to decreased cooling rates, we did not assign a direction of effect to SVF because previous studies have had mixed results for daytime temperatures (53–60). One study noted that lower air temperatures in the downtown core of Vancouver were correlated with higher SVF, but did not indicate whether this relationship applied in other greater Vancouver areas (58). We used SVF data generated from previous studies (120).

We also tested the use of Distance to Major Road as a predictive variable. Roads indicate urban development and vehicular traffic, and they experience relatively high storage heat flux, whereby heat is stored throughout the day and then released at night (51,52). Distance to Major

Road is expected to have a negative relationship with air temperature, because the further away from a major road, the lower the intensity of urbanization and vehicular density, which should result in lower air temperatures. Road network data were taken from the DMTI Spatial CanMap

2013 data (121).

Distance to Large Water Body, with large water bodies defined as the Salish Sea and the

Burrard Inlet, was calculated because large water bodies have a cooling effect on nearby areas 37

during the day. In particular, the Salish Sea is responsible for most cooling in greater Vancouver

(43,62). A positive relationship was expected because the cooling effect of the ocean decreases with increasing distance. Although distance variables are not usually buffered, the 200m buffered version of Distance to Large Water Body was used in model selection as the buffer accounts for the potential air temperature effects of being surrounded by water in multiple directions. The

Distance to Large Water Body variable was generated from CanMap Water, v2010.3 data (122).

Finally, Elevation was considered due to the established negative relationships between elevation and air temperature at larger scales in the atmosphere and in previous models

(57,58,63). Digital elevation model (DEM) data of Area 92G from DMTI Spatial Inc were used

(123). Maps of all variables can be found in Appendix B (Figures A-2 to A-7).

3.2.3 Land Use Regression Modelling

The first step was to evaluate the univariate relationship between air temperature measurements and all potentially predictive variables separately for each of the 42 runs. For variables with multiple buffer sizes, the buffer that had the strongest and most consistent correlation with the air temperature data throughout all runs was used for model fitting. The route-specific models were then constructed by forward selection. The variable with the highest r2 value in the expected direction was used as the foundation. Independent variables were added to the model one at a time, and variables that (1) had the expected direction of effect and (2) increased the r2 value by at least 0.01 were retained until these criteria were no longer met or all variables had been added. These criteria are similar to what has been used in other studies in

Vancouver (52,81,83,124). All routes had one model for each of two or three runs, but the model from the run with the highest r2 values was kept for further analyses.

Two types of models were constructed for this study: the best individual route models, and an overall model. The individual route models described each of the 20 routes, and the purpose was to assess whether there was consistency in the relationships between model variables and air temperature across the study area. The overall model covered the entire extent of the greater Vancouver study area routes. It was created by merging the route air temperature raster data from the strongest 20 route replicates, which were chosen based on their individual route model r2.

3.2.4 Model Evaluation

A spatial cluster leave-ten-out cross-validation (LTOCV) r2 was calculated using the sperrorest package in R, and this value was used to evaluate the models (125). For each model, ten spatially clustered observations were omitted from the dataset, and the remaining observations were used to parameterize the model and to predict the ten removed observations.

This was repeated until all model observations had been predicted (126,127). These predicted datasets were then used with the observed mobile air temperature values to calculate the LTOCV r2. Cross validation r2 has been used in previous studies to demonstrate that models are stable and not overfitting their datasets to the point that they lose true predictive power (58,126,128,129).

The between-route consistency of variables was evaluated by calculating the coefficients and the relative importance of each variable in all of the individual route models (130). The relative importance (the percentage contribution of each variable to the model r2) was calculated with the lmg metric using the relaimpo package in R (131). The lmg metric uses the methods proposed by Lindeman, Merenda, and Gold (132) to calculate the sequential r2 for the variable of interest, or the difference in r2 between a model with that variable and the same model without that variable (131,133). However, variable sum of squares allocation, and therefore sequential r2, 39

depends on the order in which variables are introduced to the model. The lmg calculation controls for the effect of variable introduction order by calculating and averaging sequential r2 of a variable for all possible orderings of all variables in the model.

The mean of the relative importance was calculated for each variable to evaluate its overall utility as a predictor of microscale urban air temperature. We also summed the number of models (out of 20) in which each variable appeared. The locations of routes that were and were not well-described by the potentially predictive variables were visually compared with respect to proximity to each other and geographic location in greater Vancouver.

3.3 Results

3.3.1 Independent Variable Selection

Four variables were retained as potential predictors for the individual route models and the overall model (Table 3-1). The NDVI was dropped in favour of NDWI due to the high correlation (r = 0.904) between the variables and the higher univariate correlations between

NDWI and the air temperature measurements. The Elevation variable was also dropped from consideration because it was expected to have a negative average univariate correlation with air temperatures, but the actual value was positive (Table 3-1).

3.3.2 Individual Route Models

There were 221 to 385 cells per route (excluding the route with the 1-minute data interval malfunction, which had 126 cells), with an average of 271 cells per route. On average, the best individual route models explained 38.6% of the variation in microscale air temperatures for the twenty routes (Figure 3-1 and Table A-3, Appendix A). Model r2 values ranged from 2.4-69.4%.

The individual route model average LTOCV r2 was 31.2%, ranging from 0.1-66.5%. The average

difference between the r2 and LTOCV r2 values was 7.4%, suggesting that the models were not highly overfitted to their respective datasets. We did not observe any strong geographical trend in the amount of variation in air temperature explained by each of the individual route models

(Figure 3-1). Qualitatively, the weaker models were more often located further inland from the ocean and major rivers.

Table 3-1: Variables that were evaluated as potential predictors of microscale air temperature. Bold indicates that the variable was chosen for use in in the land use regression modeling. Most Expected predictive Average direction of buffer univariate Variable effect Data source Data manipulation size (m) correlation Normalized Difference Landsat 8 Raster Math in R to calculate Water Index reflectance NDWI using established methods (NDWI) Negative (97) (61,97,119). 50 -0.170 Normalized Difference Landsat 8 Raster Math in R to calculate NDVI Vegetation reflectance using established methods Index (NDVI) Negative (97) (61,97,119). 50 -0.166 Data from Sky-View two previous Used as provided by authors of Factor (SVF) None study (120) previous study (120). 200 -0.052 Created a 30m resolution raster in greater Vancouver, and the Created from distance of each cell from the CanMap Salish Sea and Burrard Inlet in Distance to Water, the raster was calculated, then Large Water v2010.3 data averaged within different buffer Body Positive (122) sizes (122). 200 0.258 A 10m grid of points was created for greater Vancouver, the minimum distance of major roads and highways (as defined by DMTI Spatial) from each point was calculated, and the points were converted to a 10m Created from resolution raster. The 10m raster Distance to CanMap values were averaged to create a Major Road Negative 2013 (121) 30m raster (121). - -0.068 DMTI 2002 Digital Elevation Model (DEM) Combined four sections of the DEM, Area 92G and then cropped the area for greater Elevation Negative (123). Vancouver (123). 200 0.067

Four models had one independent variable, nine models had two independent variables, five models had three independent variables, and two models used all four independent variables.

There was a weak positive correlation between the number of variables in a model and the r value of that model (r =0.47). Out of the variables used to create the individual route models,

Distance to Large Water Body provided the most information. It had the highest average relative importance of 69.0% for the 13 of 20 models in which it appeared. For the remaining three variables SVF had an average relative importance of 40.6% and also appeared in 13 models;

NDWI had an average relative importance of 32.6% and appeared in 11 models; and Distance to

Major Road had an average relative importance of 30.5% and appeared in eight models (Figure

3-1). Generally, there was large variation in the model coefficients between the 20 routes, and coefficients varied spatially without clearly discernable patterns.

Distance to Large Water Body was more likely to be present and generally had higher relative importance values for routes closer to the Salish Sea (Figure 3-1). On the other hand,

NDWI and Distance to Major Road was more likely to be present in the models for inland routes, and its relative importance values increased as Distance to Large Water Body increased. The

SVF importance values did not appear to show any geographic patterns.

Figure 3-1: Map showing r2 and relative importance (variable percentage contribution to model r2) for the best individual route models. The model r2 is represented by pie chart size, with the variable relative importance represented by the split between different categories inside each pie chart, where NDWI is the Normalized Difference Water Index and SVF is the Sky-View Factor.

3.3.3 Overall Route Model

The overall air temperature model was created using Distance to Large Water Body,

NDWI, and SVF variables and all temperature measurements adjusted for short- and long-term trends from the strongest individual route models (Figure 3-2). Distance to Major Road was omitted from consideration due to its variable coefficient being in the wrong direction. The overall route model was weak, as it only explained 10.0% of air temperature variation, with

Distance to Large Water Body contributing the most to the model r2 (Table 3-2).

Figure 3-2: Plot showing raw air temperature metadata of individual run models used to create overall air temperature. Blue shows the range of raw air temperatures measured at each route, while the black dash shows the mean air temperature at YVR during the time the route was traversed.

Table 3-2: Overall air temperature model variable coefficients and percentage contribution to r2 value.

Relative Variable Coefficient Importance (%) Distance to Large Water Body, 200m Buffer 0.000165 71.9% Distance to Major Road, Unbuffered - - NDWI, 100m Buffer -0.786 6.6% SVF, 200m Buffer -3.059 21.5%

3.4 Discussion

Distance to Large Water Body, Distance to Major Road, NDWI, and SVF were selected as the most predictive independent variables within the individual route models and the overall model. The individual route model r2 values ranged from 2.4-69.4% with an average of 38.6%, while the overall model had a model r2 value of 10.0%. Distance to Large Water Body provided the most information out of all variables and was more likely to be present and to have higher

relative importance values for routes closer to the coastline. Likewise, the NDWI and Distance to

Major Road variables were more likely to be present in the model and to have higher relative importance values for routes further inland.

Six potentially predictive variables were initially selected based on our a priori understanding of the relationship between air temperature and the surrounding environment. Two of these variables were dropped from consideration in the final models for different reasons. The

NDVI was omitted in favour of NDWI, due to the high similarity and covariance between them.

Elevation was omitted because it had an unexpectedly negative relationship with air temperature.

This may have been due to coastal, lower elevation areas having lower temperatures than higher elevation, inland areas. As such, Elevation and Distance to Large Water Body would both capture the same mechanism. As well, the elevation differences within the routes was small and may not have provided enough variability for elevation to be a useful variable at a microscale resolution.

Results for the individual routes have shown that it is possible to model air temperature at the microscale resolution for areas of up to 2.5km2. Direct comparison of the individual route models with other studies is difficult because most models in the literature are developed for entire cities or regions. However, the average r2 of 38% for all routes is comparable with the r2 of local region-wide heat models (57,58). For example, a 60m x 60m resolution air temperature model for greater Vancouver included elevation, land surface temperature, NDWI, SVF, and solar radiation variables and explained 34% of the variation in the response (58). A similar model for Hong Kong predicted daytime air temperature at a 90m resolution using ASTER land surface temperature and 10m resolution emissivity data, and it explained 75% of the variation in the response (36), which is comparable with our strongest individual route model. As previously 46

mentioned, an explanation for this may be the fact that the air temperature data used to construct the Hong Kong model had higher temperatures and generally had a wider range (approximately

5oC) compared to our individual route models (mean range = 2.8oC).

Overall, the final r2 values for individual routes did not appear to be driven by any geographic pattern. However, the relative importance for three of the four variables did appear to have patterns in geographic distribution among the individual route models. One likely explanation for the higher Distance to Large Water Body relative importance values closer to the coast could be that the influence of the ocean and sea breezes is strongest by the coast, and weaker as one goes further inland (63). The further routes could have been sufficiently far away that any change in distance from the ocean had a negligible effect on microclimate air temperature. Conversely, NDWI relative importance values were stronger and more likely to be present inland than by the coast. A possible explanation for the weaker and lower presence of

NDWI relative importance values by the coast could be that the influence of the ocean and sea breezes is much stronger than the effects of vegetation. This is consistent with further work using these data, where greenness was extracted from video footage (134). The greenness video data were not used in this study because they are not widely available and cannot be used to develop

LUR models over a large area.

Another variable that generally had stronger values inland was Distance to Major Road.

Aside from the influence of the ocean and sea breezes, a possible explanation is that the more urbanized areas have higher major road density. Areas with higher major road density may not have enough variability for the Distance to Major Road variable to be used to predict air temperature. This might be the case for Routes 1, 2, 3, 4, and 20, which have high major road densities according the variable map (Figure A-5). Similarly, the homogeneity of SVF within 47

most sample routes except in the downtown Vancouver area also limits the variability and therefore predictive power of SVF (Figure A-4).

One possible reason for the high relative importance of the Distance to Large Water Body variable is the fact the climatic conditions, such as vigorous atmospheric mixing, mean that microscale influences on temperature are weakened compared with larger scale processes such as sea breezes. Furthermore, these climatic factors are often strongest during the hottest hours of the day, when our measurements were made. This may be especially important for a variable such as

NDWI, which we expected to have a stronger effect because of its influence on daylight and evapotranspiration. Variables such as Distance to Major Road and SVF may have less predictive power for our routes because their effects are strongest in the early morning, evening, and overnight, when mixing is reduced, and when radiative trapping and anthropogenic heat (such as from engines) would play a larger microscale role. Overall, evening and nocturnal microscale temperatures may be more strongly correlated with the local land use than daytime temperatures.

Chapter 4: Conclusion

4.1 Summary of Key Findings

There is limited evidence on the relationships between microscale urban air temperature measurements and other existing methods for measuring or modelling air temperature.

Comparisons between mobile and fixed site air temperatures indicated that mobile measurements were generally higher than time-synchronized fixed site measurements. Comparison between mobile air temperatures and LST derived from remote sensing showed that the two variables were positively correlated, although the relationships were weak. Finally, comparison between mobile air temperatures and the GVHM suggested that the GVHM captured spatial variability in temperatures within many routes, but that it underestimated and overestimated variation in some areas when compared with measured data.

Major Road variables were more likely to be present in the model and to have higher relative importance values for routes further inland. The presence and strength of large scale processes such as the land-sea breeze and atmospheric air mixing during the day may be a factor in the reduced strength of our microscale models and variables.

4.2 Contribution to Body of Knowledge

This is one of the first studies to collect air temperature and other related data at an ambulatory pace within an urban environment. Not only are the collected data useful for assessing spatial variability in air temperatures, the described methods will be useful to others planning to conduct similar monitoring campaigns. This study has demonstrated the potential utility of microscale air temperature data for evaluating other sources of air temperature data. For example, the collected microscale air temperature data showed that the GVHM is more representative of spatial distributions of relative air temperatures than LST and fixed weather station data. Furthermore, the observed spatial variability in the collected microscale air temperature data provides further evidence that epidemiological studies on heat-related health outcomes could benefit from microscale heat maps, because fixed station air temperatures rarely represent the actual temperature exposure in their surrounding areas. This could result in reduced resolution for observing the ‘true’ air temperature thresholds which result in mortality/morbidity outcomes and could also make comparisons of thresholds between studies difficult due to the differences between reference air temperature stations.

As well, this study has demonstrated that microscale air temperature can be modelled for areas of up to 2.5km2 using microscale air temperature data collected in urban environments.

However, the results suggest that using microscale air temperature data to construct a regional microscale air temperature model may not be possible. Finally, the strength of the Distance from

Large Water Body variable, particularly closer to the coast, suggests that distance from the coast is one of the largest drivers of daytime air temperature variability in greater Vancouver, as the

Distance to Major Road, NDWI, and SVF variables had relatively little influence in models close to the coast but were stronger inland. This may be due to the influence of large scale processes 50

such as regional circulation patterns, including the land-sea breeze, and atmospheric mixing.

This observation has implications for the generalizability of our work to other contexts, where land-sea breezes may not be present.

4.3 Strengths

One important strength of this study was the large volume of air temperature data it generated within a 102-day period. In addition, other information was also collected, including wind speed, relative humidity, and continuous video data. Not all of these data were used in the analyses presented here, but all have the potential to be useful in future research. For example, a study using the video data to generate an urban greenness index has shown promising results

(134). Another strength is the methods used to select the 20 routes, which focused on areas of particular heat vulnerability due either to environmental conditions or social risk factors. This was particularly important for evaluation of previous heat-related mapping in the region, because it meant that the GVHM could be assessed in the areas of high risk. Indeed, the existence of so much previous heat-related work in greater Vancouver was another strength of this study because we were able to leverage and build on other studies. Two final strengths are that results from this study will be useful for guiding similar studies in future, and for directing resources related to the mitigation of heat-related health impacts.

4.4 Limitations

One limitation of our study is that land-sea and mountain-valley wind system shifts are common in greater Vancouver, which could generate within-run spatially-independent variability in the mobile data. This could be assessed through analysis of wind speed and direction patterns in the greater Vancouver region during route times, which we conducted during preliminary

analyses. Results suggested that the processes stayed fairly consistent during all mobile runs

(Appendix C). Another limitation is the within-run correction for temporal trend, which assumes that the rate of change at fixed stations is the same as the rate of change in mobile measurements.

2 However, given that the r values between LST and the mobile air temperatures increased after the correction equation was applied to the mobile data, we assume that the correction was at least partially effective.

A further limitation is the between-day and time-of-day variability present in the mobile air temperature data, despite all sampling being conducted during summer months at approximately the same time of day (as indicated in Figure 3-2). We attempted to minimize this time-of-day variability using the within-run air temperature correction, but it would still be ideal to have more instantaneous measurements across the whole route.

We could not make effective corrections for between-day variability for the purposes of comparison with the GVHM. This raised concern about the true average microscale air temperature at each route and necessitated treating all runs as independent samples. The between-day variability was highlighted by the large differences between run replicates and the fact that the overall model only explained 10.0% of the variation in air temperature across the study routes. This is not surprising, given that the combined measurements from the 20 routes spanned 102 days, which introduced temporal variation at the daily, weekly, and even monthly scales. We attempted to address this by restricting the overall model inputs to models within narrower range of average route air temperatures at YVR, but this actually decreased the overall model r2 to less than 5%. This may be due to the fact that the generally weaker route models were included when the temperature restrictions were applied.

We know that the crude temperature correction for the long-term temporal trend helped with the overall model fit, because the overall model r2 value was only 2.1% when using the data that were corrected for short-term trends only. Nonetheless, long-term temporal variability between routes is one of the major study limitations, because it affects our ability to compare results with other data sources and air temperature models that cover the entire region.

4.5 Future Studies

Future studies that attempt to model microscale air temperature for larger areas should collect data over a much shorter time period to reduce temporal variability. Reducing the route traverse time with multiple thermometers or traversing the routes with vehicles or bicycles would help reduce within-route temporal variability. Likewise, reducing the overall campaign time span would help reduce the between-route temporal variability, although this may reduce the number of routes that can be measured with one thermometer. A best-case scenario would be using multiple thermometers at multiple routes simultaneously, which would drastically reduce within- route temporal variability, and almost entirely eliminate between-route temporal variability; however, this would be very resource intensive. Keeping consistent protocols for mobile monitoring would also help ensure consistency in how air temperature is being measured, especially if personnel and multiple thermometers are deployed at multiple routes. Finally, future studies should prioritize route replicates over the number of routes monitored, as having more replicate measurements will provide more certainty that the true spatial and temporal variability of air temperature is captured for that area.

Future microscale air temperature model studies should also consider incorporating municipal street tree data, because many cities now digitize this important information. As well, solar radiation data would be another useful variable to include in future models, although no 53

widely available high-resolution data are known to be available for this variable. Also, future research should consider taking regional wind circulation patterns into account, as that is suspected to be a major driver for the strength of the Distance to Large Water Body variable reported here.

4.6 Conclusions

The study in Chapter 2 was conducted to assess the GVHM and two of its inputs with respect to their ability to reflect microscale variability in air temperature. Our results suggested that: (1) the limitations of fixed site air temperature measurements should be considered when assessing public health risk; (2) the correlations between microscale air temperature and LST may be highly variable; and (3) microscale data can be valuable for assessing heat maps developed with local and macroscale data. Comparisons between the microscale air temperature measurements and the GVHM showed that the GVHM provided a reasonable spatial representation of relative air temperature for most of the routes. Furthermore, it provided more useful spatial information than LST measurements or fixed weather station data. Future microscale mobile monitoring campaigns should be specifically designed to evaluate these types of modelling tools.

The results of Chapter 3 suggest that microscale mobile air temperature measurements can be used to model microscale air temperature variations with LUR in some spatially limited locations, although the current study has only demonstrated this for individual routes with areas of 0.5-2.5km2. However, the regional microscale air temperature model could only explain

10.0% of the air temperature variation. Indeed, differences between variables and variable coefficients in the individual route models suggests that in may be impossible to develop a regional model using microscale air temperature measurements. Future studies intending to use 54

microscale modelling for LUR should collect data within a restricted time range to reduce short- and long-term temporal variability. They could also focus on areas of particular vulnerability or exposure to address specific, policy-relevant questions. For example, it might be possible to use very targeted microscale measurements and LUR to inform decisions about where to plant street trees for urban heat mitigation.

Bibliography

1. Luber G, McGeehin M. Climate Change and Extreme Heat Events. Am J Prev Med. 2008 Nov;35(5):429–35.

2. Lugo-Amador NM, Rothenhaus T, Moyer P. Heat-related illness. Emerg Med Clin North Am. 2004 May;22(2):315–27.

3. Kosatsky T, Henderson SB, Pollock SL. Shifts in Mortality During a Hot Weather Event in Vancouver, British Columbia: Rapid Assessment With Case-Only Analysis. Am J Public Health. 2012 Dec;102(12):2367–71.

4. Semenza JC, Rubin CH, Falter KH, Selanikio JD, Flanders WD, Howe HL, et al. Heat- Related Deaths during the July 1995 Heat Wave in Chicago. N Engl J Med. 1996 Jul 11;335(2):84–90.

5. Harlan SL, Brazel AJ, Prashad L, Stefanov WL, Larsen L. Neighborhood microclimates and vulnerability to heat stress. Soc Sci Med. 2006 Dec;63(11):2847–63.

6. Meehl GA, Tebaldi C. More Intense, More Frequent, and Longer Lasting Heat Waves in the 21st Century. Science. 2004 Aug 13;305(5686):994–7.

7. Robine J-M, Cheung SLK, Le Roy S, Van Oyen H, Griffiths C, Michel J-P, et al. Death toll exceeded 70,000 in Europe during the summer of 2003. C R Biol. 2008 Feb;331(2):171–8.

8. García-Herrera R, Díaz J, Trigo RM, Luterbacher J, Fischer EM. A Review of the European Summer Heat Wave of 2003. Crit Rev Environ Sci Technol. 2010 Mar 5;40(4):267–306.

9. Sardon JP. The 2003 heat wave. Euro Surveill Bull Eur Sur Mal Transm Eur Commun Dis Bull. 2007 Mar;12(3):226.

10. Centers for Disease Control and Prevention. Prevention Guide to Promote Personal Health and Safety (Part 1) [Internet]. 2012 [cited 2014 Aug 29]. Available from: http://www.bt.cdc.gov/disasters/extremeheat/heat_guide.asp

11. Henderson SB, Kosatsky T. A data-driven approach to setting trigger temperatures for heat health emergencies. Can J Public Health Rev Can Santé Publique. 2012 Jun;103(3):227–30.

12. Kalkstein L, Davis R. Weather and Human Mortality: An Evaluation of Demographic and Interregional Responses in the United States. Ann Assoc Am Geogr. 1989 Mar 1;79(1):44–64.

13. Kalkstein L, Greene JS. An evaluation of climate/mortality relationships in large U.S. cities and the possible impacts of a climate change. Environ Health Perspect. 1997 Jan;105(1):84–93.

14. Kjellstrom T, Briggs D, Freyberg C, Lemke B, Otto M, Hyatt O. Heat, Human Performance, and Occupational Health: A Key Issue for the Assessment of Global Climate Change Impacts. Annu Rev Public Health. 2016;37(1):97–112.

15. O’Malley PG. Heat waves and heat-related illness. JAMA. 2007 Aug 22;298(8):917–9.

16. Sheridan SC. A survey of public perception and response to heat warnings across four North American cities: an evaluation of municipal effectiveness. Int J Biometeorol. 2007 Oct 1;52(1):3–15.

17. City of Vancouver. Stay safe in the summer heat [Internet]. 2014 [cited 2014 Sep 5]. Available from: http://vancouver.ca/people-programs/hot-weather.aspx

18. Bailey A, Bankay D. City of Vancouver helping homeless find relief from heat. News1130 [Internet]. 2014 Jul 15 [cited 2014 Sep 5]; Available from: http://www.news1130.com/2014/07/14/city-of-vancouver-helps-homeless-find-relief- from-heat/

19. Watkins R, Palmer J, Kolokotroni M. Increased Temperature and Intensification of the Urban Heat Island: Implications for Human Comfort and Urban Design. Built Environ 1978-. 2007 Jan 1;33(1):85–96.

20. Millward AA, Torchia M, Laursen AE, Rothman LD. Vegetation Placement for Summer Built Surface Temperature Moderation in an Urban Microclimate. Environ Manage. 2014 Jun 1;53(6):1043–57.

21. Intergovernmental Panel on Climate Change. Climate change 2013: The physical science basis [Internet]. Stocker TF, Qin D, Plattner G-K, Tignor M, Allen SK, Boschung J, et al., editors. 2013 [cited 2014 Aug 31]. Available from: http://www.climatechange2013.org/images/report/WG1AR5_Frontmatter_FINAL.pdf

22. Peng RD, Bobb JF, Tebaldi C, McDaniel L, Bell ML, Dominici F. Toward a Quantitative Estimate of Future Heat Wave Mortality under Global Climate Change. Environ Health Perspect. 2011 May;119(5):701–6.

23. Salathé EP, Leung LR, Qian Y, Zhang Y. Regional climate model projections for the State of Washington. Clim Change. 2010 Sep;102(1–2):51–75.

24. Li G, Zhang X, Cannon AJ, Murdock T, Sobie S, Zwiers F, et al. Indices of Canada’s future climate for general and agricultural adaptation applications. Clim Change. 2018 May 1;148(1–2):249–63.

25. Oke TR. The distinction between canopy and boundary‐layer urban heat islands. Atmosphere. 1976 Dec 1;14(4):268–77.

26. Oke TR, Maxwell GB. Urban heat island dynamics in Montreal and Vancouver. Atmospheric Environ 1967. 1975 Feb;9(2):191–200.

27. Stewart ID, Oke TR, Krayenhoff ES. Evaluation of the “local climate zone” scheme using temperature observations and model simulations. Int J Climatol. 2014 Mar 1;34(4):1062– 80.

28. Oke TR. Initial Guidance to Obtain Representative Meteorological Observations at Urban Sites. Vol. 81. World Meteorological Organization Geneva; 2004.

29. Nakamura Y, Oke TR. Wind, temperature and stability conditions in an east-west oriented urban canyon. Atmospheric Environ 1967. 1988;22(12):2691–700.

30. Benali A, Carvalho AC, Nunes JP, Carvalhais N, Santos A. Estimating air surface temperature in Portugal using MODIS LST data. Remote Sens Environ. 2012 Sep;124:108–21.

31. Borgoni R. A Quantile Regression Approach to Evaluate Factors Influencing Residential Indoor Radon Concentration. Environ Model Assess. 2011 Jun;16(3):239–50.

32. Eliasson I, Svensson MK. Spatial air temperature variations and urban land use — a statistical approach. Meteorol Appl. 2003 Jun 1;10(2):135–49.

33. Georgi NJ, Zafiriadis K. The impact of park trees on microclimate in urban areas. Urban Ecosyst. 2006 Sep 1;9(3):195–209.

34. Ho HC, Kundby A, Sirovyak P, Xu Y, Hodul M, Henderson S. Mapping Maximum Urban Air Temperature on Hot Summer Days. Simon Fraser University; 2014.

35. Leconte F, Bouyer J, Claverie R, Pétrissans M. Using Local Climate Zone scheme for UHI assessment: Evaluation of the method using mobile measurements. Build Environ [Internet]. 2014 [cited 2014 Aug 15]; Available from: http://www.sciencedirect.com/science/article/pii/S0360132314001413

36. Nichol JE, To PH. Temporal characteristics of thermal satellite images for urban heat stress and heat island mapping. ISPRS J Photogramm Remote Sens. 2012 Nov;74:153– 62.

37. Rinner C, Hussain M. Toronto’s Urban Heat Island—Exploring the Relationship between Land Use and Surface Temperature. Remote Sens. 2011 Jun 21;3(6):1251–65.

38. Stewart ID, Oke TR. Methodological concerns surrounding the classification of urban and rural climate stations to define urban heat island magnitude. Prepr ICUC6 Göteb. 2006;431.

39. Stewart ID, Oke TR. Local Climate Zones for Urban Temperature Studies. Bull Am Meteorol Soc. 2012 May 25;93(12):1879–900.

40. Xu Y, Qin Z, Shen Y. Study on the estimation of near-surface air temperature from MODIS data by statistical methods. Int J Remote Sens. 2012;33(24):7629–43.

41. Kuhn KG, Campbell-Lendrum DH, Davies CR. A continental risk map for malaria mosquito (Diptera: Culicidae) vectors in Europe. J Med Entomol. 2002 Jul;39(4):621–30.

42. Saaroni H, Ziv B. Estimating the Urban Heat Island Contribution to Urban and Rural Air Temperature Differences over Complex Terrain: Application to an Arid City. J Appl Meteorol Climatol. 2010 Jun 7;49(10):2159–66.

43. Richards K. Urban and rural dewfall, surface moisture, and associated canopy-level air temperature and humidity measurements for Vancouver, Canada. Bound-Layer Meteorol. 2005 Jan 1;114(1):143–63.

44. Oke TR. Boundary Layer Climates, Second Edition. Methuen Lond. 1987;435.

45. World Meteorological Organization. Guide to Meterological Instruments and Methods of Observation. 7th Edition. Secretariat of the World Meteorological Organization; 2008.

46. Ho HC, Knudby A, Walker BB, Henderson SB. Delineation of Spatial Variability in the Temperature-Mortality Relationship on Extremely Hot Days in Greater Vancouver, Canada. Environ Health Perspect [Internet]. 2016 Jun 27 [cited 2016 Jul 3]; Available from: http://ehp.niehs.nih.gov/EHP224

47. Hajat S, Sheridan SC, Allen MJ, Pascal M, Laaidi K, Yagouti A, et al. Heat-health warning systems: a comparison of the predictive capacity of different approaches to identifying dangerously hot days. Am J Public Health. 2010 Jun;100(6):1137–44.

48. Parker JD, Woodruff TJ, Basu R, Schoendorf KC. Air Pollution and Birth Weight Among Term Infants in California. Pediatrics. 2005 Jan 1;115(1):121–8.

49. Krstic N, Yuchi W, Ho HC, Walker BB, Knudby AJ, Henderson SB. The Heat Exposure Integrated Deprivation Index (HEIDI): A data-driven approach to quantifying neighborhood risk during extreme hot weather. Environ Int. 2017 Dec 1;109:42–52.

50. Leech JA, Nelson WC, Burnett RT, Aaron S, Raizenne ME. It’s about time: A comparison of Canadian and American time–activity patterns†. J Expo Sci Environ Epidemiol. 2002 Nov;12(6):427–32.

51. Xu H. Analysis of Impervious Surface and its Impact on Urban Heat Environment using the Normalized Difference Impervious Surface Index (NDISI). Photogramm Eng Remote Sens. 2010;76(5):557–65.

52. Wang R, Henderson SB, Sbihi H, Allen RW, Brauer M. Temporal stability of land use regression models for traffic-related air pollution. Atmos Environ. 2013 Jan;64:312–9. 59

53. Chen L, Ng E, An X, Ren C, Lee M, Wang U, et al. Sky view factor analysis of street canyons and its implications for daytime intra-urban air temperature differentials in high- rise, high-density urban areas of Hong Kong: a GIS-based simulation approach. Int J Climatol. 2012 Jan 1;32(1):121–36.

54. Ha J, Lee S, Park C. Temporal Effects of Environmental Characteristics on Urban Air Temperature: The Influence of the Sky View Factor. Sustainability. 2016 Sep 5;8(9):895.

55. Giridharan R, Lau SSY, Ganesan S, Givoni B. Lowering the outdoor temperature in high- rise high-density residential developments of coastal Hong Kong: The vegetation influence. Build Environ. 2008 Oct;43(10):1583–95.

56. Giridharan R, Lau SSY, Ganesan S, Givoni B. Urban design factors influencing heat island intensity in high-rise high-density environments of Hong Kong. Build Environ. 2007 Oct;42(10):3669–84.

57. Ho HC, Knudby A, Xu Y, Hodul M, Aminipouri M. A comparison of urban heat islands mapped using skin temperature, air temperature, and apparent temperature (Humidex), for the greater Vancouver area. Sci Total Environ. 2016 Feb;544:929–38.

58. Ho HC, Knudby A, Sirovyak P, Xu Y, Hodul M, Henderson SB. Mapping maximum urban air temperature on hot summer days. Remote Sens Environ. 2014 Nov;154:38–45.

59. Holmer B, Thorsson S, Eliasson I. Cooling Rates, Sky View Factors and the Development of Intra-Urban Air Temperature Differences. Geogr Ann Ser Phys Geogr. 2007 Jan 1;89(4):237–48.

60. Svensson MK. Sky view factor analysis – implications for urban air temperature differences. Meteorol Appl. 2004;11(3):201–11.

61. Gao B. NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens Environ. 1996 Dec 1;58(3):257–66.

62. Oke TR, Hay JE. The Climate of Vancouver. Second Edition. University of British Columbia; 1994. 91 p. (BC Geographical Series).

63. Aguado E, Burt J. Understanding Weather and Climate. 7th Edition. 7th ed. Prentice-Hall; 2014.

64. Sohrabinia M, Zawar-Reza P, Rack W. Spatio-temporal analysis of the relationship between LST from MODIS and air temperature in New Zealand. Theor Appl Climatol. 2014 Mar 7;119(3–4):567–83.

65. Gallo K, Hale R, Tarpley D, Yu Y. Evaluation of the Relationship between Air and Land Surface Temperature under Clear- and Cloudy-Sky Conditions. J Appl Meteorol Climatol. 2010 Oct 4;50(3):767–75.

66. Nichol JE, Fung WY, Lam K, Wong MS. Urban heat island diagnosis using ASTER satellite images and “in situ” air temperature. Atmospheric Res. 2009 Oct;94(2):276–84.

67. Mostovoy GV, King RL, Reddy KR, Kakani VG, Filippova MG. Statistical Estimation of Daily Maximum and Minimum Air Temperatures from MODIS LST Data over the State of Mississippi. GIScience Remote Sens. 2006 Mar 1;43(1):78–110.

68. Emmanuel R, Krüger E. Urban heat island and its impact on climate change resilience in a shrinking city: The case of Glasgow, UK. Build Environ. 2012 Jul;53:137–49.

69. Oke TR. The energetic basis of the urban heat island. Q J R Meteorol Soc. 1982 Jan 1;108(455):1–24.

70. Oke TR. City size and the urban heat island. Atmospheric Environ 1967. 1973 Aug 1;7(8):769–79.

71. Unger J, Sümeghy Z, Zoboki J. Temperature cross-section features in an urban area. Atmospheric Res. 2001 Jul;58(2):117–27.

72. Graham DA, Vanos JK, Kenny NA, Brown RD. The relationship between neighbourhood tree canopy cover and heat-related ambulance calls during extreme heat events in Toronto, Canada. Urban For Urban Green. 2016 Dec 1;20:180–6.

73. Schowengerdt RA. Remote Sensing: Models and Methods for Image Processing. Academic Press; 2006. 559 p.

74. Vogt JV, Viau AA, Paquet F. Mapping regional air temperature fields using satellite- derived surface skin temperatures. Int J Climatol. 1997 Nov 30;17(14):1559–79.

75. NASA. MODIS Website [Internet]. 2014 [cited 2014 Sep 3]. Available from: http://modis.gsfc.nasa.gov/about/design.php

76. USGS. Landsat Enhanced Thematic Mapper Plus (ETM+) [Internet]. 2012 [cited 2014 Sep 3]. Available from: https://lta.cr.usgs.gov/LETMP

77. USGS. Landsat Thematic Mapper (TM) [Internet]. 2012 [cited 2014 Sep 3]. Available from: https://lta.cr.usgs.gov/TM

78. Bishop-Williams KE, Berke O, Pearl DL, Kelton DF. A spatial analysis of heat stress related emergency room visits in rural Southern Ontario during heat waves. BMC Emerg Med. 2015;15:17.

79. Ryan PH, LeMasters GK. A Review of Land-use Regression Models for Characterizing Intraurban Air Pollution Exposure. Inhal Toxicol. 2007;19(Suppl 1):127–33.

80. Ghassoun Y, Ruths M, Löwner M-O, Weber S. Intra-urban variation of ultrafine particles as evaluated by process related land use and pollutant driven regression modelling. Sci Total Environ. 2015 Dec 1;536:150–60. 61

81. Abernethy RC, Allen RW, McKendry IG, Brauer M. A Land Use Regression Model for Ultrafine Particles in Vancouver, Canada. Environ Sci Technol. 2013 May 21;47(10):5217–25.

82. Hoek G, Beelen R, Kos G, Dijkema M, Zee SC van der, Fischer PH, et al. Land Use Regression Model for Ultrafine Particles in Amsterdam. Environ Sci Technol. 2011 Jan 15;45(2):622–8.

83. Henderson SB, Beckerman B, Jerrett M, Brauer M. Application of Land Use Regression to Estimate Long-Term Concentrations of Traffic-Related Nitrogen Oxides and Fine Particulate Matter. Environ Sci Technol. 2007 Apr 1;41(7):2422–8.

84. Beelen R, Hoek G, Vienneau D, Eeftens M, Dimakopoulou K, Pedeli X, et al. Development of NO2 and NOx land use regression models for estimating air pollution exposure in 36 study areas in Europe – The ESCAPE project. Atmos Environ. 2013 Jun;72:10–23.

85. Hagler GSW, Thoma ED, Baldauf RW. High-Resolution Mobile Monitoring of Carbon Monoxide and Ultrafine Particle Concentrations in a Near-Road Environment. J Air Waste Manag Assoc. 2010 Mar 1;60(3):328–36.

86. Tunno BJ, Shields KN, Lioy P, Chu N, Kadane JB, Parmanto B, et al. Understanding intra-neighborhood patterns in PM2.5 and PM10 using mobile monitoring in Braddock, PA. Environ Health. 2012;11:76.

87. Adams MD, DeLuca PF, Corr D, Kanaroglou PS. Mobile Air Monitoring: Measuring Change in Air Quality in the City of Hamilton, 2005–2010. Soc Indic Res. 2012;108(2):351–64.

88. Zwack LM, Hanna SR, Spengler JD, Levy JI. Using advanced dispersion models and mobile monitoring to characterize spatial patterns of ultrafine particles in an urban area. Atmos Environ. 2011 Sep;45(28):4822–9.

89. Stisen S, Sandholt I, Nørgaard A, Fensholt R, Eklundh L. Estimation of diurnal air temperature using MSG SEVIRI data in West Africa. Remote Sens Environ. 2007 Sep 28;110(2):262–74.

90. Statistics Canada Government of Canada. Census Profile, 2016 Census - Greater Vancouver, Regional district [Census division], British Columbia and British Columbia [Province] [Internet]. 2017 [cited 2018 Apr 15]. Available from: http://www12.statcan.gc.ca/census-recensement/2016/dp- pd/prof/details/page.cfm?Lang=E&Geo1=CD&Code1=5915&Geo2=PR&Code2=59&Da ta=Count&SearchText=greater%20vancouver&SearchType=Begins&SearchPR=01&B1= All&TABID=1

91. Statistics Canada. Census geography - Geographic Attribute File [Internet]. 2008 [cited 2016 Mar 11]. Available from: https://www12.statcan.gc.ca/census- recensement/2011/geo/ref/att-eng.cfm

92. Lesack P. Sauder School of Business Custom Demographic Maps [Internet]. 2010 [cited 2014 May 19]. Available from: http://abacus.library.ubc.ca.ezproxy.library.ubc.ca/handle/10573/42375

93. Metro Vancouver. Station Information: Lower Fraser Valley Air Quality Monitoring Network. 2012.

94. Met One Instruments, Inc. MODEL 064-1, 064-2 TEMPERATURE SENSOR OPERATION MANUAL Document No 064-9800. 2005.

95. Metro Vancouver. About Us [Internet]. 2015 [cited 2015 Oct 25]. Available from: http://www.metrovancouver.org/about/Pages/default.aspx

96. NASA. About « Landsat Science [Internet]. Landsat Science- About- Landsat Then and Now. 2014 [cited 2014 Nov 6]. Available from: http://landsat.gsfc.nasa.gov/?page_id=2

97. U.S. Geological Survey. EarthExplorer [Internet]. 2015 [cited 2015 Jul 10]. Available from: http://earthexplorer.usgs.gov/

98. USGS. What are the band designations for the Landsat satellites? [Internet]. 2016 [cited 2016 Sep 18]. Available from: http://landsat.usgs.gov/band_designations_landsat_satellites.php

99. Irons J, Taylor M. Landsat 7 Data Science Users Handbook [Internet]. 2011 [cited 2015 Jul 10]. Available from: http://landsathandbook.gsfc.nasa.gov/

100. U.S. Geological Survey. Using the USGS Landsat 8 Product [Internet]. 2013 [cited 2014 Dec 30]. Available from: http://landsat.usgs.gov/Landsat8_Using_Product.php

101. R Core Team. R: A language and environment for statistical computing [Internet]. Vienna, Austria: R Foundation for Statistical Computing; 2014. Available from: http://www.R-project.org

102. Van de Griend AA, Owe M. On the relationship between thermal emissivity and the normalized difference vegetation index for natural surfaces. Int J Remote Sens. 1993;14(6):1119–1131.

103. Barsi JA, Barker JL, Schott JR. An Atmospheric Correction Parameter Calculator for a single thermal band earth-sensing instrument. In IEEE; 2003 [cited 2015 Aug 11]. p. 3014–6 vol.5. Available from: http://ieeexplore.ieee.org/articleDetails.jsp?arnumber=1294665

104. Markham B, Barsi J. Atmospheric Correction Parameter Calculator [Internet]. 2014 [cited 2015 Aug 11]. Available from: http://atmcorr.gsfc.nasa.gov/ 63

105. U.S. Geological Survey. SLC-off Products: Background [Internet]. 2015 [cited 2016 Jan 7]. Available from: http://landsat.usgs.gov/products_slcoffbackground.php

106. USGS. Landsat 8 (L8) Data Users Handbook [Internet]. 2016 [cited 2016 Jun 24]. Available from: http://landsat.usgs.gov/l8handbook_appendixa.php

107. Gislason PO, Benediktsson JA, Sveinsson JR. Random Forests for land cover classification. Pattern Recognit Lett. 2006 Mar;27(4):294–300.

108. Esri. ArcGIS 10.2. Redlands, CA: Environmental Science Research Institute; 2014.

109. Metro Vancouver. Metro Vancouver Weather Station Data: May-Sept 2014. 2014.

110. Straka JM, Rasmussen EN, Fredrickson SE. A Mobile Mesonet for Finescale Meteorological Observations. J Atmospheric Ocean Technol. 1996 Oct 1;13(5):921–36.

111. Hedquist BC, Brazel AJ. Urban, Residential, and Rural Climate Comparisons from Mobile Transects and Fixed Stations: Phoenix, Arizona. J Ariz-Nev Acad Sci. 2006;38(2):77–87.

112. Kaiser R, Le Tertre A, Schwartz J, Gotway CA, Daley WR, Rubin CH. The Effect of the 1995 Heat Wave in Chicago on All-Cause and Cause-Specific Mortality. Am J Public Health. 2007 Apr;97(Suppl 1):S158–62.

113. Shaposhnikov D, Revich B, Bellander T, Bedada GB, Bottai M, Kharkova T, et al. Mortality Related to Air Pollution with the Moscow Heat Wave and Wildfire of 2010. Epidemiol Camb Mass. 2014 May;25(3):359–64.

114. Barriopedro D, Fischer EM, Luterbacher J, Trigo RM, García-Herrera R. The Hot Summer of 2010: Redrawing the Temperature Record Map of Europe. Science. 2011 Apr 8;332(6026):220–4.

115. Schmid HP. Footprint modeling for vegetation atmosphere exchange studies: a review and perspective. Agric For Meteorol. 2002 Dec 2;113(1–4):159–83.

116. Schmid HP. Experimental design for flux measurements: matching scales of observations and fluxes. Agric For Meteorol. 1997 Nov 15;87(2):179–200.

117. Nichol JE, Wong MS. Spatial variability of air temperature and appropriate resolution for satellite‐derived air temperature estimation. Int J Remote Sens. 2008 Dec 1;29(24):7213– 23.

118. Tsin PK, Knudby A, Krayenhoff ES, Ho HC, Brauer M, Henderson SB. Microscale mobile monitoring of urban air temperature. Urban Clim. 2016 Dec;18:58–72.

119. Jackson T. Vegetation water content mapping using Landsat data derived normalized difference water index for corn and soybeans. Remote Sens Environ. 2004 Sep;92(4):475–82. 64

120. Hodul M, Knudby A, Ho HC. Estimation of Continuous Urban Sky View Factor from Landsat Data Using Shadow Detection. Remote Sens. 2016 Jul 6;8(7):568.

121. DMTI Spatial Inc. CanMap 2013.

122. DMTI Spatial Inc. CanMap Water, v2010.3, 2010 - Abacus Dataverse Network [Internet]. 2011 [cited 2017 Apr 2]. Available from: http://hdl.handle.net.ezproxy.library.ubc.ca/11272/HITUI

123. DMTI Spatial Inc. Digital Elevation Models Area 92G, [2002] [Internet]. 2009 [cited 2017 Apr 2]. Available from: http://hdl.handle.net.ezproxy.library.ubc.ca/11272/YUFSQ

124. Brauer M, Henderson SB, Marshall J. A land use regression road map for the Burrard Inlet Area Local Air Quality Study. 2008 Jun 9 [cited 2016 Sep 15]; Available from: https://open.library.ubc.ca/cIRcle/collections/facultyresearchandpublications/304/items/1. 0048202

125. Brenning A. Spatial cross-validation and bootstrap for the assessment of prediction rules in remote sensing: The R package sperrorest. In: 2012 IEEE International Geoscience and Remote Sensing Symposium. 2012. p. 5372–5.

126. Basagaña X, Rivera M, Aguilera I, Agis D, Bouso L, Elosua R, et al. Effect of the number of measurement sites on land use regression models in estimating local air pollution. Atmos Environ. 2012 Jul;54:634–42.

127. Hastie T, Tibshirani R, Friedman J. Springer Series in Statistics : Elements of Statistical Learning : Data Mining, Inference, and Prediction (2) [Internet]. New York, NL: Springer New York; 2009 [cited 2016 Sep 15]. Available from: http://site.ebrary.com/lib/alltitles/docDetail.action?docID=10289757

128. Abernethy R. A land use regression model for ultrafine particles in Vancouver, Canada [Internet]. University of British Columbia; 2012 [cited 2016 Dec 15]. Available from: https://open.library.ubc.ca/cIRcle/collections/ubctheses/24/items/1.0072650

129. Brauer M, Hoek G, van Vliet P, Meliefste K, Fischer P, Gehring U, et al. Estimating long- term average particulate air pollution concentrations: application of traffic indicators and geographic information systems. Epidemiol Camb Mass. 2003 Mar;14(2):228–39.

130. Muller KE, Peterson BL. Practical methods for computing power in testing the multivariate general linear hypothesis. Comput Stat Data Anal. 1984 Aug 1;2(2):143–58.

131. Gromping U. Relative Importance for Linear Regression in R: The Package relaimpo. J Stat Softw [Internet]. 2006 Sep 1 [cited 2017 Dec 7];17(1). Available from: https://doaj.org

132. Lindeman RH, Merenda PF, Gold RZ. Introduction to bivariate and multivariate analysis. Glenview, Ill: Scott, Foresman; 1980. 444 p.

133. Kruskal W. Relative Importance by Averaging Over Orderings. Am Stat. 1987;41(1):6– 10.

134. Hong KY, Tsin PK, van den Bosch M, Brauer M, Henderson SB. Urban greenness extracted from pedestrian video and its relationship with surrounding air temperatures. Manuscr Prep. 2018;

Appendices

Appendix A - Supplemental Tables

Table A-1: Summary statistics for each of the 42 mobile air temperature monitoring runs and their corresponding nearest fixed site. The run number indicates the route that was monitored, while the letter indicates the replicate of that route. The r column shows the run correlation between the mobile monitoring air temperature data and the time-matched fixed site data from the nearest station.

Mean Route Mean Fixed Difference in Route Route Distance from Date Start Time Run Temp in °C Site Temp Temp Means r Area Length Nearest Fixed (dd/mm/yyyy) Time (Min) (SD) in °C (SD) in °C (SD) (km²) (km) Station (km) 1A 02/07/2014 4:32 PM 25.2 (0.5) 24.3 (0.6) 0.9 (0.8) -0.09 0.8 6.7 9.1 113 1B 24/08/2014 3:40 PM 22.8 (0.5) 23.0 (0.3) -0.2 (0.6) -0.15 0.8 6.7 9.1 114 2A 17/06/2014 3:41 PM 20.6 (0.5) 18.7 (0.3) 1.9 (0.6) -0.02 0.8 7.4 8.0 127 2B 27/07/2014 3:55 PM 25.1 (0.4) 24.4 (0.6) 0.7 (0.6) 0.28 0.8 7.4 8.0 116 3A 11/07/2014 3:52 PM 27.5 (1.0) 26.8 (1.1) 0.8 (1.8) -0.43 0.6 7.2 6.2 127 3B 08/08/2014 3:10 PM 24.2 (0.4) 22.3 (0.2) 2.0 (0.5) -0.02 0.6 7.2 6.2 122 4A 27/05/2014 3:08 PM 18.7 (0.4) 16.7 (0.4) 2.0 (0.5) 0.23 0.6 8.2 3.8 142 4B 15/07/2014 3:25 PM 27.2 (0.5) 26.9 (0.6) 0.2 (0.6) 0.50 0.6 8.2 3.8 138 4C 27/08/2014 3:45 PM 25.5 (0.4) 25.4 (0.7) 0.1 (0.9) -0.31 0.6 8.2 3.8 150 5A 26/06/2014 3:02 PM 23.4 (0.6) 21.8 (0.4) 1.6 (0.4) 0.62 0.6 7.2 4.5 133 5B 04/08/2014 4:01 PM 28.9 (0.9) 27.5 (0.3) 1.4 (0.9) 0.14 0.6 7.2 4.5 121 6A 21/06/2014 3:34 PM 21.3 (0.5) 18.5 (0.3) 2.8 (0.6) -0.03 0.5 6.4 6.3 101 6B 01/08/2014 3:48 PM 27.3 (0.5) 26.0 (0.3) 1.3 (0.6) -0.43 0.5 6.4 6.1 111 7A 01/07/2014 3:57 PM 26.9 (1.5) 26.9 (1.7) 0.1 (0.5) 0.97 0.8 7.9 0.6 127 7B 19/08/2014 3:25 PM 24.6 (0.4) 23.9 (0.2) 0.7 (0.4) 0.29 0.8 7.9 0.6 129 8A 09/07/2014 4:02 PM 24.6 (0.5) 21.8 (0.3) 2.8 (0.5) 0.36 0.8 7.0 3.8 113 8B 21/07/2014 3:45 PM 22.8 (0.9) 21.5 (0.5) 1.3 (0.7) 0.56 0.8 7.0 3.9 124 9A 05/06/2014 3:41 PM 20.4 (0.8) 17.3 (0.3) 3.1 (0.9) -0.28 0.5 7.2 9.7 129 9B 31/07/2014 3:40 PM 27.0 (0.7) 23.9 (0.2) 3.1 (0.7) -0.11 0.5 7.2 9.7 122 10A 30/06/2014 4:00 PM 24.0 (0.3) 21.4 (0.3) 2.5 (0.4) 0.09 1.3 7.8 7.8 105 10B 03/08/2014 3:45 PM 28.1 (0.5) 26.0 (0.2) 2.1 (0.6) -0.37 1.3 7.8 7.8 103 11A 02/06/2014 4:05 PM 22.9 (0.4) 20.9 (0.2) 2.0 (0.4) 0.24 0.9 9.0 3.8 136 11B 30/07/2014 3:50 PM 27.8 (0.4) 26.5 (0.2) 1.3 (0.5) 0.03 0.9 9.0 3.8 113 12A 28/05/2014 4:10 PM 17.5 (0.7) 17.0 (0.2) 0.5 (0.7) 0.07 0.9 7.5 5.9 109 12B 29/07/2014 3:36 PM 26.9 (0.5) 26.4 (0.2) 0.5 (0.6) -0.24 0.9 7.5 5.9 106 12C 06/09/2014 4:19 PM 26.8 (0.4) 24.0 (0.4) 2.8 (0.7) -0.59 0.9 7.5 5.9 110 13A 10/07/2014 3:35 PM 27.7 (0.4) 25.9 (0.3) 1.8 (0.5) -0.08 1.5 8.8 5.0 131 13B 07/08/2014 3:48 PM 24.5 (0.5) 22.6 (0.5) 1.9 (0.5) 0.51 1.5 8.8 5.0 130 14A 07/07/2014 3:32 PM 26.2 (0.7) 24.5 (0.7) 1.7 (0.8) 0.36 1.8 9.2 2.3 165 14B 26/08/2014 3:49 PM 27.1 (0.6) 26.5 (0.3) 0.7 (0.5) 0.47 1.8 9.2 2.3 134 15A 03/06/2014 4:19 PM 20.1 (0.3) 18.5 (0.3) 1.6 (0.4) -0.04 0.7 7.8 3.1 140 15B 28/07/2014 3:46 PM 26.9 (0.4) 24.6 (0.5) 2.4 (0.8) -0.30 0.7 7.8 3.1 154 16A 06/06/2014 3:56 PM 21.2 (0.5) 18.0 (0.3) 3.2 (0.6) -0.05 2.5 9.2 2.8 129 16B 26/07/2014 3:33 PM 24.9 (0.6) 22.6 (0.6) 2.3 (0.8) 0.17 2.5 9.2 2.6 125 17A 30/05/2014 4:02 PM 20.0 (0.4) 18.6 (0.3) 1.4 (0.3) 0.63 2.4 9.9 2.6 124 17B 14/07/2014 4:11 PM 27.9 (0.5) 24.6 (0.3) 3.3 (0.4) 0.41 2.4 9.9 2.6 148 18A 09/06/2014 3:55 PM 19.8 (0.4) 18.4 (0.2) 1.4 (0.4) 0.24 1.9 7.6 5.9 114 18B 13/07/2014 4:01 PM 31.8 (1.0) 27.0 (0.3) 4.8 (0.9) 0.53 1.9 7.6 5.8 112 19A 08/07/2014 4:06 PM 25.7 (0.8) 25.4 (0.4) 0.3 (0.6) 0.69 0.6 7.1 6.9 109 19B 17/08/2014 3:35 PM 25.3 (0.5) 24.2 (0.2) 1.0 (0.5) 0.22 0.6 7.1 7.0 114 20A 25/06/2014 3:50 PM 22.2 (0.6) 21.6 (0.2) 0.6 (0.6) 0.32 0.8 6.9 8.7 112 20B 11/08/2014 3:46 PM 27.1 (0.6) 29.4 (0.3) -2.2 (0.5) 0.46 0.8 6.9 8.7 111 --- Average 3:47 PM 24.7 23.1 1.5 0.14 1.1 7.8 5.3 124

Table A-2: Z-score differential (GVHM z-score minus mobile air temperature data z-score) distribution for raster cell values from overall sampling campaign and all runs. The run number indicates the route that was monitored, while the letter indicates the replicate of that route. 33 runs had over 50% of raster cells with z-scores between -1 and 1.

Run z > 1 (%) z < - 1 (%) -1 < z < 1 (%) All Runs 23.3 19.3 57.4 1A 17.3 14.6 68.1 1B 16.9 18.8 64.4 2A 17.5 15.3 67.2 2B 29.6 23.6 46.8 3A 20.8 24.7 54.5 3B 17.7 16.1 66.3 4A 26.2 25.4 48.4 4B 17.6 25.4 57.0 4C 26.8 26.4 46.8 5A 21.8 23.0 55.2 5B 23.4 20.2 56.5 6A 18.6 21.7 59.7 6B 29.2 30.5 40.3 7A 16.3 12.2 71.5 7B 19.2 15.9 64.9 8A 32.9 27.9 39.2 8B 32.5 30.5 37.0 9A 21.3 25.0 53.7 9B 20.3 21.5 58.2 10A 16.6 30.9 52.5 10B 4.7 21.0 74.3 11A 18.8 31.4 49.8 11B 13.8 23.3 62.9 12A 19.5 25.8 54.7 12B 16.1 19.1 64.8 12C 15.9 18.6 65.5 13A 12.3 26.3 61.4 13B 16.9 30.3 52.9 14A 17.9 12.5 69.6 14B 19.9 14.7 65.3 15A 13.1 29.9 56.9 15B 16.4 29.1 54.5 16A 25.8 32.3 41.9 16B 23.3 26.6 50.2 17A 18.3 25.8 55.9 17B 18.5 25.6 55.9 18A 19.0 27.2 53.8 18B 5.6 17.9 76.5 19A 23.3 25.3 51.4 19B 20.9 23.8 55.3 20A 19.7 20.9 59.4 20B 13.6 20.4 66.0

Table A-3: Descriptive statistics for all individual route models, including variable coefficients, variable percentage contribution to r2, model r2, and spatial leave-10-out-cross-validation r2. Best models are in bold, and the air temperature data from these models were used to build the overall model. Distance to Large Water Body, 200m Buffer Distance to Major Road NDWI, 100m Buffer SVF, 200m Buffer Spatial Route Percentage Percentage Percentage Percentage Model r² Leave 10 Variable Contribution Variable Contribution Variable Contribution Variable Contribution Out CV r² Coefficient to r² Coefficient to r² Coefficient to r² Coefficient to r² 1A - - -0.00193 0.50 -5.63 0.50 - - 0.32 0.29 1B 0.00102 0.25 -0.00103 0.20 -2.73 0.12 -2.23 0.44 0.43 0.36 2A - - - - -2.72 0.74 -2.74 0.26 0.08 0.01 2B ------5.60 1.00 0.12 0.08 3A 0.00422 0.43 - - - - 9.14 0.57 0.28 0.20 3B 0.00124 0.48 - - - - -2.57 0.53 0.26 0.20 4A - - -0.00093 0.36 -2.65 0.64 - - 0.35 0.30 4B 0.00088 0.52 - - -2.61 0.25 3.14 0.23 0.23 0.14 4C - - - - -3.73 0.43 - - 0.43 0.39 5A - - -0.00082 1.00 - - - - 0.02 0.00 5B - - - - -1.64 1.00 - - 0.01 0.00 6A 0.00045 1.00 ------0.03 0.00 6B 0.00102 0.85 - - -2.17 0.13 3.29 0.02 0.34 0.25 7A 0.00107 0.42 -0.00253 0.20 -8.19 0.07 36.65 0.32 0.54 0.46 7B 0.00139 0.91 - - - - 5.89 0.09 0.42 0.40 8A - - - - -5.04 1.00 - - 0.26 0.24 8B 0.00316 0.66 -0.00220 0.24 - - -25.37 0.09 0.37 0.31 9A 0.00326 0.96 - - -4.91 0.04 - - 0.69 0.67 9B 0.00066 0.36 - - - - -4.06 0.64 0.39 0.24 10A 0.00053 0.83 - - - - -3.20 0.17 0.25 0.11 10B ------6.86 1.00 0.15 0.05 11A - - -0.00110 0.12 - - 9.00 0.88 0.23 0.12 11B - - - - -3.50 0.69 -2.92 0.31 0.29 0.18 12A - - - - -5.52 0.88 3.61 0.12 0.18 0.08 12B 0.00166 0.33 - - -6.61 0.55 9.36 0.12 0.61 0.53 12C 0.00068 0.22 - - -3.34 0.41 -6.02 0.38 0.49 0.39 13A - - -0.00056 0.10 - - 4.02 0.90 0.28 0.22 13B 0.00107 0.89 - - - - -3.86 0.11 0.30 0.23 14A - - -0.00210 0.26 -5.07 0.74 - - 0.46 0.41 14B 0.00056 0.23 - - - - -5.48 0.77 0.36 0.30 15A 0.00075 0.80 -0.00130 0.20 - - - - 0.48 0.45 15B 0.00113 1.00 ------0.26 0.22 16A - - - - -2.19 0.28 3.95 0.72 0.34 0.23 16B 0.00128 1.00 ------0.44 0.38 17A 0.00012 1.00 ------0.04 0.00 17B 0.00073 0.73 - - -3.65 0.16 1.13 0.11 0.50 0.42 18A - - -0.00089 0.40 - - -3.91 0.60 0.18 0.06 18B - - - - -3.31 1.00 - - 0.02 0.08 19A - - -0.00164 0.11 -10.42 0.80 10.09 0.09 0.55 0.42 19B 0.00071 0.26 -0.00092 0.16 -2.24 0.58 - - 0.26 0.19 20A 0.00187 0.41 -0.00310 0.20 -5.56 0.39 - - 0.40 0.31 20B 0.00241 0.81 - - -7.31 0.19 - - 0.46 0.42

Appendix B - Supplemental Figures

Figure A-1: Scatterplots and correlation coefficients (r) of z-score differentials (GVHM z-score minus mobile air temperature data z-score) between all replicate runs of each route. Each scatterplot represents two replicate measurements of a route, with the replicate z-scores represented on the x-axes and y-axes.

Figure A-2: Map of the 50m buffer Normalized Difference Water Index (NDWI) data used to build the individual route models and the overall model.

Figure A-3: Map of the 50m buffer Normalized Difference Vegetation Index (NDVI) data considered for inclusion into the individual route models and the overall model.

Figure A-4: Map of the 200m buffer Sky-View Factor (SVF) data used to build the individual route models and the overall model.

Figure A-5: Map of the unbuffered Distance to Major Roads data used to build the individual route models and the overall model. The road network of the region is clearly visible in the map.

Figure A-6: Map of the 200m buffer Distance to Large Water Body data used to build the individual route models and the overall model.

Figure A-7: Map of the 200m buffer Elevation data considered for inclusion into the individual route models and the overall model.

Appendix C - Wind Speed and Direction Analysis, YVR Station

Time Series of WS and WD for 27/05/2014

350

250

(km/h)

Speed 150

Degrees

00:00 05:00 10:00 15:00 20:00 Time

Time Series of WS and WD for 28/05/2014

300

15 (km/h)

200

Speed 10

Degrees

100

0 01:00 06:00 11:00 16:00 21:00 Time

Time Series of WS and WD for 30/05/2014

300

15 (km/h)

200

Speed 10

Degrees

100

0 03:00 08:00 13:00 18:00 23:00 Time

Time Series of WS and WD for 02/06/2014

300

(km/h)

200

Speed

Degrees 10

100

0 01:00 06:00 11:00 16:00 21:00 Time

Time Series of WS and WD for 03/06/2014

250

20 (km/h)

Speed

Degrees 150

50 02:00 07:00 12:00 17:00 22:00 Time

Time Series of WS and WD for 05/06/2014

300

(km/h)

200

Speed

Degrees

100 5

0 04:00 09:00 14:00 19:00 00:00 Time

Time Series of WS and WD for 06/06/2014

300

(km/h)

200 10

Speed

Degrees

100

0 00:00 05:00 10:00 15:00 20:00 Time

Time Series of WS and WD for 09/06/2014

350

250 30

(km/h)

20 Speed

Degrees 150

0 03:00 08:00 13:00 18:00 23:00 Time

Time Series of WS and WD for 17/06/2014

150

(km/h) 15

Speed

Degrees

50 5

0 01:00 06:00 11:00 16:00 21:00 Time

Time Series of WS and WD for 21/06/2014

300

(km/h)

200

Speed

Degrees

100 5

0 00:00 05:00 10:00 15:00 20:00 Time

Time Series of WS and WD for 25/06/2014

300

(km/h)

200

Speed

Degrees

100 5

0 04:00 09:00 14:00 19:00 00:00 Time

Time Series of WS and WD for 26/06/2014

150

(km/h)

Speed

Degrees 100

50 00:00 05:00 10:00 15:00 20:00 Time

Time Series of WS and WD for 30/06/2014

300

15 (km/h)

200

10 Speed

Degrees

100

0 04:00 09:00 14:00 19:00 00:00 Time

Time Series of WS and WD for 01/07/2014

300

15 (km/h)

200

Speed 10

Degrees

100

0 00:00 05:00 10:00 15:00 20:00 Time

Time Series of WS and WD for 02/07/2014

300

(km/h)

200 15

Speed

Degrees

100

0 0 01:00 06:00 11:00 16:00 21:00 Time

Time Series of WS and WD for 07/07/2014

300

(km/h)

200

Speed

Degrees

100

0 0 01:00 06:00 11:00 16:00 21:00 Time

Time Series of WS and WD for 08/07/2014

300

(km/h)

200

Speed

Degrees

100 5

0 02:00 07:00 12:00 17:00 22:00 Time

Time Series of WS and WD for 09/07/2014

320 30

280

(km/h) 20

Speed

Degrees

240

5 200 03:00 08:00 13:00 18:00 23:00 Time

Time Series of WS and WD for 10/07/2014

300

(km/h)

200

Speed

Degrees 10

100

0 04:00 09:00 14:00 19:00 00:00 Time

Time Series of WS and WD for 11/07/2014

300

(km/h)

200 10

Speed

Degrees

100 5

0 00:00 05:00 10:00 15:00 20:00 Time

Time Series of WS and WD for 13/07/2014

300

(km/h)

200

Speed

Degrees 10

100

0 02:00 07:00 12:00 17:00 22:00 Time

Time Series of WS and WD for 14/07/2014

300

(km/h)

200 10

Speed

Degrees

100

0 03:00 08:00 13:00 18:00 23:00 Time

Time Series of WS and WD for 15/07/2014

250

(km/h)

Speed

150

Degrees

50 04:00 09:00 14:00 19:00 00:00 Time

Time Series of WS and WD for 21/07/2014

200

(km/h)

Speed

Degrees

100 5

50 00:00 05:00 10:00 15:00 20:00 Time

Time Series of WS and WD for 26/07/2014

00 20

15 (km/h)

200

10 Speed

Degrees

100

0 00:00 05:00 10:00 15:00 20:00 Time

Time Series of WS and WD for 27/07/2014

300

(km/h)

200 10

Speed

Degrees

100

0 01:00 06:00 11:00 16:00 21:00 Time

Time Series of WS and WD for 28/07/2014

300

(km/h)

10 200

Speed

Degrees

100

0 02:00 07:00 12:00 17:00 22:00 Time

Time Series of WS and WD for 29/07/2014

300

(km/h)

200

Speed

Degrees

100

0 03:00 08:00 13:00 18:00 23:00 Time

Time Series of WS and WD for 30/07/2014

300

(km/h)

10 200

Spe

Degrees

100

0 04:00 09:00 14:00 19:00 00:00 Time

Time Series of WS and WD for 31/07/2014

350

250

15 (km/h)

10 Speed

Degrees

150

50 00:00 05:00 10:00 15:00 20:00 Time

Time Series of WS and WD for 01/08/2014

300

(km/h) 10

200

Speed

Degrees

100

0 01:00 06:00 11:00 16:00 21:00 Time

Time Series of WS and WD for 03/08/2014

300

(km/h)

200

Speed

Degrees

5 100

0 03:00 08:00 13:00 18:00 23:00 Time

Time Series of WS and WD for 04/08/2014

300

(km/h)

10 200

Speed

Degrees

100

0 04:00 09:00 14:00 19:00 00:00 Time

Time Series of WS and WD for 07/08/2014

300

200 (km/h)

Speed

Degrees

100

0 02:00 07:00 12:00 17:00 22:00 Time

Time Series of WS and WD for 08/08/2014

300

(km/h)

200

Speed

Degrees

100

0 03:00 08:00 13:00 18:00 23:00 Time

Time Series of WS and WD for 11/08/2014

300

(km/h) 10

200

Speed

Degrees

100

0 01:00 06:00 11:00 16:00 21:00 Time

Time Series of WS and WD for 17/08/2014

300

(km/h)

200

Speed

Degrees

5 100

0 02:00 07:00 12:00 17:00 22:00 Time

Time Series of WS and WD for 19/08/2014

300

(km/h)

200 10

Speed

Degrees

5 100

0 04:00 09:00 14:00 19:00 00:00 Time

Time Series of WS and WD for 24/08/2014

300

(km/h)

200 10

Speed

Degrees

100 5

0 04:00 09:00 14:00 19:00 00:00 Time

Time Series of WS and WD for 26/08/2014

300

(km/h)

200 10

Speed

Degrees

5 100

0 01:00 06:00 11:00 16:00 21:00 Time

Time Series of WS and WD for 24/08/2014

250

15 (km/h)

150

Speed

Degrees

50 02:00 07:00 12:00 17:00 22:00 Time

Time Series of WS and WD for 06/09/2014

300

10 (km/h)

200

Speed 6

Degrees

100

0 02:00 07:00 12:00 17:00 22:00 Time