UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Spatial assessment of forest fire smoke exposure and its health impacts in Southeastern British Columbia… Henderson, Sarah B. 2009

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

Item Metadata

Download

Media
24-ubc_2009_fall_henderson_sarah.pdf [ 4.1MB ]
Metadata
JSON: 24-1.0067764.json
JSON-LD: 24-1.0067764-ld.json
RDF/XML (Pretty): 24-1.0067764-rdf.xml
RDF/JSON: 24-1.0067764-rdf.json
Turtle: 24-1.0067764-turtle.txt
N-Triples: 24-1.0067764-rdf-ntriples.txt
Original Record: 24-1.0067764-source.json
Full Text
24-1.0067764-fulltext.txt
Citation
24-1.0067764.ris

Full Text

 SPATIAL ASSESSMENT OF FOREST FIRE SMOKE EXPOSURE AND ITS PUBLIC HEALTH IMPACTS IN SOUTHEASTERN BRITISH COLUMBIA DURING THE SUMMER OF 2003   by  Sarah B. Henderson  B.A.Sc, The University of British Columbia, 2000    A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF   DOCTOR OF PHILOSOPHY   in   THE FACULTY OF GRADUATE STUDIES   (Health Care and Epidemiology)      THE UNIVERSITY OF BRITISH COLUMBIA  (Vancouver)  October 2009        © Sarah Henderson, 2009  ii ABSTRACT Forest fires are a significant source of episodic air pollution resulting in elevated ambient concentrations of inhalable particulate matter (PM).  Although PM from fossil fuel combustion has been conclusively associated with respiratory and cardiovascular morbidity and mortality, the health effects of fire-related PM are not clearly understood. Air quality monitoring is sparse in many fire-affected areas, so it is challenging to apply epidemiologic methods that require individual-level exposure assessment.  Data from dispersion models and remote sensors are spatially extensive and may provide viable exposure estimation alternatives.  Firestorms across southeastern British Columbia during the summer of 2003 produced a unique opportunity to compare rigorous epidemiologic results based on new exposure assessment methods to those based on air quality monitoring data.  A population-based cohort of ~280 000 subjects was identified from administrative health data and three daily smoke exposure estimates were assigned for each individual according to residential location: TEOM averaged PM concentrations measured by the nearest of six air quality monitors; SMOKE indicated the presence of a plume over the area in satellite imagery; and CALPUFF averaged PM concentrations estimated by a dispersion model.  The latter was initialized and run for this project using remote sensing data to simplify the model as much as possible.  For example, emissions were calculated with the radiative power of satellite-detected fires and were comparable to those estimated by much more complex methods.  Overall performance of the model was moderate when evaluated using PM measurements, satellite imagery and atmospheric aerosol measurements.  Longitudinal logistic regression was used to examine the independent effects of each exposure over the 92-day study period. Respiratory outcomes were associated with smoke-related PM, but no cardiovascular effects were detected.  While odds ratios for the TEOM metric were consistent with other reports, those for the CALPUFF metric were biased towards the null.  Results for SMOKE tracked with those for TEOM, but with much wider confidence intervals.  This study (1) highlights the potential of new smoke exposure assessment methods, (2) demonstrates that plume dispersion models can be simplified with remote sensing data, and (3) confirms the respiratory health effects of forest fire smoke.  iii TABLE OF CONTENTS Abstract ........................................................................................................................... ii Table of contents ............................................................................................................ iii List of tables ................................................................................................................... vi List of figures ................................................................................................................. vii Acknowledgements........................................................................................................viii Dedication....................................................................................................................... ix Co-authorship statement ................................................................................................. x 1 Chapter 1: Introduction and literature review .........................................1 1.1 Overview...............................................................................................................1 1.2 Background...........................................................................................................1 1.2.1 Introduction ....................................................................................................1 1.2.2 Study area .....................................................................................................2 1.2.3 Air quality during the study period..................................................................3 1.2.4 Study population and health data ..................................................................6 1.3 Rationale and objectives.......................................................................................6 1.3.1 Chapter 2: Estimating forest fire properties from remote sensing data ..........8 1.3.2 Chapter 3: Dispersion modeling for smoke exposure assessment ................8 1.3.3 Chapter 4: Health effects of forest fire smoke................................................9 1.4 Review of the pertinent forest fire smoke literature ...............................................9 1.4.1 Particulate matter in forest fire smoke............................................................9 1.4.2 Smoke plume dispersion modeling ..............................................................10 1.4.3 Remote sensing of fires and smoke by MODIS ...........................................11 1.4.4 Smoke exposure assessment for public health studies ...............................12 1.5 Review of the pertinent health literature..............................................................14 1.5.1 Health effects of urban particulate matter ....................................................14 1.5.2 Comprehensive review on the health effects of woodsmoke .......................15 1.5.3 Physiologic effects of woodsmoke ...............................................................16 1.5.4 Epidemiologic studies on the population effects of forest fire smoke...........17 1.6 On the use of administrative data for health research.........................................19 1.7 Summary.............................................................................................................19 1.8 References..........................................................................................................21 2 Chapter 2: Validity and utility of MODIS data for simple estimation of area burned and aerosols emitted by wildfire events .................................29 2.1 Introduction .........................................................................................................29 2.2 Methods ..............................................................................................................30 2.2.1 MODIS fire detection data............................................................................30 2.2.2 Fire event identification ................................................................................31 2.2.3 Burned area estimation................................................................................32 2.2.4 Aerosol emissions rate estimation ...............................................................32  iv 2.3 Results ................................................................................................................34 2.3.1 Fire event identification ................................................................................34 2.3.2 Burned area estimation................................................................................36 2.3.3 Aerosol emissions........................................................................................39 2.4 Discussion and conclusions ................................................................................41 2.5 References..........................................................................................................43 3 Chapter 3: Use of MODIS products to simplify and evaluate a forest fire plume dispersion model for PM10 exposure assessment ...........................47 3.1 Introduction .........................................................................................................47 3.2 Methods ..............................................................................................................49 3.2.1 The CALMET/CALPUFF modeling system..................................................49 3.2.2 Meteorological inputs ...................................................................................50 3.2.3 Fire location, size and emissions rates ........................................................50 3.2.4 CAPLUFF PM10 concentration vs. TEOM measurements ...........................52 3.2.5 CALPUFF PM10 concentration estimates vs. MODIS AOT ..........................52 3.2.6 CALPUFF plume areas vs. MODIS true color images .................................53 3.3 Results and discussion .......................................................................................54 3.3.1 CALPUFF PM10 concentration estimates vs. TEOM measurements ...........54 3.3.2 CALPUFF PM10 concentration estimates vs. MODIS AOT ..........................58 3.3.3 CALPUFF vs. MODIS plumes......................................................................60 3.4 Conclusions ........................................................................................................62 3.5 References..........................................................................................................63 4 Chapter 4: Three measures of forest fire smoke exposure and their association with respiratory and cardiovascular health outcomes in a dynamic population-based cohort ..............................................................66 4.1 Introduction .........................................................................................................66 4.2 Methods ..............................................................................................................68 4.2.1 Administrative health data............................................................................68 4.2.2 Cohort identification and geolocation...........................................................68 4.2.3 Event definition ............................................................................................69 4.2.4 Exposure metrics .........................................................................................71 4.2.5 Other covariates ..........................................................................................72 4.2.6 Statistical analyses ......................................................................................72 4.3 Results ................................................................................................................73 4.3.1 Data summary .............................................................................................73 4.3.2 Adjustment and time lags.............................................................................74 4.3.3 Stratification by age and previous physician visits .......................................78 4.3.4 Specific diagnoses.......................................................................................81 4.4 Discussion...........................................................................................................82 4.5 References..........................................................................................................86    v 5 Chapter 5: Retrospective .....................................................................89 5.1 Review ................................................................................................................89 5.2 Conception (2003 – 2004) ..................................................................................91 5.3 Wind modeling (2004 – 2005) .............................................................................91 5.4 Emissions modeling (2005 – 2006).....................................................................93 5.5 Smoke dispersion modeling (2006 – 2007).........................................................96 5.6 Exposure assignment (2007 – 2008) ................................................................100 5.7 Epidemiology (2008 – 2009) .............................................................................104 5.8 Conclusions ......................................................................................................110 5.9 References........................................................................................................112 Appendix 1: Ethics certificate ...................................................................115 Appendix 2: Sample CALPUFF parameter file.........................................116 Appendix 3: Event counts for the summers of 2002, 2003 and 2004.......122 Appendix 4: CALPUFF estimates from different emissions models .........124 Appendix 5: Notes on cohort identification ...............................................129 Appendix 6: Notes on cohort geolocation.................................................131 Appendix 7: ADDRESS vs. POSTAL results............................................137  vi LIST OF TABLES   Table 2.1 – Diurnal distribution of fire emissions ...........................................................33 Table 2.2 – Summary of fire size comparison................................................................35  Table 3.1 – Summary of modeled fires ..........................................................................55 Table 3.2 – Summary of CALPUFF versus TEOM results.............................................55 Table 3.3 – AOT measurements versus CALPUFF and TEOM.....................................59 Table 3.4 – Summary of plume evaluation results .........................................................61  Table 4.1 – Summary of outcome rates.........................................................................75 Table 4.2 – Overall cohort results for TEOM exposure..................................................80 Table A4.1 – The CANFB and USEPM fuel classification systems .............................125 Table A6.1 – A sample of raw and cleaned addresses for geolocation .......................131  vii LIST OF FIGURES     Figure 1.1 – Fires in the study area .................................................................................3 Figure 1.2 – Smoke in the study area ..............................................................................4 Figure 1.3 – Measurements of particulate matter in the study area.................................5 Figure 1.4 – Satellite image of fire near Kelowna ............................................................7  Figure 2.1 – Fires in the study area ...............................................................................36 Figure 2.2 – Comparison between measured and estimated fire areas.........................37 Figure 2.3 – Correspondence between MODIS and AWIS fire detection ......................38 Figure 2.4 – The relationship between elevation and fire detection...............................39 Figure 2.5 – Scatter plots comparing emissions models................................................40  Figure 3.1 – Plume modeling domains. .........................................................................54 Figure 3.2 – Time series of CALPUFF estimates. .........................................................56  Figure 4.1 – The cohort selection process.....................................................................70 Figure 4.2 – CALPUFF concentrations vs. distance-to-fire............................................74 Figure 4.3 – ORs after adjustment and distance constraint. ..........................................76 Figure 4.4 – ORs with exposure lagged by 0, 1 and a combination of 0 and 1..............77 Figure 4.5 – ORs for physician visits stratified by age. ..................................................79 Figure 4.6 – ORs for hospital admissions stratified by age............................................80 Figure 4.7 – Age-stratified results for specific respiratory diagnoses.............................81  Figure 5.1 – Particulate matter measured at TEOM and home locations ......................92 Figure 5.2 – Potential misclassification when modeling smoke plumes.........................98 Figure 5.3 – Correlation between CALPUFF and AOT. ...............................................100 Figure 5.4 – Geolocation results. .................................................................................103 Figure A3.1 – Daily count of respiratory physician visits..............................................122 Figure A3.2 – Daily count of cardiovascular physician visits........................................122 Figure A3.3 – Daily count of respiratory hospital admissions ......................................123 Figure A3.4 – Daily count of cardiovascular hospital admissions ................................123 Figure A4.1 – Three CALPUFF estimates vs. TEOM measurements at Kamloops.....126 Figure A4.2 – Three CALPUFF estimates vs. TEOM measurements at Kelowna .......126 Figure A4.3 – Three CALPUFF estimates vs. TEOM measurements at Vernon .........127 Figure A4.4 – Three CALPUFF estimates vs. TEOM measurements at Creston ........127 Figure A4.5 – Three CALPUFF estimates vs. TEOM measurements at Revelstoke ...128 Figure A4.6 – Three CALPUFF estimates vs. TEOM measurements at Golden .........128 Figure A7.1 – Address vs. postal exposure for the TEOM metric ................................137 Figure A7.2 – Address vs. postal exposure for the SMOKE metric .............................138 Figure A7.3 – Address vs. postal exposure for the CALPUFF metric ..........................139  viii ACKNOWLEDGEMENTS There are so many people to thank for their support on this work.  Foremost is my co- supervisor Michael Brauer, whose generous mentorship and encouragement have taught me to practice good science in good humour and with good grace.  Next is my co-supervisor Susan Kennedy, whose leap of faith several years ago landed me here in 2009.  I am also grateful to the researchers Paul Demers, Charles Ichoku, Peter Jackson, Brian Klinkenberg, Mieke Koehoorn, Ying MacNab and Gavin Shaddick for their help, guidance and technical support in completing this work.  This project would not have been possible without the help of Benjamin Burkholder, Ellen Larcombe and Jason Su.  Ben did the wind field modeling under the supervision of Peter Jackson, and taught me most of what I know about the mechanics of CALMET and CALPUFF.  Ellen did the computer processing necessary to produce analysis- ready versions of all the remote sensing data.  Jason wrote the code necessary to clean the very messy address files we received for geolocation.  It was a pleasure working with all of them.  Colleagues at several agencies provided their time, data and technical advice for this project, and I would like to acknowledge them all at: the Center for Health Services and Policy Research; the BC Ministry of Heath; the BC Ministry of Forests; the BC Ministry of Environment; the US Forest Service; the US National Aeronautics and Space Administration; and the US National Oceanic and Atmospheric Administration.  Financial support for my training was generously provided by the Canadian Institutes of Health Research and the Michael Smith Foundation for Health Research.  Financial support for the project was provided by the Canadian Institutes of Health Research and the BC Lung Association.  I would like to thank all funding sources for making this work possible.  Last but not least I would like to thank David, the rest of my family and my most excellent friends for their encouragement, understanding and keen ears when needed. I know that I am lucky to be blessed with their presence in my life.   ix DEDICATION For my father, whose picture on the desk has been a motivation and whose memory will always be an inspiration.  x CO-AUTHORSHIP STATEMENT With the guidance of my committee I independently developed the plan for this thesis and was responsible for its implementation.  I conducted all data analysis and prepared the presented manuscripts in consultation with the co-authors of chapters 2 through 4. The contributions of each co-author are summarized below in order of appearance.  Benjamin Burkholder: Much of the preliminary wind field modeling was done by Mr. Burkholder, who was hired as an employee on the project in the first year.  His thoughts and ideas significantly shaped my own approach to dispersion modeling, and he was included as an author on Chapters 2 and 3 as a result.  Michael Brauer: This thesis was co-supervised by Dr. Brauer, who was generous with his time and perspective throughout its completion.  Many hours of our conversations about the methods and their results are reflected in these pages.  Peter L. Jackson: Early in this work Dr. Jackson was identified an expert in pollution dispersion modeling and was recommended for the thesis committee.  His experience with CALMET, CALPUFF and emissions estimation was a valuable resource for me and Mr. Burkholder while preparing Chapters 2 and 3.  Charles Ichoku: As described in Chapters 2 and 3 we used an approach developed by Dr. Ichoku and his colleagues to estimate aerosol emissions rates from MODIS- detected fires.  His support for this project was enthusiastic, and his knowledge of remote sensing was essential to ensure our correct use of those data.  Ying C. MacNab: Early in this work Dr. MacNab was identified as an expert in spatial statistics and was recommended for the thesis committee.  Her collaboration was integral to structuring and running the models presented in Chapter 4.  Susan M. Kennedy: This thesis was also co-supervised by Dr. Kennedy, who was always ready with keen methodological insight when needed.  Without her ideas about reducing the cohort size while amplifying the event signal the multiple analyses presented in Chapter 4 would have been impossible.  1 1 Chapter 1: Introduction and literature review 1.1 Overview This manuscript-based thesis comprises five chapters and seven appendices.  The first chapter provides an introduction to the work and a review of the literature necessary for readers to understand and critically evaluate the remaining chapters.  Each of Chapters 2 through 4 was prepared as an independent manuscript suitable for publication, and has either been published or submitted for peer review as noted on the first pages. Chapter 2 demonstrates how satellite data can be used to simply and reasonably estimate the area burned and aerosols emitted by wildfire events.  These results were used in Chapter 3 to simplify the forest fire plume dispersion model we used to estimate daily particulate matter exposure for subjects in a large epidemiologic study on the health effects of fire smoke.  The results of this study are presented in Chapter 4, which associates the risk of respiratory and cardiovascular physician visits and hospital admissions with three different measures of smoke exposure.  Chapter 5 provides a critical review of the work in context of the relevant literature and methodological information not appearing elsewhere in the document.  The appendices provide readers with more information about the different fire emissions estimates, cohort identification, cohort geolocation and epidemiologic results.  1.2 Background 1.2.1 Introduction Forest fires are a phenomenon of national and international concern.  While discrepancies in reporting make it difficult to assess the global impact of wildfires, losses between 5 and 20 million hectares are recorded annually in North America and Europe.  Regions of South America, Asia, Africa and Australia are also heavily affected1.  Canada contains approximately 10% of the global forest cover, with average annual (1970-1997) fire losses of 2.1 million hectares2. During the summer of 2003 more than 1900 of Canada’s 6900 wildfires burned in British Columbia3 destroying more than 260 000 hectares and 343 homes4.  Over the next 50 years climate change analysts predict that forest fires are going to become more frequent and intense in British Columbia, Canada and worldwide5.  2  Beyond the obvious damage to local environments and private property, the air quality impacts of forest fire events must be recognized.  Forest fire smoke contributes to elevated ambient concentrations of multiple air pollutants, many of which have been associated with adverse health effects.  Several studies have reported that the greatest increase is observed in fine particle concentrations6.  While the health effects of particle air pollution are well recognized in the urban context, forest fires do not often impact areas sufficiently populated to facilitate epidemiologic analyses.  As a result, the literature surrounding the acute health impacts of forest fire smoke is sparse, with very few large studies and none yet conducted in Canada.  Furthermore, those studies that demonstrate a significant relationship between forest fire smoke and health tend to rely on crude exposure estimates based on subject self-report or limited air quality monitoring data.  Here we build upon previous work by (1) identifying a large population-based cohort from administrative health records and (2) assigning multiple measures of forest fire smoke exposure at the residential locations of cohort members to strengthen the epidemiologic analyses.  1.2.2 Study area Mountainous terrain and dry valleys characterize the southern interior region of British Columbia.  The study area shown in Figure 1.1A measures 153 400 km2 and is heavily forested with lodgepole pine, aspen, white spruce, and Douglas fir.  This ecologic region is known as the Okanagan Dry Forest.  Tinder-dry conditions in July and August render this region especially susceptible to annual forest fires sparked by lightning or human activity.  More than 260 000 hectares of forest burned in BC during the 2003 fiscal year, approximately 75% of which were in the southern interior region3.  Figure 1.1B shows the spatial correspondence between these fires and the study area.  The 2003 season compares to provincial losses of 8 581 hectares in 2002 and 76 574 hectares in 1998, which was the worst fire season of the previous decade. According to climate change analysts, the frequency and severity of forest fires in this region is expected to increase over the next fifty years5.   3 1.2.3 Air quality during the study period The fire season of 2003 started in July, with activity lasting through the beginning of October4.  All analyses described in this thesis focus on 92-days between July 1st and September 30th.    During this period of unprecedented activity several fires burned in close proximity to densely populated parts of the southern interior, resulting in wide scale human exposure to potentially harmful levels of air pollution.  Satellite imagery (see Figure 1.2) indicates that smoke from fires in BC and the neighbouring United States covered much of the study area, especially the fertile valleys where most of the population resides.   Figure 1.1 – Fires in the study area Shows (A) the study area and (B) the location of all wild fires during the 2003 study period. A B  4  Figure 1.2 – Smoke in the study area Smoke dispersion over the study area from fires burning in BC and the neighboring United States.  Red areas indicate fire perimeters and black arrows indicate the locations of air quality monitors.  Image captured by the Moderate Resolution Imaging Spectrometer (MODIS) aboard NASA’s Terra satellite on August 20th, 2003.   Continuous measurements of ambient particulate matter concentrations (PM10, particulate matter less than 10 µg/m3) were taken by six TEOM (tapered element oscillating microbalance) instruments in Kamloops, Kelowna, Vernon, Creston, Revelstoke and Golden, as indicated in Figure 1.2.  Measurements of fine particulate matter (PM2.5) were also taken at Kamloops, Kelowna, Vernon and Golden. Measurements are averaged on an hourly basis and reported in real time by the BC Ministry of Environment (formerly the Ministry of Water, Land and Air Protection). Figure 1.3 shows all measurements over the course of the study period.  Clear elevations in concentration are seen on fire-affected days, with most particulate in the fine range, as evidenced by the ratio of PM10 to PM2.5. Kamloops  Kelowna  Vernon  Creston  Revelstoke Golden   5  Figure 1.3 – Measurements of particulate matter in the study area Daily 24-hour average PM10 and PM2.5 (µg/m3) measurements from the study area over the course of the study period (July 1st through September 30th, 2003).   6 1.2.4 Study population and health data The 2001 population of the study area was approximately 638 8007.  The weighted average 5-year (1995-2001) growth rate for these eight census districts was 2.1%8, suggesting a small population increase between the last census and the study period. Administrative health data are available from the BC Linked Health Database (BCLHD) for all members of the population covered by the provincial health insurance system, which is nearly universal.  This unique set of databases contains person-specific health information on all residents of the province, including hospital discharges and physician visits9.  Records for individuals can be linked over time and across different databases using the Personal Health Number (PHN) as a unique identifier.  We first used this database to help identify a geographically-stable cohort for the study, and we generated information about relevant health outcomes from (1) physician billings in the Medical Services Plan (MSP) database, and (2) hospital admissions in the Hospital Separations database.   The following work examines how forest fire smoke pollution was associated with these outcomes in the study area during the 92-day study period, and it was approved by the UBC Behavioral Research Ethics Board (the most recent certificate is included as Appendix 1).  1.3 Rationale and objectives Given the southern interior’s susceptibility to large forest fires, the size of its population, and the ready availability of health and environmental data, the conditions are excellent for meaningful study of the health effects associated with forest fire smoke.   If the predictions of climate change analysts are accurate, communities in this region and elsewhere stand to benefit from the knowledge generated by such work.  One commonality between previous epidemiologic studies on forest fire smoke is their reliance on exposure estimates based on limited air quality data from regulatory monitoring stations.  While these may provide accurate measurements for nearby populations, sparse distribution prevents them from capturing the true spatial variability in smoke distribution, which results in certain misclassification when these data are used to represent exposure over large areas.  Also, the filter-based particle sampling instruments used in most air quality networks are subject to failure under conditions of heavy smoke10.  Consider Figure 1.4 which shows the Okanagan Mountain Park fire  7 plume and the locations of air quality monitoring stations in Kelowna and Vernon.  The 24-hour average PM10 concentrations on this day were 249 and 174 µg/m3 at the Kelowna and Vernon stations, respectively.  However, this figure shows that the Vernon measurement may significantly under-represent the exposure for residents living to the east of the station, which falls more directly under the smoke plume. It is also obvious that neither value would give an appropriate estimate for Penticton, which does not have an air quality monitoring station but has a population of 38,000 people.  This suggests that residents of communities such as Penticton, which could be impacted by smoke in the event of a wind change or a fire to the south, would be eliminated from or grossly misclassified in any analysis based on available air quality data.  This could result in reduction of power to detect important associations and/or bias in the measure of effect.   Figure 1.4 – Satellite image of fire near Kelowna Image captured by the Advanced Spaceborne Thermal Emissions Reflectometer (ASTER) aboard NASA’s Terra satellite on August 21st, 2003.  This demonstrates that data from the air quality monitoring station in Vernon may under represent the exposure of residents living directly beneath the plume’s path.   Our premise is that smoke dispersion modeling and/or remote sensing data might provide inexpensive and widely applicable means of reducing exposure misclassification and increasing the power of population-based studies.  The primary goal of this research was to improve exposure analysis methodology for the purpose of  8 improving practical understanding of the health effects associated with forest fire smoke.   With global forest fire activity expected to become more frequent and intense, the results of this study could be valuable to any health or environmental agency working to understand and reduce the burden of smoke-related disease.  At a more local level this work also has the potential to bring discussion about the health effects of forest fire smoke into the arena of provincial forest management policy, where the current focus is on minimization of environmental and property losses.  1.3.1 Chapter 2: Estimating forest fire properties from remote sensing data The primary objective of this chapter is to develop and test some simple, data-driven methods for providing much of the information needed to model forest fire smoke dispersion.  As seen in Figure 1.2, air quality in the study area can be affected by fires in BC and Alberta, and in the states of Washington, Idaho and Montana.  Because fire records and forest catalogues are maintained by different provincial, state and federal agencies, conventional methods would require considerable time resources to gather and standardize the necessary information.  Instead we use remotely-sensed data from the Moderate Resolution Imaging Spectroradiometer (MODIS) to (1) identify fire locations, (2) estimate fire size, (3) simulate fire growth and decay, and (4) estimate aerosol emissions rates.  The performance of the methods developed is evaluated using fire data from BC, and emissions models used by the Canadian and American government agencies.  1.3.2 Chapter 3: Dispersion modeling for smoke exposure assessment The primary objective of this chapter is to generate exposure estimates for an epidemiologic assessment on the health effects of fire smoke.  We did this by estimating ambient concentrations of smoke-related PM10 across the study area with a plume dispersion model.  In the simplest terms such models require three inputs: (1) locations of the fires for which dispersion will be modeled; (2) emissions profiles for those fires; and (3) information about the winds that will carry emissions away from the fire locations.  In this case (1) and (2) come from Chapter 2, and (3) comes from work done by Benjamin Burkholder, as described elsewhere11, 12.  Of course such models are subject to considerable uncertainty, so it was necessary to evaluate the output as best as possible.  To do this we compare (1) estimated concentrations to those measured in  9 the six locations shown in Figure 1.3, (2) the birds-eye-view of plume profiles to those visible in satellite images of the area, and (2) MODIS-measured aerosol optical thickness (AOT) to TEOM-measured and CALPUFF-estimated PM10.  1.3.3 Chapter 4: Health effects of forest fire smoke The primary objective of this chapter is to assess how forest fire smoke exposure was associated with the risk of physician visits and hospital admissions with respiratory or cardiovascular diagnoses in British Columbia during the summer of 2003.  A population- based cohort was identified from the BCLHD and three different exposure metrics were assigned at the residential coordinates of each subject.  Estimates from the Chapter 3 dispersion model provide one of these metrics.  The two others are derived from (1) the TEOM measurements of PM10 taken at the six sites shown in Figure 1.2 and Figure 1.3 and (2) satellite imagery of smoke over the study area.  We estimate the fixed effects of all three metrics using logistic regression with repeated measures, adjusted for covariates such as age, sex and socioeconomic status.  The historical health records of each subject are used to create variables reflecting the number of respiratory and/or cardiovascular physician visits in the year prior to the study as a potential measure of pre-sensitivity to smoke exposure.  1.4 Review of the pertinent forest fire smoke literature 1.4.1 Particulate matter in forest fire smoke Wildfire smoke is a complex mixture of gases and solids that result from the incomplete combustion of biomass fuels. It contains a number of organic and inorganic compounds such as carbon monoxide, aldehydes, polycyclic aromatic hydrocarbons and trace elements.  Many of these pollutants are known to elicit human health effects at concentrations typically found in fire smoke and some are potential human carcinogens13.  The composition of solids in wildfire smoke is largely dependant on the fuel and fire type.  In the hot, flaming stages of a crown forest fire the fuel is burned efficiently and it produces fine organic particles with varying concentrations of elemental black carbon14.  Combustion conditions during the cool, smoldering stages of the same fire are less efficient, such that more particles are produced per unit of fuel burned, but their black carbon concentration is very low15.  This distinction may be important when  10 considering the health effects of fire smoke because the composition of respired particles is associated with their toxicity16.  The size distribution of the particles is another important consideration when assessing their potential health effects.  Particles with median aerodynamic diameter less than 10 microns (PM10) can be inhaled into the respiratory tract, those smaller than 2.5 microns (PM2.5) can penetrate into the deep lung.17.  Most particles from forest fires are in the respirable range with a peak diameter between 0.15 and 0.4 microns18, 19, which is similar to the size distribution observed in particles generated by fossil fuel combustion. Such particles have been implicated in multiple respiratory and cardiovascular health outcomes20, 21.  Moreover, particles in this size range do not exhibit rapid gravitational settling and can be transported over long distances, potentially affecting individuals far from the smoke source22.  This is supported by several studies that report the air quality impacts of near and distant fires being most evident in elevated ambient concentrations of particulate matter23-25.  Given the epidemiologic precedent regarding the deleterious health effects of fine particulate matter, this pollutant is the primary focus when studying the population health effects of forest fire smoke.  1.4.2 Smoke plume dispersion modeling Atmospheric dispersion modeling is the mathematical simulation of how air pollutants disperse after leaving their source26.  In this work the CALPUFF dispersion model is used to estimate ambient concentrations of particulate matter (PM10) due to fire emissions within the study area.  Large fires differ from other point sources of particulate pollution because the buoyancy of their emissions is widely variable27. Smoke from a hot crown fire can be injected high into the atmosphere (i.e. the lower stratosphere), meaning that it will be dispersed by the winds aloft and that it may affect air quality thousands of kilometers from its source28.  Smoke from a cool smoldering fire may not rise at all and its air quality effects will typically be localized29.  Therefore modeling the dispersion of forest fire smoke requires information about the particle emissions rates, the plume rise, the horizontal wind profile and the vertical wind profile.  Although many different dispersion modeling packages are available, most modeling of forest fire plumes is done within the framework of smoke forecasting, whereby current  11 fire information and prognostic wind forecasts are used to predict future smoke scenarios.  When this thesis was initiated BlueSky (US Forest Service, Seattle, WA) was the only operational smoke forecasting framework in North America and it was using the CALPUFF (TRC Solutions, Lowell, MA) dispersion model.  In this type of “puff” model the pollutant source regularly releases spherical puffs of particles that expand in a Gaussian concentration distribution until they exceed the size of the domain grid cell, at which point they continue to split into new puffs until concentrations approach zero30.  We chose to use CALPUFF for this thesis because of a proposal to extend BlueSky into British Columbia and Alberta, but we could have used multiple other models.  Unlike the BlueSky framework we used retrospective interpolation of already-measured meteorological data in CALMET (TRC Solutions, Lowell, MA) to provide the wind fields, not prognostic estimates from weather forecasting models.  1.4.3 Remote sensing of fires and smoke by MODIS The human eye can typically perceive information from wavelengths ranging between 0.4 and 0.7 µm.  The two MODIS electromagnetic sensors launched in 1999 and 2002 can, however, record information from 36 spectral channels with wavelengths ranging between 0.4 and 14.4 µm31.  These instruments are in polar orbit aboard the Aqua and Terra satellites, and were designed to measure a wide array of land, ocean and atmospheric properties.  As such, special channels were included for daily global observation of forest fires and smoke aerosol at a horizontal resolution of 1 km32.  Fires are mainly detected using the 4 and 11 µm wavelengths, which can sense surface temperatures greater than 500 and 400 degrees Kelvin, respectively33 (typical smoldering and flaming fire temperatures are 500 and 1200 degrees Kelvin, respectively34).  Special algorithms applied to the thermal signatures of these channels allow fires to be separated from the background signal32, 35.  Furthermore, differences between these and other channels allow MODIS to distinguish between cool, smoldering and hot, flaming fires based on its measurements of fire radiative power (FRP)36.  The validity of MODIS fire detection has been demonstrated in multiple contexts using aircraft and data from other remote sensing instruments33.  On cloudless days smoke from large fires is clearly observed in MODIS satellite images (as evident in Figure 1.2), but remote measurements of smoke-specific aerosol in the  12 atmosphere are still challenging.  One important reason for this is the difference in black carbon concentrations between emissions from flaming and smoldering fires37. Because black carbon is light-absorbing the smoke from flaming fires is easier to detect in the atmosphere, but the smoke from smoldering fires may have more impact on surface air quality.  Like many other remote sensing instruments, MODIS estimates the aerosol optical thickness (AOT) based on a measure of light extinction within cloudless columns of atmosphere between MODIS and the earth (705 km)38.  The horizontal resolution of this product is currently 10 km.  Over land the AOT is computed using the 0.47 and 0.66 µm channels, and values are interpolated to derive a standardized measure of AOT at 0.55 µm39.  Although MODIS produces a single AOT value to reflect all types of aerosol in the atmosphere (dust, salt, anthropogenic pollution, smoke, etc.), it is theoretically possible to distinguish smoke from other aerosols based on differences between particle size and absorption properties37.  However, validation of the algorithms developed to reflect these distinctions has proved that their correlation with surface air quality measurements is moderate at best38, 40-42.  1.4.4 Smoke exposure assessment for public health studies In the absence of air quality measurements sufficient for epidemiologic exposure assessment several studies have simply compared health outcomes during defined “fire periods” to those during periods not affected by fire smoke.  In a brief report Sorensen et al. compared emergency room visits and hospital admissions during a 1998 fire period in Florida to those from the same period in 1997.  The air pollution impacts of fire smoke were not quantified.  Kunii et al.43 interviewed subjects about fire-related symptoms while measuring PM10 at three sites for eight days of a much longer Indonesian haze event.  Concentrations upwards of 1800 µg/m3 were recorded.  During the same haze event in Singapore Emmanuel44 used PM10 data from a single monitor to compare emergency room visits, hospital admissions and mortality to the previous year, with concentrations up to 110 µg/m3 reported.  Again during the same haze event Mott et al.45 conducted another study in Malaysia, comparing hospital admissions during a three month fire period to those during pre- and post-fire periods, but no quantitative measure of fire smoke were used.  In a similar Californian study Mott et al.46 compared hospital admissions during two months of fire-elevated PM10 measurements in 1999 to those during the same period of the previous year when no fires were burning.  13 Concentrations greater than 350 µg/m3 were reported.  Finally, Moore et al. used TEOM data from Kelowna and Kamloops during the summer of 2003 (see Figure 1.2 and Figure 1.3) to compare physician visits counts to averages from the previous ten years47.  Concentrations reaching 250 µg/m3 were reported.  Because wildfires are temporally and spatially episodic, it is challenging to implement studies that use personal monitoring or a dense network of temporary ambient instruments.  Where monitors already exist it is possible to use their measurements for crude exposure estimates, but Figure 1.4 clearly shows how this can lead to considerable misclassification in large populations.  Even so, a few epidemiologic studies have used such measurements to quantify the health effects of fire smoke exposure in terms of the risks associated with specific elevations in PM2.5 or PM10 concentrations.  Sastry48 used data from a single PM10 monitor in Kuala Lumpur to examine the mortality effects of the 1997 southeast Asian haze in a time series design, with concentrations up to 400 µg/m3 recorded.  To study how bushfire smoke was associated with emergency room visits and hospital admissions in Australia, Johnston et al.49 (Darwin), Chen et al.50 (Brisbane) and Tham et al.51 (Victoria) all used data from a single TEOM monitor in a time series design.  In one of the few studies outside of Asia, Australia or North America, Hanninen et al.52 investigated smoke-related mortality in eleven Finnish communities using PM2.5 or PM10 data from the most representative of eight air quality monitoring stations, with a maximum PM10 concentration of 180 µg/m3 recorded.  Only two other epidemiologic studies have attempted to improve smoke estimation by modeling PM concentrations.  In the first, Hanigan et al.53 used data from a previously- developed model of visibility to predict PM10 concentrations at a single location in Darwin, Australia because air quality was only sporadically measured during the fire period.  Resulting values were partially evaluated with remote sensing data, and used in a time series analysis of hospital admissions.  In the second, Delfino et al.54 used work done by Wu et al.10 to assign daily PM2.5 exposure estimates for every postal code in Los Angeles over the fire season of 2003.  This is the only other study that investigated the potential of remote sensing data to inform forest fire smoke exposure assessment. During the 2003 fire season several monitors in the city’s dense regulatory network  14 sampled intermittently or failed altogether.  Satellite images were used to assign each sampler location a daily smoke density value of “no smoke”, “light smoke” or “heavy smoke”, and known surface concentrations were regressed on smoke density and other potentially-predictive variables.  Regression predictions were then spatially smoothed to provide the daily exposure maps used in a very large Poisson analysis of hospital admissions. 1.5 Review of the pertinent health literature 1.5.1 Health effects of urban particulate matter Although forest fires routinely result in episodes of poor air quality there is relatively little epidemiologic work focused on their population health impacts.  In contrast, there is a large volume of information associating urban particulate matter (primarily from vehicle emissions) with adverse respiratory and cardiovascular effects.  A summary of the most informative studies is provided here in order to contextualize the health risks associated with particulate matter, but it is important to remember that the time scale for forest fire smoke exposure may be quite different from that of urban air pollution exposure.  In urban environments exposure to particulate matter is typically chronic with occasional elevations due to unique events or meteorological conditions.  In areas affected by forest fires the exposures are generally acute with a few very high peak concentrations. Also recall that the composition of smoke particles is highly variable, and also different from that of the particles in urban air pollution, which generally result from fossil fuel combustion.  The physiologic effects of urban pollution particles have been assessed in several studies that consistently associate exposure with significantly increased white blood cell release from the bone marrow – a marker of systemic inflammatory response55.  Long- term exposure to ambient airborne particulate matter is associated with increased mortality, largely due to cardiovascular disease56-58.  In a comparison between cities with the lowest and highest particulate concentrations the average life expectancy was reduced by about 1.5 years59.  Time series studies have also linked increased cardiovascular mortality and hospital admissions with air pollution levels on the days preceding the event60, 61.  In fact, urban pollution has been estimated to be responsible for approximately 3% of global mortality from cardiopulmonary diseases62, with the risk  15 most strongly related to ambient concentrations of fine particles less than 2.5 µm in diameter63, 64.  Short-term exposure to elevated concentrations of urban particulate matter has been associated with multiple respiratory outcomes including mortality, hospital admissions, asthma exacerbations and decreased lung function65.  For example, a recent study including 29 European cities found that 10 µg/m3 increases in PM10 and Black Smoke (strongly correlated with elemental carbon) were associated with 0.58% and 0.84% increases in mortality, respectively66.  In a nation-wide study of American Medicare users Dominici et al.67 showed that a 10 µg/m3 increase in PM2.5 was associated with increases of approximately 1.0% and 0.9% in hospital admissions for chronic obstructive pulmonary disease (COPD) and respiratory infections, respectively. Similarly Peel et al.68 report significant increases in emergency room visits for asthma and COPD for a 10 µg/m3 increase in PM10, and they detect increased visits for pneumonia at PM2.5 elevations as low as 2 µg/m3.  In a joint study on children and adults Trenga et al.69 measured personal indoor, outdoor and ambient PM10 concentrations for 10 days and followed their forced expiratory volume (FEV) over three years.  Increased PM was consistently associated with decreased FEV in adults diagnosed with COPD and in asthmatic children, even at very low concentrations.  The overall results from this large body of literature suggest that children, the elderly and people with pre-existing disease are at highest risk of being affected by airborne particulate matter.  1.5.2 Comprehensive review on the health effects of woodsmoke In 2007 Naeher et al.13 published a comprehensive, 40-page review on the health effects of woodsmoke.  This exhaustive document discussed relevant literature from toxicological assays to occupational exposure in wildland fire fighters to the population health effects of agricultural burning.  In general the authors concluded that: (1) fine particulate matter is currently the best metric by which to assess woodsmoke exposure; (2) the respiratory hazards presented by woodsmoke particles appear similar to those presented by urban particulate pollution; (3) there is insufficient evidence to address the cardiovascular and carcinogenic effects of woodsmoke pollution; and (4) further epidemiologic studies are needed to better understand the health risks of woodsmoke.  16 The following sections do not repeat the work of Naeher et al., but provide a summary of the material most relevant to this thesis in addition to discussion of the most recently published work.  1.5.3 Physiologic effects of woodsmoke Several studies of controlled or carefully-measured exposure indicate that the particulate matter in biomass smoke may have health impacts that are similar to those of the particulate matter in urban air pollution.  Barregard et al.70 exposed 13 healthy subjects to particle concentrations of 240-280 µg/m3 derived from wood combustion (typical of those reported in areas impacted by fire smoke, but produced under different conditions), and found deleterious effects on markers of inflammation, coagulation and lipid peroxidation.  Tan et al. examined immunological response to ambient vegetation fire smoke exposure (up to 216 µg/m3) in 30 healthy volunteers and observed that the immature white blood cell count tracked with PM concentrations71.  Studies on acute and chronic exposures in laboratory animals showed generally consistent results and provide further insight into mechanisms by which woodsmoke first affects human health.  As summarized by Naeher et al.13, most studies indicate that short-term exposure to woodsmoke primarily compromises the immunological defenses of the lung.  This is consistent with the epidemiologic evidence presented in the next section, where the respiratory effects of forest fire smoke exposure are demonstrated but no cardiovascular effects are reported.  However, the lack of cardiovascular effect is inconsistent with the aforementioned studies indicating a systemic inflammatory response to woodsmoke similar to that observed for urban air pollution.  Although this discrepancy requires further study, some recent evidence suggests that woodsmoke may be less toxic than particulate pollution from fossil fuel combustion.  First, Kocbacha et al.72 showed that the inflammatory potential of woodsmoke particles is less than that of urban particles for exposures greater than 12 hours.  Second, Londahl et al.73, 74 showed that biomass smoke particles deposit less efficiently than urban (traffic-related) particles in the human respiratory tract.  However, given the weight of evidence suggesting some relationship between smoke exposure and cardiovascular health13, the plausibility of this relationship cannot be dismissed without further investigation.   17 1.5.4 Epidemiologic studies on the population effects of forest fire smoke A growing body of epidemiologic literature on the health impacts of forest fire smoke consistently associates smoke-related particles with increased risk of acute respiratory events, but not with cardiovascular effects.  During a two-week fire period in California with measured PM10 concentrations as high as 237 µg/m3, Duclos et al.75 demonstrated that emergency room (ER) visits for asthma and chronic obstructive pulmonary disease (COPD) were increased by 40 and 30%, respectively. Subjects with pre-existing respiratory and cardiovascular conditions were at higher risk.  Another California wildfire produced daily PM10 averages exceeding 150 µg/m3 fifteen times and 500 µg/m3 twice during a 10-week period.  Mott et al.46 report that the medical clinic for a nearby Indian reservation experienced a 52% increase in visits for respiratory outcomes over the same period for the previous year, and that weekly PM10 averages were strongly correlated with weekly visits during the fire year (r = 0.74) compared to the previous year (r = -0.63).  In a community survey of 289 respondents, more than 60% reported respiratory symptoms during the smoke episode and 20% reported symptoms persisting at least two weeks after the smoke cleared. Individuals with pre-existing cardiopulmonary diseases reported significantly more symptoms before, during, and after the fire than those without such illnesses.  More recently in California the aforementioned work by Wu et al.10 was used by Delfino et al.54, who reported that a 10 µg/m3 increase in the two-day mean of PM2.5 was positively associated with the relative rate of respiratory hospital admissions in Los Angeles (RR = 1.028; 95%CI = 1.014– 1.041), with higher risk for children and the elderly.  Major forest fire activity in Southeast Asia in 1997 and 1998 resulted in several studies demonstrating health impacts associated with particulate pollution. In Singapore, for example, a time series analysis indicated that a PM10 increase of 100 µg/m3 was associated with 12%, 19% and 26% increases in cases of upper respiratory tract illness, asthma and rhinitis, respectively, though no significant increase in hospital admissions or mortality was observed44.  Brauer & Hisham-Hashim76 and Leech et al.77 reported similar findings for Malaysia.  During the same episode Hisham-Hashim et al.78 measured lung function in 107 school-aged children and found that it was significantly reduced when compared to similar measurements taken in the previous year.  One study in Kuala Lumpur evaluated the relationship between fire-related particles and  18 mortality and reported that a 10 µg/m3 increase in PM10 was associated with 0.7% (all ages) and 1.8% (ages 65-74) increases in adjusted relative risk of non-traumatic death48.  Vedal et al.79 and Hanninen et al.52 also explored the relationship between smoke exposure and mortality in Denver and Finland, respectively, but report no significant associations, possibly due to the small number of deaths in both studies.  While results from California and Malaysia are mutually consistent, three early studies assessing the health impacts of 1994 bushfires burning near to Sydney, Australia reported no increased effect despite 1-hour PM10 concentrations as high as 250 µg/m3 80-82. Possible reasons for this discrepancy include differences in study design, sample size, the chemical composition of the particles, and the relative toxicity of the specific particle mixture. A more recent study in Darwin, Australia evaluated the association between daily asthma ER visits and measured PM10 over a 7-month period which included two bushfire smoke episodes. Johnston et al.49 reported that increased asthma presentation was positively associated with PM10 concentrations, especially for days in which PM10 concentrations exceeded 40 µg/m3. Hanigan et al.53 later expanded this study and continued to observe the same increase, as well as an increased (yet insignificant) association for the indigenous subpopulation.  In summary, recent studies demonstrate that particulate pollution from forest fire smoke is associated with increased respiratory symptoms, increased risk of respiratory illness and decreased lung function in both children and adults, with evidence that those with pre-existing respiratory conditions are particularly sensitive.  A limited number of studies also indicate a relationship between smoke exposure and ER visits, though the implications for hospitalization and mortality outcomes have not been sufficiently studied.  Despite the limited data and the exception of the Australian bushfire investigations, the epidemiologic literature indicates a generally consistent relationship between forest fire smoke exposure and increased incidence of respiratory health effects.  Although several studies have reported null cardiovascular results13, 47, 53, 83, Naeher et al. conclude that the data have been insufficient to address the question with adequate power13.  Given the weight of evidence suggesting some hypothetical relationship between smoke exposure and cardiovascular health, further study is  19 needed, and enhanced exposure assessment tools may improve our ability to detect the effects if they actually exist.  1.6 On the use of administrative data for health research The epidemiologic analyses presented in this thesis are entirely based on administrative health data, and it is important for readers to consider the strengths and limitations of such data sources as they review the work.  Administrative data are collected by non- research organizations for administrative purposes84.  This includes data on physician visits and hospital admissions, as used in this study.  Because British Columbia has a nearly-universal public health care system, its administrative databases make it possible to conduct population-based studies that would be impossible elsewhere85. While this allows us to address many research problems that might otherwise go unstudied, it is important to remember that these data were not collected for research purposes and their accuracy is not guaranteed86.  For example, the diagnoses underlying each physician visit or hospital admission are recorded using the 9th edition of the International Classification of Disease codes (ICD9), which range between three and five digits.  The first three digits explain the general grouping (460-591 for respiratory, 390-459 for cardiovascular) while the fourth and fifth provide more specific information.  However, these diagnoses are not necessarily standardized because individual physicians decide which ICD9 code to use and what level of accuracy to record.  Medical charts and survey data have been used to test the reliability of ICD9 coding in Canada with reports indicating that the accuracy of administrative records (i.e. the notes on the chart matched the code in the database) is moderate87-91.  This suggests that specific event definitions based on ICD9 will be subject to some misclassification, so most epidemiologic analyses presented here use the general disease groupings rather than individual codes.  1.7 Summary Forest fires are a spatially and temporally episodic source of particulate air pollution, and the health effects of fire smoke are not well-understood.  It is challenging to assess smoke exposure in most fire-affected areas because air quality monitoring is spatially sparse or non-existent.  Despite this limitation several studies have already  20 demonstrated that exposure to forest fire smoke is associated with respiratory symptoms, emergency room visits and hospital admissions.  Given the similarity between woodsmoke particles and urban pollution particles there is reason to believe that forest fire smoke may also affect cardiovascular health, but this relationship has not been demonstrated thus far.  With much of the global population living in areas affected by smoke from forest and other vegetation fires, it is important to improve our understanding of the health risks associated with air pollutants from these sources.  The results of epidemiologic studies can be strongly influenced by the quality of the exposure estimates used, and new tools are needed to address the spatial limitations inherent to using air quality monitoring data for smoke exposure assessment.  Such new tools cannot be adequately tested in the absence of (1) valid measures of ambient smoke concentrations and (2) high-quality data on population health.  Large fires in British Columbia during the summer of 2003 provided a unique opportunity to develop new exposure assessment methods that could be (1) suitably evaluated and (2) applied to a large cohort.  The following chapters describe (1) the development a sophisticated exposure assessment tool that combines dispersion modeling with remote sensing data, and (2) its application to the largest population yet studied in fire smoke epidemiology.  21 1.8 References 1. World Health Organization (1998). Biregional Workshop on Health Impacts of Haze-related Air Pollution. Regional Office of the Western Pacific; Kuala Lumpur, Malaysia. 51 pages. 2. Global Forest Watch (2000). Canada's Forests at a Crossroads: An Assessment in the Year 2000. World Resources Institute; Washington, DC. 107 pages. 3. BC Ministry of Forests (2004) Forest Fire Statistics in Protection Branch.http://www.for.gov.bc.ca/pScripts/Protect/WildfireNews. Accessed on January 20th, 2004. 4. Filmon G (2004), Firestorm 2003: Provincial Review Vancouver, BC. 100 pages. 5. Flannigan MD, Stocks BJ and Wotton BM (2000). Climate change and forest fires. Sci Total Environ. 262(3): 221-229. 6. Brauer M (2000). Health impacts of air pollution from vegetation fires. Brunei International Medical Journal (invited paper). 2: 221-236. 7. Population and dwelling counts for Canada, Provinces and Territies, by Census Divisions, 2001 and 1996 censuses - 100% data. (2002). Population and dwelling counts for Canada, Provinces and Territies, by Census Divisions, 2001 and 1996 censuses - 100% data. Statistics Canada; 8. British Columbia Ministry of Management Statistics (2001) Census Profile in BC Statistics.http://www.bcstats.gov.bc.ca/data/cen01/profiles/csd_txt.htm. Accessed on February 24, 2004. 9. Centre for Health Services and Policy Research (2004) About the BC Linked Health Database in.http://www.chspr.ubc.ca/Bclhd/aboutbclhd.htm. Accessed on September 13th, 2004. 10. Wu J, Winer A and Delfino R (2006). Exposure assessment of particulate matter air pollution before, during, and after the 2003 Southern California wildfires. Atmospheric Environment. 40(18): 3333-3348. 11. Burkholder BJ (2005) Report on the spatial assessment of forest fire smoke exposure and its health effects -- Part I: Initialization of the CALMET meteorological model. School of Occupational & Environmental Hygiene, The University of British Columbia. Vancouver, BC. 46 pages. http://hdl.handle.net/2429/13646 12. Burkholder BJ (2005) Report on the spatial assessment of forest fire smoke exposure and its health effects -- Part II: CALMET initialization methodology. School of Occupational & Environmental Hygiene, The University of British Columbia. Vancouver, BC. 28 pages. http://hdl.handle.net/2429/13645  22 13. Naeher LP, Brauer M, Lipsett M, Zelikoff JT, Simpson CD, Koenig JQ and Smith KR (2007). Woodsmoke Health Effects: A Review. Inhalation Toxicology. 19(1): 67-106. 14. Ross JL, Hobbs PV and Holben B (1998). Radiative characteristics of regional hazes dominated by smoke from biomass burning in Brazil: closure tests and direct radiative forcing. Journal of Geophysical Research. 103: 31925–31941. 15. Eck TF (2001). Characterization of the optical properties of biomass burning aerosols in Zambia during the 1997 ZIBBEE field campaign. Journal of Geophysical Research. 106: 3425–3448. 16. Saldiva P, Clarke R, Coull B, Stearns R, Lawrence J, Murthy G, Diaz E, Koutrakis P, Suh H, Tsuda A and Godleski J (2002). Lung inflammation induced by concentrated ambient air particles is related to particle composition. American Journal of Respiratory and Critical Care Medicine. 165(12): 1610-1617. 17. de Kok TMCM, Driece HAL, Hogervorst JGF and Briede JJ (2006). Toxicological assessment of ambient and traffic-related particulate matter: A review of recent studies. Mutation Research/Reviews in Mutation Research. 613(2-3): 103-122. 18. Sandberg D and Martin R (1975) Particle sizes in slash fire smoke. U.S. Department of Agriculture, Forest Service, Pacific Northwest Research Station. Portland, OR. 19. Hueglin C, Gaegauf C, Kunzel S and Burscher H (1997). Characterization of Wood Combustion Particles: Morphology, Mobility, and Photoelectirc Activity. Environ. Sci. Technol. 31: 3439-3447. 20. Pope CA, 3rd (2000). Epidemiology of fine particulate air pollution and human health: biologic mechanisms and who's at risk? Environmental Health Perspectives. 108 Suppl 4: 713-23. 21. Dockery DW (2001). Epidemiologic Evidence of Cardiovascular Effects of Particulate Air Pollution. Environmental Health Perspectives. 109(Supplement 4): 483-486. 22. Echalar F, Gaudichet A, Cachier H and Artaxo P (1995). Aerosol emission by tropical forest and savanna biomass burning : characteristic trace elements and fluxes. Geophysical Research Letters. 22(22): 3039-3042. 23. Reinhardt TE, Ottmar RD and Castilla C (2001). Smoke impacts from agricultural burning in a rural Brazilian town. J Air Waste Manag Assoc. 51(3): 443-50. 24. Nichol J (1997). Bioclimatic impacts of the 1994 smoke haze event in Southeast Asia. Atmospheric Environment. 31(8): 1209-1219. 25. Hassan M, Zainal A, Yusoff M, Ghazali A, Muda A and Zakaria M (1995). Damage costs of the 1991 and 1994 haze episodes in Malaysia. in Climate and life in the Asia Pacific. Brunei Darussalam: Universiti Brunei Darussalam.  23 26. Arya PS (1999), Air Pollution Meteorology and Dispersion New York, NY: Oxford University Press. 310 pages. 27. Freitas SR, Longo KM and Andreae MO (2006). Impact of including the plume rise of vegetation fires in numerical simulations of associated atmospheric pollutants. Geophysical Research Letters. 33(L17808): doi:10.1029/2006GL026608. 28. Andreae MO (1991). Biomass burning - Its history, use, and distribution and its impact on environmental quality and global climate, in Global biomass burning - Atmospheric, climatic, and biospheric implicationsMIT Press: Cambridge, MA. p. 3-21. 29. Yokelson RJ, R. Susott, D. E. Ward, J. Reardon and Griffith DWT (1997). Emissions from smoldering combustion of biomass measured by open-path Fourier transform infrared spectroscopy. Journal of Geophysical Research. 102(D18): 18865-18877. 30. Scire J, DG S and RJ Y (2000), A users guide for the CALPUFF dispersion model Concorde, MA: Earth Tech Inc. 31. Justice CO, Vermote E, Townshend JRG, Defries R, Roy DP, Hall DK, Salomonson VV, Privette JL, Riggs G, Strahler A, Lucht W, Myneni RB, Knyazikhin Y, Running SW, Nemani RR, Zhengming W, Huete AR, van Leeuwen W, Wolfe RE, Giglio L, Muller J, Lewis P and Barnsley MJ (1998). The Moderate Resolution Imaging Spectroradiometer (MODIS): land remote sensing for global change research. IEEE Transactions on Geoscience and Remote Sensing. 36(4): 1228 - 1249. 32. Justice CO, Giglio L, Korontzi S, Owens J, Morisette JT, Roy D, Descloitres J, Alleaume S, Petitcolin F and Kaufman Y (2002). The MODIS fire products. Remote Sensing of Environment. 83(1-2): 244-262. 33. Kaufman YJ, Ichoku C, Giglio L, Korontzi S, Chu DA, Hao WM, Li R-R and Justice CO (2003). Fire and smoke observed from the Earth Observing System MODIS instrument--products, validation, and operational use. International Journal of Remote Sensing. 24(8): 1765-1781. 34. Pyne SJ, Andrews PL and Laven RD (1996), Introduction to Wildland Fire New York, NY: Wiley and Sons. 35. Giglio L, Descloitres J, Justice CO and Kaufman YJ (2003). An enhanced contextual fire detection algorithm for MODIS. Remote Sensing of Environment. 87(2-3): 273-282. 36. Ichoku C and Kaufman YJ (2005). A method to derive smoke emission rates from MODIS fire radiative energy measurements. IEEE Transactions on Geoscience and Remote Sensing. 43(11): 26-36. 37. Kaufman YJ, Tanre D and Boucher O (2002). A satellite view of aerosols in the climate system. Nature. 419(6903): 215-223.  24 38. Wang J and Christopher SA (2003). Intercomparison between satellite-derived aerosol optical thickness and PM2.5 mass: Implications for air quality studies. Geophysical Research Letters. 30(21): doi:10.1029/2003GL018174. 39. Chu DA, Kaufman YJ, Ichoku C, Remer LA, Tanre D and Holben BN (2002). Validation of MODIS aerosol optical depth retrieval over land. Geophysical Research Letters. 29(12). 40. Hutchison KD, Smith S and Faruqui S (2004). The use of MODIS data and aerosol products for air quality prediction. Atmospheric Environment. 38(30): 5057-5070. 41. Koelemeijer RBA, Homan CD and Matthijsen J (2006). Comparison of spatial and temporal variations of aerosol optical thickness and particulate matter over Europe. Atmospheric Environment. 40(27): 5304-5315. 42. Gupta P, Christopher SA, Box MA and Box GP (2007). Multi year satellite remote sensing of particulate matter air quality over Sydney, Australia. International Journal of Remote Sensing. 28(20): 4483-4498. 43. Kunii O, Kanagawa S, Yajima I, Hisamatsu Y, Yamamura S, Amagai T and Ismail IT (2002). The 1997 haze disaster in Indonesia: its air quality and health effects. Arch Environ Health. 57(1): 16-22. 44. Emmanuel SC (2000). Impact to lung health of haze from forest fires: the Singapore experience. Respirology. 5(2): 175-82. 45. Mott JA, Mannino DM, Alverson CJ, Kiyu A, Hashim J, Lee T, Falter K and Redd SC (2005). Cardiorespiratory hospitalizations associated with smoke exposure during the 1997 Southeast Asian forest fires. International Journal of Hygiene and Environmental Health. 208(1-2): 75-85. 46. Mott JA, Meyer P, Mannino D, Redd SC, Smith EM, Gotway-Crawford C and Chase E (2002). Wildland forest fire smoke: health effects and intervention evaluation, Hoopa, California, 1999. West J Med. 176(3): 157-62. 47. Moore D, Copes R, Fist R, Joy R, Chan K and Brauer M (2006). Population health effects of air quality changes due to forest fires in British Columbia in 2003. Canadian Journal of Public Health. 97(2): 105-108. 48. Sastry N (2002). Forest fires, air pollution, and mortality in southeast Asia. Demography. 39(1): 1-23. 49. Johnston FH, Kavanagh AM, Bowman DM and Scott RK (2002). Exposure to bushfire smoke and asthma: an ecological study. Med J Aust. 176(11): 535-8. 50. Chen L, Verrall K and Tong S (2006). Air particulate pollution due to bushfires and respiratory hospital admissions in Brisbane, Australia. International Journal of Environmental Health Research. 16(3): 181-191.  25 51. Tham R, Erbas B, Akram M, Dennekamp M and Abromson MJ (2009). The impact of smoke on respiratory hospital outcomes during the 2002–2003 bushfire season, Victoria, Australia. Respirology. 14: 69-75. 52. Hänninen OO, Salonen RO, Koistinen K, Lanki T, Barregard L and Jantunen M (2009). Population exposure to fine particles and estimated excess mortality in Finland from an East European wildfire episode. Journal of Exposure Science and Environmental Epidemiology. 19: 414-422. 53. Hanigan IC, Johnston FH and Morgan GG (2008). Vegetation fire smoke, indigenous status and cardio-respiratory hospital admissions in Darwin, Australia, 1996–2005: a time-series study. Environmental Health. 7(42): doi:10.1186/1476-069X-7-42. 54. Delfino RJ, Brummel S, Wu J, Stern H, Ostro B, Lipsett M, Winer A, 6 DHS, Zhang L, Tjoa T and Gillen DL (2009). The relationship of respiratory and cardiovascular hospital admissions to the southern California wildfires of 2003. Occupational and Environmental Medicine. 66: 189-197. 55. van Eeden SF and Hogg JC (2002). Systemic inflammatory response induced by particulate matter in air pollution: The importance of bone marrow stimulation. Journal of Toxicology and Environmental Health, Part A: Current Issue. 65(20): 1597-1613. 56. Dockery DW, Pope CA, X. Xu, Spengler JD, Ware JH, Fay ME, Ferris BG and Speizer FE (1993). An association between air pollution and mortality in six U.S. cities. New England Journal of Medicine. 329(24): 1753-1759. 57. Pope C, Thun M, Namboodiri M, Dockery D, Evans J, Speizer F and Health C (1995). Particulate air pollution as a predictor of mortality in a prospective study of U.S. adults. American Journal of Respiratory and Critical Care Medicine. 151: 669-674. 58. Pope C, Burnett R, Thun M, Calle E, Krewski D, Ito K and Thurston G (2002). Lung cancer, cardiopulmonary mortality, and long-term exposure to fine particulate air pollution. Journal of the American Medical Association. 287: 1123- 1141. 59. Brunekreef B (1997). Air pollution and life expectancy: is there a relation? Occupational and Environmental Medicine. 54: 781-784. 60. Burnett RT, Dales R, Krewski D, Vincent R, Dann T and Brook JR (1995). Associations between ambient particulate sulfate and admissions to Ontario hospitals for cardiac and respiratory diseases. American Journal of Epidemiology. 142(1): 15-22. 61. Schwartz J (1999). Air pollution and hospital admissions for heart disease in eight U.S. counties. Epidemiology. 10(1): 17-22. 62. Cohen A, Ross-Anderson H, Ostro B, Pandey K, Krzyzanowski M, Künzli N, Gutschmidt K, Pope A, Romieu I, Samet J and Smith K (2005). The global  26 burden of disease due to outdoor air pollution. Journal of Toxicology and Environmental Health, Part A. 68(13-14): 1301-1307. 63. Laden F, Neas L, Dockery D and Schwartz J (2000). ssociation of fine particulate matter from different sources with daily mortality in six U.S. cities. Environmental Health Perspectives. 108(10): 941-947. 64. Samet JM, Dominici F, Curriero FC, Coursac I and Zeger SL (2000). Fine particulate air pollution and mortality in 20 U.S. cities, 1987-1994. New England Journal Of Medicine. 343(24): 1742-1749. 65. Dockery DW and Pope CA (1994). Acute respiratory effects of particulate air pollution. Annual Review of Public Health. 15: 107-132. 66. Analitis A, Katsouyanni K, Dimakopoulou K, Samoli E, Nikoloulopoulos AK, Petasakis Y, Touloumi G, Schwartz J, Anderson HR, Cambra K, Forastiere F, Zmirou D, Vonk JM, Clancy L, Kriz B, Bobvos J and Pekkanen J (2006). Short- Term Effects of Ambient Particles on Cardiovascular and Respiratory Mortality. Epidemiology. 17(2): 230-233. 67. Dominici F, Peng RD, Bell ML, Pham L, McDermott A, Zeger SL and Samet JM (2006). Fine Particulate Air Pollution and Hospital Admission for Cardiovascular and Respiratory Diseases. Journal of the American Medical Association. 295(10): 1127-1134. 68. Peel JL, Tolbert PE, Klein M, Metzger KB, Flanders WD, Todd K, Mulholland JA, Ryan PB and Frumkin H (2005). Ambient Air Pollution and Respiratory Emergency Department Visits. Epidemiology. 16(2): 164-174. 69. Trenga CA, Sullivan JH, Schildcrout JS, Shepherd KP, Gail G. Shapiro, Liu L-JS, Kaufman JD and Koenig JQ Effect of Particulate Air Pollution on Lung Function in Adult and Pediatric Subjects in a Seattle Panel Study. 70. Barregard L, Sällsten G, Gustafson P, Andersson L, Johansson L, Basu S and Stigendal L (2006). Experimental Exposure to Wood-Smoke Particles in Healthy Humans: Effects on Markers of Inflammation, Coagulation, and Lipid Peroxidation. Inhalation Toxicology. 18(11): 845-853. 71. Tan WC, Qiu D, Liam BL, Ng TP, Lee SH, van Eeden SF, D'Yachkova Y and Hogg JC (2000). The human bone marrow response to acute air pollution caused by forest fires. Am J Respir Crit Care Med. 161(4 Pt 1): 1213-7. 72. Kocbacha A, Hersethb JI, Låga M, Refsnesa M and Schwarzea PE (2008). Particles from wood smoke and traffic induce differential pro-inflammatory response patterns in co-cultures. Toxicology and Applied Parmicology. 232(2): 317-326. 73. Londahl J, Massling A, Swietlicki E, Brauner EV, Ketzel M, Pagels J and Loft S (2009). Experimentally Determined Human Respiratory Tract Deposition of Airborne Particles at a Busy Street. Environmental Science and Technology. DOI: 10.1021/es803029b.  27 74. Londahl J, Pagels J, Boman C, Swietlicki E, Massling A, Rissler J, Blomberg A, Bohgard M and Sandstrom T (2008). Deposition of biomass combustion aerosol particles in the human respiratory tract. Inhalation Toxicology. 20(10): 923-933. 75. Duclos P, Sanderson L and Lipsett M (1990). The 1987 Forest Fire Disaster in California: Assessment of Emergency Room Visits. Archives of Environmental Health. Jan, 45(1): 53-58. 76. Brauer M and Hisham-Hashim J (1998). Indonesian fires: Crisis and reaction. Env Sci Technol. 32: 404A-407A. 77. Leech J, Burnett R, Cakmak S, Arif M and Chang G (1998). The Sarawak September haze episode. Am J Resp Crit Care Med. 157(3): A260. 78. Hisham-Hashim J, Hashim Z, Jalaludin J, Lubis S and Hashim R (1998). Respiratory function of elementary school children exposed to the 1997 Kuala Lumpur haze. Epidemiology. 9(4 (S1)): S1. 79. Vedal S and Duttonb SJ (2006). Wildfire air pollution and daily mortality in a large urban area. Environmental Research. 102(1): 29-35. 80. Cooper C, Mira M, Danforth M, Abraham K, Fasher B and Bolton P (1994). Acute exacerbations of asthma and bushfires. The Lancet. 343(June 11): 1509. 81. Smith M, Jalaludin B, Byles J, Lim L and Leeder S (1996). Asthma presentations to emergency departments in western Sydney during the January 1994 Bushfires. Int J Epidemiol. 25(6): 1227-1236. 82. Jalaludin B, Smith M, O'Toole B and Leeder S (2000). Acute effects of bushfires on peak expiratory flow rates in children with wheeze: a time series analysis. Aust N Z J Public Health. 24(2): 174-7. 83. Johnston FH, Bailie RS, Pilotto LS and Hanigan IC (2007). Ambient biomass smoke and cardio-respiratory hospital admissions in Darwin, Australia. BioMed Central Public Health. 7(240): doi:10.1186/1471-2458-7-240. 84. Spasoff RA (1999), Epidemiologic Methods for Health Policy New York, New York: Oxford University Press. 85. Chamberlayne R, Green B, Barer ML, Hertzman C, Lawrence WJ and Sheps SB (1998). Creating a population-based linked health database: a new resource for health services research. Canadian Journal of Public Health. 89: 270-273. 86. Coster CD, Quan H, Finlayson A, Gao M, Halfon P, Humphries KH, Johansen H, Lix LM, Luthi J-C, Ma J, Romano PS, Roos L, Sundararajan V, Tu JV, Webster G and Ghali WA (2006). Identifying priorities in methodological research using ICD- 9-CM and ICD-10 administrative data: report from an international consortium. BMC Health Services Research. 6(77): doi:10.1186/1472-6963-6-77.  28 87. Hsia DC, Krushat WM, Fagan AB, J.A. T and Kusserow RP (1988). Accuracy of diagnostic coding for Medicare patients under the prospective-payment system. New England Journal of Medicine. 318: 352-355. 88. Levy AR, Tamblyn RM, Fitchett D, McLeod PJ and Hanley JA (1999). Coding accuracy of hospital discharge data for elderly survivors of myocardial infarction. Canadian Journal of Cardiology. 15(1277-1282). 89. Hux JE, Ivis F, Flintoft V and Bica A (2002). Diabetes in Ontario: determination of prevalence and incidence using a validated administrative data algorithm. Diabetes Care. 25: 512-516. 90. Birman-Deych E, Waterman AD, Yan Y, Nilasena DS, Radford MJ and Gage BF (2005). Accuracy of ICD-9-CM Codes for Identifying Cardiovascular and Stroke Risk Factors. Medical Care. 43(5): 480-485. 91. Aronsky D, Haug PJ, Lagor C and Dean NC (2005). Accuracy of Administrative Data for Identifying Patients With Pneumonia. American Journal of Medical Quality. 20(6): 319-328.   29 2 Chapter 2: Validity and utility of MODIS data for simple estimation of area burned and aerosols emitted by wildfire events* 2.1 Introduction Several climate change scenarios predict that wildfires will become more frequent and intense over the next 50 years1-4, reinforcing international concern about the environmental and public health impacts of smoke from these events.  Atmospheric scientists are working to understand how aerosols emitted by wildfires contribute to the atmospheric changes associated with global warming5-7, and how local, regional and global patterns of precipitation can be affected by smoke particles8-10.  Epidemiologists are devising new tools to assess how acute and chronic exposure to elevated concentrations of smoke-related particulate matter can affect human respiratory and cardiovascular health11-13.  Despite this multi-disciplinary interest in the widespread effects of fire smoke, quantifying the aerosols emitted during wildfire events remains challenging and inexact.  Opportunities to make direct measurements on wildfires are rare due to the spatially and temporally sporadic nature of events.  Studying controlled burns is expensive and logistically complex.  As such, most estimates of emissions from wildfire events come from complex models that are evaluated and revised with available field and laboratory data14-16.  Where the fire size is known, such models use parameters like fuel type, fuel load, fuel moisture content, fire type, elevation, slope, relative humidity and wind speed to estimate the combustion fraction of biomass and the rate at which gases and aerosols are released.  Where only the fire location is known, further modeling is required to simulate event growth and decay.  These emissions and consumption models are constantly being improved, but most remain difficult and time-consuming for non-experts to employ.  Furthermore, the data needed to initialize these models are often differentially measured and recorded between political boundaries, making it challenging to estimate burned area and pollutant emission rates for multiple fires over  * A version of this chapter has been submitted for publication. Henderson SB, Burkholder B, Brauer M, Jackson PL and Ichoku C. Validity and utility of MODIS data for simple estimation of area burned and aerosols emitted by wildfire events.  30 large areas.  A simple and reliable method for making such estimates from one data source could minimize the human and computational resources necessary to estimate wildfire burned areas and emissions rates.  Several satellites carry remote sensing instruments that can detect thermal anomalies on the surface of the earth.  The Moderate Resolution Imaging Spectroradiometer (MODIS) sensors aboard the Aqua and Terra polar-orbiting satellites map fires with a horizontal pixel resolution of 1 km at nadir four times daily (at approximately 01:30, 10:30, 13:30, and 22:30 local time).  Processed attributes for the level 2 fire product (MOD14 for Terra, MYD14 for Aqua) include a confidence rating (where 0 could be a false positive and 100 is certainly a fire) and a fire radiative power measurement (FRP, in MW)17. Here we demonstrate how this product can be used to simply and reasonably estimate the area burned and aerosols emitted by wildfire events, and we assess the utility of the approach by comparing these results to those from already-validated measures and models.  Our study is focused on a 165 000 km2 region of southeastern British Columbia, the westernmost province in Canada.  During the summer of 2003 approximately 2600 km2 of this area was burned by wildfire18.  Several large events occurred in close proximity to densely populated areas, and more than 600 000 residents19 may have been exposed to elevated concentrations of potentially-harmful particulate matter.  Because air quality monitoring in this region is too sparse to support rigorous epidemiological analyses, we modeled individual exposure to elevated aerosol concentrations by simulating smoke production and dispersion throughout the study area20.  The methods described here were developed and tested to optimize and facilitate that process.  2.2 Methods 2.2.1 MODIS fire detection data Data products from MODIS can be requested online through the US National Aeronautics and Space Administration (NASA).  We received a file containing the central latitude and longitude of all MODIS active fire pixels (referred to herein as “MODIS detects”) recorded between July and September 2003, inclusive.  These data  31 were plotted as points in the ArcMap 9.0 (ESRI, Redlands, California) geographic information system (GIS) and projected to a Lambert conformal conic centered over the British Columbia study area to facilitate visualization and geospatial analyses.  We extracted MODIS detects within the area bounded by (-122.0, 53.0) and (-114.0, 48.5) degrees latitude/longitude pairs.  This 325 000 km2 domain includes the British Columbia study area and parts of Alberta, Washington, Idaho and Montana (Figure 2.1). All detects with FRP > 0 were retained for our analyses, regardless of their detection confidence.  2.2.2 Fire event identification Clusters of MODIS detects were manually grouped into discrete fire events (referred to herein as “MODIS events”) with a maximum distance of three kilometers between detects.  Where detects were spatially eligible for inclusion in multiple MODIS events, they were assigned to the event in which other detects most closely matched their date of detection.  Lone detects were kept in the analyses as isolated MODIS events, and all events were assigned a unique identification number.  To assess the validity of this method we compared the MODIS events in British Columbia to records maintained by the provincial Ministry of Forests (MOF), which includes all fires greater than four hectares (referred to herein as “MOF events”).  The MOF database provided start location, burn duration, and final size as estimated by fire crews for all events observed in the province, excluding those in national parks.  Assuming this to be already quality controlled, we manually matched MODIS events to MOF events by location and burn period.   A maximum distance of two kilometers was permitted between MOF start locations and the nearest MODIS detect in a matched event.  An overlap of at least one day was required between MOF and MODIS events.  Analyses were stratified for MOF event size categories of (1) ≥4 to <20 ha, (2) ≥20 to <50 ha, (3) ≥50 to <100 ha, (4) ≥100 to <500 ha, and (5) ≥500 ha.  The number, size and duration of MOF events detected by MODIS were compared to those not detected by MODIS for each event size category.   32 2.2.3 Burned area estimation Simple estimates of the total area burned during each MODIS event were made in GIS. Because the resolution of the MODIS fire product is 1X1 km at nadir, we estimate that each MODIS detect represents 1 km2 (100 ha) of burning area.  In reality the ground sampling area represented by each detect increases with the instrument scan angle (to a maximum of 9.7 km2 at 55 degrees).  However, the median per-pixel fire radiative power (FRP) measurement is relatively consistent across all scan angles, suggesting that burning area is not proportional to sampled area.   By assuming a uniform burned area of 100 ha we avoid falsely inflating the area contributed by off-nadir detects.  A circular buffer with radius 564 meters (1 km2 = 100 ha) was drawn around each detect, and the boundaries between buffers within the same MODIS event were dissolved to yield a single shape whose area was taken as the burned area estimate for that event. Matched MOF and MODIS events were used to compare MOF-recorded and MODIS- estimated burned areas using simple linear regression.  In addition to maintaining its fire database, the MOF employs an Airborne Wildfire Intelligence System (AWIS, Calgary, Alberta) to monitor a subset of fires under aggressive management.  Remote sensing instruments aboard low flying aircraft are used to measure surface temperatures and to produce highly resolved (pixel size less than 5 m) fire perimeter maps for a subset of MOF events (referred to herein as “AWIS events”).  All corresponding MODIS and AWIS events were overlaid to examine how effectively MODIS can detect fire shape and extent.  Topographic differences between areas where MODIS did and did not delineate AWIS perimeters were characterized by analyzing the elevation and slope values at points laid over the AWIS polygons in a 100 meter grid.  2.2.4 Aerosol emissions rate estimation Ichoku and Kaufman21 recently demonstrated that FRP measurements on MODIS detects can be multiplied against fuel-specific coefficients to estimate aerosol emission rates from fire events.  Using the MODIS aerosol product (MOD04 and MYD04 for Terra and Aqua, respectively) and a wind field model they: (1) attributed pixels containing smoke to their source fires, (2) compared the total mass of aerosol in the affected  33 columns to the total FRP of detects in the source fires, and (3) repeated the process for multiple fires worldwide to establish coefficients for different fuel types.  Coefficient values reported for savannah, tropical forests, boreal forests and agricultural material had average values of 61, 65, 29 and 95 grams of aerosol per MJ of FRP, respectively. An average value of 20 g/MJ was reported for Canadian forests.  We applied this finding to our study by estimating an instantaneous emission rate for every MODIS detect as shown in Equation 2.1, where FRPMD is the radiative power of the MODIS detect.  To evaluate the utility of this approach (referred to herein as the “MODIS method”) we estimated emissions rates for all detects with one of the two models described below. For both models we apportioned daily emissions according to the diurnal distribution used for models run by the Western Regional Air Partnership22 as shown in Table 2.1.  The emissions rate for each individual detect was calculated according to its hour of detection. Equation 2.1         20(g/MJ)(MJ/s)FRP(g/s) Rate  Emission MD ×=  Table 2.1 – Diurnal distribution of fire emissions Diurnal distribution (%) of daily particulate matter emissions for the CANFB and USEPM methods. Hour 1 2 3 4 5 6 7 8 9 10 11 12 % in Hour 0.57 0.57 0.57 0.57 0.57 0.57 0.57 0.57 0.57 2.00 4.00 7.00 Hour 13 14 15 16 17 18 19 20 21 22 23 24 % in Hour 10.00 14.00 16.00 17.00 12.00 7.00 4.00 0.57 0.57 0.57 0.57 0.57   Version 4.3 of the Canadian Fire Behavior Prediction System (referred to herein as the “CANFB method”) was used to estimate emissions from all detects within Canada according to the methods described by Li et al.23  This system was developed by the Canadian Forest Service (CFS) as part of the Canadian Forest Fire Danger Rating System.  It estimates consumption rates for 16 fuel categories based on the Fine Fuel Moisture Code (FFMC), local topography and meteorological conditions24.  Category-  34 specific emissions factors are then applied to estimate particle emissions rates.  High resolution maps of the fuel categories in British Columbia and Alberta were provided by the Ministry of Forests (MOF).  Daily maps of interpolated FFMC values at 16:00 (local time) were provided by the CFS and were diurnally adjusted according to the CANFB field guide24.  Hourly meteorological data from 99 weather stations were provided by the MOF and Environment Canada.  Version 1.0 of the Emissions Production Model (referred to herein as the “USEPM method”) was used to estimate emissions from all detects within the USA.  The USEPM was developed by the United States Forest Service25 as part of the BlueSky smoke forecasting framework26.  It estimates aerosol emission rates using fuel type, fuel load and fuel moisture content as its primary inputs.  The USEPM method relies on the US Fuel Characteristic Classification System, which specifies 113 unique fuel codes with density and moisture estimates under multiple conditions.  Maps of the values used by BlueSky for Washington, Idaho and Montana during the study period were provided by the AirFire research branch of the US Forest Service.  To compare the MODIS emissions to the CANFB emissions, instantaneous estimates from each method were summed for all MODIS detects within the Canadian portion of the modeling domain on each of the 92 days in the study period (some days did not have any detections).  The same was done to compare MODIS emissions to USEPM emissions for fires in the American portion of the modeling domain.  Because these methods were tested for use in a model outputting 24-hour averages20, evaluating the comparability of the daily, area-wide emissions is more relevant than evaluating the comparability of point-by-point estimates.  2.3 Results 2.3.1 Fire event identification Figure 2.1 shows 242 MODIS (175 in British Columbia, 67 in neighboring areas) composed of 11004 MODIS detects.  All MOF events are depicted by their size strata. Table 2.2 compares the MOF events in British Columbia detected by MODIS to those not detected by MODIS in terms of their recorded burned area and duration.  MODIS  35 was able to detect 98% of fires greater than 500 hectares, but only 50% of fires less than 500 hectares.  For the one MOF event >500 ha (50 km2) that was not detected, an unmatched MODIS event of similar size and duration was located approximately 10 km away, possibly indicating an error in the MOF start location coordinates.  Of the 85 MODIS events in British Columbia that were not matched to MOF events, 11 were located in national parks beyond the MOF jurisdiction.  Relaxation of our duration and location criteria for clustering MODIS detects would absorb 19 of these cases into 14 already-matched MODIS events. Four further cases could be matched to previously- undetected MOF events if the search radius was expanded from 2 to 5 km and the requirement for a 1-day overlap was relaxed to allow a 2-day time gap.  Of the remaining 51 cases, 35 were isolated, one-day events containing ≤4 detects (sum of detects from all events = 54) with an average confidence value of 62 and a mean FRP measurement of 19 MW.  Assuming the detections to be valid, these events might reflect small, rapidly suppressed wildfires or non-recorded open burning operations. The final 15 cases are multi-day events containing ≥5 detects (sum of detects from all events = 332).  When estimated by our method the mean burned area is 928 ha (range = 172 – 3726 ha).  We assume these events to be wildfires that went unrecorded by the MOF due to their remote locations.  Table 2.2 – Summary of fire size comparison The number, mean size (SD=standard deviation) and mean duration (SD) of MOF events detected by MODIS compared to those not detected MODIS for five MOF size strata.  Detected by MODIS Not detected by MODIS MOF Size Stratum N Mean Size (SD) in ha Duration (SD) in days N Mean Size (SD) in ha Duration (SD) in days ≥ 500 40 4219 (5935) 29 (11) 1 505 7 100 to <500  18 238 (102) 29 (16) 5 211 (80) 15 (13) 50 to <100 9 77 (8) 13 (12) 4 67 (13) 17 (9) 20 to <50 13 33 (9) 10 (8) 8 30 (7) 3 (2) 4 to <20 10 11 (5) 15 (20) 32 8 (4) 7 (10)     36 ## # ### ## ## ## # # ## ### # # # # # # # ## #### ## # # # ## ## # # # ## ## # # # # # ## # # # # # # # ## # # # ## # # ################# ## # ### # # # # # # ## # # # # # ## # # # # ## # # # # # # # ## ### # ## # ###### ## # ### # # #### # # # #### # ##### # # ### ### # # # # # # # ## # ## # ## # # # ## ## # ### # ### # ## # # # # # # ## ### # # # # ## ## # # # # # # ## # ## # # # # # # # ## # # # # # # ### ## # # # # # # # # # ## # ## # # # # # # # # # # # # ## ### # # ## # # # ## ## # # # # # # # ### # # # # # # # # # # # # # # # # # # ## # # # # ## # # # ## # # # ### # # # # # # # # # ## # # ## # # # # # # # # # # #### ## ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # ## # # # # ## # # # # # # # # # # # # # # # # # ## # ### ## # # # # # # # # # # # # # ## # ## # # ## # # # # ## # ## # ## # # # # # # ## # # # # # # ## ## ## # # # # # # # # # ## # ### ## # ## # ### # # ### # # # # ### # # ## # ## ## ### ## # ## ## # # # ## # # # # # # # # # ### ## # # # ## ### # # # ## # # # # # ## # # # # # # # # # # ## # ## # # # # ## # # ### # # ## ## # # # # # # # # # # ## # # # # # # # # # ## # ### ##### # # ### # # # ## #### # ### ### # # # # # # # # ## # ## # # # # ### ## # # # # # # # ## # # # ## # ## # ## # # ### # # # ### # # # # # ## # # # ## ## # # ## # ## # # # # # # # ##### ##### #### ##### # ## ## ## # # # # # ## # # ## # # ### ### # # ## # # # ### # # # # ## ## ## ## #### # # # # # # # ### # # # ### # # # ##### ## ## # # ## # # # # ## ### # # ### ## # ## # # # # ## ### # ## # # # # ## # # # # ## # # # ## ## # ### # # # # # # # # # ## # # # # # # # # # ### ## # # # # ## # # # # # ## ## ## # # # # # # ## # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # ## # # # # # # # # ### # # ## # # # # # #### ## # # # # # ## # # # # # # # #### ## # # # # # ## # # ## ## # # # # # # # # ## # # # # # # # # # # ## # # # # # ### ### # # # # ### ## ## # ## # # # # # # ## # # ## # ## # # # ### # # # # # # # ## ### ## # ## # # #### ## ## # ## ## ## ### ## # # ## # # ### # # ### ## ## # # # # ## # # ## # # ## # # # # ## ## # # # ## # # # # ## ## # # # ## ## # ## # # #### # # # # # ## # ### # # ## ### # #### # # # # # ### # ## ### # # # # # ## # ### ### # # # # # ## # ## # # #### # # ## ## # # # ###### # # # ## # # # ## # # # # ## # # # # # # ## # # ### # # # ### # # ## # # # # # # # ## # # # # ## # ## # # # ## # # ## ### ## ## ## # # ## # ###### # # # # # # ## ### ## ## # # N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N NN N N N N N N N N N N N N N N N N N N N N N N N NN N N N N N N N NN N N N N N N N N N N N N N N N N NN N N N N N N N N N N NN NN N N N N N N N N N N N N N N N N N N N N N N N N 0 50 Kilometers # BC study area MODIS events MOF events N 4 - 20 ha N 20 - 50 ha N 50 - 100 ha N 100 - 500 ha N >500 ha N  Figure 2.1 – Fires in the study area Black areas show the 242 MODIS events in southeastern British Columbia (dark gray), southwestern Alberta, and northern Washington, Idaho and Montana.  White crosses show the start locations of all 140 MOF events recorded in domain.   2.3.2 Burned area estimation Simple linear regression on MODIS burned area estimates versus MOF records for the 90 matched events produces the relationship summarized in Equation 2.2 (R2 = 0.96) and shown in Figure 2.2.  Given that our method assumes a minimum burned area of 100 ha, the overestimation for smaller fires is expected.  For fires larger than 500 ha the mean MOF area burned per MODIS detect is 30.0 ha (SD = 9.3) while the mean MODIS area burned per detect is 37.8 ha (SD = 7.3). Equation 2.2          MODIS area = MOF area X 0.933 + 478 ha Figure 2.3 shows that MODIS burned areas spatially overlap with AWIS perimeters for the 16 events under surveillance in British Columbia.  Figure 2.4 shows the probability distributions of the surface elevations of fire locations where: (1) MODIS and AWIS burned areas overlap, and (2) AWIS burned areas are not captured by the MODIS.  The means (1296 and 1087 meters, respectively) are significantly different (t=-73.2,  37 p<0.0001).  Because fires under AWIS surveillance were also under active suppression, this could reflect suppressed temperatures (below MODIS detection) at the more heavily populated lower elevations.  It could also reflect the ecology of British Columbia’s semi-arid steppe, where quick-burning bunchgrasses are found at the lower elevations27.  The mean surface slope in both cases was 15 degrees, with similar probability distributions (not shown).  The mean confidence values of MODIS detects (1) overall, (2) belonging to AWIS events, and (3) within AWIS perimeters are 78.1, 78.9, and 80.1, respectively, giving little evidence to support the exclusion of detects based on this attribute.   Figure 2.2 – Comparison between measured and estimated fire areas Simple linear regression between estimated and measured burned area for the 90 MODIS events matched to MOF events.  As shown in Equation 2.1: slope = 0.933 (standard error = 0.02) and intercept = 478 (standard error = 100).     38 A B C D # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # N detects: 84 N detects: 55 N detects: 500 N detects: 333 MODIS area: 3693 MODIS size: 2163 MODIS area: 19686 MODIS area: 10984 MOF area: 5730 MOF area: 3300 MOF area: 25912 MOF area: 11395 E F G H # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  ## ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # #  # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  N detects: 33 N detects: 163 N detects: 152 N detects: 949 MODIS area: 1638 MODIS area: 7766 MODIS area: 5077 MODIS area: 27689 MOF area: 1620   MOF area: 7636 MOF area: 4836 MOF area: 26345 I J K L # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  # # # # # # # # # # # # # # # # # # # # # # # # # #  N detects: 286 N detects: 124 N detects: 258 N detects: 26 MODIS area: 11028 MODIS area: 4554 MODIS area: 9334 MODIS area: 1316 MOF area: 10084 MOF area: 4000 MOF area: 7808 MOF area: 1087 N O P Q # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # ## ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #  N detects: 506 N detects: 121 N detects: 155 N detects: 179 MODIS area: 14486 MODIS area: 5597 MODIS area: 8141  MODIS area: 4862 MOF area: 11882 MOF area: 3981 MOF area: 5536 MOF area: 3086 Figure 2.3 – Correspondence between MODIS and AWIS fire detection Black outlines show the fire perimeters last captured for all AWIS events in British Columbia (at different scales).  Underlying grey regions show the estimated burned areas.  Fires A through Q are ranked by the relative agreement between the MODIS and MOF burned area estimates.  Three discrete events have been concatenated in P.  Fires A, B and C burned in close proximity to densely populated areas and were very aggressively suppressed, likely reducing the number of MODIS detects for these events.  39  Figure 2.4 – The relationship between elevation and fire detection Probability distribution of surface elevations in areas where (1) the MODIS and AWIS burned areas agree and (2) the AWIS burned area is not characterized by MODIS.  Mean elevation values (1296 and 1087 meters, respectively) are significantly different (t=-73.2, p<0.0001). Mean slope was 15 degrees in both cases with both probabilities having similar distribution.   2.3.3 Aerosol emissions For the 8221 detects in Canada the MODIS method produced emissions rates with more variability (range =111–35024 , SD = 3035 g/s) than the CANFB method (range = 46–10434, SD = 1329 g/s), but median values were similar (600 and 579 g/s, respectively) for both approaches.  For the 2783 detects in the USA the MODIS method produced more variability and a much higher median rate (range = 107– 34352, SD = 3090 and median = 585 g/s) than the USEPM method (range = 21–2404, SD = 421, and median = 382 g/s). This may be due to the fact that the coefficient used to derive emission rates from the FRP measurements (Equation 2.1) is applicable to the Canadian forests but not necessarily to those in the USA.  Figure 2.5 shows scatter plots comparing daily sums of the emissions rates output by each method in Canada and the USA.   40  Figure 2.5 – Scatter plots comparing emissions models Scatter plots of daily summed emissions rates (in kg/s) from the MODIS method compared to (5A) the CANFB method for 8221 detects in Canada and (5B) the USEPM method for 2783 detects in the USA.  Regression lines are solid and dotted lines show hypothetical 1:1 relationship.   41 2.4 Discussion and conclusions Data from MODIS instruments identified and delineated large wildfires in complex, predominantly wooded terrain.  Areas burned at lower elevation were not as well characterized as those burned at higher elevation, possibly due to fire suppression efforts.  For fires larger than 500 ha, simple estimates of burned area from MODIS data had a strong one-to-one relationship with MOF records.  Wiedinmyer et al.28 used a similar but more complex method for estimating North American fire sizes by limiting the burnable area within each fire pixel based on the land cover classifications it contained. For three fires >25 000 ha they report one case of agreement within 2%, one underestimation by 64% and one overestimation by 82%.  For the two British Columbia fires within the same size range we overestimated one by 5% and underestimated the other by 24%.  In addition to detecting and sizing fires documented by the MOF, MODIS also identified several large events that (1) occurred beyond provincial jurisdictions or (2) went otherwise unrecorded.  This emphasizes the utility and simplicity of using MODIS data for cross-boundary applications and in remote or developing areas without extensive fire monitoring.  Although MODIS detected several sub-pixel (<100 ha) fires, a method for estimating the actual size of within-pixel fires using MODIS data has not been developed.  However, overestimation of the area burned by small fires has negligible impact on the results of dispersion modeling based on the methods used here20.  Emissions estimates from the USEPM method had a lower median and narrower distribution than those from the CANFB and MODIS methods.  A tendency for the USEPM to underestimate emissions has also been documented in validation studies on the BlueSky modeling framework29, 30.  Emissions rates from the CANFB and MODIS methods had similar distributions, but 145 of the 8134 MODIS values were in excess of the 10433 g/s maximum estimated by the CANFB method.  While the MODIS maximum rate of 35024 g/s is realistic for intense fires, it is only reflective of instantaneous conditions at the time of detection.  To simulate time-varying conditions some diurnal distribution of FRP should be assumed, as discussed in detail by Ichoku et al.31 Nonetheless, we have demonstrated that a simple calculation on the FRP attribute of MODIS detects in Canadian forests can produce emissions rates that are comparable  42 to those derived from more complex models.  Extension of these methods to other land cover types could result in a simple and broadly applicable tool for quickly and simply assessing the impacts of smoke from biomass fires.  Several studies have calculated bulk emissions from forest fires by multiplying estimates of burned area from satellite products against fuel-specific emissions factors32-34.  While such approaches may be valuable for improving emission inventories, burned area products created from aggregated data are not available in near real-time for smoke forecasting.  Although the methods described here use retrospective clustering of MODIS detects to estimate the final area burned by MODIS events, we propose that the mean area burned per MODIS detect could be useful (if not ideal) for near real time emissions modeling (MODIS data are available immediately following collection).  In this example MOF fires larger than 500 ha had a burned area of 30.0 ha (SD = 9.3) per each MODIS detect in the fire, and values could be randomly sampled from a similar distribution for smoke forecasting applications.  Modelers worldwide are working to integrate satellite-detected forest fires into air quality predictions26, 34, 35, and the methods described here could easily be adapted for such purposes.                        43 2.5 References 1. Hoelzemann JJ, Schultz MG, Brasseur GP, Granier C and Simon M (2004). Global Wildland Fire Emissions Model (GWEM): Evaluating the use of global area burnt satellite data. Journal of Geophysical Research. 109(D14S04): doi:10.1029/2003JD003666. 2. Flannigan MD, Stocks BJ and Wotton BM (2000). Climate change and forest fires. Science of the Total Environment. 262(3): 221-229. 3. Dale VH, Joyce LA, McNulty S, Neilson RP, Ayres MP, Flannigan MD, Hansom PJ, Irland LC, Lugo AE, Peterson CJ, Simberloff D, Swanson FJ, Stocks BJ and Wotton MB (2001). Climate Change and Forest Disturbances. BioScience. 51(9): 723-734. 4. Williams AAJ, Karoly DJ and Tapper N (2001). The Sensitivity of Australian Fire Danger to Climate Change. Climatic Change. 49(1): 171-191. 5. Kaufman YJ, Tanre D and Boucher O (2002). A satellite view of aerosols in the climate system. Nature. 419(6903): 215-223. 6. Randerson JT, Liu H, Flanner MG, Chambers SD, Jin Y, Hess PG, Pfister G, Mack MC, Treseder KK, Welp LR, Chapin FS, Harden JW, Goulden ML, Lyons E, Neff JC, Schuur EAG and Zender CS (2006). The Impact of Boreal Forest Fire on Climate Warming. Science. 314: 1130-1132. 7. Roeckner E, Stier P, Feichter J, Kloster S, Esch M and Fischer-Bruns I (2006). Impact of carbonaceous aerosol emissions on regional climate change. Climate Dynamics. 27(6): 553-571. 8. Andreae MO, Rosenfeld D, Artaxo P, Costa AA, Frank CP, Longo KM and Silva- Dias MAF (2004). Smoking Rain Clouds over the Amazon. Science. 303(5662): 1337-1342. 9. Kaufman YJ and Koren I (2006). Smoke and Pollution Aerosol Effect on Cloud Cover. Science. 313: 655-658. 10. Koren I, Kaufman YJ, Remer LA and Martins JV (2004). Measurement of the Effect of Amazon Smoke on Inhibition of Cloud Formation. Science. 303: 1342- 1345. 11. Wu J, Winer A and Delfino R (2006). Exposure assessment of particulate matter air pollution before, during, and after the 2003 Southern California wildfires. Atmospheric Environment. 40(18): 3333-3348. 12. Naeher LP, Brauer M, Lipsett M, Zelikoff JT, Simpson CD, Koenig JQ and Smith KR (2007). Woodsmoke Health Effects: A Review. Inhalation Toxicology. 19(1): 67-106.  44 13. Slaughter JC, Koenig JQ and Reinhardt TE (2004). Association Between Lung Function and Exposure to Smoke Among Firefighters at Prescribed Burns. Journal of Occupational and Environmental Hygiene. 1(1): 45-49. 14. Hays MD, Geron CD, Linna KJ, Smith ND and Schauer JJ (2002). Speciation of Gas-Phase and Fine Particle Emissions from Burning of Foliar Fuels. Environmental Science and Technology. 36(11): 2281-2295. 15. Brown D, Dunn W, Lazaro M and Policastro A (1999). The FIREPLUME model: Tool for eventual application to prescribed burns and wildland fires. in Joint Fire Science Conference and Workshop. Boise, ID. 16. Delmas R, Lacaux JP and Brocard D (1995). Determination of biomass burning emissions factors: Methods and results. Environmental Monitoring and Assessment. 38(2-3): 181-204. 17. Giglio L, Descloitres J, Justice CO and Kaufman YJ (2003). An enhanced contextual fire detection algorithm for MODIS. Remote Sensing of Environment. 87(2-3): 273-282. 18. Filmon G (2004) Firestorm 2003, provincial review. Report to the Provincial Government of British Columbia. Firestorm 2003, provincial review. Report to the Provincial Government of British Columbia. Vancouver, BC. 100 pages. 19. Population and dwelling counts for Canada, Provinces and Territies, by Census Divisions, 2001 and 1996 censuses - 100% data., 93F0050XCB01003., Editor. 2002, Statistics Canada. 20. Henderson SB, Burkholder BJ, Jackson PL, Brauer M and Ichoku C (2008). Use of MODIS products to simplify and evaluate a forest fire plume dispersion model for PM10 exposure assessment. Atmospheric Environment. 42(36): 8524-8532. 21. Ichoku C and Kaufman YJ (2005). A method to derive smoke emission rates from MODIS fire radiative energy measurements. IEEE Transactions on Geoscience and Remote Sensing. 43(11): 26-36. 22. WRAP (2007) Inter-RPO National 2002 Wildfire Emissions Inventory - Final Report Inter-RPO National 2002 Wildfire Emissions Inventory - Final Report. Air Sciences & EC/R Project No. 178-7. 23. Li Z, Jin J and Fraser RH (2000). Mapping burned areas and estimating fire emissions over the Canadian Boreal forest. in The Role of Fire in the Control and Spread of Invasive Species. Tallahassee, FL. 24. Taylor SW, Pike RG and Alexander ME (1996) Field guide to the Canadian Forest Fire Behavior Prediction (FBP) System Field guide to the Canadian Forest Fire Behavior Prediction (FBP) System. Canadian Forest Service. Victoria, BC. 83 pages.  45 25. Sandberg DV and Peterson CJ (1984). A source strength model for prescrived fire in coniferous logging slash. Air Pollution Control Association, Pacific Northwest Section. Portland, OR. 26. O'Neill SM, Ferguson SA, Peterson J and Wilson R (2003). The BlueSky Smoke Modeling Framework. in Fifth symposium on fire and forest meteorology, American Meteorological Society. Orland, FL. 27. Bai Y, Broersma K, Thompson D and Ross TJ (2004). Landscape-level dynamics of grassland-forest transitions in British Columbia. Rangeland Ecology & Management. 51: 66-75. 28. Wiedinmyer C, Quayle B, Geron C, Belote A, McKenzie D, Zhang X, O'Neill S and Wynne KK (2006). Estimating emissions from fires in North America for air quality modeling. Atmospheric Environment. 40(19): 3419-3432. 29. Fusina L, Zhong S, Koracin J, Brown T, Esperanza A, Tarney L and Preisler H (2007). Validation of BlueSky Smoke Prediction System Using Surface and Satellite Observations During Major Wildland Fire Events in Northern California. in The fire environment—innovations, management, and policy; conference. Destin, FL. 30. Adkins JW, O’Neill SM, Rorig M, Ferguson SA, Berg CA and L. HJ (2003). Assessing Accuracy of the BlueSky Smoke Modeling Framework During Wildfire Events. in Joint meeting of the International Wildland Fire Ecology and Fire Management Congress and the Symposium on Fire and Forest Meteorology. Orlando, FL. 31. Ichoku C, Giglio L, Wooster MJ and Remer LA (2008). Global characterization of biomass-burning patterns using satellite measurements of radiative energy. Remote Sensing of Environment. in press. 32. Roy B, Pouliot GA, Gilliland A, Pierce T, Howard S, Bhave PV and Benjey W (2007). Refining fire emissions for air quality modeling with remotely sensed fire counts: A wildfire case study. Atmospheric Environment. 41(3): 655-665. 33. Isaev AS, Korovin GN, Bartalev SA, Ershov DV, Janetos A, Kasischke ES, Shugart HH, French NHF, Orlick BE and Murphy TL (2002). Using Remote Sensing to Assess Russian Forest Fire Carbon Emissions. Climatic Change. 55(1): 235-249. 34. Pouliot GA, Pierce T, Benjey W, O'Neill SM and Ferguson SA (2005). Wildfire Emission Modeling: Integrating BlueSky and SMOKE. in International Emission Inventory Conference. Las Vegas, NV.     46 35. Reid JS, Prins EM, Westphal DL, Schmidt CC, Richardson KA, Christopher SA, Eck TF, Reid EA, Curtis CA and Hoffman JP (2004). Real-time monitoring of South American smoke particle emissions and transport using a coupled remote sensing//box-model approach. Journal of Geophysical Research. 31(L06107): doi:10.1029/2003GL018845.    47 3 Chapter 3: Use of MODIS products to simplify and evaluate a forest fire plume dispersion model for PM10 exposure assessment* 3.1 Introduction Human exposure to potentially-harmful forest fire smoke pollution is challenging to assess with data from air quality networks.  Monitors tend to be sparsely distributed in remote areas, so measurements are rarely made in small, heavily-impacted communities.  Still, several studies have used air quality networks to measure ambient particulate matter concentrations during forest fire events.  Here we focus on PM10 (particles less than 10 microns in aerodynamic diameter) as the most commonly- measured pollutant, though most particles from biomass fires fall into the PM2.5 fraction1.  Peak 24-hour averages of 250, 270, 285, 375, 440, 700, and 930 µg/m3 have been reported for British Columbia2, Lithuania3, Singapore4, California5, Brunei6, the Amazon Basin7, and Malaysia8, respectively.  Such high concentrations are likely to have measurable public health impacts, and simple tools are needed to improve epidemiological exposure estimation, to facilitate risk assessment, and to support public preparedness.  In recent years there has been interest in using quantitative and qualitative remote sensing products to supplement air quality network data9-12.  Unlike surface monitors, satellite borne sensors can provide gridded arrays of measurements over vast geographic areas.  Of particular interest are the aerosol products generated by Moderate Resolution Imaging Spectroradiometer (MODIS) sensors aboard the US National Aeronautics and Space Administration’s (NASA) Aqua and Terra Earth Observing Satellites (EOS).  The MODIS aerosol optical thickness (AOT) is a measure of the light extinction in cloudless atmospheric columns with a horizontal resolution of 10 km (at nadir).  By modeling the relationship between surface concentrations and columnar coefficients, statistical downscaling can be used to estimate surface conditions wherever MODIS measurements are made.  While this approach shows  * A version of this chapter has been published. Henderson SB, Burkholder B, Jackson PL, Brauer M and Ichoku C (2008). Use of MODIS products to simplify and evaluate a forest fire plume dispersion model for PM10 exposure assessment. Atmospheric Environment.  42: 8524-8532.   48 promise with further refinement, preliminary studies report only moderate correlation (r = 0.31 – 0.70) between AOT and surface measurements9, 11.  Furthermore, the spatial resolution of estimates remains limited to 10 km which, in forest fire applications, is wider than some plumes.  The true color images captured by MODIS instruments are available at much finer resolution (up to 250 m), but they provide no quantitative information about atmospheric aerosols.  However, when not masked by cloud these images clearly show the spatial extent of smoke plumes.  Color variation between smoke-affected pixels can be used to estimate whether the surface air quality is heavily or lightly impacted.  In one novel use of these images Wu et al.13  were able to fill data missing (due to failure or intermittence) from 37 particulate matter samplers in Southern California during the 2003 fire season.  Each pixel underlying the sampler locations was assigned a daily smoke density value of “no smoke”, “light smoke” or “heavy smoke” in a geographic information system (GIS).  Known surface concentrations were then regressed on smoke density and other potentially-predictive variables.  Once missing values were filled, exposure estimates were made for each zip code by spatial interpolation between the samplers.  Although results were favorable, this use of MODIS data remains qualitative and the method is limited to areas with relatively dense monitoring networks.  Pollution dispersion models can provide concentration estimates at very high spatial resolution over large areas.  Although commonly used for air quality forecasting, such models require complex inputs and parameterizations beyond the scope of most public health research.  Furthermore, simulations of forest fire plume dispersion require output from equally-complex sub-models of meteorological conditions, fire spread, and pollutant emission rates.  If the domain spans political boundaries, modeling can be further complicated by inputs that are differentially measured by various authorities. Once a dispersion model is running its performance is challenging to evaluate when adequate data for validation are not available.  Where few monitors exist, the relationship between model output and surface measurements may appear (1) weak due to lack of data; (2) weak due to poor model performance; or (3) deceptively strong due to specific attributes of the monitored locations.  If exposure estimates based on dispersion model output cannot be assessed with data from other validated sources,   49 they may be no more accurate than crude estimates derived only from surface measurements.  Here we describe an approach that makes plume dispersion models more accessible for pubic health applications by simplifying and evaluating them with MODIS fire detection, aerosol, and true color products.  The study area is a mountainous 165 000 km2 region of southeastern British Columbia, in Canada, which experienced the worst fire season in provincial records during the summer of 200314.  More than 640 000 residents were exposed to potentially-harmful smoke pollution, but only six of the thirty- five communities (approximately 370 000 of the residents) were monitored by continuous Tapered Element Oscillating Microbalance (TEOM) PM10 samplers (see Figure 3.1).  The following methods were developed to estimate ambient daily PM10 concentrations for all communities so that the entire population can contribute information to subsequent epidemiological analyses.  We use three simple approaches, each with strengths and weaknesses, to evaluate the overall quality of exposure estimates output by the model.  3.2 Methods All MODIS products were processed with the HDF-EOS to GIS (HEG) conversion tool (Atmospheric Science Data Center, Langley, VA).  Spatial analyses were conducted in ArcGIS 9.0 (ESRI, Redlands, CA), and statistical analyses were done with S-PLUS 7.0 (Insightful Corp, Seattle, WA).  Dispersion of smoke-related PM10 was modeled in hourly time steps with CALPUFF for the period of July 1st through September 30th, 2003.  A sample of the model parameter file is provided in Appendix 2.  Strategies used to provide CALPUFF input and to evaluate CALPUFF output are described in the following sections.  3.2.1 The CALMET/CALPUFF modeling system The CALMET/CALPUFF system (TRC Companies, Lowell, MA) is an advanced non- steady-state meteorological and air quality modeling framework developed by atmospheric scientists. According to guidelines established by the US Environmental Protection Agency it is the preferred system for assessing long range transport of air   50 pollutants.  The CALMET component is a meteorological interpolator that uses observed data and/or output from other models to calculate hourly temperature and wind fields for a three-dimensionally gridded domain15.  The CALPUFF component uses CALMET fields to advect Lagrangian puffs of gases and/or aerosols emitted from modeled sources while simulating dispersion and transformation processes throughout the domain16.  3.2.2 Meteorological inputs Mountainous terrain can produce localized winds that are not reflected in the output from coarse resolution numerical weather prediction models.  To better account for the complex topography of the study area we ran CALMET at a 1 km horizontal resolution, with 12 vertical steps (20 to 5000 m).  As described elsewhere17 the CALMET initialization was tested with: (1) surface measurements; (2) 12 km output from the US National Meteorological Center’s ETA model18 run by SENES Consultants (Richmond Hill, ON); and (3) a combination of both inputs.  All three scenarios performed poorly under low wind conditions, but method (3) produced the best results based on our evaluation criteria, and was used to produce the wind fields for this study19.  3.2.3 Fire location, size and emissions rates Evaluation of the methods used for identifying fires, estimating burned areas and calculating particle emissions rates are reported in Chapter 2.  In brief, the 11004 MODIS fire pixels detected in the study area during the study period (referred to herein as “MODIS detects”) were grouped into 242 discrete fires (referred to herein as “MODIS events”) based on their spatial and temporal proximity.  Assuming that each MODIS detect represents 1 km2 of burned area we drew circular buffers of radius 564 m (area = 0.99 km2) around them and merged buffers within the same MODIS event to obtain polygons approximating the 242 burn scars.  Polygon areas were divided by the number of MODIS detects they contained to calculate the area burned per each detect.  By modeling the daily sequence of MODIS detects, the growth and decay of all events was simulated without the need for fire propagation and fuel consumption models.    51 Based on the 90 MODIS events matched to British Columbia fire size records we overestimate the total burned area by ~32 000 ha (about 20%).  Of this, only ~7000 ha (about 4%) results from overestimation of the area burned by fires >1470 ha (the 75th percentile).  Based on all 242 MODIS events, fires >1470 ha contain 91% of the 11004 detects.  Changes to model output are marginal when fires <1470 ha are excluded from the simulation (not shown), but we retain the smaller fires to ensure that their local effects on air quality are represented.  The particulate emission rate for each MODIS detect was estimated using its fire radiative power (FRP) measurement, which is the radiant component of the total heat energy released by fire within a pixel20, 21.  The same FRP value might represent an intense fire burning a small fraction of the pixel or a cool fire burning a large fraction of the pixel.  Either way, Wooster et al.20 used a large set of measurements on small fires to show a linear relationship between time-integrated FRP and total biomass consumed.  Roberts et al.21 used that result to show that FRP-based calculations of burned biomass agree with in situ measurements.  Given that particles emitted are directly related to biomass consumed, and given that biomass consumed is directly related to FRP, it follows that particle emissions are directly related to FRP.  Ichoku and Kaufman22 demonstrated this by using wind vectors from a global meteorological model to associate smoke-affected columns with specific fires.  The total mass of aerosol in those columns (derived from their AOT) was then regressed on the summed FRP of all MODIS detects in the source fires.  Application of the algorithm to multiple fire events worldwide yields emission coefficients for different fuel types and regions.  Reported values ranged from 20 g/MJ for all Canadian forests to 107 g/MJ for Russian croplands20.  Using this result we calculate emission rates for each MODIS detect as follows. Equation 3.1          Mass (g/s) = 20 (g/MJ) × (ha) 100 (MJ/s) FRPMD  × AMD (ha) Where AMD is the area burned by the MODIS detect and FRPMD is its fire radiative power.  This approach yields emissions estimates that are consistent with the output from complex models.  The Emissions Production Model (US Forest Service, Seattle,   52 WA), was used to estimate heat release rates (used to calculate plume rise) for this study.  3.2.4 CAPLUFF PM10 concentration vs. TEOM measurements Output from dispersion models is traditionally evaluated by comparing concentration estimates to actual measurements, but the spatial coverage of these “gold standard” data may be limited.  For example, surface concentrations of PM10 were measured at only six stations in the study area.  We placed a CALPUFF receptor at each site and simulated PM10 dispersion from all 11004 MODIS detects over the 92-day study period. Mean 24-hour concentration estimates were compared to the daily averages measured at each station using time-series plots and summary statistics (see Appendix 3 for scatter plots).  The Pearson correlation coefficient (r), mean error (ME), mean absolute error (MAE) and index of agreement (IOA) were calculated for each site as suggested by Willmott23.  The latter is defined in Equation 3.2, where Ci is the CALPUFF estimate for day i, and Mi is the corresponding TEOM measurement. Equation 3.2          IOA = ( ) ( )  −+−−− ∑ ∑= = N i N i iiii MMMCMC 1 1 22 /1  3.2.5 CALPUFF PM10 concentration estimates vs. MODIS AOT To better assess model performance around communities without TEOM instruments we use the MODIS AOT measurements.  These values do not reflect surface PM10 concentrations, but they provide a spatially complete measure of the magnitude of atmospheric aerosol.  We split the domain with a 10 km2 grid (the resolution of the AOT product) and placed CALPUFF receptors in 70 of the 141 cells (such that none were adjacent) with a population density of more than 2 persons per km2.  One MODIS AOT raster (from the MYD04 and MOD04 products) showing the entire study area was identified for each day of the study period, and values at the 70 receptors were extracted.  All MODIS events were modeled, and CALPUFF estimates for (1) the hours corresponding to MODIS AOT time stamps and (2) the 24-hour means were modeled at each receptor.  Correlations between AOT values and CALPUFF estimates were calculated with all available pairs (null AOT values due to cloud render the data   53 temporally incomplete) over the 92-day study period.  To put these results in context we also compared AOT values to 1-hour and 24-hour PM10 measurements and CALPUFF estimates at the six TEOM locations.  3.2.6 CALPUFF plume areas vs. MODIS true color images While TEOM and AOT measurements can be used to assess the magnitude of PM10 concentrations output by CALPUFF, they cannot be used to evaluate the spatial accuracy of plume trajectories.  To do this we established test areas around the towns of Kamloops (140 X 100 km), Kelowna (100 X 125 km) and Golden (145 X 125 km) in British Columbia.  Only emissions from the MODIS detects within each test area were simulated, and surface concentrations of PM10 were estimated at 1 km resolution.  Test durations were 21, 40 and 40 days, respectively.  One MODIS true color image (from the MYD02 and MOD02 products, 500 m resolution) showing the entire study domain was identified for each test day (40 in total).  Image colors were inverted to better distinguish smoke from cloud, and visible plumes originating from fires in the test areas were manually traced in GIS (referred to herein as “MODIS plumes”).  Contours of test area CALPUFF estimates >10 µg/m3 were plotted in Surfer 8.0 (Golden Software, Golden, CO) and exported to GIS for hours corresponding to the time stamps of MODIS plumes (referred to herein as “CALPUFF plumes”).  Comparison between the CALPUFF and MODIS plume shapes was made by calculating the sensitivity and specificity of the horizontal overlap between them.  Assuming MODIS plumes to be the “gold standard”, areas covered by both MODIS and CALPUFF plumes were called “truly positive” (or TP), and those covered by neither were called “truly negative” (or TN).  It follows that areas with CALPUFF but no MODIS plumes were “falsely positive” (or FP) and the opposite were “falsely negative” (or FN).  Sensitivity and specificity are defined as TP/(TP+FN) and TN/(TN+FP), respectively24.  Results close to 1.0 for both measures indicate a highly accurate plume, so we define the spread between measures as discrepancy = sensitivity + specificity – 1 to assess this.  Values were calculated for all days on which MODIS plumes were not obscured by cloud.    54 3.3 Results and discussion 3.3.1 CALPUFF PM10 concentration estimates vs. TEOM measurements Figure 3.1 shows the model domain, study area, three plume test areas, locations of six TEOM instruments, 70 CALPUFF receptors for the AOT comparison, and 242 MODIS events made up of 11004 MODIS detects.  Table 3.1 summarizes the attributes of the MODIS detects and MODIS events.  Figure 3.2 shows how CALPUFF estimates compare to the 24-hour average PM10 concentrations in Kamloops, Kelowna, Vernon, Revelstoke, Golden and Creston over the 92-day study period.  In the latter five cases, estimates track well with TEOM measurements, and peak concentrations are realistic. These results are further supported by the correlation, error and agreement measures summarized in Table 3.2.  Given that small increases in ambient particulate matter can be associated with measurable public health effects25, the CALPUFF exposure estimates can be classified into discrete categories like “background”, “low smoke”, moderate smoke” and “high smoke” prior to use in epidemiologic analyses or health risk assessment.  As such, we place greater emphasis on the measures of agreement (r, IOA) than on measures of error (ME, MAE).  Study Area MODIS Detects/Fires Plume Test Areas N Receptors for AOT Comparison #Y TEOM Locations # ### # ### ## ## ## # # # # # #### # # ## # # # ### #### # ## # # # ## ## # # # ## ## # # # ## # # # # # ## # # ## # ### ##### ####### ## ### # # # # # ## # # # # # # # # ### # # # # ## # # ### # # ## ## ###### ## # ## # # #### # # # #### # #### # # ## ### # # # # # # # # ## # ## # # # # ## # ## # # ### ## ### # ## # # # # # # # # # # ## # ### # # ## # # # # # # # ## # ## # # # # # # # # # # # # # # ## # ### ## ## # # # # # # # # ## # ### # # # # # # # # ## ## ## # # # ## ### # # # # ## ### ## # # # # # # # # # # # # # # # # # # # # # # # ## # # # # # # # ## # # # ## # # # # ## ## # # # # # # ## # ## # # # # # # # # # # # # # ##### # # ## # # # ## ## # # # # # # # # # # # # # # # ## ## # # # # # # # # # ## # # # ## # # ## # ## # # # ## # # # # # # # # ## #### # # # # # # # # # # #### # # # ## # # ## ## # ## # ## ## # # # # ## # ## # # ## ### ## ## # # # # # # ## # ## ## ### ##### # # # ## # # # # # # # # # # # ## # # ## # # # ## # # # # # # # # # # # ## # # # # ### ## # # # ## ### # # # # # # # # # ## # # # # # # # # # ## # # ## ##### # # # ## ## # ## # # # # # # ## # ## # # # # # # # # # # ### ###### # # ### # # ## ## # ### ## ### # # # # # # # ## # # # # # # # # # # # ## ## # # # # # # # # # # # # ## ## # ## ## # # ##### # # # # ## ## # # # ## ## # ## # # ### # # # # # # ## #### ## #### ## #### # ## ## ## # ## # # # # # # # # ### ## # ## # # ## # ### # # # ## # # # ## #### ## # # # # # # ### # # # ## # # # ### ## ## ## # # ## # # # # # # ### # # ### ## # ## # # # # ## #### ## # # # # ### # # # ## # # ### ## # ## # # # # # # # ## # # # # ## # # # # # # ## # # # # # # # # # # ## # ## # # # # # # # ## ## # # # # # # # # # # # # # # # # # # # # # # # # # # # # ## # ## # # # ## # # # # # # # ### # ### # # # # # #### ## ## # # ## # # # # # # # # ### # # # # # # # # # #### ## # # # # # # # # ## # # # # # # # # # ## # # # # # ### ## ## # # ### ## ## # ### # # # # # # ## # # ### # # # # ## # # # # # # ## # # ### ### # # ## # ## ##### ## # # ## # ### ## # # ### ## # # ### # # # ### # ## # ## # # # ## # ## # # # # # # # # # ## # # # # ## # # # ## # # # # ## ## # # # # # ##### # ## # # # # # ## # # ### ### ## # # # # # # # # # ## ### # # # # ## # # ### ## ## ## ## # # # # # # # # # # # # # # ## # # ## # # # # # #### ## # # ## ## # # # # # # # # # # # ## # ## ## # # # # # # ### # ### ## ## # # # # # # ### # ### # # # # # # ## # # ## # # # ## # # # # # # # ##### ### # ## # # # #### ## ## # N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N N Y #Y #Y #Y #Y #Y 2 1 3 6 5 4 N 50 0 50 Kilometers  Figure 3.1 – Plume modeling domains. Modeling domain including the study area in southeastern British Columbia, and parts of Alberta, Washington, Idaho and Montana.  Towns with TEOM instruments are (1) Kamloops, (2) Kelowna, (3) Vernon, (4) Revelstoke, (5) Golden and (6) Creston.    55  Table 3.1 – Summary of modeled fires Attributes of fire events, MODIS detects and CALPUFF fires included in our model. Attribute Mean SD* Min Median Max IQR** MODIS event duration (days) 10 13 1 3 68 1 – 19 MODIS event size (ha) 1831 4444 100 400 33178 136 – 1470 MODIS detects per fire event 62 141 1 8 973 2 – 64 MODIS detect FRP (MW) 69.8 152.4 5.4 29.8 1751.2 18.0 – 58.9 Area burned per detect (ha) 40.1 12.3 23.6 37.2 100.0 31.1 – 43.4 Particle emission rate (g/s) 540.4 1204.4 33.36 234.0 18567.4 141.3 – 459.8 Heat release rate (MBTU/s) 180 56 106 166 451 139 – 195 * SD = standard deviation **IQR = interquartile range   Table 3.2 – Summary of CALPUFF versus TEOM results Agreement between 24-hour average CALPUFF estimates and TEOM measurements at six sites over 92 days. Receptor Measure* Kamloops (N=92) Kelowna (N=92) Vernon (N=92) Revelstoke (N=90) Golden (N=86) Creston (N=92) R 0.377 0.731 0.762 0.529 0.530 0.474 ME -9.43 3.21 -0.75 -4.75 -4.00 -5.03 MAE 12.3 20.0 10.5 14.5 12.6 12.2 Central IOA 0.382 0.730 0.859 0.676 0.699 0.683 *R = Pearson’s correlation coefficient.  ME = mean error.  MAE = mean absolute error.  IOA = index of agreement, as defined in Equation 3.2.    56  Figure 3.2 – Time series of CALPUFF estimates. Time series plots showing 24-hour average PM10 concentrations (in µg/m3) as (1) measured by TEOM instruments and (2) estimated by the CALPUFF model in six locations.      57 In a similar study Langmann and Heil26 used the REMO dispersion model to simulate atmospheric transport of aerosols from the Indonesian and Australian fires in 1997/1998.  They estimated weekly burned area from ATSR (Along Track Scanning Radiometer) fire count data, and emissions were calculated for each fire according to the Levine27 estimates of biomass density and combustion efficiency for three fuel categories.  Six months of model output was compared to mean 24-hour measurements of particulate matter at seven sites in Malaysia.  Ambient concentrations were considerably higher than those reported here for British Columbia and the model consistently underestimated measured values.  Time series plots show that estimates tracked well with measurements and reported correlations ranged from 0.44 – 0.73 (as compared to 0.37 – 0.76 in this study).  Among the five sites where CALPUFF performed well in British Columbia, systematic overestimation of surface concentrations is seen only in Kelowna, largely due to one powerful fire that contains 18% of MODIS detects in the 99th percentile of FRP values. Concentrations are underestimated at all sites between August 23rd and September 1st, 2003.  During this period MODIS detected only 1140 fire pixels, compared to 3546 in the preceding ten days and 1882 in the following six days, after which it began to rain throughout the study area.  Furthermore, the mean FRP during this period was only 54.0 MW while mean FRP during the other periods were 69.1 and 64.0 MW, respectively.  This combination of fewer detects and less powerful fires resulted in reduced particle emissions within CALPUFF.  True color images from these days reveal a period of moderate to heavy cloud, which can lead to detection suppression and reduced FRP measurements28, 29.  Researchers continue to develop new algorithms for detecting smaller and less powerful fires with MODIS data29-31, so this limitation may be improved or resolved as enhanced methods become operational.  Concentrations at the Kamloops station are consistently underestimated.  According to administrative records, the Strawberry Hill fire near to Kamloops burned 5730 ha between August 1st and 10th, but our methodology (described in Chapter 2 and Section 3.2.3) estimated only 4073 ha.  When MODIS data are spatially compared to the recorded fire perimeter, detection failure on the down slopes of this topographic feature is evident.  Furthermore, the mean FRP for MODIS detects in this event is only 41.5   58 MW.  Given its close proximity to a large population, this fire was aggressively managed, and we hypothesize that suppression activities resulted in reduced MODIS detection and FRP.  The constant rate of particulate emissions from each MODIS detect (Equation 3.1) is a simplification that may have contributed to the discrepancies observed at Kamloops and elsewhere.  In reality fire intensity varies considerably throughout the diurnal cycle, with fuel consumption and emissions rates generally peaking in the mid-afternoon and slowing during the night32.  In preliminary tests we imposed such hourly variation on PM10 emissions, but found that peak CALPUFF estimates were vastly inflated while low concentrations remained unchanged.  We therefore generalized emissions by representing each MODIS event as a composite of detects from the four daily overpasses.   More recent work by Ichoku et al.33 explicitly characterizes diurnal variation in FRP flux (W/m2), which is a product of the number MODIS detects and the total FRP measured by each overpass.  Plots for most regions show maximum flux occurring at the 13:30 (local time) overpass, though the variability between passes was relatively low throughout Canada33.  3.3.2 CALPUFF PM10 concentration estimates vs. MODIS AOT Table 3.3 shows how 92 days of CALPUFF estimates compare to AOT values at 70 sites in the study area.  Correlations were not sensitive to latitude, longitude, elevation, distance to the nearest large fire, or the number of AOT values available for comparison (N ranges from 38 to 65 out of 92 days).  Although correlations are moderate, they are comparable to those reported elsewhere for the relationship between MODIS AOT and surface concentrations of particulate matter.  For hundreds of sites across the United States, Engel-Cox et al.9 found mean 1-hour and 24-hour correlations of 0.40 and 0.43 (compared to 0.35 and 0.37 here), respectively, and showed increasing strengths of association from west to east.  A single pocket of weak and negative values was observed in the Pacific Northwest along the British Columbia border.      59 Table 3.3 – AOT measurements versus CALPUFF and TEOM Summary of the correlations between MODIS AOT values and (1) CALPUFF estimates at 70 receptors distributed over populated areas, (2) CALPUFF estimates at the locations of six TEOM instruments, and (3) TEOM measurements at those locations. Values for AOT are only available for cloudless pixels, so the number of daily AOT/CALPUFF pairs used in correlation calculations (N pairs) was less than 92 in all cases.  70 Sites CALPUFF 6 Sites CALPUFF 6 Sites TEOM  1-hour* 24-hour** 1-hour 24-hour 1-hour 24-hour N pairs 38 to 65 (out of 92 days) 48 to 60 (out of 92 days) 46 to 60 (out of 92 days)*** Weakest -0.090 -0.074 0.321 0.312 0.448 0.435 Mean 0.349 0.371 0.505 0.478 0.500 0.578 St Dev 0.211 0.187 0.132 0.153 0.042 0.088 Strongest 0.848 0.730 0.688 0.713 0.537 0.673 *Hour nearest to the time of daily AOT capture **Average for the calendar day of AOT capture ***Missing TEOM measurements reduced the number of pairs from 48 to 46 at one site.   Table 3.3 also shows how MODIS AOT values compare to CALPUFF estimates and PM10 measurements at the locations of six TEOM instruments (N ranges from 46 to 60 out of 92 days).  Mean correlations are similar for the CALPUFF and TEOM data, and the ranking from weakest to strongest site is identical in both cases.  Although the sample size is small, this trend association between AOT values, CALPUFF estimates and TEOM measurements suggests that correlation gradients may be useful for qualitative spatial evaluation of model performance.  Furthermore, TEOM measurements are similarly correlated with CALPUFF estimates (r=0.57) and AOT values (r=0.58), which suggests that AOT might also be useful for exposure assessment.  However, the incompleteness of the AOT data renders them less than ideal applications where complete temporal information is required.  Furthermore, the spatial resolution of the AOT product is 10 km2 whereas the resolution of CALPUFF is established by the user (1 km2 in this case).  These issues highlight limitations of AOT for public health applications, though we do intend to investigate their utility as a measure of smoke exposure for epidemiological research.     60 3.3.3 CALPUFF vs. MODIS plumes Table 3.4 summarizes the spatial comparison between CALPUFF and MODIS smoke plumes around Kamloops, Kelowna and Golden.  Weak sensitivities suggest that our model has only a moderate probability of predicting smoke where smoke is actually observed.  Performance is poorest on days with low winds and small plumes.  If all sensitivity values are multiplied by the size of the truly positive plume, summed, and divided by the total truly positive plume area (i.e. giving large, correct plumes more weight) the averages become 0.70, 0.59 and 0.55 for Kamloops, Kelowna and Golden, respectively.  Strong specificities suggest that CALPUFF reliably does not predict smoke in unaffected areas, though false positives are large compared to true positives and false negatives.  This is partially due to the 10 µg/m3 cutoff resulting in CALPUFF plumes that are systematically larger than their MODIS counterparts.  We established this value by comparing 10 test MODIS plumes to CALPUFF plumes with 5, 10, 20 and 50 µg/m3 contours and assessing which concentration best approximated the manually- traced plume sizes.  Although the 10 µg/m3 cutoff resulted in minimal estimation error, the test cases were not randomly selected – MODIS plumes from cloudless days were preferentially chosen because we assumed they were more accurate.  While color manipulation can help to differentiate smoke from cloud, areas with lower aerosol concentrations may not be clearly visible on partially cloudy days.  Therefore the size of each MODIS plume may be inversely proportional to the degree of cloudiness at its time of capture, resulting in a systematically flawed “gold standard”.  Analysts at NOAA use imagery from GOES (geostationary operational environmental satellites) to manually identify smoke plumes for the North American fire hazard mapping system† (referred to herein as “HMS plumes”).  When qualitatively compared to our MODIS plumes the HMS plumes were generally larger, confirming that our method underestimated plume size.  To assess CALPUFF performance against these HMS plumes we dissolved the individual polygons into a single daily shape, and we output 24-hour plumes from CALPUFF with a 5 µg/m3 cutoff.  Limited availability of the HMS data reduced the number of comparable dates to 22 of 40, 12 of 21 and 16 of 40 for Kamloops, Kelowna and Golden, respectively.  Specificities were marginally  † http://www.firedetect.noaa.gov/viewer.htm   61 improved in all cases (Kamloops = 0.93, Kelowna = 0.93, Golden = 0.94) due to an overall decrease in falsely positive plume area.  Sensitivity was increased to 0.61 and 0.49 in Kelowna and Golden, respectively, but remained virtually unchanged at 0.58 for Kamloops.  It appears that daily averaging smoothed directionality errors for CALPUFF plumes in the complex terrain of the first two test areas, but had less impact in the lower elevations around Kamloops.  Regardless, the incidence of falsely negative plume area remained high in all test areas suggesting error between the observed and modeled plume trajectories.  Table 3.4 – Summary of plume evaluation results Summary of the spatial comparison between MODIS and CALPUFF plumes. Numbers in parentheses are standard deviations with N samples. Mean Measure* (SD) Kamloops (N=30/40) Kelowna (N=16/21) Golden (N=24/40) MODIS plume area (km2) 1014 (834) 1605 (1149) 1169 (1127) CALPUFF plume area (km2) 1712 (1484) 1957 (1368) 1948 (1787) Truly Positive area (km2) 615 (661) 821 (780) 534 (770) Falsely Negative area (km2) 399 (328) 784 (541) 635 (616) Falsely Positive area (km2) 1096 (1170) 1137 (1047) 1414 (1178) Truly Negative area (km2) 11894 (1556) 9768 (1516) 14790 (3756) Sensitivity 0.57 (0.29) 0.46 (0.23) 0.42 (0.24) Specificity 0.91 (0.09) 0.90 (0.09) 0.91 (0.08) Discrepancy 0.53 (0.25)  0.64 (0.20) 0.74 (0.16) * Sensitivity = TP/(TP+FN); Specificity = TN/(TN+FP); Discrepancy = Sensitivity + Specificity – 1   Some error is attributable to the CALPUFF plumes being based on estimates of surface concentration while the MODIS and HMS plumes are traced from a birds-eye view that is impossible to replicate with model output.  However, poor overlap between plume trajectories is most obvious during times of low and moderate wind, when CALMET performance is known to be poor.  Initializing CALMET with output from a numerical weather model running at higher resolution (2 or 4 km instead of 12 km) may produce better results in complex terrain, though surface measurements alone might be adequate in areas with simpler topography.   62  Results still compare well to those reported elsewhere.  The Air Resources Laboratory at NOAA models wildfires detected by its hazard mapping system (from which the HMS plumes were extracted) with the HYSPLIT dispersion model to make daily estimates of smoke-related elevations in PM2.5 across the United States.  Emissions estimates are derived from the BlueSky framework for modeling smoke from prescribed burns. Output from the model is evaluated daily using a statistical coefficient called the Figure of Merit in Space (FMS) as described by Klug et al.34, which is similar to the sensitivity and specificity calculations used here.  At multiple concentration contours the overlapping area between two plumes is divided by the total area covered by both plumes, with values ranging from 0 (no agreement) to 1 (perfect agreement).  Mean FMS values for July through September 2006 at the 1, 5, 20 and 100 µg/m3 contours were 0.14, 0.11, 0.06 and 0.02, respectively‡.  In contrast, the mean FMS values for Kamloops, Kelowna and Golden are 0.27, 0.26 and 0.17, respectively.  3.4 Conclusions Dispersion models of fire smoke can produce estimates at the spatial and temporal resolution necessary for public health applications, but this complex process is challenging to simulate and model output needs rigorous evaluation.  The simplified approach presented here produces results comparable to those found in more sophisticated work, but all studies report considerable error between observed and output data under some conditions.  Using the measurement, spatial and temporal strengths of different data sets allows for straightforward and holistic evaluation of model performance.  Our methods can provide public health researchers with simpler, more accessible and globally applicable tools for enhancing smoke exposure assessment and, therefore, advancing understanding of health risks associated with forest fire smoke.     ‡ http://www.arl.noaa.gov/smoke/MonthlyFMS.pdf   63  3.5 References 1. Andreae MO and Merlet P (2001). Emission of trace gases and aerosols from biomass burning. Global Biogeochemical Cycles. 15(4): 955-966. 2. Moore D, Copes R, Fisk R, Joy R and al e (2006). Population Health Effects of Air Quality Changes Due to Forest Fires in British Columbia in 2003: Estimates from Physician-visit Billing Data. Canadian Journal of Public Health. 97(2): 105. 3. Ovadnevaite J, Kvietkus K and Marsalka A (2006). 2002 summer fires in Lithuania: Impact on the Vilnius city air quality and the inhabitants health. Science of The Total Environment. 356(1-3): 11-21. 4. Nichol J (1997). Bioclimatic impacts of the 1994 smoke haze event in Southeast Asia. Atmospheric Environment. 31(8): 1209-1219. 5. Phuleria HC, Fine PM, Zhu Y and Sioutas C (2005). Air quality impacts of the October 2003 Southern California wildfires. J. Geophys. Res. 110: D07S20. 6. Radojevic M and Hassan H (1999). Air quality in Brunei Darussalam during the 1998 haze episode. Atmospheric Environment. 33(22): 3651-3658. 7. Artaxo P, Gerab F, Yamasoe M and Martins J (1994). Fine mode aerosol composition at three long-term atmospheric monitoring sites in the Amazon Basin. Journal of Geophysical Research. 99(D11): 22857-22868. 8. Brauer M and Hisham-Hashim J (1998). Indonesian fires: Crisis and reaction. Env Sci Technol. 32: 404A-407A. 9. Engel-Cox JA, Holloman CH, Coutant BW and Hoff RM (2004). Qualitative and quantitative evaluation of MODIS satellite sensor data for regional and urban scale air quality. Atmospheric Environment. 38(16): 2495-2509. 10. Hutchison KD (2003). Applications of MODIS satellite data and products for monitoring air quality in the state of Texas. Atmospheric Environment. 37(17): 2403-2412. 11. Wang J and Christopher SA (2003). Intercomparison between satellite-derived aerosol optical thickness and PM2.5 mass: Implications for air quality studies. Geophysical Research Letters. 30(21): doi:10.1029/2003GL018174. 12. Al-Saadi J, Szykman J, Pierce RB, Kittaka C, Neil D, Chu DA, Remer L, Gumley L, Prins E, Weinstock L, MacDonald C, Wayland R, Dimmick F and Fishman J (2005). Improving national air quality forecasts with satellite aerosol observations. Bulletin of the American Meteorological Society. 86(9): 1249-1261.   64 13. Wu J, Winer A and Delfino R (2006). Exposure assessment of particulate matter air pollution before, during, and after the 2003 Southern California wildfires. Atmospheric Environment. 40(18): 3333-3348. 14. Filmon G (2004), Firestorm 2003: Provincial Review Vancouver, BC. 100 pages. 15. Scire J, Robe F, Fernau ME and RJ Y (2000), A users guide for the CALMET meteorological model Concord, MA: Earth Tech Inc. 16. Scire J, DG S and RJ Y (2000), A users guide for the CALPUFF dispersion model Concorde, MA: Earth Tech Inc. 17. Burkholder BJ (2005) Report on the spatial assessment of forest fire smoke exposure and its health effects -- Part I: Initialization of the CALMET meteorological model. School of Occupational & Environmental Hygiene, The University of British Columbia. Vancouver, BC. 46 pages. http://hdl.handle.net/2429/13646 18. Black TL (1994). The New NMC Mesoscale Eta Model: Description and Forecast Examples. Weather and Forecasting. 9(2): 265-278. 19. Burkholder BJ (2005) Report on the spatial assessment of forest fire smoke exposure and its health effects -- Part II: CALMET initialization methodology. School of Occupational & Environmental Hygiene, The University of British Columbia. Vancouver, BC. 28 pages. http://hdl.handle.net/2429/13645 20. Wooster MJ, Roberts G, Perry GLW and Kaufman YJ (2005). Retrieval of biomass combustion rates and totals from fire radiative power observations: FRP derivation and calibration relationships between biomass consumption and fire radiative energy release. Journal of Geophysical Research. 110(D24311): doi:10.1029/2005JD006318. 21. Roberts G, Wooster MJ, Perry GLW, Drake N, Rebelo LM and Dipotso F (2005). Retrieval of biomass combustion rates and totals from fire radiative power observations: Application to southern Africa using geostationary SEVIRI imagery. Journal of Geophysical Research. 110(D21111): doi:10.1029/2005JD006018. 22. Ichoku C and Kaufman YJ (2005). A method to derive smoke emission rates from MODIS fire radiative energy measurements. IEEE Transactions on Geoscience and Remote Sensing. 43(11): 2636-2649. 23. Willmott CJ (1982). Some Comments on the Evaluation of Model Performance. Bulletin of the American Meteorological Society. 63(11): 1309-1313. 24. Fielding AH and Bell JF (1997). A review of methods for the assessment of prediction errors in conservation presence/absence models. Environmental Conservation. 24: 38-49. 25. Dockery DW and Pope CA (1994). Acute Respiratory Effects of Particulate Air Pollution. Annual Review of Public Health. 15: 107-132.   65 26. Langmann B and Heil A (2004). Release and dispersion of vegetation and peat fire emissions in the atmosphere over Indonesia 1997/1998. Atmospheric Chemistry and Physics. 4: 2145-2160. 27. Levine JS (1999). The 1997 fires in Kalimantan and Sumatra, Indonesia: Gaseous and particulate emissions. Geophysical Research Letters. 26(7): 815- 818. 28. Kaufman YJ, Ichoku C, Giglio L, Korontzi S, Chu DA, Hao WM, Li R-R and Justice CO (2003). Fire and smoke observed from the Earth Observing System MODIS instrument--products, validation, and operational use. International Journal of Remote Sensing. 24(8): 1765-1781. 29. Giglio L, Descloitres J, Justice CO and Kaufman YJ (2003). An enhanced contextual fire detection algorithm for MODIS. Remote Sensing of Environment. 87(2-3): 273-282. 30. Morisette JT, Giglio L, Csiszar I, Setzer A, Schroeder W, Morton D and Justice CO (2005). Validation of MODIS Active Fire Detection Products Derived from Two Algorithms. Earth Interactions. 1-25. 31. Li Y, Vodacek A, Kremens RL, Ononye A and Tang C (2005). A Hybrid Contextual Approach to Wildland Fire Detection Using Multispectral Imagery. Geoscience and Remote Sensing, IEEE Transactions on. 43(9): 2115-2126. 32. Beck A, Alexander ME, Harvey SD and Beaver AK (2002). Forecasting diurnal variation in fire intensity to enhance wildland firefighter safety. International Journal of Wildland Fire. 11(33): 173-182. 33. Ichoku C, Giglio L, Wooster MJ and Remer L (2008). Global characterization of biomass burning patterns using satellite measurements of radiative energy. Remote Sensing of Environment. 112(6): 2950-2962. 34. Klug W, Graziani G, Grippa G, Pierce D and Tassone C (1992), Evaluation of long term atmospheric transport models using environmental radioactivity data from Chernobyl accident. The ATMES Report: Elsevier. 366.   66 4 Chapter 4: Three measures of forest fire smoke exposure and their association with respiratory and cardiovascular health outcomes in a dynamic population-based cohort* 4.1 Introduction Forest fire smoke is a globally significant source of particulate matter (PM) pollution1, but its health effects are challenging to assess epidemiologically.  Prospective study designs are infeasible because fires are spatially and temporally sporadic, and retrospective designs are limited by crude exposure estimates based on sparse air quality measurements and/or subject self-report.  Regardless, studies using such data have consistently detected some association between forest fire smoke and respiratory health based on physician/emergency room visits, hospital admissions and symptom questionnaires2-14.  Less consistency has been observed for cardiovascular outcomes. Although we know that (1) acute exposure to woodsmoke triggers a systemic inflammatory response15-17 and (2) chronic exposure to urban fine particulate matter increases cardiovascular morbidity and mortality18-20, observational studies of acute exposure to ambient fire smoke have generally reported null findings for cardiovascular outcomes4, 6, 21, 22.  Enhanced exposure assessment may improve our ability to detect the effects of fire smoke on cardiovascular health if a relationship actually exists.  Recently Wu et al.23 used remote sensing data to enhance smoke exposure assessment for residents of Los Angeles, California.  During the 2003 fire season several monitors in the city’s dense regularity network sampled intermittently or failed altogether.   Satellite images were used to assign each sampler location a daily smoke density value of “no smoke”, “light smoke” or “heavy smoke”, and known surface concentrations were regressed on smoke density and other potentially-predictive variables.  Significant associations between PM2.5 (PM less than 2.5 microns in aerodynamic diameter) and hospital admissions for several respiratory outcomes were reported, while no relationship with cardiovascular outcomes was observed24.  * A version of this chapter has been submitted for publication. Henderson SB, Brauer M, MacNab YC and Kennedy SM. Three measures of forest fire smoke exposure and their association with respiratory and cardiovascular health outcomes in a dynamic population-based cohort.  67 During the same 2003 fire season more than 2600 km2 of forest were consumed in the southern interior region of British Columbia (BC), Canada.  A total of 343 homes were destroyed25 and approximately 640 000 residents were potentially exposed to smoke pollution26.  In particular, the cities of Kelowna (population ~105 000) and Kamloops (population ~80 000), were both affected by large and sustained fires burning on their outskirts.  Moore et al.6 used air quality data from the TEOM (tapered element oscillating microbalance) PM monitors in each city to investigate associations with respiratory and cardiovascular physician billing records.  A smoke-related increase in respiratory visits for Kelowna was reported, but no other associations were found.  We extended this analysis and took advantage of extensive linked administrative data records to employ more rigorous epidemiologic methods.  Health care in Canada is publicly funded, and researchers can access individually- linked administrative records for publicly valuable health studies27.  Here we identify a population-based cohort of individuals in the southern interior region who (1) regularly use the health care system and (2) had geocodable residential addresses on file for the summer of 2003.   We explore how three measures of forest fire smoke exposure are associated with physician visits and hospital admissions for respiratory and cardiovascular disease: (1) 24-hour average PM10 concentration at the nearest of six TEOM instruments in the 150 000 km2 study area.  These measurements are accurate but their spatial scarcity results in exposure misclassification for distant subjects; (2) An indicator of smoke plume coverage based on two daily satellite images.  This dichotomous variable is spatially extensive but provides no quantitative information about surface PM10 concentrations; (3) 24-hour average PM10 concentration estimated from a smoke dispersion model28.  The model output is quantitative and spatially extensive, but uncertain due to input errors and simulation assumptions.  We hypothesize that simultaneous consideration of epidemiologic results based on all three metrics may develop a more comprehensive understanding about the relationship between forest fire smoke and public health.  68 4.2 Methods 4.2.1 Administrative health data The BC Ministry of Health (MOH) maintains records of all health care charges incurred by every personal health number (PHN) in the province and researchers are able to apply for access to anonymized information from a series of linked databases. Approval is granted by the MOH on a case-by-case basis after the public health merits of the application have been assessed.  The only spatial attribute held in the health care billings file is a 6-digit postal code, but a historical record of residential addresses is retained in the MOH client registry.  Health care users are asked to confirm their last known address every time they use their PHN, and changes are flagged for amendment in this master file.  When people move residences their addresses are incorrect in the client registry until (1) they contact the MOH directly to update their records, (2) they update their address when they next use their PHN, or (3) their employers update their address when paying annual fees for premium insurance.  When none of these actions are taken the addresses in the client registry will remain indefinitely incorrect.  Because exposure assignment for this study is based on residential address, we endeavored to minimize misclassification by restricting the cohort to PHNs with regular billings from the same 6-digit postal code. Only PHNs with at least one billing from that postal code in (1) the year before the study period and (2) they year after the study period were eligible, thereby maximizing our confidence that the correctness of the postal code had been confirmed by method (2) above,  4.2.2 Cohort identification and geolocation The study population was initially restricted to postal codes in the southern interior of BC.  An individual PHN was eligible for inclusion in the cohort only if the 6-digit postal code associated with its last use (birth, MSP billing or hospitalization) in the year before the study period (Jul 1, 2002 – Jun 30, 2003) matched that associated with its first use (death, MSP billing or hospitalization) during the study period (Jul 1 – Sep 31, 2003) or in the following year (Oct 1, 2003 – Sep 31, 2004).  All babies born during the study period were eligible.  Once eligibility was established an individual PHN was only included in the cohort if it was associated with a reliably geocodable (i.e. the highest  69 accuracy ranking on Google’s geocoding utility) residential address in the MOH client registry during the summer of 2003.  Many people in the study area live in rural and semi-rural areas where one 6-digit postal code can cover thousands of square kilometers, so we placed this restriction on the cohort to further protect against exposure misclassification.  To protect the personal privacy of cohort members their spatial information was never linked directly to their health information.  Instead the MOH provided a list of all street addresses in the client registry, and we used the batch geocoding capability of Google Maps to precisely locate the addresses.  Any PHN associated with an address that could not be precisely geolocated (i.e. a Google geocoding score of 2) was dropped. Details of this process are outlined in Figure 4.1.  We then used ArcGIS 9.1 (ESRI, Redlands, California) to impose a 1 km2 grid over the 500×650 km study area, and all geolocated points were aggregated to the center of the 2538 grid cells into which they fell (referred to herein as “exposure cell”).  Exposure estimates for each address in the cell were assigned at its central coordinates, and the list of all addresses was returned to the MOH with corresponding variables.  These included (1) daily values for the three exposure metrics described in the next section, (2) a daily distance-to-fire value based on a straight-line drawn between the cell and the edge of the nearest fire and (3) a single distance-to-nearest-TEOM measurement.  The MOH used the address field to link study PHNs with their exposure variables, and then stripped all spatial information from the file before returning it to us for linkage with the health data.  4.2.3 Event definition Fields in the provincial Medical Services Plan (MSP) physician visit file include the PHN, date of service, service code and one ICD9 code.  For these analyses we retained only those records with (1) a service code indicating that the billing was for a physician visit, and (2) an ICD9 code in the respiratory (460-519) or cardiovascular (390-459) categories.  For each PHN in the cohort we (1) collapsed any visits occurring on the same day and within the same ICD9 category into a single event, and (2) deleted any event occurring within seven days of a preceding event to protect against over-counting a single episode.  A binary status indicator (1 = event, 0 = no event) was assigned to every PHN for each of the 92 days in the study period.  Hospital admissions are  70 recorded in a different file with fields including the PHN, date of admission and one or more associated ICD9 of codes.  We treated these data in exactly the same way as the MSP data based on the primary ICD9 code.  Figure 4.1 – The cohort selection process. Outline of the cohort selection and geolocation process.  Note that the cohort used here (N = 281 711) is the subset of the total cohort (N = 439 942) for which we were able to precisely geocode a residential address.  Exposures based on 6-digit postal code are not used in this study.  71 4.2.4 Exposure metrics TEOM: Provincial and federal environmental authorities maintain PM10 TEOM instruments in the towns of Kelowna, Kamloops, Vernon, Creston, Revelstoke, and Golden.  Fine particulate matter (PM2.5) was only measured in Kelowna and Kamloops, but accounted for ~75% of PM10 concentrations on fire days6.  We received hourly measurements from the BC Ministry of Environment and used them to calculate midnight-to-midnight 24-hour average concentrations at each site.  All subjects were assigned daily concentrations from the TEOM nearest to their exposure cell. Calculations were made using the “path distance” function of ArcGIS, which accounts for the vertical and horizontal aspects of the topography between points.  For example, an exposure cell in a valley would be considered nearer to a TEOM in the same valley than to a TEOM on the other side of a mountain, even if the Euclidean distance (as the bird flies) was longer.  Although TEOM measurements reflect ambient particulate matter from all sources, distinctive peaks are evident during fire episodes6.  Effects associated with an increase of 30 µg/m3 (approximately one standard deviation in the complete set of values for 92 days × 281711 subjects) are reported for most regression analyses to facilitate comparison between metrics.  SMOKE: The US National Oceanic and Atmospheric Administration (NOAA) Hazard Mapping System (HMS) provides GIS-ready polygon files of fire smoke plumes based on multiple reviews and manual tracing of satellite images each day†. These files were overlaid on the study area.  For days on which the centre of an exposure cell fell within the plume outline all subjects within that cell received a value of 1, and otherwise received a value of 0.  While these polygons are spatially extensive they only provide a birds-eye-view of atmospheric conditions and they contain no information about aerosol concentrations.  Effects associated with the difference between smoke/no-smoke are reported for all regression analyses.  CALPUFF: We previously employed fire detection data from MODIS (Moderate Resolution Imaging Spectroradiometer) instruments to estimate particulate emissions in a CALMET/ CALPUFF smoke dispersion model28.  Simulated ambient 24-hour concentrations were used to estimate smoke-related PM10 exposure at the center of  † http://www.osdpd.noaa.gov/ml/land/hms.html  72 each exposure cell for all subjects within that cell.  Evaluation of model performance established that output was most reliable under conditions of high wind28, but all estimates are uncertain due to errors in input (emissions estimates, wind vectors output from the CALMET meteorological model, etc.) and simulation assumptions.  Effects associated with an increase of 60 µg/m3 (approximately one standard deviation in the complete set of 92×281711 values) are reported for all regression analyses.  4.2.5 Other covariates In addition to the time-varying exposure estimates, health outcome indicators and distance-to-fire variable, our data set included the following non-time-varying covariates for each cohort member: age and sex, as recorded in the client registry; socioeconomic quintile of their residential census dissemination area; distance to the nearest TEOM in km; and number of physician visits (0, 1-2, 3-5, 6+) in each ICD9 category (one variable for respiratory, another for cardiovascular) in the year prior to the study period (July 1st 2002  - June 30th 2003).  The latter was designed to control for people who visit their physicians multiple times annually, which is a potential indicator of pre-existing sensitivity to fire smoke.  A day-of-week indicator (0=weekend/holiday, 1=Monday, etc.) was also assigned to all outcomes.  4.2.6 Statistical analyses Logistic regression with repeated measures29 was used to estimate the independent fixed effects of TEOM, SMOKE and CALPUFF on (1) all respiratory physician visits, (2) all respiratory hospital admissions, (3) all cardiovascular physician visits and (4) all cardiovascular hospital admissions.  More specific analyses were conducted on physician visits for asthma (ICD9 = 493), acute upper respiratory infections (ICD9 = 465 and 466), and non-hypertensive cardiovascular diagnoses (ICD9 = 410-459). Coefficients were calculated with generalized estimating equations (GEE) using the GENMOD procedure in SAS 9.1 (SAS Institute, Cary, North Carolina) and assuming a uniform correlation structure.  Equation 4.1 outlines the generalized model, where: Pit = Pr(Yit =1); Yit is the binary outcome for individual i on day t; β0 is a single, fixed intercept; Z1 through ZQ are non-time-varying explanatory variables; X is a time-varying exposure variable.  73 Equation 4.1         Logit(Pit) = β0 + (β1Z1i + … + βQZQi) + β2Xit Because models including the entire cohort took approximately 48 hours to run (quad- core Intel® Xeon® 3500) many analyses were restricted to the subset of persons with three or more respiratory/cardiovascular physician visits in the year before the study period.  This allowed us to maximize the event rate in each model (see Table 4.1) while testing it on a potentially sensitive sub-population.  In the following section we show how the estimated odds ratios are affected by (1) adjustment for the aforementioned covariates, (2) stratification on the distance-to-fire variable, (3) consideration of multiple time lags (day t, day t-1, and a combination of days t and t-1), and (4) stratification on 10-year age categories.  4.3 Results 4.3.1 Data summary Figure 4.1 summarizes the methods used to define the cohort and to identify the 281 711 members PHNs included in this study.  Table 4.1 summarizes the overall and stratified outcome rates for respiratory and cardiovascular physician visits and hospital admissions in the cohort.  The weekday mean number (standard deviation) of physician visits was 484 (80) and 692 (45) for respiratory and cardiovascular ICD9 codes, respectively.  Weekend and holiday (3 during study period) numbers were significantly lower at 147 (32) and 84 (18), respectively.  The mean number of weekday hospital admissions was 6.6 (2.7) and 14.7 (3.4) while the weekend numbers were 4.9 (2.2) and 9.8 (3.0), respectively.  Within week days there was a decreasing gradient in event frequency between Monday and Friday for all outcomes.  Measured 24-hour PM10 concentrations ranged from 5.1 to 248.4 µg/m3 between the six TEOMs.  Mean exposure within the population was 29.4 (30.7) µg/m3 with an interquartile range between 14.1 and 31.0 µg/m3.  Days of SMOKE coverage ranged from 1 to 24 (out of 92) within the cohort, with a mean of 13.5 (2.4).  Modeled PM10 concentrations from CALPUFF ranged from 0 to 2497 µg/m3, with a population mean of 11.4 (61.1) µg/m3 and an interquartile range between 0 and 3.5 µg/m3.  This skewed distribution is due to the fact that most exposure cells had values close to zero (i.e. no  74 smoke) on most days.  Figure 4.2 shows a smooth curve fitted to the complete set of CALPUFF estimates plotted against the distance-to-fire variable.  Concentrations taper to less than 5 µg/m3 at approximately 50 km, and we used this as the value to stratify analyses for all exposure metrics based on the distance-to-fire variable.  Exposure days within 50 km of a fire ranged from 1 to 58 (out of 92) with a mean of 21.3 (7.9).  Figure 4.2 – CALPUFF concentrations vs. distance-to-fire Smooth spline fit to the complete set of CALPUFF estimates (3366 exposure cell receptors over 92 days).  The dotted line shows the 50 km cutoff used for stratification of epidemiologic analyses for all three exposure metrics.  4.3.2 Adjustment and time lags The results presented in Figure 4.3 and 3 are based on the subset of cohort members with three or more respiratory or cardiovascular physician visits in the year prior to the study period.  Figure 4.3 shows (1) the effect of adjustment for covariates (age, sex, socioeconomic status, day-of-week,  and distance-to-TEOM for the TEOM metric) and (2) how the OR is increased for many combinations when analyses are limited to cohort members living within 50 km of a fire on day t (lag 0).  Because we expect the accuracy of the SMOKE and CALPUFF metrics to decrease with increasing distance from source fires, all further analyses were restricted by this 50 km cutoff on the distance-to-fire variable.  Misclassification of SMOKE is more likely in areas where the plumes are  75 dispersed (i.e. less visible in satellite images), and estimates from CALPUFF are less certain with increasing distance from the plume sources when wind fields are inaccurate.  Table 4.1 – Summary of outcome rates Outcome rates (events per 100 000 person-days) for the cohort during the 92-day study period, presented by categories of the ordinal non-time-varying covariates.  Respiratory Cardiovascular Variable Categories % of 281711 Physician Visits Hospital Admissions Physician Visits Hospital Admissions  Overall 100 134.2 (n=34771) 2.2 (n=557) 177.6 (n=46037) 4.7 (n=1208) Sex Male Female 44.6 55.4 129.4 141.6 2.6 1.8 184.6 172.0 5.8 3.8 Age Category  0 < 5 5 < 10 10 < 20 20 < 30 30 < 40 40 < 50 50 < 60 60 < 70 70 < 80 80+ 4.3 5.0 12.3 9.5 11.0 15.4 14.3 11.8 10.3 6.3 251.1 176.6 144.7 141.8 132.4 106.4 100.4 115.4 141.5 160.5 4.6 2.0 0.8 0.8 0.7 1.0 1.0 2.3 4.6 9.1 4.7 2.4 4.7 17.1 34.8 79.8 180.3 335.5 512.7 657.8 0.2 0.1 0.0 0.3 0.6 1.5 3.6 7.4 14.7 22.8 Socio- economic quintile (lowest) 1 2 3 4 (highest) 5 NA 22.7 18.6 18.8 18.4 18.1 2.3 159.6 138.8 133.8 119.9 114.0 130.4 2.9 2.5 2.0 2.0 1.2 2.2 220.1 194.7 168.8 153.4 142.8 183.7 6.1 5.1 4.3 4.1 3.5 4.4  R  C 68.3 78.6 20.5 8.9 7.2 6.0 4.0 6.5 *Number of physician visits in prior year in same ICD9 category  0 1-2 3-5 6+   67.2 163.3 324.4 739.5   1.0 1.6 4.6 20.4   49.7 146.7 529.1 1907.8   2.7 7.4 11.2 30.3  *Respiratory outcome rates are stratified by previous respiratory visits (R) and cardiovascular outcome rates are stratified by previous cardiovascular visits (C).     76   Figure 4.3 – ORs after adjustment and distance constraint. Models include only those cohort members who had three or more physician visits for respiratory or cardiovascular ICD9 codes during the year prior to the study.  Plots show the exponentiated β1 values from the simplified equations below (EXP = exposure, SES = socioeconomic quintile; DOW = day of week; TDIS = distance to TEOM; FDIS = distance to fire; PRE = previous visits). ■ logit(Pit | PREi ≥ 3) = β1EXPit × logit(Pit | PREi ≥ 3) = β1EXPit + β2SEXi + β3AGEi + β4SESi + β5DOWt (+β6TDISi when EXP=TEOM) U logit(Pit | PREi ≥ 3 and FDISit ≤ 50km) = β1EXPit + β2SEXi + β3AGEi + β4SESi + β5DOWt (+β6TDISi) The final model is restricted to persons within 50km of a fire on day t. The ORs reflect (1) a 30 µg/m3 increase in TEOM-measured PM10, (2) the difference between SMOKE and no SMOKE and (3) a 60 µg/m3 increase in CALPUFF-estimated PM10.   77    Figure 4.4 – ORs with exposure lagged by 0, 1 and a combination of 0 and 1.. Models for different time lags include (1) only those cohort members who had three or more physician visits for respiratory or cardiovascular ICD9 codes during the year prior to the study (2) on days when they were within 50km of a fire.  Plots show the exponentiated β1 values from the simplified equations below (EXP = exposure, SES = socioeconomic quintile; DOW = day of week; TDIS = distance to TEOM; FDIS = distance to fire; PRE = previous visits). ■ logit(Pit | PREi ≥ 3 and FDISit ≤ 50km) = β1EXPit + β2SEXi + β3AGEi + β4SESi + β5DOWt (+β6TDISi) × logit(Pit | PREi ≥ 3 and FDISit ≤ 50km) = β1EXPit-1 + β2SEXi + β3AGEi + β4SESi + β5DOWt (+β6TDISi) U logit(Pit | PREi ≥ 3 and FDISit ≤ 50km) = β1EXPi(t,t-1) + β2SEXi + β3AGEi + β4SESi + β5DOWt (+β6TDISi) Highlighted bars indicate which lag was chosen for all further analyses. The ORs reflect (1) a 30 µg/m3 increase in TEOM-measured PM10, (2) the difference between SMOKE and no SMOKE and (3) a 60 µg/m3 increase in CALPUFF-estimated PM10.   78 Figure 4.4 shows the effect of lag specification on the adjusted, distance-restricted odds ratios.  Highlighted bars indicate which lags were selected for further analyses.  Other lags (2 to 7 days) were tested in the exploratory data analyses, but none were associated with any outcome.  Respiratory outcomes were most strongly associated with the two-day mean (days t and t-1) for the TEOM and CALPUFF metrics, and with lag zero (day t) for the SMOKE metric.  All three of the SMOKE lags were binary with 1 indicating coverage on (1) day t, (2) day t-1 and (3) day t and/or day t-1.  Although many of the cardiovascular ORs were less than one, the most positive association was with lag 0 (day t) in five out of six cases.  4.3.3 Stratification by age and previous physician visits The adjusted, distance-restricted results presented in Figure 4.5, Figure 4.6 and Table 4.2 include all cohort members.  Figure 4.5 shows physician visit ORs for the TEOM, SMOKE and CALPUFF metrics stratified by age category.  For respiratory physician visits the largest effect estimates were observed in the 20-70 age categories.  People aged less than 30 have been omitted from the cardiovascular plot due to low event rates (see Table 4.1), and no significant associations were evident.  Figure 4.6 shows the hospital admission results split for cohort members under 20 (respiratory only), between 20 and 60, and over 60 (this split is also due to the low event rates in Table 4.1).  Table 4.2 presents ORs for a 30 µg/m3 increase in TEOM-measured PM10 for (1) people with 3 or more respiratory/cardiovascular physician visits in the year before the study period (a potentially-sensitive subpopulation), (2) people less than three physician visits in the previous year, and (3) the entire cohort.  The ORs for respiratory outcomes were generally greater than one while those for the cardiovascular outcomes were below one.  In both cases there was no difference in the TEOM effects across previous visit categories.     79   Figure 4.5 – ORs for physician visits stratified by age. Physician visit models stratified on age include all members of the cohort within 50km of a fire on day t.  on days when they were within 50km of a fire.  Plots show the exponentiated β1 values from the simplified equations below (SES = socioeconomic quintile; DOW = day of week; TDIS = distance to TEOM; FDIS = distance to fire; PRE = respiratory/ cardiovascular physician visits in year prior to study). ■ logit(Pit | FDISit ≤ 50km) = β1respTEOMi(t,t-1) / β1cardioTEOMit + β2SEXi + β3PREi + β4SESi + β5DOWt +β6TDISi × logit(Pit | FDISit ≤ 50km) = β1SMOKEit + β2SEXi + β3PREi + β4SESi + β5DOWt U logit(Pit | FDISit ≤ 50km) = β1respCALPUFFi(t,t-1) / β1cardioCALPUFFit + β2SEXi + β3PREi + β4SESi + β5DOWt The ORs reflect (1) a 30 µg/m3 increase in TEOM-measured PM10, (2) the difference between SMOKE and no SMOKE and (3) a 60 µg/m3 increase in CALPUFF-estimated PM10.  80   Figure 4.6 – ORs for hospital admissions stratified by age. Hospital admission models stratified on age include all members of the cohort within 50km of a fire on day t.  Plots show the exponentiated β1 values from the simplified equations below (SES = socioeconomic quintile; DOW = day of week; TDIS = distance to TEOM; FDIS = distance to fire; PRE = respiratory/ cardiovascular physician visits in year prior to study). The ORs reflect (1) a 30 µg/m3 increase in TEOM-measured PM10, (2) the difference between SMOKE and no SMOKE and (3) a 60 µg/m3 increase in CALPUFF-estimated PM10. ■ logit(Yit | FDISit ≤ 50km) = (β1respTEOMi(t,t-1) / β1cardioTEOMit) + β2SEXi + β3PREi + β4SESi + β5DOWt +β6TDISi × logit(Yit | FDISit ≤ 50km) = β1SMOKEit + β2SEXi + β3PREi + β4SESi + β5DOWt U logit(Yit | FDISit ≤ 50km) = (β1respCALPUFFi(t,t-1) / β1cardioCALPUFFit) + β2SEXi + β3PREi + β4SESi + β5DOWt   Table 4.2 – Overall cohort results for TEOM exposure The ORs (95% confidence intervals) associated with a 30 µg/m3 increase in TEOM- measured PM10 for all outcomes (1) stratified by previous physician visits and (2) overall. ICD9 Outcome 3 or more  previous visits Less than 3 previous visits Entire cohort Physician Visits 1.05 (1.02 – 1.08) 1.05 (1.03 – 1.07) 1.05 (1.03 – 1.07) Resp. Hospital Admissions 1.18 (0.94 – 1.49) 1.22 (1.01 – 1.46) 1.17 (1.00 – 1.39) Physician Visits 0.98 (0.96 – 1.01) 0.98 (0.97 – 1.00) 0.98 (0.97-1.00) Cardio. Hospital Admissions 0.95 (0.83 – 1.07) 0.83 (0.71 – 0.96) 0.90 (0.81 – 1.00)     81 4.3.4 Specific diagnoses Of the 34771 general respiratory physician visits (see Table 4.1) included in the overall analyses, approximately 16% were coded as “asthma” (ICD9 = 493) and 21% were coded as “acute bronchitis” or “acute upper respiratory infection” (ICD9 = 465 and 466, respectively).  Age-stratified results for these diagnoses based on the TEOM metric are shown in Figure 4.7.  Asthmatic visits are significantly associated with increased PM10 concentrations across most age categories, with ORs highest for the groups between 20 and 70 years.  The overall asthma-specific OR for a 10 µg/m3 increase in TEOM- measured PM10 is 1.08 (95% CI = 1.04 – 1.12).  Observed effects are highest in children less than five years of age, where the OR for a 10 µg/m3 increase is 1.12 (95% CI = 1.07 – 1.18).  There was no association between PM10 and upper respiratory infections.  Similarly, approximately 50% of the 46037 general cardiovascular visits included in the overall analyses were coded as “essential hypertension” (ICD9 = 401), which is unlikely to be associated with smoke exposure.  When analyses for cardiovascular physician visits are repeated excluding ICD9 diagnoses 390 through 409 (acute rheumatic fever, chronic rheumatic heart disease and hypertensive disease), the effects of TEOM, SMOKE and CALPUFF remain null (not shown).   Figure 4.7 – Age-stratified results for specific respiratory diagnoses Results reflect 5199 physician visits for asthma (ICD9 = 493) and 7336 physician visits for acute upper respiratory infections (ICD9 = 465 and 466).  82 4.4 Discussion In general our results are consistent with previous reports, most of which use 10 µg/m3 as a standard elevation in exposure (following most urban air pollution studies).  Overall we found that a 10 µg/m3 increase in the two-day mean of TEOM-measured PM10 was significantly associated with increased odds of respiratory physician visits (OR = 1.007; 95%CI = 1.004–1.009) and hospital admissions (OR = 1.023; 95%CI = 1.000–1.049). Larger effects were observed for asthma, but not for upper respiratory infections, when respiratory analyses were focused on more specific ICD9 diagnoses.  Exposure to PM10 was not associated with any general or specific cardiovascular outcomes.  .In the most similar study published to date,  Delfino et al. also reported that a 10 µg/m3 increase in the two-day mean of PM2.5 was more strongly associated with the relative rate of respiratory hospital admissions in Los Angeles (RR = 1.028; 95%CI = 1.014–1.041) than with cardiovascular admissions (RR = 1.008; 95%CI = 0.999–1.018)24.  Although we are using PM10 and Delfino et al. are using PM2.5, previous work has shown that most PM10 in our study area during the fire season was composed of PM2.56.  Likewise, hospital admissions were increased by 5% for a 10 µg/m3 increase in PM2.5 in Delfino et al.24 while a 10 µg/m3 increase in PM10 increased the odds of an asthma-specific physician visit by 8% in this study population. If the relative risk of an event is equivalent to the change in odds, we estimate that approximately 760 of the 8683 asthma visits (~9%) during the study period were attributable to smoke exposure by assuming a background rate of 80 visits per day (the average for the summers of 2002-2004) and the same number/magnitude of PM10 elevations observed in Kelowna.   More specifically: (1) one day with PM10 elevation of 200 µg/m3 (OR = 4.41, events = 273); (2) two days with elevations of 100 µg/m3 (OR = 2.20, events = 192); (3) ten days with elevations of 30 µg/m3 (OR = 1.25, events = 200); and (4) twelve days with elevations of 10 µg/m3 (OR = 1.08, events = 96).  Given that visit counts in 2002 and 2004 (when fire activity was low) were 6919 and 6388, respectively, this estimate may be too low.  While several previous studies suggest that specific subpopulations are sensitive to the respiratory effects of fire smoke, our results suggest that these effects (1) are similar across the entire the 20-70 age categories and (2) do not differ considerably across subpopulations stratified by potential sensitivity to increased exposure.  Others have  83 found that children24, the elderly14, 24 and people with pre-existing respiratory and/or cardiovascular conditions7, 11, 14 are at higher risk of respiratory outcomes.  This disparity may result from different study designs.  In our dynamic cohort, each member contributed exposure profiles for all eligible person-days (i.e. those spent within 50 km of a fire) to the analyses, regardless of their event status.  Most other studies only count persons experiencing an event.  Given that children, the elderly and other potentially- sensitive individuals have higher baseline event rates (see Table 4.1) such study designs are, by definition, disproportionately weighted towards these subpopulations. By grouping individuals according to their baseline number of physician visits we were able to adjust our effect estimates for these underlying trends (see Appendix 4 for plots showing the daily number of events).  We assume that steep elevations in TEOM-measured PM10 were caused by forest fire smoke.  Although there are no other significant sources of particulate matter in the study area the TEOM metric does reflect ambient concentrations from all sources whereas the SMOKE and CALPUFF metrics reflect forest fire smoke alone.  Based on this and their improved spatial resolution our a priori expectation was that SMOKE and CALPUFF would be more significantly associated with the outcomes than TEOM, but they are not.  In the case of the CALPUFF estimates we know the model performs poorly under low wind conditions28, which would result in non-differential exposure misclassification on such days (when fires are typically smoldering, and concentrations are high).  This is consistent with Figures 4.2 through 4.5, all of which show that the ORs for CALPUFF track with those for TEOM, but that they are more null30.  In the case of SMOKE we expected less spatial misclassification of exposure;  the absolute values of the OR estimates tend to be higher than those for TEOM and CALPUFF, but the confidence intervals are much wider.  Like the CALPUFF estimates they generally track with the TEOM estimates, suggesting that purely satellite-derived smoke exposure estimates may be valuable with further refinement.  Although several studies report no association between fire smoke and cardiovascular outcomes, a recent comprehensive review by Naeher et al. concluded that the available data were insufficient to address such hypotheses21.   In our study area Moore et al. limited their analyses to physician visits6, but most acute cardiovascular outcomes  84 would present as an emergency room rather than outpatient visits.  In their study on Australian bush fires Johnston et al. used a case-crossover design on 992 cardiovascular hospital admissions, and found no overall association with PM10, but that ORs were slightly increased for indigenous people4.  Subsequent to the Naeher review, Hanigan et al. expanded the Australian study to 3443 events and continued to observe no overall relationship, but an increased (yet insignificant) association for the indigenous subpopulation22.  In California Delfino et al. included nearly fifty thousand cardiovascular admissions in their analyses, and reported a small but insignificant increase in events during the fire period24.  Here we include 46307 cardiovascular physician visits and 1208 hospital admissions, but we see no elevated effect estimates for any of the three smoke exposure metrics.  Despite the limited epidemiologic evidence, toxicological studies indicate that biomass smoke is associated with acute outcomes that pose some cardiovascular health risk. Barregard et al. exposed 13 healthy subjects to particle concentrations of 240-280 µg/m3 derived from wood combustion (typical of those reported in areas impacted by fire smoke, but produced under different conditions), and found adverse effects on markers of inflammation, coagulation and lipid peroxidation15.  Tan et al. examined immunological response to ambient vegetation fire smoke exposure (up to 216 mg/m3) in 30 healthy volunteers and observed that the immature white blood cell count tracked with PM concentrations16.  However, other work suggests (1) that the inflammatory potential of woodsmoke particles is less than that of urban particles for exposures greater than 12 hours31, and (2) that biomass smoke particles deposit less efficiently than urban (traffic-related) particles in the human respiratory tract32, 33.  These results may help to explain the epidemiologic discrepancy between the cardiovascular effects of biomass smoke compared to urban pollution.  Taken in a global context, our study population is small and its air quality monitoring network is dense, with one instrument per every ~50 000 cohort members.  Despite the distance between instruments (250 km maximum) we were able to use the TEOM metric to detect significant relationships between acute fire smoke exposure and respiratory health.  Given the weight of evidence suggesting a relationship between particle exposure and cardiovascular health21,the impacts of fire smoke on  85 cardiovascular health still requires further investigation.  Weak associations with acute exposure may be impossible to detect with relatively crude metrics, but may have large public health consequences.  Likewise, the effects of chronic exposure to wildfire smoke have yet to be studied but could pose a considerable health risk in areas that are affected annually, especially where slash burning is practiced.  In either case further use of remote sensing data may produce quick, accurate and spatially-resolved exposure metrics, especially in heavily-impacted areas where air quality monitoring is practically non-existent.                                       86 4.5 References 1. Andreae MO and Merlet P (2001). Emission of trace gases and aerosols from biomass burning. Global Biogeochemical Cycles. 15(4): 955-966. 2. Tham R, Erbas B, Akram M, Dennekamp M and Abramson MJ (2009). The impact of smoke on respiratory hospital outcomes during the 2002-2003 bushfire season, Victoria, Australia. Respirology. 14(1): 69-75. 3. Wiwanitkit V (2008). PM10 in the atmosphere and incidence of respiratory illness in Chiangmai during the smoggy pollution. Stochastic Environmental Research and Risk Assessment. 22(3): 437-440. 4. Johnston FH, Bailie RS, Pilotto LS and Hanigan IC (2007). Ambient biomass smoke and cardio-respiratory hospital admissions in Darwin, Australia. BioMed Central Public Health. 7(240): doi:10.1186/1471-2458-7-240. 5. Kuenzli N, Avol E, Wu J, Gauderman WJ, Rappaport E, Millstein J, Bennion J, McConnell R, Gilliland FD, Berhane K, Lurmann F, Winer A and Peters JM (2009). Health Effects of the 2003 Southern California Wildfire on Children. Am. J. Respir. Crit. Care Med. doi:10.1164/rccm.200604-519OC. 6. Moore D, Copes R, Fist R, Joy R, Chan K and Brauer M (2006). Population health effects of air quality changes due to forest fires in British Columbia in 2003. Canadian Journal of Public Health. 97(2): 105-108. 7. Mott JA, Meyer P, Mannino D, Redd SC, Smith EM, Gotway-Crawford C and Chase E (2002). Wildland forest fire smoke: health effects and intervention evaluation, Hoopa, California, 1999. West J Med. 176(3): 157-62. 8. Mott JA, Mannino DM, Alverson CJ, Kiyu A, Hashim J, Lee T, Falter K and Redd SC (2005). Cardiorespiratory hospitalizations associated with smoke exposure during the 1997 Southeast Asian forest fires. International Journal of Hygiene and Environmental Health. 208(1-2): 75-85. 9. Reinhardt TE, Ottmar RD and Castilla C (2001). Smoke impacts from agricultural burning in a rural Brazilian town. J Air Waste Manag Assoc. 51(3): 443-50. 10. Emmanuel SC (2000). Impact to lung health of haze from forest fires: the Singapore experience. Respirology. 5(2): 175-82. 11. Duclos P, Sanderson L and Lipsett M (1990). The 1987 Forest Fire Disaster in California: Assessment of Emergency Room Visits. Archives of Environmental Health. Jan, 45(1): 53-58. 12. Johnston FH, Kavanagh AM, Bowman DM and Scott RK (2002). Exposure to bushfire smoke and asthma: an ecological study. Med J Aust. 176(11): 535-8. 13. Lipsett M, Hurley S and B O (1997). Air Pollution and Emergency Room Visits for Asthma in Santa Clara  County, California. Environmental Health Perspectives. 105(2): 216-222.  87 14. Kunii O, Kanagawa S, Yajima I, Hisamatsu Y, Yamamura S, Amagai T and Ismail IT (2002). The 1997 haze disaster in Indonesia: its air quality and health effects. Arch Environ Health. 57(1): 16-22. 15. Barregard L, Sällsten G, Gustafson P, Andersson L, Johansson L, Basu S and Stigendal L (2006). Experimental Exposure to Wood-Smoke Particles in Healthy Humans: Effects on Markers of Inflammation, Coagulation, and Lipid Peroxidation. Inhalation Toxicology. 18(11): 845-853. 16. Tan WC, Qiu D, Liam BL, Ng TP, Lee SH, van Eeden SF, D'Yachkova Y and Hogg JC (2000). The human bone marrow response to acute air pollution caused by forest fires. Am J Respir Crit Care Med. 161(4 Pt 1): 1213-7. 17. Swiston J, Davidson W, Attridge S, Li G, Brauer M and van Eeden S (2008). Woodsmoke exposure induces a pulmonary and systemic inflammatory response in firefighters. European Respiratory Journal. 32(1): 129-138. 18. Pope CA, Burnett RT, Thurston GD, Michael J. Thun M, Calle EE, Krewski D and Godleski JJ (2004). Cardiovascular Mortality and Long-Term Exposure to Particulate Air Pollution: Epidemiological Evidence of General Pathophysiological Pathways of Disease. Circulation. 109(1): 71-77. 19. Dominici F, Peng RD, Bell ML, Pham L, McDermott A, Zeger SL and Samet JM (2006). Fine Particulate Air Pollution and Hospital Admission for Cardiovascular and Respiratory Diseases. Journal of the American Medical Association. 295: 1127-1134. 20. Dockery DW (2001). Epidemiologic Evidence of Cardiovascular Effects of Particulate Air Pollution. Environmental Health Perspectives. 109(Supplement 4): 483-486. 21. Naeher LP, Brauer M, Lipsett M, Zelikoff JT, Simpson CD, Koenig JQ and Smith KR (2007). Woodsmoke Health Effects: A Review. Inhalation Toxicology. 19(1): 67-106. 22. Hanigan IC, Johnston FH and Morgan GG (2008). Vegetation fire smoke, indigenous status and cardio-respiratory hospital admissions in Darwin, Australia, 1996–2005: a time-series study. Environmental Health. 7(42): doi:10.1186/1476-069X-7-42. 23. Wu J, Winer A and Delfino R (2006). Exposure assessment of particulate matter air pollution before, during, and after the 2003 Southern California wildfires. Atmospheric Environment. 40(18): 3333-3348. 24. Delfino RJ, Brummel S, Wu J, Stern H, Ostro B, Lipsett M, Winer A, 6 DHS, Zhang L, Tjoa T and Gillen DL (2009). The relationship of respiratory and cardiovascular hospital admissions to the southern California wildfires of 2003. Occupational and Environmental Medicine. 66: 189-197.  88 25. Filmon G (2004) Firestorm 2003, provincial review. Report to the Provincial Government of British Columbia. Firestorm 2003, provincial review. Report to the Provincial Government of British Columbia. Vancouver, BC. 100 pages. 26. Population and dwelling counts for Canada, Provinces and Territies, by Census Divisions, 2001 and 1996 censuses - 100% data., 93F0050XCB01003., Editor. 2002, Statistics Canada. 27. Chamberlayne R, Green B, Barer ML, Hertzman C, Lawrence WJ and Sheps SB (1998). Creating a population-based linked health database: a new resource for health services research. Canadian Journal of Public Health. 89: 270-273. 28. Henderson SB, Burkholder B, Jackson PL, Brauer M and Ichoku C (2008). Use of MODIS products to simplify and evaluate a forest fire plume dispersion model for PM10 exposure assessment. Atmospheric Environment. 42: 8524-8532. 29. Allison PD (2005), Fixed effects regression methods for longitudinal data using SAS Cary, NC: SAS Press. 30. Copeland KT, Checkoway H, McMichael AJ and Holbrook RH (1977). Bias due to misclassification in the estimation of relative risk. American Journal of Epidemiology. 105(5): 488-495. 31. Kocbacha A, Hersethb JI, Låga M, Refsnesa M and Schwarzea PE (2008). Particles from woodsmoke and traffic induce differential pro-inflammatory response patterns in co-cultures. Toxicology and Applied Parmicology. 232(2): 317-326. 32. Londahl J, Pagels J, Boman C, Swietlicki E, Massling A, Rissler J, Blomberg A, Bohgard M and Sandstrom T (2008). Deposition of biomass combustion aerosol particles in the human respiratory tract. Inhalation Toxicology. 20(10): 923-933. 33. Londahl J, Massling A, Swietlicki E, Brauner EV, Ketzel M, Pagels J and Loft S (2009). Experimentally Determined Human Respiratory Tract Deposition of Airborne Particles at a Busy Street. Environmental Science and Technology. DOI: 10.1021/es803029b.   89 5 Chapter 5: Retrospective 5.1 Review The previous three chapters describe a cohesive body of work that started in September of 2003, when the province of British Columbia was still reeling in the aftermath of recent firestorms.  Given the size of the exposed population and the density of the air quality network, we recognized a unique opportunity to study the health effects of forest fire smoke in a relatively large (when compared to previous studies) and well-monitored population.  Based on the premise that remote sensing data and output from dispersion models would be more spatially representative of smoke exposure than data from a few air quality monitors, we proposed an epidemiologic study that would compare respiratory and cardiovascular health effects using multiple exposure estimation methods.  By improving the spatial resolution of smoke exposure estimates we expected to minimize misclassification bias in the epidemiologic results, thereby clarifying the effect of fire smoke on population health. Furthermore, by demonstrating that epidemiologic results based on remote sensing data and dispersion modeling were comparable to those based on air quality data we wanted to provide a foundation for similar studies in fire-affected areas where air quality monitoring is sparse or non-existent.  Large-area dispersion modeling of forest fire smoke requires information about (1) the location, (2) the growth/decay and (3) the emissions rates for all fires in the domain. Most conventional approaches use further models to generate the data necessary for (2) and (3), making it challenging and time-consuming for non-experts to take advantage of this potentially-valuable approach.  In Chapter 2 we described methods to rapidly derive requirements (1), (2) and (3) from MODIS remote sensing data.  These products are freely-available and globally consistent, so they transcend the political boundaries that might otherwise impact data collection and model implementation. Evaluation exercises showed that MODIS reliably detected and delineated large fires, and that its radiative power measurement could be used to calculate emissions rates comparable to those estimated by more complex models.    90 The results of Chapter 2 were used in Chapter 3 as input for the CALPUFF dispersion model (along with the CALMET wind fields) to estimate ambient concentrations of PM10 for use the epidemiologic study.   However, output from any dispersion model is uncertain because any input errors (in the wind fields, fire location, emissions rates, etc.) will propagate through to the concentration estimates.  It was therefore important to evaluate the CALPUFF output as thoroughly as possible, so we combined qualitative, quantitative and spatial approaches to yield a holistic picture of model performance in the absence a single ‘gold standard’.  Measurements from the six TEOM instruments provided a quantitatively accurate benchmark, but their usefulness was limited by their spatial scarcity.  Conversely, tracings of plume footprints were spatially extensive, but provided no quantitative information about surface concentrations. The MODIS AOT measurements seemed like a good compromise because they are both quantitative and spatially extensive, but their temporal availability was inconsistent and their correlation with surface measurements was highly variable.  Even so, the evaluation framework developed in Chapter 3 allowed us to conclude that CALPUFF performance was moderate at best, and unlikely to add value to the epidemiologic analyses.  In Chapter 4 we assigned three daily measures of smoke exposure to 281 711 individual cohort members based on their residential addresses.  The measures were: (1) the 24-hour average PM10 concentration of the nearest TEOM instrument; (2) an indicator of SMOKE coverage based on satellite images; and (3) the 24-hour average PM10 concentration estimate from CALPUFF.  We then followed respiratory and cardiovascular physician visits and hospital admissions in the cohort over a 92-day period spanning the 2003 forest fire season.  Like several other studies we found that TEOM-measured PM10 concentrations were significantly associated with respiratory outcomes (both physician visits and hospital admissions) but not with cardiovascular outcomes.  As expected from the results of Chapter 3 the effects of the CALPUFF metric were biased towards the null.  The effects of the SMOKE metric were also null, but tracked quite well with those of the TEOM metric, suggesting that exposure estimates derived solely from satellite data may be valuable with further refinement.    91 This thesis took almost six years to complete, and Chapters 2 through 4 only summarize the work done over this period.  Although there is some chronological overlap between the following sections, they provide a structure for discussing the problems we faced, solutions we devised, contributions we made, and ideas we had in light of some information not previously provided.  5.2 Conception (2003 – 2004) Late in August of 2003 we started a short field sampling campaign to measure particulate matter concentrations around smoke-affected homes in Kelowna.  Figure 5.1 confirms what Figure 1.4 suggested – namely that PM10 concentrations measured by a community TEOM (in this case 2 km away) do not capture local variability in PM10 concentrations and therefore must bias epidemiologic results when used for smoke exposure assessment.  This observation prompted (1) conversation about the limitations of conventional approaches to fire smoke epidemiology and (2) grant applications proposing methods by which we could build on previous work.  Where many studies had been limited to quantitative exposure estimates based on monitoring data, we secured the resources necessary to improve the spatial resolution of smoke exposure estimates using dispersion modeling and satellite data.  And where many other studies only included individuals who experienced certain health events, we were able to identify and follow a large, population-based cohort through the unique British Columbia Linked Health Database (BCLHD).  5.3 Wind modeling (2004 – 2005) In the very beginning we decided to test the usefulness of particulate matter (PM) dispersion modeling as a tool for smoke exposure assessment.   Although air quality data are reliable, monitors are too spatially discrete to capture the full range of concentrations experienced by a disperse population.  Conversely, a dispersion model can estimate concentrations at any location, but its output cannot be considered reliable due to uncertainty in its inputs.  At the simplest level all dispersion models only have two inputs: meteorology and emissions.  Both must generally be derived from secondary and tertiary models, and then entered into the dispersion model where the  92 wind fields are used to simulate spread of the emissions based on a complex set of conditions and equations.    Figure 5.1 – Particulate matter measured at TEOM and home locations Particulate matter concentrations as measured by the TEOM in Kelowna and a light- scattering sampler outside of a home two kilometres away.  Note that TEOM concentrations do not reflect those near to the home after ~10:00 pm.   We originally chose to use the CALPUFF dispersion model because it was part of the US Forest Service BlueSky smoke forecasting framework, which we hoped to have extended into western Canada.  Although CALPUFF can accept meteorological input from a variety of sources, the weather models for British Columbia were being run at spatial resolutions of 10 and 4 km2, neither of which is fine enough to fully resolve the complex terrain of the study area.  To improve prediction of the wind fields through the mountain valleys we chose to run the CALMET model at a resolution of 1 km2.  In the simplest terms CALMET is an interpolator that estimates meteorological conditions between points while accounting for the surrounding land use and topography.  Initial information at those points can be provided by (1) measurements from weather stations and radiosondes, (2) output from some other meteorological model, or (3) both.  Prior to our study, Barna and Lamb did similar modeling (at a resolution of 5 km2)  in the same  93 geographic region and found that using both sources of initialization data produced the best results1.  After experimenting with multiple options we similarly initialized CALMET by combining measurements from 99 weather stations and one radiosonde with prognostic output from the ETA weather model at a resolution of 12 km2.  Much of this work was done by Benjamin Burkholder, who has considerable expertise in initializing and running meteorological models.  Very little of it is presented in the preceding chapters, but it is summarized thoroughly elsewhere2, 3.  Although these reports were not published in a peer-reviewed journal, they were distributed widely amongst the meteorological modeling community in BC and the north-western states.  It took almost one year to gather the data needed for CALMET, and to get the model producing satisfactory results such that wind field estimates provided a reasonable approximation of actual measurements under most conditions.  Even so, we know that wind direction was poorly predicted when wind speed was low.  The poorness of the estimates under such conditions is the biggest weakness of the whole project, with errors in the CALMET output propagating through the dispersion model and into the subsequent epidemiologic exposure estimates.  Although we had CALMET performing well in the context of meteorological modeling over complex terrain, we must also evaluate the exercise in the context of its public health value.  We used CALMET/CALPUFF in an attempt to reduce exposure misclassification, and it failed to reliably fulfill this primary objective.  In retrospect the time and human resources necessary to initialize, test and tweak the CALMET model provided no added value for the epidemiologic analyses, but they do highlight the simplicity of some alternative, satellite-based approaches explored in Chapter 4 and suggested as the focus of future work.  5.4 Emissions modeling (2005 – 2006) After producing the CALMET output for CALPUFF our next step was to locate fires and specify their growth and decay so that we could produce emissions estimates. Conventional approaches require considerable expertise in fire behaviour modeling, but we hoped to simplify the method by negating the need to identify, initialize, run and evaluate further simulations.  Furthermore, we wanted to use globally-available data  94 that were not restricted by political record-keeping so that the methods could be applied in other fire-affected areas.  This prompted the initial investigation of simplifications based on remote sensing data.  As we became more familiar with products from the Moderate Resolution Imagine Spectroradiometer (MODIS) instruments maintained by NASA (US National Aeronautics and Space Administration), the richness of their potential became more apparent.  The MODIS fire product (MOD14) detects thermal anomalies at a spatial resolution of 1 km.  Any pixel containing a probable fire is flagged and several of its attributes are recorded.  As described in Chapter 2 we used the latitude and longitude attributes to map the growth and decay of large fires over multiple days.  We also found that circular buffers drawn around these locations provided a quick-and-dirty delineation of the fire burn scars, thus yielding our method for simply estimating burned area with good accuracy.  While developing the size estimation methodology we learned about the work of Charles Ichoku and colleagues at NASA, who were refining a method to estimate the emissions from individual fire pixels based on their radiative power (FRP) attribute4.  Where conventional models require information about fuel type, fuel load, fuel moisture content, fire type, elevation, slope, relative humidity and wind speed, Ichoku’s method only requires an FRP value and a coefficient specific to the biomass type.  With support and encouragement from Dr. Ichocku we decided to evaluate the usefulness of this simple FRP method by comparing its emissions estimates to those generated by two well-established models.  This work is partially presented In Chapter 2 where we used the Canadian Wildfire Behaviour Prediction System (CANFB) to model emissions from fires in Canada, and the American Emissions Production Model (USEPM) to estimate emissions from fires in the north-western states.  Simple regression shows good agreement between the FRP and CANFB methods, with higher emissions estimated by the FRP method (R2 = 0.71, slope = 0.71).  There is a larger discrepancy between the FRP and USEPM methods (R2 = 0.28, slope = 0.68).  In Appendix 5 we compare CALPUFF concentration estimates based on the FRP, CANFB and USEPM methods to actual PM10 measurements at the six TEOM sites.  Plots clearly show how the USEPM method consistently under-predicted concentrations, while the CANFB and FRP methods were  95 more realistic. Based on the results in Chapter 2 and Appendix 5 we chose to continue the project with the quick and simple FRP method.  Chapter 2, which describes simple methods for estimating fire properties from satellite data, is a unique contribution to the specialized body of literature where remote sensing and wildfire science intersect.  Although the methods presented in Chapter 2 were developed for a specific public health application, their potential usefulness is widespread and we consider them to be the most important general outcome of the thesis project.  Under all climate change scenarios forest fires are expected to become more frequent and intense5, 6, so any activity that requires information about fire size and emissions rates stands to benefit from these simple, rapid and reasonable estimates.  Consider, for example, that many fire-prone countries do not record their annual burned area, or they do not keep publicly-accessible records.  If our sizing approach proves to be globally generalizable this information could be estimated within minutes with an annual aggregation of the global MOD14 data and a Geographic Information System (GIS).  Furthermore, the annual mass of particulate emissions from these fires could be estimated with a few quick calculations.  Unlike the greenhouse gases, particulate matter absorbs ultraviolet radiation, making it challenging to understand how the overall emissions from forest fires feed back into the global climate change mechanisms.  The FRP method developed by Ichoku et al. and tested by us could help to minimize the resources necessary to estimate particulate emissions, thereby maximizing the resources available for addressing bigger, more important questions.  Of course there are other, more modest applications.  We have made Chapter 2 widely available to smoke forecasting operations in Canada and the US with the hope that it provides a strong foundation for developing quick, data-driven alternatives to some of the cumbersome, resource-intensive models that are currently in use and under constant revision.  Where smoke forecasting frameworks already exist, conventional models for estimating fire size and particulate emissions could be augmented by and evaluated with the methods outlined here.  Where frameworks are under development these methods could be used instead of conventional models.  From working with the BlueSky team we know that vast computational and human resources are committed to  96 running, maintaining and improving smoke forecasts with only moderate success7, 8.  If the fire size and emissions modules could be simplified with reasonable, data-driven methods, more resources would be available to improve the all-important wind inputs and dispersion outputs.  This has tangible implications for public health because forecasting plays an important role in communication about the risks associated with fire smoke.  Quality predictions in near-real-time can be made available through websites like www.firesmoke.ubc.ca (created as part of this thesis), and they allow people and authorities in affected areas to make informed decisions about minimizing exposure. Furthermore, output from reliable forecasts would be valuable for future epidemiologic research in areas where air quality monitoring is sparse or non-existent.  5.5 Smoke dispersion modeling (2006 – 2007) With the meteorological and emissions inputs prepared for dispersion modeling, our next step was to run CALPUFF and to evaluate its performance in the context of smoke exposure assessment.  The problem of evaluation is complex due to its spatial, temporal and quantitative nature.  On one day CALPUFF might do a good job of estimating the magnitude of particulate matter concentrations within a smoke plume while miscalculating its trajectory.  The quantitative distribution of the resulting exposure estimates might be correct, but their spatial assignment will be incorrect.  For the example shown in Figure 5.2, most subjects within the actual plume will be misclassified, either by (1) receiving a zero concentration instead of an elevated value or (2) receiving an exposure estimate that is higher or lower than the actual concentration.  Likewise, most subjects within the predicted plume will be misclassified, either by (1) receiving some elevated exposure instead of a zero value, or (2) receiving an exposure estimate that is higher or lower than the actual concentration.  On the next day the modeled plume trajectory might correctly align with the true smoke trajectory, but estimated concentrations could be uniformly too low or too high meaning that exposures for all subjects within the plume should be quantitatively misclassified. Of course neither scenario leads to exposure misclassification if the plume affects an unpopulated area, which is an important consideration when evaluating a dispersion model for a public health application.   If the model performs well in populated areas,  97 the quality of predictions throughout the rest of the domain is trivial.  Likewise, a simulation with adequate general performance by conventional dispersion modeling standards may not be suitable for reliable exposure assessment.  In Chapter 3 we use three methods to evaluate CALPUFF performance around populated areas of the modeing domain by comparing: (1) CALPUFF-estimated PM10 concentrations to TEOM-measured PM10 concentrations at 6 sites; (2) CALPUFF- estimated plume profiles to MODIS-depicted plume profiles in three test areas; and (3) CALPUFF-estimated PM10 concentrations to MODIS-measured AOT at 70 sites.  None provides a single gold standard against which to compare the model output, but together they provide valuable insight into its spatial, temporal and quantitative accuracy.  First we estimate 24-hour average PM10 concentrations at the sites of six TEOM instrusments located in Kamloops, Kelowna, Vernon, Revelstoke, Golden and Creston.  Together these towns account for approximately 60% of the study area population9.  By comparing 92 days of CALPUFF estimates to their corresponding TEOM measurements by (1) correlation, (2) mean and absolute errors and (3) an index of agreement, we found that model performance varied by location.  The temporal and quantitative agreement between values was good around Kelowna and Vernon, moderate around Revelstoke, Golden and Creston, and quite poor around Kamloops.  We further investigated the spatial and temporal accuracy CALPUFF output around Kelowna, Golden and Kamloops using plume shapes traced from satellite images.  By calculating the horizontal overlap between measured and modeled plumes we expressed their agreement in the epidemiologic terms of sensitivity (probablity of predicting smoke where smoke acutally exists) and specificity (probability of not predicting smoke where smoke does not actually exist).  Results close to 1.0 for both measures would indicate a highly accurate plume.  We further defined the daily spread between measures as discrepancy = sensitivity + specificity - 1, where an average value close to 1.0 indicates greater temporal stability (i.e. more days with sensitivity and specificity both close to 1).   The overall specificity was good (~0.90) for all three test areas.  Sensitivity was generally poor, but suprisingly highest for the area around Kamloops (mean = 0.53, SD = 0.29), where CALPUFF estimates performed most poorly compared to the TEOM measurements.  However, the discrepancy was best for  98 the area around Kelowna (mean = 0.74, SD = 0.16), suggesting that a few very poor days were responsible for its low sensitivity.       Figure 5.2 – Potential misclassification when modeling smoke plumes. The figure on the left shows a hypothetical real plume pointing straight up while its CALPUFF-estimated plume is angled to the right.  Plumes are shaded to represent three categories of exposure – heavy in the center, moderate in the middle and light at the outside. The figure on the right illustrates how exposure misclassification would occur in this scenario assuming that concentrations in the CALPUFF-estimated plume were correct, but that their trajectory was incorrect.    Finally we attempted to evaluate the spatial, temporal and quantitative quality of CALPUFF output around all populated areas using the aerosol optical thickness (AOT) data available from MODIS at a resolution of 10 km2.  The AOT is a measure of light extinction in a cloudless atmospheric column due to the light-absorbing particles in contains.  The data are spatially and temporally intermittent because the presence of any cloud invalidates the measurements, but several studies have used long time series to show varying strengths of association (r = 0.31 – 0.70) between AOT and surface PM concentrations10-13.  We used the population density within the study area to establish the 70 points at which CALPUFF estimates were compared to AOT measurements, Actual plume Predicted plume No misclassification Overexposed by 1 category Overexposed by 2 categories Overexposed by 3 categories Underexposed by 1 category Underexposed by 2 categories Underexposed by 3 categories  99 resulting in varying strengths of association (r = -0.08 – 0.73).  Figure 5.3 shows three low values clustered around the TEOM in Kamloops, while correlations appear higher around Kelowna, Vernon and Golden, and more moderate around Creston and Revelstoke.  These results show similar variability to those from the other evaluation exercises, which does lend some confidence to their credibility.  However, studies on the relationship between AOT and surface PM in North America report a gradient across the continent with its weakest values in the northwestern United States12, 13, possibly due to the reflectance from the arid and semi-arid landscapes12.  Therefore correlations might be low even if the CALPUFF estimates reflected surface PM10 concentrations with perfect accuracy.  As such we cannot have complete confidence in the usefulness of this exercise.  Regardless of concerns about the utility of the AOT evaluation, the overall results from Chapter 3 suggest that we cannot have much confidence in the CALPUFF estimates for the purposes of exposure assessment.  Once again our methods have performed well when compared to similar work14-16, but the CALPUFF output remains too erroneous in populated areas to provide added value for epidemiologic research.  Although this result was disappointing in terms of the remaining thesis objectives, the process of developing the satellite-based evaluation tools started us thinking about the potential of using smoke exposure metrics derived purely from remote sensing data.    100  Figure 5.3 – Correlation between CALPUFF and AOT. The correlation between CALPUFF PM10 estimates and MODIS AOT measurements across varying population densities in the study area.  Yellow dots indicate the location of six TEOM instruments.   5.6 Exposure assignment (2007 – 2008) At the beginning of the project we decided that it would be beneficial in terms of avoiding exposure misclassification to assign daily smoke exposure estimates to each cohort member based on residential address.  While the idea is simple, the reality is more challenging for technical and ethical reasons.  Technically speaking, the structure of the data maintained by the BC Ministry of Health (MOH) in the BCLHD is the most 0 persons/km2 1-10 persons/km2 11-100 persons/km2 100-1000 persons/km2 1000+ persons/km2 Population Density < 0.10 0.10 < 0.30 0.30 < 0.50 0.50 < 0.70 0.70+ CALPUFF vs AOT Correlation  101 important consideration.  Every individual in the health care system has a personal health number (PHN), and the residential address history for that number is recorded in a master client registry.  Under ideal conditions the cohort would have included all individuals who appeared to be living in the study area during the study period, but addresses in the registry can be out-of-date for several reasons, as explained in Chapter 4.   To reduce potential exposure misclassification due to this uncertainly we restricted cohort eligibility to people who had multiple contacts with the healthcare system from the same 6-digit postal code spanning the study period.  Specifically, the code associated with the last health care contact in the year before the study period (July 1st 2002 through June 30th 2003) had to match the code associated with the first contact in the year following the study period (October 1st 2003 through September 30th 2004).  A detailed methodology is summarized in Appendix 6.  Thus the study population was limited to (1) geographically stable individuals who (2) make regular use of the health care system.  In the hypothetical scenario where (1) transient populations are less healthy than stable populations, and (2) people who use the health care system are healthier than those who do not, it would follow that the people not included in the cohort are at higher risk of the health effects we are studying.  Therefore the decision to minimize misclassification in the exposure assignment may have been balanced against the generalizability of our epidemiologic results.  Ethically speaking, the privacy of health care users must be balanced against the public health benefits of research projects like this one.  In the past the BCLHD has allowed researchers to spatially locate study subjects using their postal codes.  The first three digits could generally be used, and the full six digits could be used with special precautions when there was a demonstrable need for greater spatial accuracy.  In urban and sub-urban areas one 6-digit postal code may serve several houses that are close together, but in rural areas a single postal code can cover hundreds of square kilometers.  Postal codes are mapped to spatial locations with a conversion file (PCCF) that uses population-weighted probabilistic modeling to estimate their coordinates. Although the estimates are reasonable in densely populated areas, they can be hundreds of kilometers away from an actual address in rural areas.  Because fires predominantly affect rural areas, assigning smoke exposure by 6-digit postal code alone would result in unavoidable misclassification.  Based on this argument we approached  102 the MOH with a proposed method to assign exposure estimates by residential address while protecting the spatial anonymity of cohort members.  After multiple revisions over many months we agreed on the approach summarized in Chapter 4.  Address records provided by the MOH needed considerable cleaning before they could be geocoded.  The single ADDRESS field contained an inconsistent combination of the following of information: (1) street number; (2) street name; (3) rural route number; (4) suite or apartment number; (5) road type; (6) road direction; (7) PO box; (8) legal lot descriptor; and (9) town/city name.  Our first step was to develop an algorithm for parsing these records into geocodable addresses, which was done mostly by Dr. Jason Su.  A description of the code and a sample of its output are found in Appendix 7.  After cleaning the records we tested multiple methods for geocoding the addresses and found the Google Maps (Mountain View, California) index had a success rate of ~85% while the national index provided by DMTI Spatial (Markham, Ontario) was at ~70%. Google Maps does allow users to batch geocode up to 50 000 records daily from a single internet protocol (IP) address on a single application programming interface (API) key.  Thus we split the cleaned data (562 434 records) into ~20 files and ran them on multiple machines over the course of several days.  For each record Google Maps either (1) assigned latitude, longitude and spatial confidence values or (2) failed the geolocation.  Our final product was a file containing only those addresses geolocated with the highest spatial accuracy.  At the end of this process we had (1) estimated latitude and longitude coordinates for all 562 434 records based on the PCCF output for their associated 6-digit POSTAL codes, and (2) precise coordinates for 460 608 (82%) of the ADDRESS records based on the Google Maps output.  The addresses that we could not geolocate with Google Maps were predominantly rural.  Our next step was to map these coordinates onto the study domain in a 1X1 km grid so that we could assign exposure estimates at a reasonably high resolution while maintaining the spatial anonymity of the cohort members.  The POSTAL coordinates fell into 1244 cells while the ADDRESS coordinates fell into 2538 cells (Figure 5.4).  For 253 201 of the records, the POSTAL and ADDRESS coordinates fell into the same cell and for 207 404 records they fell into different cells (101 826 had POSTAL coordinates only).  For the complete set of 3366 cells we assigned daily  103 exposure estimates for the TEOM (from regulatory monitoring data), SMOKE (from satellite imagery) and CALPUFF (from the dispersion model) metrics, as described in Chapter 4.  The records were returned to the MOH with the following fields: (1) the original address; (2) the cleaned addresses; (3) the POSTAL cell identifier; and (4) the ADDRESS cell identifier, where applicable.  Cohort PHNs were linked to the exposure estimates based on the original address field, and then returned to us for the epidemiologic analyses.   Figure 5.4 – Geolocation results. Showing the agreement and discrepancy between spatial location of POSTAL and ADDRESS coordinates.   Although we managed to geolocate 82% of the records in the historical address file, only 64% of the cohort (281 711 of 439 887) ended up with POSTAL and ADDRESS exposure estimates.  In other words, a disproportionate number of the cohort members were living at the ungeocodable rural addresses.  If we assume that rural dwellers are less transient than urban and suburban dwellers, this is probably a function of the criterion that restricted cohort eligibility by geographic stability.  Given that we initiated this whole process to better estimate exposure for people in rural areas, it seems that the time and effort were not fully justified by the outcome.  In the remainder of the POSTAL coordinates only ADDRESS coordinates only Both types of coordinates  104 cohort the POSTAL and ADDRESS exposure estimates were different for 36% of members (101 416 out of 281 711).  Therefore only 23% of the full cohort (101 416 out of 439 887) ended up with exposure estimates that were different from what they would have received if we had only assigned exposure by 6-digit postal code.  A comparison of the age-stratified epidemiologic results for respiratory physician visits using POSTAL versus ADDRESS exposure estimates is found in Appendix 8.  The plots show very little difference between the methods, so we cannot provide the MOH with evidence that the process was worthwhile for this application.  Although this information is valuable for public health research in BC, it was disappointing to find that spatial improvements to smoke exposure assessment did not translate into improved understanding of its health effects.  Of course it is important to remember that such an administrative approach to exposure assignment is prone to considerable error because (1) people spend varying amounts of time away from their residences and (2) the majority of time is spent indoors, with indoor penetration of outdoor air pollution depending on several variables (primarily the air exchange rate) that we could not model for these analyses. Furthermore, there is no way to address the likelihoods that (1) some subjects simply were not in the study area (due to vacation, voluntary of mandatory evacuation, etc.) during the study period, and (2) some people who reside outside of the study area had fire-related events within the study area (parts of which are popular vacation destinations) that were not included.  5.7 Epidemiology (2008 – 2009) In the funding applications for this project we initially proposed to use Bayesian hierarchical modeling for all epidemiologic analyses because Bayesian models make use of information from previous studies and can accommodate latent variables for unobserved and potentially-confounding factors17.  For example, we could not use administrative health data to reliably assess the presence of pre-existing cardiopulmonary diseases for this study, but within a Bayesian framework we could have included some reasonable assumptions about the effects of this unmeasured variable based on previous work.  Likewise, a hierarchical structure could have accounted for spatial differences in the exposure effects by grouping subjects within the same area into analysis clusters.  Unfortunately our switch to conventional fixed effects  105 logistic regression was motivated by the size of the longitudinal data set – 281 711 subjects X 92 days yields nearly 26 million observations to be processed.  Given that Bayesian methods require many thousands of iterations to converge, it would have been impossible to generate effect estimates based on the complete data set.  Although we discussed options for randomly subsetting the data set and averaging the resulting Bayesian estimates, it was clear that such an approach was worthy of another graduate thesis in statistics.   Getting the logistic regression models running in SAS (Cary, North Carolina) was challenging enough, with only PROC GENMOD able to handle the whole data set.  Once we had the unadjusted exposure models running our next step was to establish a set of other potentially explanatory variables.  For every cohort member we received a record of all medical services plan (MSP) billings and hospitalizations between January 2002 and December 2004.  All information found in these records was available for constructing the outcome and explanatory variables described in Chapter 4.  While adjusting epidemiologic relationships for AGE and SEX and socio-economic status (SES) is standard practice, we also wanted to create a variable that reflected (1) the baseline rate of cardiovascular and respiratory outcomes for individuals and (2) their potential sensitivity to fire smoke exposure.  Although there is no way to establish clear disease diagnoses from administrative records, we were able to assess a historical pattern of cardiovascular and respiratory physician visits for each person.  The resulting PREresp and PREcardio variables indicate the number of visits (0, 1-2, 3-5 and 5+) in each outcome category during the year prior to the study period.  Given the strong trend in coefficients across the PRE categories for all models we feel confident that these variables adjusted for the probability of an outcome during the study period based on precedent alone.  If the PRE variables also successfully reflected potential sensitivity to fire smoke we may have over-adjusted the effects of exposure by including them, but Table 4.2 shows that stratification on the PRE variables yields no change in the effect estimates.  Given that multiple other studies have shown that populations with pre-existing disease are more sensitive to smoke exposure18-20 it could be that the PRE variables are a poor indicator for pre-existing disease – they may simply reflect individual relationships with the health care system.  However, it is also possible that the PRE variables actually do  106 reflect disease status, but that pre-existing disease does not render subjects more susceptible to forest fire smoke despite rendering them more susceptible to urban pollution21-23.  Recall that this is the first study on the population health effects of forest fire smoke to use a cohort design, meaning that all subjects contribute exposure profiles to the analyses regardless of their outcome status.  Other studies on outpatient visits and hospital admissions have only included people who actually experienced the outcome of interest, and were therefore weighted towards populations with higher underlying outcome rates (e.g. children, the elderly, and people with pre-existing diseases).  Although the resulting effect estimates are not necessarily biased (i.e. they may internally valid), their generalizability to a wider population must be considered (i.e. they may not be externally valid)24.  For example, the results of studies that are conducted on people admitted to hospital may be representative of a hospital-attending population, but the majority of people are rarely hospitalized.  In contrast, this study was done on subjects drawn from an entire population, so the results can be generalized back to a much wider group.  This is important to consider when interpreting the epidemiologic results from a public health policy perspective.  Another variable that had significant impact in sensitivity analyses was the daily distance-to-fire (DIST) measurement assigned at the same time as the TEOM, SMOKE and CALPUFF metrics.  We originally created this variable because it seemed plausible that people closer to fires would be exposed to more smoke, and we felt it would be valuable for boosting the exposure signal in the models.  In fact, when the crude models were adjusted for DIST alone there was a significant increase in the effect of all exposures metrics, but the coefficients for DIST were always non-intuitive, with the odds of an outcome increasing with increasing distance from the nearest fire.  To investigate this phenomenon we plotted the relationship between DIST and CALPUFF for each day of the study period and found that exposures estimates generally tailed off after distances of more than 50 km (see Figure 4.2).  When analyses were stratified on this cut off we found that (1) effect estimates were higher in the <50 km group, and (2) the direction of the DIST coefficients was intuitive when included in the models.  As discussed in Chapter 4, it is possible that DIST is acting as a proxy for the error in exposure metrics, where the validity of assigned TEOM, SMOKE and CALPUFF values is positively associated with their distance to the nearest fire.  We also found that effect  107 estimates were consistently higher for the >50 km group when exposures were lagged by one day, possibly capturing the impacts of long-range smoke transport.  Overall we feel that variables like DIST may be important to generate for future studies, and that we did not explore its full potential in this study.  For example, it is possible that including a DIST*EXPOSURE interaction term in the regression models would strengthen the effect estimates, but the results would be challenging to interpret.  Analyses were also sensitive to the day-of-week (DOW) indicator, which was originally overlooked in the Chapter 4 analyses.  In fact, we were reminded of its importance when checking the results for cardiovascular and respiratory physician visits against those for gastrointestinal physician visits, which are theoretically unrelated to smoke exposure.  An unexpected association prompted us to notice that (1) PM10 concentrations tended to be higher early in the week when physician visits are most frequent (see below) and (2) the SMOKE was more frequently null on weekends than on weekdays, possibly due to NOAA technicians taking time off.  The subsequent inclusion of DOW in all models (1) eliminated the association between smoke exposure and gastrointestinal effects and (2) significantly adjusted the cardiovascular and respiratory effect estimates for all exposure metrics, with the most pronounced changes observed for the SMOKE coefficients.  Again it is possible that inclusion of the DOW variable might over-adjust for the exposure effects.  Consider that some of the heaviest fire activity occurred on Friday August 21st and Saturday August 22nd.  It follows that anyone feeling mild or moderate discomfort from the fire smoke might put off a visit with the family doctor until regular office hours on Monday.  Given that Monday is always the busiest day of the week for outpatient visits, these additional visits may get lost in the noise when models are adjusted for DOW, or they may contribute too much signal when DOW is not included.  Our decision to include a fully-specified set of DOW categories (instead of a simple weekday/weekend indicatory) is generally conservative, so we expect that any bias in the resulting effect estimates is towards the null.   Finally, the construction of the SES variable deserves some discussion.  Income data are collected as part of the Canadian census, and every dissemination area (the smallest geographic division) is assigned a median neighbourhood income.  These median values can then be ranked into quintiles across any given geographic area.  108 Therefore the SES variable does not reflect the income of individual study subjects, but it reflects the median income for their neighbourhood.  Of course, the only way to match study subjects to their residential locations is the PCCF, the limitations of which have already been discussed.  Furthermore, postal code boundaries do not match one-to-one with census boundaries, so not only does the SES variable fail to reflect individual income, it may also fail to reflect actual neighbourhood income.  Even so, Table 4.1 shows a strong gradient in outcome rates between the quintiles.  This is what we expect given well-documented association between economic status and diseases25 and it suggests that the variable, although imperfect, does reflect SES to some degree.  In reality this variable did not provide much adjustment to exposure effects, but in theory its inclusion adds accuracy to the estimates.  Once the models were running and appropriately adjusted our results for the TEOM exposure metric turned out to be very consistent with those from other studies, as discussed in Section 4.4.  The results from the other two metrics are, however, more interesting in terms of the originally-proposed thesis.  Based on the limitations and uncertainties presented in Section 5.4 we expected the CALPUFF metric to perform poorly in the epidemiologic analyses.  Under the hypothetical situation in which the CALPUFF metric was a perfect reflection of actual exposure, we would expect the CALPUFF effect estimates to be higher than those for TEOM, and with tighter confidence intervals.  Figures 4.2 through 4.5 clearly show that the odds ratios for CALPUFF are lower than those for TEOM, but their confidence intervals are tighter because the estimates are more spatially resolved (drawn from 2538 locations instead of just 6).  We conclude that dispersion modeling cannot perform well enough the in complex terrain of the study area to provide reliable estimates for forest fire smoke exposure assessment.  Given the uncertainties still surrounding the health effects of fire smoke it would be interesting to try a similar study in an area where CALPUFF could perform more successfully.  Remote sensing measurements are not affected by topography or the underlying infrastructure for air quality monitoring.  The decision to include the SMOKE was made when we were using the same data to evaluate CALPUFF performance work, and its general agreement with the TEOM results suggests that satellite-derived metrics have  109 the potential to provide truly valuable exposure estimates with further refinement.  For example, we know that the FRP of a fire pixel is directly proportional to its aerosol emissions26, so combining this information with the visible presence of smoke may yield reasonable concentration estimates.  We also found that it is possible to identify smoky pixels in satellite images based on their red, green and blue (RGB) colour values, so the degree of deviation from some baseline value might also provide information that could be used to establish concentration gradients.  In the global context, smoke from vegetation fires predominantly affects populations in tropical areas where air quality monitoring is minimal or non-existent.  If remote sensing data can be used to provide exposure information for these areas, fire smoke epidemiology could be advanced by the opportunity to study (1) high exposures in very large populations and (2) chronic exposure as opposed to acute exposure.  With real uncertainly still surrounding (1) the global health impacts and (2) the cardiovascular effects of fire smoke, development of satellite-based methods could help to answer such lingering research questions.  Overall Chapter 4 is one of the largest and most methodologically rigorous studies yet done on the population health effects of forest fire smoke – only the work of Delfino et al.27 is comparable in terms of scale and scope.  As summarized in Chapter 4, Wu et al.28 used MODIS images to help fill missing 24-hour average PM2.5 values in a dense network of 37 monitoring stations around Los Angeles, California.  Next, they spatially interpolated the ambient daily concentrations across the entire city at a resolution of 250 meters, assigning exposures estimate to each postal code in the study area.  Delfino and colleagues used these results to examine how exposure was associated with hospital admission rates for respiratory and cardiovascular outcomes.  The postal codes of hospitalized individuals were used to estimate their exposure in the days immediately preceding their admissions.  Estimates were then used in time series analyses to examine the effects of exposure on relative rates of admission for pre-, during- and post-fire periods.  Because Los Angeles is much more densely populated than southeastern British Columbia, the study included 21 019 respiratory and 27 170 cardiovascular hospital admissions.  In this study we had 557 and 1208 respiratory and cardiovascular hospital admissions, respectively, as well as 34 771 and 46 037 physician visits.  Even so, our odds ratio for the risk of respiratory hospital admission associated with one 10 µg/m3 increase in PM10 (1.03, 95%CI = 1.00 to 1.05) is almost  110 the same as the rate ratio reported by Delfino et al. for one 10 µg/m3 increase in PM2.5 (1.03, 95%CI = 1.01 to 1.04).  Likewise, neither of these large, rigorous studies detected any cardiovascular effects. Together in context of other, smaller studies reporting similar results they build on a growing body of evidence suggesting that acute exposure to forest fire smoke poses a significant risk to respiratory health, but does not elicit the cardiovascular effects associated with urban air pollution.  5.8 Conclusions Every research project wends along a unique path, often deviating from the route originally envisioned by its investigators.  While too wide a deviation can muddy the initial hypotheses, narrow adherence to a set plan limits the creativity with which the work can be approached.  This thesis has fulfilled its original objectives with both faithfulness and innovation, producing valuable methods and results.  From Chapter 2 we know that publicly- and globally- available fire detection data from MODIS can be used to reasonably estimate (1) the location of large fires, (2) their size and shape, and (3) their aerosol emissions.  These results are potentially valuable for any fire modeling application.  In Chapter 3 we developed a systematic framework for the quantitative and spatial evaluation of models that simulate fire smoke behavior.  These results are valuable for any smoke modeling application, be it operational forecasting or smaller- scale exposure assessment.  Finally in Chapter 4 we examined how three different measures of smoke exposure were associated with respiratory and cardiovascular outcomes in the first cohort study yet done on the population health effects of fire smoke.  Although dispersion modeling was not appropriate for exposure assessment in our study area, we did (1) demonstrate the potential usefulness of purely satellite- derived metrics while (2) confirming the respiratory risks associated with acute smoke exposure.  We hope that all of these results will contribute to future studies on important questions still surrounding the effects of fire smoke on cardiovascular health, and on populations in highly-exposed, under-monitored areas.  The methods described here set the stage for creative new approaches to future work on fire smoke exposure assessment and its related epidemiology.  For example, we know that (1) FRP measurements from MODIS are directly proportional to the aerosol  111 emitted by detected fires, (2) AOT measurements from MODIS are moderately correlated with surface PM10 concentrations, and (3) remote sensing imagery clearly shows smoke plumes.  These data could be used in combination with other fixed and time-varying spatial variables (air shed characteristics, elevation, visibility, wind direction, etc.) to develop regression equations for predicting daily TEOM-measured PM10 concentrations, similar to the methods used for land use regression29.  Further data for creating or evaluating such an approach could be collected over large areas with the same mobile monitoring technology that has been developed for measuring residential woodsmoke30.  This would eliminate the need for complex meteorological and dispersion models, especially in terrain where they are unlikely to produce reliable results.  It would also allow analysts to focus explicitly on populated areas so that resources are not committed to optimizing methods for regions that are epidemiologically irrelevant.  Because epidemiologic results based on new methods can be compared to those based on conventional exposure assessment (i.e. the TEOM metric in monitored areas), we can test and revise new tools with the objective of developing something spatially, temporally and quantitatively reliable.  Furthermore, estimates from such tools could be combined with conventional metrics to take full advantage of their strengths.  In this case, for example, we could have used air quality monitoring data for all subjects living within 5 kilometers of a TEOM, and then switched to CALPUFF estimates for all other subjects who were covered by visible SMOKE plumes.  Given the successes presented here, the potential of future work in this field is promising and exciting.  At the time of writing in July 2009 there is another large, uncontained wildfire burning on the outskirts of Kelowna.  Although it feels like an unhappy bookend to this story, it also highlights the need for ongoing work in this challenging field.  As the global climate changes we need to better understand how increased smoke exposure can impact the health of the billions of people living in smoke-affected areas.         112 5.9 References 1. Barna M and Lamb B (2000). Improving ozone modeling in regions of complex terrain using observational nudging in a prognostic meteorological model. Atmospheric Environment. 34: 4889-4906. 2. Burkholder BJ (2005) Report on the spatial assessment of forest fire smoke exposure and its health effects -- Part I: Initialization of the CALMET meteorological model. School of Occupational & Environmental Hygiene, The University of British Columbia. Vancouver, BC. 46 pages. http://hdl.handle.net/2429/13646 3. Burkholder BJ (2005) Report on the spatial assessment of forest fire smoke exposure and its health effects -- Part II: CALMET initialization methodology. School of Occupational & Environmental Hygiene, The University of British Columbia. Vancouver, BC. 28 pages. http://hdl.handle.net/2429/13645 4. Ichoku C and Kaufman YJ (2005). A method to derive smoke emission rates from MODIS fire radiative energy measurements. IEEE Transactions on Geoscience and Remote Sensing. 43(11): 2636-2649. 5. Dale VH, Joyce LA, McNulty S, Neilson RP, Ayres MP, Flannigan MD, Hansom PJ, Irland LC, Lugo AE, Peterson CJ, Simberloff D, Swanson FJ, Stocks BJ and Wotton MB (2001). Climate Change and Forest Disturbances. BioScience. 51(9): 723-734. 6. Flannigan MD, Stocks BJ and Wotton BM (2000). Climate change and forest fires. Sci Total Environ. 262(3): 221-229. 7. Adkins JW, O’Neill SM, Rorig M, Ferguson SA, Berg CA and L. HJ (2003). Assessing Accuracy of the BlueSky Smoke Modeling Framework During Wildfire Events. in Joint meeting of the International Wildland Fire Ecology and Fire Management Congress and the Symposium on Fire and Forest Meteorology. Orlando, FL. 8. Fusina L, Zhong S, Koracin J, Brown T, Esperanza A, Tarney L and Preisler H (2007). Validation of BlueSky Smoke Prediction System Using Surface and Satellite Observations During Major Wildland Fire Events in Northern California. in The fire environment—innovations, management, and policy; conference. Destin, FL. 9. Population and dwelling counts for Canada, Provinces and Territories, by Census Divisions, 2001 and 1996 censuses - 100% data., 93F0050XCB01003., Editor. 2002, Statistics Canada. 10. Hutchison KD, Smith S and Faruqui S (2004). The use of MODIS data and aerosol products for air quality prediction. Atmospheric Environment. 38(30): 5057-5070.  113 11. Kumar N, Chu A and Foster A (2007). An empirical relationship between PM2.5 and aerosol optical depth in Delhi Metropolitan. Atmospheric Environment. 41(21): 4492-4503. 12. Engel-Cox JA, Holloman CH, Coutant BW and Hoff RM (2004). Qualitative and quantitative evaluation of MODIS satellite sensor data for regional and urban scale air quality. Atmospheric Environment. 38(16): 2495-2509. 13. Hu Z (2009). Spatial analysis of MODIS aerosol optical depth, PM2.5, and chronic coronary heart disease. International Journal of Health Geographics. 8: doi:10.1186/1476-072X-8-27. 14. Klug W, Graziani G, Grippa G, Pierce D and Tassone C (1992), Evaluation of long term atmospheric transport models using environmental radioactivity data from Chernobyl accident. The ATMES Report: Elsevier. 366. 15. McKenzie D, O'Neill SM, Larkin NK and Norheim RA (2006). Integrating models to predict regional haze from wildland fire. Ecological Modelling. 199: 278-288. 16. Peffers L, Fuelberg H and Rao A (2009). Evaluation of smoke plume dispersion in complex terrain using a Lagrangian particle dispersion model driven by WRF output. in 11th Conference on Atmospheric Chemistry. Pheonix, AZ. 17. Dunson D (2001). Practical advantages of Bayesian analysis of epidemiologic data. American Journal of Epidemiology. 153(12): 1222-1226. 18. Mott JA, Meyer P, Mannino D, Redd SC, Smith EM, Gotway-Crawford C and Chase E (2002). Wildland forest fire smoke: health effects and intervention evaluation, Hoopa, California, 1999. West J Med. 176(3): 157-62. 19. Duclos P, Sanderson L and Lipsett M (1990). The 1987 Forest Fire Disaster in California: Assessment of Emergency Room Visits. Archives of Environmental Health. Jan, 45(1): 53-58. 20. Kunii O, Kanagawa S, Yajima I, Hisamatsu Y, Yamamura S, Amagai T and Ismail IT (2002). The 1997 haze disaster in Indonesia: its air quality and health effects. Arch Environ Health. 57(1): 16-22. 21. Sunyer J, Schwartz J, Tobías A, Macfarlane D, Garcia J and Antó JM (2000). Patients with Chronic Obstructive Pulmonary Disease Are at Increased Risk of Death Associated with Urban Particle Air Pollution: A Case-Crossover Analysis. American Journal of Epidemiology. 151: 50-56. 22. Hiltermann T, Stolk J, van der Zee S, Brunekreef B, de Bruijne C, Fischer P, Ameling C, Sterk P and Hiemstra P (1998). Asthma severity and susceptibility to air pollution. European Respiratory Journal. 11(3): 686-693. 23. Lagorio S, Forastiere F, Pistelli R, Iavarone I, Michelozzi P, Fano V, Marconi A, Ziemacki G and Ostro B (2006). Air pollution and lung function among susceptible adult subjects: a panel study. Environmental Health: A Global Access Science Source. doi:10.1186/1476-069X-5-11.  114 24. Rothman K and Greenland S (1998). Precision and Validity, in Modern EpidemiologyLippincott-Raven Publishers: Philadelphia, PA. p. 115-134. 25. Marmot M, Kogevinas M and Elston M (1987). Social/economic status and disease. Annual Review of Public Health. 8: 111-135. 26. Ichoku C and Kaufman YJ (2005). A method to derive smoke emission rates from MODIS fire radiative energy measurements. IEEE Transactions on Geoscience and Remote Sensing. 43(11): 26-36. 27. Delfino RJ, Brummel S, Wu J, Stern H, Ostro B, Lipsett M, Winer A, 6 DHS, Zhang L, Tjoa T and Gillen DL (2009). The relationship of respiratory and cardiovascular hospital admissions to the southern California wildfires of 2003. Occupational and Environmental Medicine. 66: 189-197. 28. Wu J, Winer A and Delfino R (2006). Exposure assessment of particulate matter air pollution before, during, and after the 2003 Southern California wildfires. Atmospheric Environment. 40(18): 3333-3348. 29. Briggs DJ, de Hoogh C, Gulliver J, Wills J, Elliott P, Kingham S and Smallbone K (2000). A regression-based method for mapping traffic-related air pollution: application and testing in four contrasting urban environments. Sci. Total Environ. 253(1-3): 151-167. 30. Larson T, Su J, Baribeau A-M, Buzzelli M, Setton E and Brauer M (2007). A Spatial Model of Urban Winter Woodsmoke Concentrations. Environmental Science and Technology. 41(7): 2429-2436.    115 Appendix 1: Ethics certificate   116 Appendix 2: Sample CALPUFF parameter file c---------------------------------------------------------------------- c --- PARAMETER statements                                      CALPUFF c---------------------------------------------------------------------- c --- Specify model version       character*12 mver, mlevel, mmodel       parameter(mver='5.724',mlevel='041013')       parameter(mmodel='CALPUFF') c c --- Specify parameters       parameter(mxpuff=1000000)       parameter(mxspec=20)       parameter(mxnx=650,mxny=500,mxnz=12)       parameter(mxnxg=650,mxnyg=500,mxrec=10000)       parameter(mxrfog=40)       parameter(mxss=250,mxus=99,mxps=300)       parameter(mxpt1=200,mxpt2=200,mxarea=6000,mxvert=5)       parameter(mxlines=200,mxlngrp=1,mxvol=200)       parameter(mxrise=50)       parameter(mxpdep=9,mxint=9)       parameter(mxoz=50,mxaq=50)       parameter(mxhill=20,mxtpts=25,mxrect=1000,mxcntr=21)       parameter(mxprfz=50)       parameter(mxent=10,mxntr=50,mxnw=5000)       parameter(mxvalz=10)       parameter(mxcoast=10,mxptcst=5000)       parameter(mxbndry=10,mxptbdy=5000)   ! keep mxbndry LE 20       parameter(mxmetdat=92, mxemdat=92)       parameter(mxmetsav=2)       parameter(mxsg=30)       parameter(io3=3,io4=4,io5=1,io6=2,io7=7,io8=8,io9=9)       parameter(io10=10,io11=11,io12=12,io15=15,io19=19)       parameter(io20=20,io22=22,io23=23,io24=24)       parameter(io25=25,io28=28,io29=29,io30=30,io31=31,io32=32)       parameter(io35=35,io36=36,io37=37)       parameter(iomesg=0)       parameter(iox=99)       parameter(iopt2=100)       parameter(ioar2=iopt2+mxemdat)       parameter(iovol=ioar2+mxemdat) c c c --- Compute derived parameters       parameter(mxbc=2*mxnx+2*mxny)       parameter(mxnzp1=mxnz+1)       parameter(mxvertp1=mxvert+1)       parameter(mxnxy=mxnx*mxny)       parameter(mxnxyg=mxnxg*mxnyg)  117       parameter(mxgsp=mxnxg*mxnyg*mxspec)       parameter(mxrsp=mxrec*mxspec)       parameter(mxcsp=mxrect*mxspec)       parameter(mx2=2*mxspec,mx5=5*mxspec,mx7=7*mxspec)       parameter(mxp2=2+mxspec,mxp3=3+mxspec)       parameter(mxp4=4+mxspec,mxp6=6+mxspec)       parameter(mxp7=7+mxspec,mxp8=8+mxspec,mxp14=mxspec+14)       parameter(mxpuf6=6*mxpuff)       parameter(mxlev=mxprfz)       parameter(mxprfp1=mxprfz+1)       parameter(mxentp1=mxent+1)       parameter(mxgrup=mxspec)       parameter(mxq12=mxspec*(mxpt1+mxarea)*2)       parameter(mxspar=mxspec*mxarea,mxspln=mxspec*mxlines)       parameter(mxsppt1=mxspec*mxpt1,mxspvl=mxspec*mxvol)       parameter(mxspbc=mxspec*mxbc) c c --- Specify parameters for sizing GUI:       parameter(mxavar=0)       parameter(mxlvar=0)       parameter(mxpvar=0)       parameter(mxvvar=0) c c --- GENERAL PARAMETER definitions: c        MXPUFF - Maximum number of active puffs allowed on the c                 computational grid at one time c        MXSLUG - Maximum number of active slugs allowed on the c                 computational grid at one time (can be set to c                 one if the slug option is not used) c        MXSPEC - Maximum number of chemical species.  N.B.: Changes c                 to MXSPEC may also require code changes to BLOCK DATA c                 and READCF. c        MXGRUP - Maximum number of Species-Groups.  Results for grouped c                 species are added together and reported using the c                 name of the group, rather than the name of one of the c                 species in the group. (MXGRUP = MXSPEC since specie c                 names are used as group names whenever group names are c                 not provided) c          MXNX - Maximum number of METEOROLOGICAL grid cells in c                 the X direction c          MXNY - Maximum number of METEOROLOGICAL grid cells in c                 the Y direction c          MXNZ - Maximum number of vertical layers in c                 the METEOROLOGICAL grid c         MXNXG - Maximum number of SAMPLING grid cells in c                 the X direction c         MXNYG - Maximum number of SAMPLING grid cells in c                 the Y direction c         MXREC - Maximum number of non-gridded receptors  118 c        MXRFOG - Maximum number of distances used when MFOG=1 c                 NOTE:  There are NPT1+NPT2 receptor 'trails', with c                        MXRFOG receptors on each, so c                        MXREC >= (NPT1+NPT2)*MXRFOG c          MXSS - Maximum number of surface meteorological stations c                 in the CALMET data c          MXUS - Maximum number of upper air stations in the CALMET c                 data c          MXPS - Maximum number of precipitation stations in the c                 CALMET data c          MXBC - Maximum number of sources used to represent boundary c                 conditions (inlux of background mass);  source c                 segments span the computational domain perimeter c         MXPT1 - Maximum number of point sources with constant c                 emission parameters c         MXPT2 - Maximum number of point sources with time-varying c                 emission parameters c        MXAREA - Maximum number of polygon area sources with constant c                 emission parameters (i.e., non-gridded area sources) c        MXVERT - Maximum number of vertices in polygon area source c        MXLINES- Maximum number of line sources c        MXLNGRP- Maximum number of groups of line sources c         MXVOL - Maximum number of volume sources c        MXRISE - Maximum number of points in computed plume rise c                 tabulation for buoyant area and line sources c        MXPDEP - Maximum number of particle species dry deposited c         MXINT - Maximum number of particle size intervals used c                 in defining mass-weighted deposition velocities c          MXOZ - Maximum number of ozone data stations (for use in the c                 chemistry module) c          MXAQ - Maximum number of Air Quality data stations (e.g. c                 H2O2 data stations for aqueous chemistry module) c        MXHILL - Maximum number of subgrid-scale (CTSG) terrain c                 features c        MXTPTS - Maximum number of points used to obtain flow c                 factors along the trajectory of a puff over the hill c        MXRECT - Maximum number of complex terrain (CTSG) receptors c        MXCNTR - Maximum number of hill height contours (CTDM ellipses) c        MXPRFZ - Maximum number of vertical levels of met. data in c                 CTDM PROFILE file c         MXLEV - Maximum number of vertical levels of met. data c                 allowed in the CTSG module (set to MXPRFZ in the c                 current implementation of CALPUFF) c         MXENT - Maximum number of perturbed entrainment coefficients c                 entered c         MXNTR - Maximum number of downwind distances for which c                 numerical plume rise will be reported c          MXNW - Maximum number of downwind distances for numerical c                 plume rise integration (should be set equal to  119 c                 SLAST/DS) c        MXVALZ - Maximum number of heights above ground at which valley c                 widths are found for each grid cell c       MXCOAST - Maximum number of coasts provided in COASTLN.DAT file c       MXPTCST - Maximum number of points used to store all coastlines c       MXBNDRY - Maximum number of boundaries provided in FLUXBDY.DAT c       MXPTBDY - Maximum number of points used to store all boundaries c      MXMETDAT - Maximum number of CALMET.DAT files used in run c       MXEMDAT - Maximum number of variable emissions files (each type) c      MXMETSAV - Maximum number of met periods for which source tables c                 (e.g. numerical rise) are saved c         MXQ12 - Maximum number of groups of 12 emission rate scaling c                 factors.  Factors come in groups of 12,24,36, or 96. c                 These are specified for source-species combinations, c                 but not all combinations will be filled.  Default c                 value of MXQ12 assumes that no more than 24 factors c                 are provided for each source-species combination for c                 point and area sources. c c --- CONTROL FILE READER definitions: c          MXSG - Maximum number of input groups in control file c c --- FORTRAN I/O unit numbers: c           IO3 - Restart file (RESTARTB.DAT)   - input  - unformatted c           IO4 - Restart file (RESTARTE.DAT)   - output - unformatted c           IO5 - Control file (CALPUFF.INP)    - input  - formatted c           IO6 - List file (CALPUFF.LST)       - output - formatted c c           IO7 - Meteorological data file      - input  - unformatted c                 (CALMET.DAT) c c           IO8 - Concentration output file     - output - unformatted c                 (CONC.DAT) c           IO9 - Dry flux output file          - output - unformatted c                 (DFLX.DAT) c          IO10 - Wet flux output file          - output - unformatted c                 (WFLX.DAT) c          IO11 - Visibility output file        - output - unformatted c                 (VISB.DAT) c          IO12 - Fog plume data output file    - output - unformatted c                 (FOG.DAT) c          IO15 - Boundary Condition file       - input  - unformatted c                 (BCON.DAT) c          IO19 - Buoyant line sources file     - input  - free format c                 (LNEMARB.DAT) with arbitrarily c                 varying location & emissions c          IO20 - User-specified deposition     - input  - formatted c                 velocities (VD.DAT) c          IO22 - Hourly ozone monitoring data  - input  - formatted  120 c                 (OZONE.DAT) c          IO23 - Hourly H2O2 monitoring data   - input  - formatted c                 (H2O2.DAT) c          IO24 - User-specified chemical       - input  - formatted c                 transformation rates c                 (CHEM.DAT) c          IO25 - User-specified coast line(s)  - input  - free format c                 for sub-grid TIBL module c                 (COASTLN.DAT) c          IO28 - CTSG hill specifications from - input  - formatted c                 CTDM terrain processor c                 (HILL.DAT) c          IO29 - CTSG receptor specifications  - input  - formatted c                 from CTDM receptor generator c                 (RECS.DAT) c          IO30 - Tracking puff/slug data       - output - formatted c                 (DEBUG.DAT) c          IO31 - CTDM "tower" data             - input  - formatted c                 (PROFILE.DAT) c          IO32 - CTDM surface layer parameters - input  - formatted c                 (SURFACE.DAT) c          IO35 - User-specified boundary lines(s)- input- free format c                 for mass flux calculations c                 (FLUXBDY.DAT) c          IO36 - Mass flux data                - output - formatted c                 (MASSFLX.DAT) c          IO37 - Mass balance data             - output - formatted c                 (MASSBAL.DAT) c         IOPT2 - 1st Pt. source emissions file - input  - unformatted c                 (PTEMARB.DAT) with arbitrarily           or free fmt c                 varying point source emissions c         IOAR2 - 1st Buoyant area sources file - input  - free format c                 (BAEMARB.DAT) with arbitrarily c                 varying location & emissions c         IOVOL - 1st Volume source file        - input  - unformatted c                 (VOLEMARB.DAT) with arbitrarily          of free fmt c                 varying location & emissions c        IOMESG - Fortran unit number for screen- output - formatted c                 output (NOTE: This unit is c                 NOT opened -- it must be a c                 preconnected unit to the screen c                 -- Screen output can be suppressed c                 by the input "IMESG" in the c                 control file) c           IOX - Fortran unit number for      - scratch - formatted c                 temporary file of "doc" records c                 written to header of output files c c  121 c --- GUI memory control parameters:  variable emissions scaling factors c     for areas, lines, points, and volumes require much memory in GUI. c     To reduce GUI memory requirement, set one or more of the c     following parameters to ZERO when such scaling is not required. c     These parameters have no effect on CALPUFF, but are read by the c     GUI at execution time. c c        MXAVAR - Using scaled area sources?   (1:yes, 0:no) c        MXLVAR - Using scaled line sources?   (1:yes, 0:no) c        MXPVAR - Using scaled point sources?  (1:yes, 0:no) c        MXVVAR - Using scaled volume sources? (1:yes, 0:no) c c c -----------------------------------------------------------------     122 Appendix 3: Event counts for the summers of 2002, 2003 and 2004   Figure A3.1 – Daily count of respiratory physician visits    Figure A3.2 – Daily count of cardiovascular physician visits  123  Figure A3.3 – Daily count of respiratory hospital admissions    Figure A3.4 – Daily count of cardiovascular hospital admissions   124 Appendix 4: CALPUFF estimates from different emissions models As described in Chapters 2 and 3 we used a simple calculation on the MODIS fire radiative power measurement (FRP) to estimate aerosol emissions for this thesis.  As discussed in Chapter 2 we tested this method against two much more complex models of fire emissions – the Canadian Fire Behavior Prediction System (CANFB) and the US Emissions Production Model (USEPM).  What did not get shown in either chapter is a comparison between CALPUFF results when these three different methods are used. National differences between fuels and their classification systems make it challenging to interchange the USEPM and CANFB methods between the US and Canada.  The USEPM method uses the US Fuel Characteristic Classification System, which specifies 113 unique fuel codes with density and moisture estimates under multiple conditions. The CANFB system specifies 16 broader fuel categories with no density or moisture information.  To address this discrepancy we simply used category descriptions, frequencies and proximities to equate all USEPM fuel codes occurring in US MODIS events to a single Canadian CANFB code, and vise versa.   A detailed breakdown is shown in Table 1.  Figure 1 through Figure 6 show CALPUFF output for all three methods compared to TEOM measurements at Kamloops, Kelowna, Vernon, Creston, Revelstoke and Golden (see Figure 1.2).  Note that that modeled estimates are offset from the TEOM measurements to better show how well they track with the peak concentrations.  Therefore the only the TEOM concentrations are accurately reflected by the y-axes.                 125   Table A4.1 – The CANFB and USEPM fuel classification systems Summary of the equalities made between Canadian (CANFB) and US (USEPM) fuel classification systems for cross-border application of the aerosol emissions models.  Where multiple EPM codes were equated to one FBP code for the Canadian model, that FBP code was equated to the most prevalent of EPM codes for the US model, as shown in gray. Canadian Classification US Classification FBP Code Description % CND Detects EPM Code Description % US Detects C2 Boreal Spruce 12.0 59 Subalpine Fir / Engelmann Spruce / Lodgepole Pine 15.7 C3/4 Jack or Lodgepole Pine 12.1 22 Lodgepole Pine 28.8 18 Douglas Fir / Ponderosa Pine 15.8 C7 Ponderosa Pine / Douglas Fir 4.7 52 Douglas Fir / Pacific Ponderosa Pine / Oceanspray 1.7 D1 Aspen 12.3 42 Trembling Aspen - 238 Pacific Silver Fir / Mountain Hemlock  11.9 M1/2 Boreal Mixed Wood 22.1 8 Western Hemlock / Douglas Fir / Western Red Cedar / Vine Maple 2.1 U Unclassified / National Park 26.8 24/28 Ponderosa Pine / Douglas Fir / Unclassified 24.0 O1 Grass 9.4 63 Showy Sedge / Alpine Black Sedge Grassland 0.1                   126    Figure A4.1 – Three CALPUFF estimates vs. TEOM measurements at Kamloops    Figure A4.2 – Three CALPUFF estimates vs. TEOM measurements at Kelowna  127  Figure A4.3 – Three CALPUFF estimates vs. TEOM measurements at Vernon    Figure A4.4 – Three CALPUFF estimates vs. TEOM measurements at Creston  128  Figure A4.5 – Three CALPUFF estimates vs. TEOM measurements at Revelstoke    Figure A4.6 – Three CALPUFF estimates vs. TEOM measurements at Golden  129 Appendix 5: Notes on cohort identification Under ideal conditions the study cohort would include all people in the Medical Services Plan (MSP) Registration and Premium Billings (R&PB) file whose postal code indicates they lived in the study area during the study period.  However, the nature of this file makes it impossible to know whether its postal codes are up-to-date and/or reflective of a residential address.  To reduce potential bias due to this uncertainly the cohort was identified using a combination of the MSP Payment Information (referred to as ‘MSP’ herein), Hospital Separations (referred to as ‘HS’ herein) and Vital Statistics (referred to as ‘VS’ herein) files.  By limiting the cohort to people who had multiple contacts with the healthcare system from matching 6-digit postal codes between July 1st 2002 and September 30th 2004 we hoped to more accurately capture each subject’s residential location, thereby maximizing the quality of our exposure estimates.  The steps used by programmers at the Center for Health Services and Policy Research (CHSPR) to identify the study cohort are outlined below:  1. Search the MSP file between July 1st 2002 and June 30th 2003 for all entries with a 6-digit postal code in the study area.  Where a unique personal health number (PHN) occurs multiple times in this period, keep the latest entry and discard all others.  Output a file containing all PHNs and 6-digit postal codes.  Label the postal code field 1A.  2. Repeat for the HS and VS Birth files.  Label the postal code fields 1B and 1C, respectively.  3. Merge the three files by PHN and add a Boolean field labelled PRE, giving all PHNs a value of TRUE.  Output a file of potential cohort members containing the fields PHN, PRE, 1A, 1B and 1C (the latter three can contain no data but the former two cannot).  Record the number of entries.  4. Search the MSP file between July 1st 2003 and September 30th 2004 for all entries with a 6-digit postal code in the study area.  Where a unique PHN occurs multiple times, keep the earliest entry and discard all others. Output a file containing all PHNs and 6-digit postal codes.  Label the postal code field 2A.  5. Repeat for the HS and VS Death files.  Label the postal code fields 2B and 2C, respectively.  6. Merge these three files by PHN and add a Boolean field labelled POST, giving all PHNs a value of TRUE.  Output a file containing the fields PHN, POST, 2A, 2B and 2C (once again the latter three can contain no data but the former two cannot).  Record the number of entries.  130   7. Merge the PRE and POST files by PHN and then:  • If a PHN does not have TRUE values for both PRE and POST, then discard the entry.  Record the number of deletions.  • If one of 1A or 1B or 1C does not match one of 2A or 2B or 2C (on five of the six digits), then discard the entry.  Record the number of deletions.  • Forward a list of remaining PHNs with the repeated 6-digit postal code to the Ministry, where a list of potentially matching residential addresses will be generated.                                      131  Appendix 6: Notes on cohort geolocation A sample of the raw address file as received from the BC Ministry of Health is compared to the cleaned addresses used for geolocation.  The cleaning algorithm written by Dr. Jason Su is described below.  Table A6.1 – A sample of raw and cleaned addresses for geolocation Raw Cleaned # B 3203 31 AVE VERNON BC 3203 31 AVE VERNON #305 1779 PANDOSY ST KELOWNA BC 1779 PANDOSY ST KELOWNA 08-460 DALGLEISH DR KAMLOOPS BC 460 DALGLEISH DR Kamloops 1 1336 COLUMBIA STREET KAMLOOPS BC 1336 COLUMBIA ST Kamloops 1 1890 ORME BOX 2277 MERRITT BC 1 1890 ORME MERRITT 1 1950 burtch rd KELOWNA BC 1950 BURTCH RD KELOWNA 1 2319 PANDOSY ST KELOWNA BC 2319 PANDOSY ST KELOWNA 10 631 FORTUNE RD. KAMLOOPS BC 631 FORTUNE RD Kamloops 1002 6TH ST CASTLEGAR BC 1002 6TH ST CASTLEGAR 10088 GREY ROAD COLDSTREAM BC 10088 GREY RD Coldstream 101 1370 TRANQUILLE ROAD KAMLOOPS BC 1370 TRANQUILLE RD Kamloops 101 2005 BOUCHERIE RD WESTBANK BC 2005 BOUCHERIE RD Westbank 101 3350 10 AVE NE SALMON ARM BC 3350 10 AVE NE Salmon Arm 1012 WINTERGREEN DRIVE KELOWNA BC 1012 WINTERGREEN DR KELOWNA 1017 CALMELS CR KELOWNA BC 1017 CALMELS CRES KELOWNA 101B 637 VAN HORNE ST PENTICTON BC 637 VAN HORNE ST Penticton 102 2275 ATKINSON ST PENTICTON BC 2275 ATKINSON ST Penticton 102 2360 SOUTH MAIN ST PENTICTON BC 2360 S MAIN ST Penticton 1023 BX ROAD VERNON BC 1023 BX RD VERNON 1027 NICOLA ST KAMLOOPS BC 1027 NICOLA ST Kamloops 102A 1375A SUMMIT DRIVE KAMLOOPS BC 1375 SUMMIT DR Kamloops   Link to spatial coordinates based on 6 digital postal codes A 2003 postal code (PC) conversion file with coordinates and 6 digital postal codes from Census of Canada was used to link the 6 digital postal codes of the addresses using a binary search algorithm. However, because a single postal code might have more than  132 one match on the PC conversion file, the linkage only reflects a proximity to the coordinates assigned.  Clean data to standard addresses for geocoding The purpose of data cleaning is to create a standard street address recognizable by geocoding software. A standard street address after cleaning should follow this sequence but not necessarily with all of them: street number, prefix direction, prefix type, street name, suffix type and suffix direction. Because prefix type is only for highway in the province of British Columbia (BC), addresses with key word ‘HWY’ are treated separately. Data cleaning involves three basic steps: a) Cleaning preparation stage, b) Remove redundant information around a known town, and c) Extract a standard street address.  Data cleaning preparation During the preparation stage, the first effort focuses on standardizing street prefix and suffix types of addresses. These include standardizing Dr, Lake, Cres, Ct, Ave, Rd, N, S, E, W, NW, NE, SW and SE. For example, suffix type ‘Dr’ was seen in raw addresses as DRI, DRIV, DRIVE and DR. Suffix type ‘Cres’ was seen as CR, CRE, CRS, CRESC, CRESCENT and CRES. These two suffix types are standardized to ‘Dr’ and ‘Cres’, respectively. The abbreviation of province BC was seen as ‘B C’, ‘BC B C’ and other formats at the end of addresses. These are corrected as a single word ‘BC’. PO Box, C/O and general delivery information are also removed. For example, the variety of a PO Box includes ‘P O Box 3080’, ‘PO Box 3080’, ‘Box 3080’, these information are removed including the numeric values. Another issue with the addresses are the single letters, for example, ‘120 B 3 A ST N Cranbrook BC’. Letter ‘B’ is an apartment number while ‘A’ is part of a street name. These single letters (i.e., A, B, C, D, E and F) are combined with their respective preceding words for later processing. Special characters like !?,;:""'()[]+-_/ are also filtered. Consecutive empty spaces are replaced with single spaces.       133 Remove redundant information around town name The town name database file is firs used to identify the position of a town name in an address. Sometimes a town name might include more than one word and part of it might be the name of another town. For example, the word ‘Fort George’ in the town named ‘Fraser Fort George’ is also an eligible town name in the province of BC. To avoid part of the town name being treated as unwanted information (i.e., Fraser), the town name data after importing into a dynamic town array are immediately sorted in a descending order based on the length of a town name.  We assume each address before cleaning should have a suffix type. A suffix type is identified by searching each word of an address from right to left against the suffix type database created from the preparation stage. Caution should be taken to avoid a true suffix type being removed as unwanted information because of existence of a fake street suffix type. For example, in the address RR2 S1A C 20 2475 SILVERY BEACH RD CHASE BC, ‘Beach’ is an eligible street suffix type as ‘Rd’ in the suffix type database. If ‘Beach’ is in front of ‘RD’ in the suffix type database, ‘Beach’ will be identified first as a street suffix type of the address. In this situation, a second search on the address right side of the identified suffix type (i.e., Beach) should be sought, also, against the suffix type database to find if another suffix type exists. If a suffix type exists, then the pointer moves to the word immediate after it to see if a suffix direction exists by checking against one of the eight suffix directions created in a database file during the preparation stage.  If a suffix direction and a town name exist for an address, or if a suffix direction does not exists but a suffix type and a town name, everything between is removed. Also, if a town name is identified, everything after the identified town name in an address is removed. If a suffix type does not exist, then nothing is done for that address. Removing redundant information around a town name is much easier than removing redundant information from a street name. Therefore, it is executed first. Cleaning data around a town name also reduces the unwanted complexity in cleaning the street addresses followed.    134 Clean data around street name and extract street address After removing redundant information around a town name, an address is further cleaned to have only street address left for geocoding.  Each address is checked against the street name database created on the data preparation stage. Like dealing with town name, to avoid treating part of the street name as a street name, the street name database is also sorted in a descending order based on the length of the street names when they are imported into a dynamic array. If an address is found to have a corresponding street name in the database, the word immediate right side of the identified street name in the address is checked to see if a suffix type exists. Again, if a suffix type is found, we need to identify if that suffix type is a real or fake suffix type. If a real suffix type is identified, the word immediate right side of the suffix type is checked to see if there is an existence of a suffix direction. The middle part (street name) and right side (suffix type/direction) of a street address is identified. An effort is then sought to identify the possibility of existence of a prefix direction left side of the identified street name against the prefix direction database. If it does, the street address position on the left side is moved to the first character of the prefix direction. A street number is immediately sought left side of the prefix direction (if it exists) or the street name (if a prefix direction does not exist) by identifying a word whose first character is numeric. Anything left side of the street number is then removed, so does everything between the street number and the identified street name. For example, in the address ‘RR SS 2800A GENERAL DELIVERY SUSAN MYER N WEGWEG VALLEY RD SUMMERLAND’. ‘Wegweg Valley’ is an eligible street name in the street name database. ‘2800A’ is a street number. After cleaning, the address changes to ‘2800A N Wegweg Valley Rd Summerland’.  If a street name does not exist in the street name database, then a town name is searched against the dynamic array of town name which is sorted by length in a descending order. A suffix type left side of the identified town name is also checked to identify its existence. A word immediate left side of the identified suffix type is checked to see if its first character is numeric. If it does, the word immediate left side of this word is also checked to see if its first character is numeric. If both do, we have a street address with those two words plus the suffix. We also need to make sure if there is an  135 existence of a suffix direction. The situations like these are, for example, 245A 3RD ST N Vernon or 245 3 ST Vernon. If the first character of a word immediate left side of the identified suffix type is not numeric, then the search goes left side, one word at a time, until a word with its first character numeric is found. The proportion from the last identified word to its suffix type/direction is identified as the street address of that entry. Everything in front of the identified street address is removed, so does everything between the identified street address and its town name.  In all of the above cleaning procedures, highway addresses are not processed and are dealt with in the last step. Basically there are two types of addresses with highway data: a) the highway number is after key word ‘HWY’ but the street number is on the left side of the key word, such as an address ‘R R 5 S 16A C 7 186C HWY 97A Vernon BC’. b) The street and highway numbers both are on the left side of the key word ‘HWY’, like ‘80B 97A HWY Vernon BC’. The algorithm first identifies the position of the word ‘HWY’, and then seeks one word immediate after it. If the first character of the sought word is numeric, then that word is identified as a highway number. Another word immediate right side of the highway number is checked against the suffix direction database. If a suffix direction exists, then that direction is added to the highway address string. The word immediate left side of the key word ‘HWY’ is also checked to see if the first character of the word is numeric. If it does, then the address is identified as a string from the word immediate before the key word ‘HWY’ to its suffix direction if that direction exists or to a highway number if a suffix direction does not exists. If a first sought from the word immediate after the key word ‘HWY’ finds that the first character of the word is not numeric, then the algorithm is shifted to the second situation: the street and highway numbers are both situated on the left side of the key word ‘HWY’. The algorithm then checks the first word left side of the key word ‘HWY’. If the first character of that word is numeric, then it is identified as a highway number. Another word immediate left side of the highway number is sought to check if its first character is numeric. If it does, then it is identified as a street number. The string from the identified street number to the key word ‘HWY’ is then treated as the street address of that entry. Everything left side and right side of the identified highway street address is removed except the town name.   136 After implementing the above three cleaning procedures, apartment number is the last redundant information left. The final cleaning procedure finds the word left side of a street name identified from the above three procedures with its first character being numeric. The last character of type string is then removed from the word left side of the identified street name and whose first character is numeric. Caution should be taken to ensure the letter being removed is an apartment number, not part of a street name. For example, for the address ‘99A 15A AVE VERNON BC’; letter ‘A’ in ‘15A’ is part of the street name ‘15A Ave’, not an apartment number. However, the letter ‘A’ in ‘99A’ should be removed.    137 Appendix 7: ADDRESS vs. POSTAL results Comparison of age-stratified respiratory physician visit results based on postal vs. address exposure assignment   Figure A7.1 – Address vs. postal exposure for the TEOM metric Compares odds ratios for the address-assigned TEOM concentrations in the address- matched cohort (281 711 subjects) to postal-assigned TEOM concentrations in the addresss- matched cohort and the full cohort (439 887 subjects).         138  Figure A7.2 – Address vs. postal exposure for the SMOKE metric Compares odds ratios for the address-assigned SMOKE indicator in the address-matched cohort (281 711 subjects) to postal-assigned SMOKE indicator in the address-matched cohort and the full cohort (439 887 subjects).   139  Figure A7.3 – Address vs. postal exposure for the CALPUFF metric Compares odds ratios for the address-assigned CALPUFF concentrations in the address- matched cohort (281 711 subjects) to postal-assigned CALPUFF concentrations in the address-matched cohort and the full cohort (439 887 subjects). 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0067764/manifest

Comment

Related Items