Getting Down to Earth INTERNATIONAL DE VELOPMENT IN FOCUS Are Satellites Reliable for Measuring Air Pollutants That Cause Mortality in Low- and Middle-Income Countries? I N T E R N AT I O N A L D E V E L O P M E N T I N F O C U S Getting Down to Earth Are Satellites Reliable for Measuring Air Pollutants That Cause Mortality in Low- and Middle-Income Countries? © 2022 International Bank for Reconstruction and Development / The World Bank 1818 H Street NW, Washington, DC 20433 Telephone: 202-473-1000; Internet: www.worldbank.org Some rights reserved 1 2 3 4 25 24 23 22 Books in this series are published to communicate the results of World Bank research, analysis, and oper- ational experience with the least possible delay. The extent of language editing varies from book to book. This work is a product of the staff of The World Bank with external contributions. The findings, interpre- tations, and conclusions expressed in this work do not necessarily reflect the views of The World Bank, its Board of Executive Directors, or the governments they represent. The World Bank does not guarantee the accuracy, completeness, or currency of the data included in this work and does not assume responsibility for any errors, omissions, or discrepancies in the information, or liability with respect to the use of or failure to use the information, methods, processes, or conclusions set forth. The boundaries, colors, denominations, and other information shown on any map in this work do not imply any judgment on the part of The World Bank concerning the legal status of any territory or the endorsement or acceptance of such boundaries. Nothing herein shall constitute, imply, or be considered to be a limitation upon or waiver of the privileges and immunities of The World Bank, all of which are specifically reserved. Rights and Permissions This work is available under the Creative Commons Attribution 3.0 IGO license (CC BY 3.0 IGO) http:// creativecommons.org/licenses/by/3.0/igo. Under the Creative Commons Attribution license, you are free to copy, distribute, transmit, and adapt this work, including for commercial purposes, under the following conditions: Attribution—Please cite the work as follows: World Bank. 2022. Getting Down to Earth: Are Satellites Reliable for Measuring Air Pollutants That Cause Mortality in Low- and Middle-Income Countries? International Development in Focus. Washington, DC: World Bank. doi:10.1596/978-1-4648-1727-4. License: Creative Commons Attribution CC BY 3.0 IGO Translations—If you create a translation of this work, please add the following disclaimer along with the attribution: This translation was not created by The World Bank and should not be considered an official World Bank translation. The World Bank shall not be liable for any content or error in this translation. Adaptations—If you create an adaptation of this work, please add the following disclaimer along with the attribution: This is an adaptation of an original work by The World Bank. Views and opinions expressed in the adaptation are the sole responsibility of the author or authors of the adaptation and are not endorsed by The World Bank. Third-party content—The World Bank does not necessarily own each component of the content con- tained within the work. The World Bank therefore does not warrant that the use of any third-party-owned individual component or part contained in the work will not infringe on the rights of those third parties. The risk of claims resulting from such infringement rests solely with you. If you wish to re-use a compo- nent of the work, it is your responsibility to determine whether permission is needed for that re-use and to obtain permission from the copyright owner. Examples of components can include, but are not limited to, tables, figures, or images. All queries on rights and licenses should be addressed to World Bank Publications, The World Bank Group, 1818 H Street NW, Washington, DC 20433, USA; e-mail: pubrights@worldbank.org. ISBN: 978-1-4648-1727-4 DOI: 10.1596/978-1-4648-1727-4 Cover photo: © Maximus256/Shutterstock.com. Used with permission; further permission required for reuse. Cover design: Debra Naylor / Naylor Design Inc. Contents Preface  ix About This Work   xi Acknowledgments  xiii Executive Summary  xv Abbreviations  xxi Glossary  xxiii CHAPTER 1 Contextual Background and Objectives   1 CHAPTER 2 Introduction  3 Report methodology  8 Outline of this report   9 Note  9 References  9  iterature Review of Approaches Used to Combine Satellite CHAPTER 3 L Observations with Ground-Level Monitoring in Urban Areas  1 1 Retrieval of AOD from satellite observations of reflected solar radiation  12 The translation of AOD to ground-level PM2.5 concentrations  16 Combining satellite PM2.5 estimates and GLM data into high-resolution grids centered on selected cities   18 Notes  18 References  19  uality-Assurance Procedures in Low- and Middle-Income CHAPTER 4 Q Countries   21 Issues affecting the availability and accuracy of GLM data in LMICs  21 Using automated scripts to flag potentially incorrect GLM data   24 Examining metadata, warning files, and operator logs   25 Note  26  sing Satellite Data for Daily, City-Scale PM2.5 CHAPTER 5 U Monitoring  27 Initial testing of satellite products in selected cities   28 Identifying the type of cities where satellites work best   34 Notes  41 References  41  iii iv | Getting Down to Earth CHAPTER 6 Conclusions and Recommendations   43 Conclusions  44 Recommendations  47 Note  50 References  51 Appendix A Literature Review 53  onverting Satellite Aerosol Optical Depth to Appendix B C Ground-Level PM2.5  97 Appendix C Evaluation of Satellite Approaches 121  uality Considerations for Ground-Level-Monitoring Data in Appendix D Q Low- and Middle-Income Countries 149 Box 4.1 Open-access web platform for air quality data   22 Figures 2.1 MODIS Aqua image of wildfire smoke being transported eastward to the Chesapeake Bay, US   4 2.2 Current and future coverage provided by available polar and geostationary weather satellites  5 2.3 How satellites can measure aerosol optical depth from reflected sunlight   6 2.4 Highly idealized example of three different aerosol vertical profiles, each with different ground-level PM2.5 concentrations, that result in the same aerosol- optical-depth value of one when measured by satellites   7 3.1 Annual average ground-level PM10 concentrations, London, UK, and Rome, Italy  12 3.2 MODIS Dark Target 10-kilometer and 3-kilometer aerosol-optical-depth products retrieved for clear land and ocean fields of view and the local 5-kilometer average derived from the products, outer circle, compared to ground-based measurements, inner circle, over Baltimore, US   14 3.3 Example of good-quality aerosol optical depth 550 nm observations from a geostationary satellite from the US GOES-R SMAOD product   15 3.4 Valid MODIS Terra aerosol-optical-depth retrievals and Lima, Peru, monitoring sites in the OpenAQ database, 2016–17    16 4.1 Example of US federal equivalent method measurement and uncertainties: Data set slope, intercept, and limits   23 5.1 Comparison of the chemical-transport-model–based estimates for daily average ground-level PM2.5 concentrations in micrograms per cubic meter over Delhi, India, using VIIRS and MODIS Terra, November 1, 2017   32 5.2 Comparison of the chemical-transport-model–based estimates for daily average ground-level PM2.5 concentrations in micrograms per cubic meter over Accra, Ghana, using SEVIRI and MODIS Aqua, December 26, 2017   33 5.3 Correlation coefficient R for the statistical and chemical-transport-model– based methods for the different low- and middle-income country cities   37 5.4 Mean normalized bias for the statistical and chemical-transport-model–based methods for the different low- and middle-income country cities   37 5.5 Mean normalized gross error for the statistical and chemical-transport-model– based methods for the different low- and middle-income country cities   38 A.1 Coverage provided by available polar and geostationary weather satellites  57 A.2 MODIS Dark Target 10-kilometer and 3-kilometer aerosol-optical-depth products retrieved for clear land and ocean fields of view and the local 5-kilometer average derived from the products, outer circle, compared to ground-based measurements, inner circle, over Baltimore, US    60 Contents | v A.3 Comparison of VIIRS 750-meter and MAIAC one-kilometer aerosol-optical- depth products during western US fires, 2013   63 A.4 Detection of temporal variations in aerosol optical depth with the SEVIRI-MSG aerosol over land product compared to AERONET measurements at Palaiseau, France, and to MODIS, July 14, 2006   66 A.5 Full disk coverage at 550 nanometers of GOES-R SMAOD product over ocean and land  67 A.6 Planetary boundary layer height in areas surrounding Washington, DC, US, on July 14, 2011, at 21:00 UTC   72 B.1 Illustration of the regions covered by a 1° by 1° box around each focus region  100 B.2 Local Ulaanbaatar, Mongolia, ground-level-monitoring network and US Diplomatic Post annual average PM2.5 concentrations as well as a gridded estimate produced via ordinary kriging   105 B.3 Example of satellite coverage showing the CTM–based estimates for daily average ground-level PM2.5 concentrations for all valid MODIS Terra and Aqua aerosol-optical-depth retrievals, Lima, Peru, May 12, 2017   107 B.4 Residual evaluation plots for Ulaanbaatar, Mongolia, using the MODIS Aqua aerosol-optical-depth product  108 B.5 Observed and GAM–predicted daily average PM2.5 values for the US Diplomatic Posts in Delhi, India, and Ulaanbaatar, Mongolia, for both MODIS Aqua and Terra   109 B.6 Observed and CTM–based estimates daily average PM2.5 values for the US Diplomatic Posts in Delhi, India, and Ulaanbaatar, Mongolia, for both MODIS Aqua and Terra   110 B.7 Monthly average PM2.5 surface concentration, 2017   112 B.8 Observed values and VIIRS-based estimates of daily average PM2.5 concentrations for the Income Tax Office in Delhi, India, using GAM and CTM–based approaches  114 B.9 A comparison of the CTM–based estimates for daily-average ground-level PM2.5 concentrations over Delhi, India, using VIIRS and MODIS Terra, November 1, 2017   115 B.10 A comparison of the CTM–based estimates for daily-average ground-level PM2.5 concentrations over Accra, Ghana, using SEVIRI and MODIS Aqua, December 26, 2017  116 C.1 Residual evaluation plots for the generalized additive model trained on the Terra satellite data, Delhi, India   123 C.2 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Terra aerosol optical depth and the MERRA planetary-boundary-layer height, Delhi, India    124 C.3 Scatterplots of the Terra satellite estimate of PM2.5 from the generalized additive model and the chemical transport model versus the OpenAQ data, US Diplomatic Post in Delhi, India   124 C.4 Monthly average PM2.5 for each method, Delhi, India   126 C.5 Annual average PM2.5 for each method, by monitor site, Delhi, India   126 C.6 Residual evaluation plots for the generalized additive model trained on the Aqua satellite data, Ulaanbaatar, Mongolia   127 C.7 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Aqua aerosol optical depth and the MERRA planetary-boundary-layer height, Ulaanbaatar, Mongolia  128 C.8 Scatterplots of the generalized-additive-model– and chemical-transport- model–based concentrations using the Aqua aerosol-optical-depth data, versus OpenAQ data, US Diplomatic Post in Ulaanbaatar, Mongolia   129 C.9 Monthly average PM2.5 for each method, Ulaanbaatar, Mongolia   130 C.10 Annual average (March–November) PM2.5 for each method by monitor site, Ulaanbaatar, Mongolia  130 C.11 Valid Terra aerosol-optical-depth retrievals for monitoring sites in the OpenAQ database, Lima, Peru, 2016–17   131 vi | Getting Down to Earth C.12 Scatterplot of the chemical-transport-model–based PM2.5 concentrations using the Terra aerosol-optical-depth data, versus ground-level-monitoring PM2.5 data, Malam Junction site in Accra, Ghana   132 C.13 Residual evaluation plots for the generalized additive model trained on the Aqua satellite data, Kathmandu, Nepal   133 C.14 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Aqua aerosol optical depth and the MERRA planetary-boundary-layer height, Kathmandu, Nepal   134 C.15 Scatterplots of the generalized-additive-model– and chemical-transport- model–based concentrations using the Aqua aerosol-optical-depth data, versus OpenAQ data, US Diplomatic Post in Kathmandu, Nepal   134 C.16 Residual evaluation plots for the generalized additive model trained on the Terra satellite data, Addis Ababa, Ethiopia   135 C.17 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Terra aerosol optical depth and the MERRA planetary-boundary-layer height, Addis Ababa, Ethiopia   136 C.18 Scatterplots of the generalized-additive-model– and chemical-transport- model–based concentrations, using the Aqua aerosol-optical-depth data, versus OpenAQ Data, US Diplomatic Post in Addis Ababa, Ethiopia   137 C.19 Scatterplot of the chemical-transport-model–based PM2.5 concentrations using the Terra aerosol-optical-depth data versus citywide average PM2.5 data, Dakar, Senegal  137 C.20 Residual evaluation plots for the generalized additive model trained on the Aqua satellite data, Kampala, Uganda   139 C.21 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Aqua aerosol optical depth and the MERRA planetary-boundary-layer height, Kampala, Uganda   140 C.22 Scatterplots of the generalized-additive-model– and chemical-transport- model–based PM2.5 concentrations, using the Aqua aerosol-optical-depth data, versus OpenAQ data, US Diplomatic Post, Kampala, Uganda   140 C.23 Residual evaluation plots for the generalized additive model trained on the Terra satellite data, Hanoi, Vietnam   141 C.24 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Terra aerosol optical depth and the MERRA planetary-boundary-layer height, Hanoi, Vietnam   142 C.25 Scatterplots of the generalized-additive-model– and chemical-transport- model–based concentrations using Terra aerosol-optical-depth data, versus OpenAQ data, US Diplomatic Post in Hanoi, Vietnam   143 C.26 Difference in base processing monthly average versus monthly averages after the quality assurance tool is applied   144 C.27 Generalized-additive-model residual evaluation plots from Terra quality assurance tool  145 C.28 Scatterplots of the statistical and chemical-transport-model–based concentrations using the Aqua aerosol-optical-depth data versus the QA filtered OpenAQ data, US Diplomatic Post in Ulaanbaatar, Mongolia   146 D.1 Picture of the Met One beta attenuation monitor unit showing the beta source and the filter tape used to collect the particulate matter samples   150 D.2 Schematic diagram of a tapered element oscillating microbalance    152 D.3 Tapered element oscillating microbalance–filter dynamics measurement system schematic  154 D.4 Time-series, seasonal, monthly, weekly, and diurnal PM2.5 mass concentration data measured from the US Diplomatic Post in Delhi, India   157 D.5 PM2.5 trends from data acquired at the Mandir Marg, New Delhi, India–DPCC monitoring site, 1.45 miles west of US Diplomatic Post   158 D.6 PM2.5 trends from data acquired at the Shadipur, New Delhi, India–CPCB monitoring site, 4.75 miles northwest of US Diplomatic Post   158 D.7 PM2.5 trends from data acquired at the Punjabi Bagh, Delhi, India–DPCC monitoring site, 6.91 miles northwest of US Diplomatic Post   159 D.8 PM2.5 trends from data acquired at the NSIT Dwarka, Delhi, India–CPCB monitoring site, 11 miles southwest of US Diplomatic Post   159 Contents | vii D.9 PM2.5 trends from data acquired at the Ramakrishna Puram, Delhi, India–DPCC monitoring site, 5.69 miles south-southwest of US Diplomatic Post   160 D.10 PM2.5 trends from data acquired at the Anand Vihar, Delhi, India–DPCC monitoring site, 4.6 miles east-northeast of US Diplomatic Post   160 D.11 Time-series, seasonal, monthly, weekly, and diurnal PM2.5 mass concentration data measured from the US Diplomatic Post in Lima, Peru   162 D.12 PM2.5 trends from data acquired at the San Borja, Peru, monitoring site, 2.7 miles southwest of US Diplomatic Post   162 D.13 PM2.5 trends from data acquired at the Campo de Marte, Peru, monitoring site, 5.4 miles northwest of US Diplomatic Post   163 D.14 PM2.5 trends from data acquired at the Santa Anita, Peru, monitoring site, 3.9 miles north of US Diplomatic Post   163 D.15 PM2.5 trends from data acquired at the Ate, Peru, monitoring site, 6.1 miles northeast of US Diplomatic Post   164 Maps A.1 Fraction of good-quality attempted retrievals from Deep Blue and Dark Target algorithms showing differences in coverage over desert regions, and showing differences in coverage due to the quality checks applied in each algorithm  61 A.2 Radiosonde launch locations for 00:00 UTC, December 8, 2017   73 A.3 Aircraft observation coverage from AMDAR and MDCRS, December 8, 2017  74 D.1 Proximity and distribution of government-run air-quality monitoring stations around the US Diplomatic Post in Delhi, India   156 D.2 Illustration of the proximity of the US Diplomatic Post in Lima, Peru, relative to a subset of the nearest government-run assurance quality monitoring stations for which OpenAQ PM2.5 data were available   164 Tables ES.1 Cities included in this analysis   xvi 3.1 Strengths and weaknesses of the different daily satellite aerosol-optical-depth products publicly available in near real-time    13 3.2 Strengths and weaknesses of different approaches to converting satellite aerosol-optical-depth data into estimates of ground-level PM2.5 concentrations  17 4.1 PM2.5 instrument manufacturers’ stated uncertainties    24 5.1 Statistics for the Delhi, India, satellite ground-level PM2.5 products tested in this work  30 5.2 Statistics for the Ulaanbaatar, Mongolia, satellite ground-level PM2.5 products tested in this work    30 5.3 Annual average PM2.5 surface concentrations for Delhi, India, 2017   31 5.4 Cities included in this work   35 5.5 Correlation coefficient, mean normalized bias, and mean normalized gross error for each city    36 5.6 Lessons learned and recommended next steps to improve monitoring of PM2.5 understand and reduce the health impacts of in each city to better ­ PM2.5  40 6.1 Summary of typical problems in using satellites in PM2.5 monitoring, the consequences, and the recommended actions to further the use of satellite data in PM2.5 monitoring to better understand and reduce the health impacts of PM2.5 in low- and middle-income countries   47 6.2 Recommendations for the use of satellite observations to supplement ground-level monitoring data in low- and middle-income countries   48 A.1 Summary of operational aerosol products   57 A.2 Case study of satellite aerosol-optical-depth applicability for three representative low- and ­ middle-income country urban areas   68 viii | Getting Down to Earth B.1 Three focus cities where OpenAQ aggregates real-time local government particulate matter data overlap with US Diplomatic Post measurements   98 B.2 Contents of the subsampled aerosol-optical-depth product files   99 B.3 Mapping between the parsed data file variable names and the aerosol-optical-depth products in the L2 files   100 B.4 Region of interest data fields in the SEVIRI netCDF output files   101 B.5 MERRA-2 data file collections and extracted variables for input into the chemical transport model and statistical approach scripts   103 B.6 Statistics for the Delhi, India, satellite ground-level PM2.5 products tested in this work   108 B.7 Statistics for the Ulaanbaatar, Mongolia, satellite ground-level PM2.5 products tested in this work   110 B.8 Annual average PM2.5 surface concentrations for Delhi, India, 2017   111 C.1 Cities included in this study   122 C.2 Coordinates and number of ground-level monitoring sites, by city   122 C.3 Delhi, India, evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods  125 C.4 Evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods, Ulaanbaatar, Mongolia   128 C.5 Evaluation statistics for the chemical-transport-model–based method using Terra data, Accra, Ghana   132 C.6 Evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods, Kathmandu, Nepal   135 C.7 Evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods, Addis Ababa, Ethiopia   136 C.8 Evaluation statistics for the chemical-transport-model–based method, Dakar, Senegal  138 C.9 Evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods, Kampala, Uganda   141 C.10 Evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods, Hanoi, Vietnam   143 C.11 Average PM2.5 concentrations for each monitor site in Delhi, India, 2016–17  144 C.12 Generalized-additive-model output statistics of quality assurance tool versus base processing, Delhi, India   145 C.13 Evaluation statistics (using the filtered OpenAQ data) for the statistical and chemical-transport-model–based methods, Delhi, India   146 C.14 Kriging and co-kriging statistics, Delhi, India   147 C.15 Kriging and co-kriging statistics, Ulaanbaatar, Mongolia   147 Preface Outdoor air pollution creates a significant health burden across the world. Exposure to outdoor airborne pollutant particles, called fine particulate matter or PM2.5, leads to an estimated 4.2 million deaths a year.1 Most of those deaths occur in low- and middle-income countries (LMICs). Understanding the sever- ity of air pollution is a fundamental step in reducing its health burden. Most LMICs, however, do not have any infrastructure for measuring air quality and monitoring air pollution. Budgetary constraints and lack of technical expertise pose additional challenges. There is increasing interest in harnessing Earth-orbiting satellite technol- ogy to fill critical gaps in air-quality data in LMICs. Satellite technology has been used with success in air-quality measuring a ­ pplications in high-income countries, which typically have well-established, operational ground-level monitoring networks. This report investigates whether satellite observations could be used to improve the monitoring of outdoor PM2.5 in LMICs and, if so, to identify path- ways for those countries to incorporate satellite data into their daily, city-scale PM2.5 monitoring. Specifically, the report evaluates the performance of satellite observations in predicting ground-level concentrations of PM2.5 in nine cities in LMICs. The report’s findings indicate that satellite-derived estimates of surface PM2.5 in LMICs are not reliable. The average errors in ­satellite-based estimates tend to be very large, and reliability was limited by cloudiness, snow cover, landscape features, satellite coverage, and topography. Consequently, satellites cannot be considered as a replacement for properly run and maintained ground-level monitoring networks in LMICs. This report shows that ground-level monitoring and satellite data are best thought of as complements to each other. For the purpose of protecting human health, govern- ments of LMICs, and institutions such as the World Bank, should prioritize establishing or strengthening ground-level monitoring networks to measure PM2.5 air pollution and other harmful air pollutants. The results of this report have been published in a special issue of the peer-­ reviewed journal Atmospheric Environment: “Emerging Strategies to Fill the Gaps in Ground-level Air Quality Data in Low- and Middle-Income Countries” (Alvarado and others 2019; Awe and Hagler 2020).  ix x | Getting Down to Earth NOTE 1. https://www.who.int/health-topics/air-pollution#tab=tab_2. REFERENCE Alvarado, Matthew, J., A. E. McVey, J. D. Hegarty, E. S. Cross, C. A. Hasenkopf, R. Lynch, E. J. Kennelly, T. B. Onasch, Y. Awe, E. Sanchez-Triana, and G. Kleiman. 2019. “Evaluating the Use of Satellite Observations to Supplement Ground-Level Air Quality Data in Selected Cities in Low- and Middle-Income Countries.” Atmospheric Environment 218: 117016. https://doi.org/10.1016/j.atmosenv.2019.117016. Awe, Yewande and Gayle Hagler, ed. 2020. “Emerging Strategies to Fill the Gaps in Ground- level Air Quality Data in Low and Middle Income Countries.” https://www.sciencedirect​ .com/journal/atmospheric-environment/special-issue/10TGZBV0GCB. About This Work The analytical work in this report builds on a growing body of evidence that the World Bank is building to inform more effective and efficient pollution manage- ment interventions and harness the transition to a circular economy in low- and middle-income countries (LMICs). This growing body of evidence focuses on strengthening the knowledge base that will prompt dedicated action to tackle the forms of pollution that cause the most significant health and social costs in LMICs. It also advances interdisciplinary approaches to assess the linkages between pollution management and the circular economy and the World Bank’s dual goals of eradicating poverty and promoting shared prosperity. Recent con- tributions from this body of work include (1) the monetary valuation of the global cost of mortality and morbidity caused by exposure to ambient fine particulate matter air pollution, (2) the bolstering of the case for establishment and strength- ening of ground-level air-quality monitoring networks in LMICs, (3) the devel- opment of a systematic framework to support analysis of health impacts from land-based pollution, and (4) economic and financial instruments to support the transition to a circular economy.  xi Acknowledgments This report was prepared by a team led by Yewande Awe with the core team comprising Gary Kleiman and Ernesto Sánchez-Triana (all from the World Bank). The background report was prepared by Matthew Alvarado, Jennifer Hegarty, Edward Kennelly, Richard Lynch, and Amy McVey (Atmospheric and Environmental Research, Inc.); Eben Cross and Timothy Onasch (Aerodyne Research, Inc.); and Christa Hasenkopf (Airglow Labs, formerly of OpenAQ). Shane Ferdinandus (World Bank) provided administrative support. The team would like to acknowledge, with thanks, the valuable advice and inputs of peer reviewers Martin Heger and Karin Shepardson. This report also benefited from comments provided by World Bank staff, includ- ing Marcelo Bortman, Hocine Chalal, Stephen Dorey, Momoe Kanada, and Fernando Loayza, as well as several members of the World Bank’s Pollution Management Global Solutions Group and other participants, during two brown bag lunch events organized to share the findings at different stages of the analytical work. The contributions of Phil Dickerson (US Environmental Protection Agency AirNow program) and Joanne Green (Ricardo Energy & Environment, formerly of the Clean Air Institute) are also appreciated. The team is grateful for ground-level monitoring data on air quality pro- vided by Emmanuel Appoh and staff at the Environmental Protection Agency of the Ghana Ministry of Environment, Science, Technology and Innovation (EPA Ghana); and André Jacques M. Ngor Dioh and Aminata Mbow Diokhané at the Center for Air Quality Management of the Senegal Ministry of the Environment and Sustainable Development. This report is a product of the Environment, Natural Resources and Blue Economy Global Practice (ENB GP) of the World Bank. This work was con- ducted under the supervision of Juergen Voegele (Vice President, Sustainable Development); Karin Kemper (Global Director, ENB GP); Richard Damania (Chief Economist, Sustainable Development); Iain Shuker (Practice Manager, East Africa, ENB GP); Julia Bucknall (Senior Adviser, Eastern and Southern  Africa); Benoit Bosquet (Regional Director for Sustainable  xiii xiv | Getting Down to Earth Development, East Asia and Pacific); and Christian Albert Peter (Practice Manager, Global Platform Unit, ENB GP). The financial support of the World Bank–administered Pollution Management and Environmental Health multidonor trust fund in the prepara- tion of this report is gratefully acknowledged. Executive Summary Outdoor air pollution accounts for an estimated 4.2 million1 deaths around the globe, caused by exposure to fine, inhalable particles with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5, also called fine particulate matter). Reducing PM2.5 air pollution is thus crucial for improving public-health outcomes. Measuring and reducing the health impacts of PM2.5 is especially challenging in low- and middle-income countries (LMICs). This is because LMICs tend to have limited funds for air quality monitors and limited local expertise in air quality. These shortcomings can undermine the effectiveness of a country’s policies to improve its air quality. Earth-orbiting satellites measure the amount of aerosols in the atmosphere with a metric called aerosol optical depth (AOD), which is based on the amount of light scattered and absorbed by all the aerosols over a given location (from the Earth’s surface to space). The global coverage of these satellites and their decades-long records can potentially complement the sparse ground-level monitoring networks in LMICs. However, the performance of the techniques used to convert the satellite observations of AOD into estimates of ground-level PM2.5 concentrations, and thus PM2.5 exposures, is not well established in LMICs, where ground-level monitoring (GLM) data on PM2.5 may be infrequent or absent. It can be difficult to accurately estimate surface PM2.5 concentration based on satellite data. The satellite is directly measuring the reflected sunlight from the Earth, which is affected not only by the aerosols in the atmosphere, but also by the amount of sunlight reflected by the ground (the surface reflectance). Furthermore, because the reflected sunlight is affected by all the aerosols in the atmosphere, determining the ground-level PM2.5 concentration requires making an estimate of how the AOD (from all levels of the atmosphere) is related to the ground-level PM2.5 concentration (that is, just the PM2.5 located at the surface). Thus, to calculate ground-level PM2.5 concentrations and exposures using satellite data, assumptions must be made regarding surface reflectance and the relationship between the AOD and ground-level PM2.5 concentration. Errors or uncertainties in these assumptions will lead to errors and uncertainties in the satellite-derived PM2.5 concentrations and exposures.  xv xvi | Getting Down to Earth TABLE ES.1  Cities included in this analysis CITY COUNTRY LOCATION INCOME GROUP Accra Ghana Coastal, low altitude Lower middle Addis Ababa Ethiopia Inland, high altitude Low Dakar Senegal Coastal, low altitude Lower middle Delhi India Inland, low altitude Lower middle Hanoi Vietnam Inland, low altitude Lower middle Kampala Uganda Inland but near lake, Low high altitude Kathmandu Nepal Inland, high altitude Lower middle Lima Peru Coastal, low altitude Upper middle Ulaanbaatar Mongolia Inland, high altitude Lower middle Source: World Bank. Note: Income groups correspond to World Bank Country Classifications by Income Level: 2021–22. Income classifications are affected by several factors and are subject to change over time. The goal of this report is to investigate if satellite observations could be used to improve PM2.5 monitoring in LMICs and, if so, to identify pathways for LMICs to incorporate satellite data into their daily, city-scale PM2.5 monitoring. A review of the scientific literature was conducted to identify different approaches used to combine satellite observations with GLM of PM2.5 and to understand their reported performance. In addition, several methods for using satellite observations from publicly available sources to predict observed ground-level PM2.5 daily average concentrations (and thus exposures) were tested for use in cities in LMICs. These methods were then applied to nine cities within LMICs (table ES.1) to identify patterns in the satellite performance with respect to city altitude, location, and other variables. LITERATURE REVIEW Many groups have successfully used different approaches to estimate ground-level PM 2.5 concentrations, and thus exposures, using satellite observations. These groups have generally looked at a continental-to-global scale (rather than the city-specific scale of interest to this work) and have predicted monthly and annual average PM2.5 concentrations and exposures (rather than the daily average estimates of interest in this work). However, the errors in these satellite estimates of PM2.5 concentrations can be large (about 50 percent). Although many satellites measure AOD, only a smaller number of satellites cover the entire globe each day and provide free data, shortly after the observation is made, that are available for use in LMICs for daily observations. In addition, many satellite data sets of AOD can be difficult to access, either because the user must pay for the data or because the download process cannot be easily automated. Because the estimated errors in the satellite AOD products are similar, the satellite product that offers the best balance of AOD coverage and fine horizontal resolution for a given city should be used. For this report, the National Aeronautics and Space Administration’s (NASA) MODIS combined Deep Blue and Dark Target AOD data set was used, since it is free, available Executive Summary | xvii shortly after the observation is made, easy to access, has been extensively validated, and has good coverage over urban and coastal areas. Previous studies have attempted to convert satellite AOD measurements to ground-level PM2.5 estimates using statistical techniques, chemical transport model (CTM)–based approaches, or hybrid approaches. • CTM-based approaches can be applied to any region of the globe and do not require any GLM data. Public global CTM data sets, such as the NASA MERRA-2 reanalysis, can be used for estimates of aerosol vertical profiles and other parameters. However, these CTM data sets can have significant errors in the CTM’s simulation of how aerosol concentrations change with altitude. • Statistical approaches do not depend on a CTM, but they require a multiyear, accurate GLM data record. Only some LMICs have sufficient GLM data to use these approaches. • Hybrid approaches that correct CTM-based approaches with statistical models tend to give the best performance. However, they also require a multiyear, accurate GLM data record over many different land-use types (for example, urban and rural sites, coastal and inland sites), which few LMICs have. Interpolation approaches, such as co-kriging and land-use regression, can be used to combine satellite PM2.5 estimates with GLM data to provide estimates of PM2.5 concentrations at a high horizontal resolution (100–500 meters) across urban areas. However, these approaches tend to work better as the satellite horizontal resolution is increased and as the GLM network covers a wider variety of sites. The highest horizontal resolution possible for current satellite AOD measurements is between 2 and 4 kilometers (for geostationary satellites) or 1 and 10 kilometers (for polar-orbiting satellites). Consequently, finer-scale predictions of ground-level PM2.5 will be possible only for areas where extensive, high-quality GLM data exist. PERFORMANCE OF SATELLITE APPROACHES IN LMICs This report found that satellites cannot be a complete replacement for a GLM network. The CTM-based and statistical approaches for converting satellite AOD to ground-level PM2.5 concentrations did a poor job of predicting the day-to-day or site-to-site variations in daily average PM2.5 values within a city. In addition, the average errors in all satellite-based estimates of the daily average PM2.5 concentration at a given location in a city tended to be very large (21–77 percent for the statistical methods, and 48–85 percent for the CTM-based methods), much larger than the 10 percent error that is usually considered acceptable for PM2.5 monitoring (Alvarado and others 2019). These large errors indicate that satellite-derived estimates of PM2.5 concentrations in LMICs are not reliable. Many cities also had significant limitations in the availability of satellite observations of aerosols throughout the year. For example, no satellite observations are available in Ulaanbaatar, Mongolia, for the high-PM2.5 winter months of December to mid-March because of persistent snow cover. Thus, even a perfect method for converting satellite AOD measurements to ground-level PM2.5 estimates would underestimate the true annual average PM2.5 concentration for Ulaanbaatar by a factor of two. xviii | Getting Down to Earth The work reported here found some patterns in the performance of satellite methods in different cities. Satellite-based methods appear to work best for low- altitude, inland cities such as Delhi, India, and Hanoi, Vietnam, but still have significant errors (43–60 percent) in predictions of daily average PM 2.5 concentrations at sites within these cities. Coastal cities (including cities near large lakes) have poor satellite coverage, either because of persistent clouds or the mixture of land and water surfaces in the satellite measurement of AOD. CTM-based satellite approaches tend to underestimate PM2.5 in high-altitude cities (except for Addis Ababa, Ethiopia, likely because of its location near the Sahara). Although this work found that, under some conditions, adding satellite data to GLM network data via co-kriging may reduce the number of GLM sites needed to characterize PM2.5 concentrations within a city, this ability varies from city to city and could result in large errors in annual average estimates for cities with a seasonal or persistent lack of satellite AOD coverage. The 2017 annual average PM2.5 concentrations in Delhi and Ulaanbaatar derived with the CTM-based method used in this project were similar to the 2016 annual averages estimated in the uncorrected2 Global Burden of Disease (GBD) 2016 data set.3 However, at the city level, results are considerably less accurate. For example, in Ulaanbaatar the CTM-based method tested in this report and the uncorrected GBD 2016 data set each underestimates the true annual average PM2.5 concentration by a factor of 10. This is likely because cities in appreciably different air quality environments than their surroundings, such as cities in mountain valleys surrounded by rural land like Ulaanbaatar, will not get accurate relationships between satellite AOD and ground-level PM2.5 from the relatively coarse resolution of global CTMs. The geographically weighted bias correction used in the GBD 2016 data set corrects for the fact that Ulaanbaatar is surrounded by a region that is much less polluted; but the corrected GBD 2016 data set still underestimates the true annual average by a factor of two. In addition, the correction requires the use of GLM data at a continental scale. If only the GLM data from a given city are used, the land-use parameters are not highly correlated with the PM 2.5 variations in the city. Estimating variations in annual average PM2.5 within a city is therefore unlikely to be possible with satellite AOD data alone, especially for cities affected by the satellite-coverage issues noted above. Estimating fine- scale geographic variations of PM2.5 in a city has only been done using extensive GLM network data (covering both urban and rural sites) and land-use regression, with the satellite-based PM2.5 estimate used as a variable in the land-use regression. RECOMMENDATIONS Based on these results, GLM and satellite data are best thought of as comple- ments to each other. Many GLM networks could be improved by considering satellite data, but all approaches using satellite data improve as the number of high-quality GLM sites is increased. Thus, it is important that LMICs strengthen support for the establishment of GLM networks to measure air pollutants that cause mortality, notably PM2.5, in LMICs in Sub-Saharan Africa and other regions. These GLM networks must have adequate quality assurance and quality control and follow standard operating procedures to ensure the data are of suffi- cient quality to be used to estimate PM2.5 exposures for health studies and to be combined with the satellite estimates. Executive Summary | xix In other words, satellite data may be useful for estimating average air quality for countries or large geographical areas. However, for the purpose of protecting human health, LMICs need to prioritize the establishment or strengthening of GLM networks where they are lacking or weak, to measure air quality at the level where human activity is typically carried out and people are exposed to air pollutants, notably PM2.5, that are h ­ armful to health and cause death. This work also provides the following recommendations for the use of satellite observations to supplement GLM data in LMICs (see proposed LMIC typology in chapter 1): • For countries with no GLM data (Type I), the only possible approach to convert AOD to ground-level PM2.5 is a CTM-based approach. The satellite-derived ground-level PM2.5 values should be assigned a high uncer- tainty that reflects not only the uncertainty in the AOD (about 20 percent) but also the estimated uncertainty in the CTM-derived AOD-to-PM2.5 relation- ship (about 50 percent). • For countries with a small amount of GLM data with variable quality (Type II), it is possible to derive a bias estimate for the CTM-based estimates of ground-level PM2.5. However, it will not be possible to ascertain whether the bias-corrected estimate is truly more accurate than the raw CTM-based esti- mate, and thus it is advisable to report and store both values. This will allow for reprocessing of the satellite estimates when more-rigorous quality assur- ance procedures are developed. • Countries currently establishing GLM networks (Type III) can explore both statistical and bias-corrected CTM-based approaches for converting AOD to PM2.5. CTM-based approaches likely would be best for these countries, but statistical approaches should also be tested before making a final decision. • Countries with reliable, comprehensive GLM networks (Types IV and V) will be able to use satellite observations to fill in the gaps of the existing GLM networks. The more extensive GLM networks in these countries will allow for more-accurate estimates of the geographical and seasonal variation in the AOD-to-PM2.5 relationship, and thus purely statistical approaches may out- perform CTM-based estimates. In addition, the results of this work suggest the following recommendations for LMICs and the US State Department for reporting the metadata for their GLM measurements to ensure the data are sufficiently accurate to be usefully combined with satellite estimates of PM2.5: • PM2.5 measurements need to be provided along with information on the instrument/technique type, estimates of measurement uncertainties, rele- vant metadata, and operational history. It is important that each PM 2.5 measurement not only identifies the type of instrument used for the measure- ment but also tracks the instrument’s relevant history of use and calibration. This includes all “meta” data from instrument data files, including “housekeeping” files. • Any recorded results from tests relating to standard operating procedures and quality-assurance protocols should be collected and disseminated along with the PM2.5 measurements to establish an instrument track record over time, providing a transparent, data-centric measure of instrument reliability and performance. • Currently, US embassies do not publish their full quality-assurance protocol metadata along with the PM2.5 levels, thereby reducing the use of these xx | Getting Down to Earth measurements as well-established and controlled reference measurements. Thus, it is recommended that US embassies publish their full quality-­ assurance protocol and relevant metadata along with the PM2.5 levels. ADDED VALUE OF THIS WORK The analytical added value of this report includes the following: • A review and synthesis of the scientific literature on using satellites to esti- mate ground-level PM2.5 concentrations, with a focus on the relevance of the approaches for city-scale, daily-average PM2.5 monitoring in LMICs • An analysis of the performance of different satellite AOD measurements and methods for converting AOD to ground-level PM2.5 concentrations from a local, city-specific perspective, including a discussion of the patterns in the performance due to city location and • Recommendations for practitioners and policy makers in LMICs on how to improve their estimates of PM2.5 exposure through expansion of their GLM networks, adoption of rigorous quality-assurance procedures, and incorpora- tion of satellite data. NOTES 1. https://www.who.int/health-topics/air-pollution#tab=tab_2. 2. The uncorrected GBD 2016 data set is the data set derived by relating satellite AOD to ground-level PM2.5 before correction for persistent errors (biases) using GLM data and geographically weighted regression. 3. Downloaded from http://fizz.phys.dal.ca/~atmos/martin/?page_id=140. REFERENCES Alvarado, M. J., A. E. McVey, J. D. Hegarty, E. S. Cross, C. A. Hasenkopf, R. Lynch, E. J. Kennelly, T. B. Onasch, Y. Awe, E. Sanchez-Triana, and G. Kleiman. 2019. “Evaluating the Use of Satellite Observations to Supplement Ground-Level Air Quality Data in Selected Cities in Low- and Middle-Income Countries.” Atmospheric Environment 218: 117016. https://doi​ .org/10.1016/j.atmosenv.2019.117016. GBD 2016 Risk Factors Collaborators. 2017. “Global, Regional, and National Comparative Risk Assessment of 84 Behavioural, Environmental and Occupational, and Metabolic Risks or Clusters of Risks, 1990–2016: A Systematic Analysis for the Global Burden of Disease Study 2016.” Lancet 390: 1345–422. Abbreviations ABI Advanced Baseline Imager AERONET Aerosol Robotic Network AHI Advanced Himawari Imager AMDAR Aircraft Meteorological Data Relay AOD aerosol optical depth AOT aerosol optical thickness AQ air quality BAM beta attenuation monitor BAMM beta attenuation mass monitor CLASS Comprehensive Large Array-Data Stewardship System (NOAA) CTM chemical transport model DAAC Distributed Active Archive Center ECMWF European Centre for Medium-Range Weather Forecasts EDGAR Emission Database for Global Atmospheric Research EDM environmental dust monitor EPA Environmental Protection Agency (US) FEM federal equivalent method FRM federal reference method GBD Global Burden of Disease GEOS Goddard Earth Observing System GFS Global Forecast System GLM ground-level monitoring GOES Geostationary Operational Environmental Satellite GPS global positioning system GWR geographically weighted regression JMA Japan Meteorology Agency JPSS Joint Polar Satellite System LAADS Level-1 and Atmospheric Archive and Distribution System (NASA) LEO low earth orbit Lidar light detection and ranging LMICs low- and middle-income countries LUR land-use regression  xxi xxii | Getting Down to Earth LUT look-up table MACC Monitoring Atmospheric Chemistry and Climate MAIAC Multi-Angle Implementation of Atmospheric Correction MERRA Modern Era Retrospective-analysis for Research and Applications MISR Multi-angle Imaging Spectroradiometer MNB mean normalized bias MNGE mean normalized gross error MODIS Moderate-Resolution Imaging Spectroradiometer MPL micro-pulse lidar MSG Meteosat Second Generation NASA National Aeronautics and Space Administration NCEP National Centers for Environmental Prediction NDVI normalized difference vegetation index NEMS NOAA Environmental Modeling System NGAC NEMS GFS aerosol component NIR near-infrared NOAA National Oceanic and Atmospheric Administration NRT near real-time NWP numerical weather prediction PBL planetary boundary layer PBLH planetary boundary layer height PM2.5/PM10 particulate matter with an aerodynamic diameter less than or equal to 2.5 microns/10 microns, respectively QA quality assurance QC quality control RMSE root-mean-square error RT radiative transfer SEDAC Socioeconomic Data and Applications Center SEVIRI Spinning Enhanced Visible and Infrared Imager SMAOD suspended matter, aerosol optical depth SMAOL SEVIRI-MSG Aerosol Over Land SPARTAN Surface Particulate Matter Network TEOM tapered element oscillating microbalance TOA top of atmosphere UN United Nations USGS United States Geological Survey UV ultraviolet VIIRS Visible Infrared Imaging Radiometer Suite VIS visible WMO World Meteorological Organization Glossary Aerosol(s). Tiny solid and liquid particles suspended in the air. Windblown dust, sea salts, volcanic ash, smoke from wildfires, and pollution from factories are all examples of aerosols. Aerosol optical depth (AOD) or aerosol optical thickness (AOT). A measure of the extinction (absorption and scattering) of light by aerosols in a column of air from the Earth’s surface up to space. Defined as the natural logarithm of the ratio of the incident light at the top of the atmosphere to the transmitted light at the surface, assuming the light travels straight down. Bias or mean bias. The average (mean) difference between a measurement and the true value of the property being measured. An ideal measurement has a mean bias of zero. If this value is large, it suggests that there are consistent (nonrandom) errors in the measurement. Chemical transport model (CTM). A computer model that simulates the trans- port and chemistry of gases and aerosols in the atmosphere and can aid in air-quality forecasting and planning. Co-kriging. An extension of kriging that uses a secondary data set to provide additional information on the spatial distribution of the air-quality variable being interpolated. Correlation coefficient (R). In statistics, the proportion of the variance in the dependent variable that is predictable from the independent variable(s). If two measurements of PM2.5 are well correlated such that they predict similar spatial and temporal patterns, then the R value will be near 1. If the values are uncor- related, the value will be near 0. Data assimilation. A process for using observations of the atmosphere (weather and chemical composition) to improve computer-model representations of the atmosphere. Doppler wind profiler (DWP). A weather instrument that uses radar to detect the wind speed and direction at various elevations above the ground. Generalized additive model (GAM). A statistical model of a system that is an extension of multiple linear regression models that allows the unknown func- tional form for each of the predictor variables to be determined from the data.  xxiii xxiv | Getting Down to Earth Geographic information system (GIS). A system designed to capture, store, manipulate, analyze, manage, and present geographic data. Usually analyzed using specialized software packages such as ArcGIS. Geographically weighted regression. A statistical model of a system that is an extension of multiple linear regression models that allows the coefficients to vary with location. Geostationary satellite. A satellite for which the orbit remains stationary over a given point on the Earth’s equator. This type of satellite can continuously view the area underneath it but provides no data on other regions of the globe. Global bias. Bias relative to a large set of observations, such as in several moni- toring networks covering a continent. Kriging. A method of data interpolation (that is, determining the value of a func- tion at a given point by computing a weighted average of the known values of the function in the neighborhood of the point) used to provide estimates of air qual- ity at locations without monitors, based on the measurements of, and distances from, surrounding air-quality monitors. Land-use regression (LUR). A multiple linear regression model that uses differ- ent land-use variables (for example, population density, percent of urban land cover, and traffic intensity) to predict the concentration of air pollutants at a high resolution across an area using a sparser measurement network. Lidar. Light detection and ranging; a measurement technique that uses reflected laser light to determine atmospheric properties such as the height of clouds and aerosol layers. Local bias. Bias relative to a small set of nearby observations, such as the moni- toring network within a city. Look-up table (LUT). An approach to speeding up computer models that precal- culates the values of many functions and stores them in a table for later use. Mean normalized bias. Also called mean fractional bias, this is the average (mean) of the normalized bias, which is the difference between a measurement and the true value of the property being measured, divided by the true value. An ideal measurement has a mean normalized bias of zero. If this value is large, it suggests that there are consistent (nonrandom) errors in the measurement. Mean normalized gross error. Similar to mean normalized bias, but the abso- lute value of the normalized bias is used so that positive and negative errors do not cancel out. Near real-time (NRT) data. Model output or observational data that are pro- cessed and distributed as quickly as possible after the time being modeled or the time of the measurement. Noise. In the context of a satellite measurement, noise refers to the random errors present in the measurement. Noise can generally be reduced by averaging over a larger area at the expense of a coarser horizontal resolution. Particulate matter. See Aerosol(s). Planetary boundary layer height (PBLH). The height of the well-mixed layer of air near the Earth’s surface. Glossary | xxv Polar-orbiting satellite. A satellite that has an orbit passing over or near the poles and thus can provide observations of the entire globe with the trade-off that each location is observed only (at most) twice a day. Root-mean-square error (RMSE). An estimate of the average error in a mea- surement or model prediction that, unlike mean bias, does not allow positive and negative errors to cancel each other out. Satellite AOD products. The estimate of the aerosol optical depth (AOD) over a given area produced from a given satellite instrument. Surface reflectance. The fraction of sunlight that is reflected by the Earth’s surface at a given location. Vegetation and ocean water tend to have low surface ­ reflectance, while snow, deserts, and many urban surfaces have high surface reflectance. 1 Contextual Background and Objectives This report was produced as part of a program of analytical work conducted under the framework of the World Bank’s multidonor-funded Pollution Management and Environmental Health program. The objectives of this report are to improve knowledge regarding the following: ow satellite measurements can best be used to enhance air-quality (AQ) • H monitoring, and thus improve human exposure assessment, in low- and middle-income countries (LMICs) and • How satellite measurements can be brought into closer agreement with ground-level monitoring data, considering the shortcomings and advantages of satellite and ground-level measurements. This report summarizes the findings of the three tasks: (1) a literature review of approaches used to combine satellite observations with ground-level moni- toring measurements of particulate matter with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5), (2) the effort to develop and evaluate meth- ods for converting satellite aerosol optical depth readings into ground-level PM2.5 estimates in individual cities (Accra, Ghana; Delhi, India; Lima, Peru; and Ulaanbaatar, Mongolia), and (3) evaluation of the application of these methods to nine LMIC cities to identify geographic patterns in the performance of the satellite estimates. Based on this work, recommendations are made for LMICs regarding how best to incorporate satellite data into their monitoring plans using the following proposed typology based on the country’s level of engagement in AQ monitoring: • Type I: Countries with no existing measurements and no history of any kind of routine measurements of atmospheric composition. Some anecdotal mea- surements or one-time sampling may have taken place. • Type II: Countries with some information on atmospheric composition avail- able (perhaps PM10—particulate matter with an aerodynamic diameter of less than or equal to 10 microns—or total suspended particulates) but of variable quality without rigorous quality assurance procedures.  1 2 | Getting Down to Earth • Type III: Countries that possess reliable information but with poor spatial or temporal coverage; for example, monitoring may exist in only one city or rou- tine monitoring may exist for a year or two but is no longer being collected because of equipment malfunction and/or lack of repairs. • Type IV: Good, reliable AQ monitoring underway or being established. • Type V: Routine, long-term AQ monitoring. 2 Introduction Ambient or outdoor air pollution has significant impacts on public health around the globe. According to the World Health Organization, an estimated 4.2 million1 people die every year worldwide from exposure to tiny solid and liquid particles— referred to as aerosols or fine particulate matter with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5)—suspended in outdoor air. Reducing these impacts through air pollution controls is thus a major goal of environmen- tal efforts to improve public health. However, quantifying and reducing the health impacts of PM2.5 is especially challenging in low- and middle-income countries (LMICs) because these countries tend to have limited air-quality (AQ) monitoring infrastructure, have insufficient quality assurance and quality con- trol of their AQ monitoring data, and have limited local technical expertise in AQ monitoring, modeling, and planning. These shortcomings related to data and expertise can undermine the effectiveness of the design and implementation of policies to improve AQ in LMICs. Earth-orbiting satellites can detect aerosols using observations of the sun- light scattered by the aerosols (figure 2.1; Bernard and others 2011; Levy and oth- ers 2013). This report examines whether the extensive spatial and temporal coverage of these satellites (figure 2.2) can potentially complement the sparse networks of ground-level monitors that typically exist in LMICs, thereby improving AQ monitoring and enforcement actions in these countries. However, satellites are not able to measure ambient ground-level concentra- tions of PM2.5 directly. Instead, satellites use the observations of the sunlight scattered by aerosols to estimate a parameter called aerosol optical depth (AOD), which represents the extinction (absorption and scattering) of light by all of the aerosols in a column of air from the Earth’s surface up to the top of the atmo- sphere (for example, Levy and others 2013). As shown in figure 2.3, the reflected sunlight measured by the satellite depends on the intensity of sunlight reaching the Earth (Io), the fraction of the sunlight reflected by the ground (α, also called the “surface reflectance” or “albedo”), the AOD, and the aerosol light scattering and absorbing properties (also called aerosol optical properties, here repre- sented by a single factor ω). Although the intensity of the incoming sunlight reaching Earth is well known, the other three properties are more uncertain.  3 4 | Getting Down to Earth FIGURE 2.1 MODIS Aqua image of wildfire smoke being transported eastward to the Chesapeake Bay, US Source: World Bank. Note: The white haze image is produced by the scattering of sunlight by the aerosols in the smoke. Image taken June 10, 2015, 18:55 UTC. MODIS = Moderate-Resolution Imaging Spectroradiometer; UTC = Coordinated Universal Time. To measure AOD, “retrieval algorithms” are designed that use tabulated esti- mates of the surface reflectance and aerosol optical properties. Radiative-transfer (RT) models enable accurate simulations of reflectance at the top of the atmosphere (TOA) in the presence of aerosol layers with a variety of optical properties. By properly accounting for surface reflectance from land and ocean backgrounds, it is possible to retrieve several important aerosol properties by comparing satellite-observed TOA reflectances to those calculated from an RT model. Typically, to optimize processing time, TOA reflectances are precalculated by the RT model for several combinations of aerosols and are stored in look-up tables accessed by the retrieval algorithm. The RT calculations are performed for a range of aerosol optical thicknesses, so each stored reflectance value in the look-up table corresponds to an aerosol optical thickness. The estimates of all three parameters (surface reflectance, AOD, and aerosol optical properties) are then refined to provide the best match with the observed reflected sunlight at multiple wavelengths (that is, the satellite measurement of the different colors in the reflected sunlight). However, uncertainties or errors in the estimated surface reflectance and aerosol optical properties can lead to similar uncertainties and errors in the measured AOD. Introduction | 5 FIGURE 2.2 Current and future coverage provided by available polar and geostationary weather satellites a. Current coverage provided by available polar b. Future coverage after planned updates of and geostationary satellites geostationary satellites Source: World Bank. Note: The three bands in each panel illustrate three consecutive swaths from Moderate-Resolution Imaging Spectroradiometer (MODIS) in a polar orbit. Green, red, turquoise, magenta, and blue points represent coverage from geostationary satellites. The exact relationship between AOD and ground-level PM2.5 concentrations depends on the vertical distribution of aerosols above the observed location and their ability to scatter and absorb sunlight, which in turn depends on their size, shape, and chemical composition (van Donkelaar, Martin, and Park 2006). Two main approaches are used to estimate the relationship between AOD and ground-level PM2.5 concentrations: statistical approaches (for example, general- ized additive models [GAMs]) and chemical transport model (CTM)–based approaches. Statistical techniques can be used to predict the relationship between AOD and ground-level PM2.5 using historical data sets for both param- eters and other variables (for example, Hu and others 2014; Sorek-Hamer and others 2013, 2015; Strawa and others 2013). Alternatively, the relationship can be estimated by using a computer model of the chemistry of the atmosphere, called a chemical transport model (CTM) (for example, Geng and others 2015; van Donkelaar and others 2006, 2010, 2011, 2015a, 2015b). Several factors can limit the ability of satellites to obtain accurate AOD mea- surements. Satellite observations of AOD generally require that the fraction of the sunlight reflected by the ground (that is, the surface reflectance) is relatively uniform across the satellite “footprint.” This “footprint” is the area covered by a single satellite AOD observation, which may include multiple satellite image pic- ture elements or “pixels.” The horizontal extent of this footprint is referred to as the horizontal resolution of the AOD observation. Thus, mixed and reflective surfaces—including deserts, persistent snow cover, and mixed land and water 6 | Getting Down to Earth FIGURE 2.3 How satellites can measure aerosol optical depth from reflected sunlight Io .54 OD =0 .5A –0 g e rin Io ωl o tte αloe–2AOD = 0.027Io sca s ol ro Ae AOD = 1 PM2.5 = 100 µg m–3 loe–AOD = 0.37Io αloe–AOD = 0.074Io Surface reflectance α = 0.2 Source: World Bank. Note: The reflected sunlight measured by the satellite includes that reflected by the aerosols (54 percent) and that reflected by the surface (2.7 percent). Here Io represents the initial intensity of the sunlight, the surface reflectance is represented by one factor (α), and the aerosol scattering and absorbing properties are represented by a single combined factor (ω). The numbers are just to illustrate relative magnitudes and are not results from a radiative transfer model simulation. AOD = aerosol optical depth; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. surfaces near coastal cities—can reduce the ability of satellites to provide accu- rate AOD observations for many cities. Clouds also prevent satellite AOD obser- vations, and thus cities with seasonal persistent clouds will have reduced satellite observations during those periods. The statistical and CTM-based approaches for relating AOD to ground-level concentrations of PM2.5 also have important limitations. The vertical profile of the aerosols, and thus the relationship between AOD and ground-level PM2.5, depends strongly on the height of the well-mixed layer of air near the Earth’s surface called the planetary boundary layer (PBL). For example, a one-kilometer PBL height with a more concentrated aerosol layer near the Earth’s surface (green line in figure 2.4) may have the same AOD as a more dilute aerosol layer with a two-kilometer PBL height (blue line in figure 2.4). In addition, if there is a concentrated aerosol layer above the PBL—as can happen when wildfire, dust, or pollution is transported a long distance from its source—the AOD may be due mainly to this concentrated aerosol layer. In this case, the AOD observation does not provide much useful information about the aerosol concentration at the Earth’s surface (red line in figure 2.4). Lack of satellite AOD products at night, or due to clouds, snow cover, and other problems, can limit the ability of satellites to be used to estimate expo- sure to PM2.5 for health studies. Satellite AOD observations are obtainable only for one specific time during the day (for polar orbiting satellites) or during the cloud-free daylight times (for geostationary satellites). However, the daily Introduction | 7 FIGURE 2.4 Highly idealized example of three different aerosol vertical profiles, each with different ground-level PM2.5 concentrations, that result in the same aerosol-optical-depth value of one when measured by satellites 16 14 12 10 Altitude (kilometers) 8 6 4 2 0 50 100 150 200 PM2.5 (µg/m3) Source: World Bank. Note: The green profile shows a 1-kilometer planetary boundary layer (PBL) height. The blue profile shows a 2-kilometer PBL height. The red profile shows a 2-kilometer PBL height and an elevated aerosol layer between 3.5 and 5.5 kilometers in altitude. AOD = aerosol optical depth; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. average PM2.5 concentration is the required metric for estimation of acute (short-term) health effects of PM2.5. Consequently, even if the relationship between satellite AOD and the ground-level PM2.5 at the satellite observation time were perfectly known, the relationship between the ground-level PM2.5 concentration at the satellite observation time and the daily average PM2.5 con- centration would have to be estimated as well. This estimation would be based on either the statistics of past ground-level PM2.5 observations or the output of a CTM. Errors in this relationship would lead to errors in the satellite-based estimate of PM2.5 exposure. In addition, investigations of the chronic (long-term) health effects of ground- level PM2.5 require estimating the annual average PM2.5 concentration. Thus, any seasonal pattern that affects the probability of a successful AOD retrieval, from either seasonal patterns in clouds or surface properties (that is, winter snow  cover), will lead to a potentially incorrect annual average of 8 | Getting Down to Earth PM2.5 from satellite observations. This is so even if the relationship between AOD and ground-level PM2.5 concentrations is known. The ability of satellite observations to estimate ground-level PM2.5 has not been established in LMICs. In addition, most previous studies have provided global or continental-scale results on monthly or annual timescales, rather than the city-scale, daily average PM2.5 concentration estimates needed for routine AQ monitoring and for public health alert programs. Most previous validation work on the use of satellite observations to estimate ground-level PM2.5 has been per- formed in developed countries with extensive, well-calibrated, long-term ground-level monitoring (GLM) observations of PM2.5 (for example, Geng and others 2015; Hu and others 2014; Sorek-Hamer and others 2013, 2015; Strawa and others 2013; van Donkelaar and others 2006, 2010, 2011, 2015a, 2015b). A few studies have evaluated the performance in some of the East Asia countries where GLM data are relatively new (for example, van Donkelaar, Martin, Brauer, and Boys 2015). However, the performance of the statistical models and CTMs used to derive these relationships has not been established in LMICs where GLM is infrequent or absent. Thus, the satellite estimates of ground-level PM2.5 in LMICs may be biased—that is, the average difference between the estimated ground-level PM2.5 and the true PM2.5 concentration may be large, suggesting consistent (nonrandom) errors in the satellite estimates. Techniques that attempt to remove these consistent errors are called bias-correction methods (for example, van Donkelaar, Martin, Spurr, and Burnett 2015), but these corrections depend on data from well-calibrated, long-term GLM networks and thus may not perform well in LMICs. Errors in the city-scale, daily average PM2.5 concentration estimates could cause significant problems for public health alert programs, such as “code red” days when self-­ protective measures are advised for sensitive populations. This could result in sig- nificant economic impacts through lost work or school (if PM2.5 is overestimated) or significant health, welfare, and productivity losses (if PM2.5 is underestimated). The lack of GLM data in many LMICs also introduces difficulties in trying to relate the spatially averaged AOD observations to variations in PM2.5 within a city (due to distance from roadways and other pollution sources). This suggests the need to combine satellite observations with GLM observations and data on pol- lution sources within a city using interpolation techniques (for example, Lee, Chatfield, and Strawa 2016; Millar and others 2010; Vienneau and others 2013). REPORT METHODOLOGY This report evaluates different approaches for using satellite AOD data in esti- mating daily average ground-level PM2.5 concentrations within selected cities in LMICs, with the goal of improving estimates of human PM2.5 exposure. The methodological approach was, first, to conduct a literature review of previous approaches used to estimate PM2.5 exposures employing satellite data, with a focus on identifying those satellite instruments and approaches for PM2.5 esti- mation best suited to provide city-scale, daily average PM2.5 concentrations and exposures. Nine LMIC cities with multiyear GLM data records were then identified to be used to test these approaches. Since the accuracy of these GLM data could affect the results of the analysis, the accuracy of the GLM instruments and the details of the quality-assurance procedures applied to these data were investigated. Finally, the identified satellite approaches Introduction | 9 (one  statistical and one CTM-based) were applied to the nine cities to determine how the usefulness of satellite observations for estimating PM2.5 concentrations and exposures varies across the geographic locations repre- sented by the different cities. OUTLINE OF THIS REPORT This report discusses the results of a review of the literature on approaches used to combine satellite observations with GLM measurements of PM2.5 and dis- cusses the strengths and limitations of these approaches (chapter 3). The quality of GLM data available in LMICs is also discussed, and recommendations for improving the GLM data are made (chapter 4). In addition, this report discusses the effort to develop and evaluate methods for converting satellite AOD into daily average, city-scale, ground-level PM2.5 estimates in nine cities in LMICs (chapter 5). Based on this work, recommendations are made for how LMICs could best incorporate satellite data into their ambient AQ monitoring efforts (chapter 6). The recommendations use the same typology proposed in chapter 1 based on a given LMIC’s level of engagement with AQ monitoring. The appen- dixes provide additional information on the literature review conducted for this report (appendix A); converting satellite AOD to ground-level PM2.5 (appendix B); evaluation of satellite approaches (appendix C); and quality considerations for GLM data in LMICs (appendix C). NOTE 1. https://www.who.int/health-topics/air-pollution#tab=tab_2. REFERENCES Bernard, E., C. Moulin, D. Ramon, D. Jolivet, J. Riedi, and J.-M. Nicolas. 2011. “Description and Validation of an AOT Product over Land at the 0.6 μm Channel of the SEVIRI Sensor Onboard MSG.” Atmospheric Measurement Techniques 4: 2543–65. Geng, G., Q. Zhang, R. V. Martin, A. van Donkelaar, H. Huo, H. Che, J. Lin, and K. He. 2015. “Estimating Long-Term PM2.5 Concentrations in China Using Satellite-Based Aerosol Optical Depth and a Chemical Transport Model.” Remote Sensing of Environment 166: 262–70. Hu, X., L. A. Waller, A. Lyapustin, Y. Wang, M. Z. Al-Hamdan, W. L. Crosson, M. G. Estes, S. M. Estes, D. A. Quattrochi, S. J. Puttaswamy, and Y. Liu. 2014. “Estimating Ground-Level PM2.5 Concentrations in the Southeastern United States Using MAIAC AOD Retrievals and a Two-Stage Model.” Remote Sensing of Environment 140: 220–32. Lee, H. J., R. B. Chatfield, and A. W. Strawa. 2016. “Enhancing the Applicability of Satellite Remote Sensing for PM2.5 Estimation Using MODIS Deep Blue AOD and Land Use Regression in California, United States.” Environmental Science & Technology 50 (12): 6546–55. Levy, R. C., S. Mattoo, L. A. Munchak, L. A. Remer, A. M. Sayer, F. Patadia, and N. C. Hsu. 2013. “The Collection 6 MODIS Aerosol Products over Land and Ocean.” Atmospheric Measurement Techniques 6: 2989–3034. Millar, G., T. Abel, J. Allen, P. Barn, M. Noullett, J. Spagnol, and P. L. Jackson. 2010. “Evaluating Human Exposure to Fine Particulate Matter Part II: Modeling.” Geography Compass 4 (7): 731–49. 10 | Getting Down to Earth Sorek-Hamer, M., I. Kloog, P. Koutrakis, A. W. Strawa, R. Chatfield, A. Cohen, W. L. Ridgway, and D. M. Broday. 2015. “Assessment of PM2.5 Concentrations over Bright Surfaces Using MODIS Satellite Observations.” Remote Sensing of Environment 163: 180–85. Sorek-Hamer, M., A. W. Strawa, R. B. Chatfield, R. Esswein, A. Cohen, and D. M. Broday. 2013. “Improved Retrieval of PM2.5 from Satellite Data Products Using Non-Linear Methods.” Environmental Pollution 182: 417–23. Strawa, A. W., R. B. Chatfield, M. Legg, B. Scarnato, and R. Esswein. 2013. “Improving Retrievals of Regional Fine Particulate Matter Concentrations from Moderate Resolution Imaging Spectroradiometer (MODIS) and Ozone Monitoring Instrument (OMI) Multisatellite Observations.” Journal of the Air & Waste Management Association 63 (12): 1434–46. van Donkelaar, A., R. V. Martin, M. Brauer, and B. L. Boys. 2015a. “Use of Satellite Observations for Long-Term Exposure Assessment of Global Concentrations of Fine Particulate Matter.” Environmental Health Perspectives 123 (2): 135–43. van Donkelaar, A., R. V. Martin, M. Brauer, R. Kahn, R. Levy, C. Verduzco, and P. J. Villeneuve. 2010. Global Estimates of Ambient Fine Particulate Matter Concentrations from Satellite- Based Aerosol Optical Depth: Development and Application.” Environmental Health Perspectives 118 (6): 847–55. van Donkelaar, A., R. V. Martin, R. C. Levy, A. M. da Silva, M. Krzyzanowski, N. E. Chubarova, E. Semutnikova, and A. J. Cohen. 2011. “Satellite-Based Estimates of Ground-Level Fine Particulate Matter during Extreme Events: A Case Study of the Moscow Fires in 2010.” Atmospheric Environment 45 (34): 6225–32. van Donkelaar, A., R. V. Martin, and R. J. Park. 2006. “Estimating Ground-Level PM2.5 Using Aerosol Optical Depth Determined from Satellite Remote Sensing.” Journal of Geophysical Research: Atmospheres 111: D21201. van Donkelaar, A., R. V. Martin, R. J. Spurr, and R. T. Burnett. 2015b. “High-Resolution Satellite- Derived PM2.5 from Optimal Estimation and Geographically Weighted Regression over North America.” Environmental Science & Technology 49 (17): 10482–91. Vienneau, D., K. de Hoogh, M. J. Bechle, R. Beelen, A. van Donkelaar, R. V. Martin, D. B. Millet, G. Hoek, and J. D. Marshall. 2013. “Western European Land Use Regression Incorporating Satellite- and Ground-Based Measurements of NO2 and PM10.” Environmental Science & Technology 47 (23): 13555–64. 3 Literature Review of Approaches Used to Combine Satellite Observations with Ground-Level Monitoring in Urban Areas The review of the scientific literature focused on identifying approaches to using satellite aerosol-optical-depth (AOD) observations that could provide city-scale, daily average PM2.5 (particulate matter with an aerodynamic diameter less than or equal to 2.5 microns) monitoring in low- and middle-income countries (LMICs). This chapter discusses the approaches used in each of the three key steps in combining satellite observations with ground-level monitoring (GLM) of PM2.5 for estimating ground-level concentrations in a given city: 1. The retrieval of AOD from satellite observations of reflected solar radiation (­figure 2.3) 2. The translation of the vertically integrated AOD observation into an estimate of the ground-level PM2.5 concentration, including any bias corrections (­figure 2.4) and 3. Simultaneously interpolating these satellite-derived PM2.5 estimates and the available GLM data to high-resolution grids centered on selected cities (figure 3.1). The major findings of the review are presented below (see also table 3.1). The full literature review is provided in appendix A.  11 12 | Getting Down to Earth FIGURE 3.1 Annual average ground-level PM10 concentrations, London, UK, and Rome, Italy a. London b. Rome 42 km 0 57 km 0 Source: Reproduced from Vienneau and others 2013. Note: These images were created by combining ground-layer-monitoring data, satellite aerosol-optical-depth observations, and land-use data in a land-use regression model. km = kilometer; PM10 = particulate matter with an aerodynamic diameter less than or equal to 10 microns. RETRIEVAL OF AOD FROM SATELLITE OBSERVATIONS OF REFLECTED SOLAR RADIATION Although many satellites provide AOD measurements using different approaches (the different estimates are referred to as “AOD products”), only a small number provide global coverage for each day free of charge shortly after the observation is made (that is, in near real-time [NRT]) at a horizontal reso- lution of 10 kilometers or finer. Table 3.1 summarizes the strengths and limita- tions of each of these identified NRT satellite AOD products with daily global coverage. The section “Available aerosol-optical-depth products” in appendix A presents a more comprehensive list of satellite products, including those not available in NRT and those that require subscription fees. For polar-orbiting satellites (that is, satellites that observe the entire globe, but at most twice a day), this includes the Moderate-Resolution Imaging Spectroradiometer (MODIS) Dark Target algorithm (10-kilometer and 3-kilometer resolution with daily global land and ocean coverage; Levy and others 2013; see figure 3.2); the MODIS Deep Blue algorithm (10-kilometer resolution with daily global land coverage; Hsu and others 2013); and the National Oceanic and Atmospheric Administration’s Visible Infrared Imaging Radiometer Suite (VIIRS) AOD product (6-kilometer and 750-meter resolution with daily global land and ocean coverage; Jackson and others 2013). Only the MODIS products have had a comprehensive review of their performance, but comparisons with MODIS suggest the VIIRS 750-meter product tends to be biased high in urban and Literature Review of Approaches Used to Combine Satellite Observations with Ground-Level Monitoring in Urban Areas | 13 TABLE 3.1  Strengths and weaknesses of the different daily satellite aerosol-optical-depth products publicly available in near real-time AOD PRODUCT (AGENCY) ORBIT STRENGTHS WEAKNESSES MODIS Dark Polar •• Extensively validated •• Coarse horizontal resolution (10-kilometer—3-kilometer Target version much noisier) •• Global coverage (NASA) •• Only two observations a day (10:30 and 13:30 local •• Easy to access standard time) •• Includes AOD over both land and •• No AOD over bright surfaces (deserts, cities) ocean MODIS Deep Polar •• Extensively validated •• Coarse horizontal resolution (10-kilometer—3-kilometer Blue version much noisier) •• Global coverage (NASA) •• Only two observations a day (10:30 and 13:30 local •• Easy to access standard time) •• Improved coverage over cities and •• Only includes AOD over land deserts MODIS Polar •• Extensively validated •• Coarse horizontal resolution (10-kilometer—3-kilometer Combined version much noisier) •• Global coverage (NASA) •• Only two observations a day (10:30 and 13:30 local •• Easy to access standard time) •• Improved coverage over cities and deserts •• Includes AOD over both land and ocean VIIRS Polar •• Global coverage •• Data difficult to access, download cannot be automated (NOAA) •• Includes AOD over both land and •• 0.75-kilometer product very noisy and not yet ocean extensively validated •• Finest horizontal resolution for polar •• Currently only one observation a day (13:30 local orbiting products (6-­kilometer and standard time) 0.75-kilometer) •• No AOD over bright surfaces (deserts, cities) MSG-SEVIRI Geostationary •• Extensively validated •• Coverage only over Africa and Europe SMAOL •• High resolution (3-kilometer) •• AOD only over land; also excludes most coastal areas (EUMETSAT) •• Multiple observations a day •• Easy to access GOES GASP Geostationary •• Extensively validated •• Coverage only over North and South America (NOAA) •• High resolution (4-kilometer) •• AOD only over land; also excludes most coastal areas •• Multiple observations a day •• Data difficult to access, download cannot be automated GOES SMAOD Geostationary •• High resolution (2-kilometer) •• Coverage only over North and South America (NOAA) •• Multiple observations a day •• AOD only over land; also excludes most coastal areas •• Data difficult to access, download cannot be automated •• New product (starting 2018), needs further evaluation Source: World Bank. Note: In the context of a satellite measurement of AOD, noise refers to the random errors present in the measurement. Noise can generally be reduced by averaging over a larger area. AOD = aerosol optical depth; EUMETSAT = European Organisation for the Exploitation of Meteorological Satellites; GASP = GOES Aerosol/Smoke Product; GOES = Geostationary Operational Environmental Satellite; MODIS = Moderate-Resolution Imaging Spectroradiometer; MSG-SEVIRI = Meteosat Second Generation–Spinning Enhanced Visible and Infrared Imager; NASA = National Aeronautics and Space Administration; NOAA = National Oceanic and Atmospheric Administration; SMAOD = suspended matter, aerosol optical depth; SMAOL = SEVIRI-MSG Aerosol Over Land; VIIRS = Visible Infrared Imaging Radiometer Suite. 14 | Getting Down to Earth FIGURE 3.2 MODIS Dark Target 10-kilometer and 3-kilometer aerosol-optical-depth products retrieved for clear land and ocean fields of view and the local 5-kilometer average derived from the products, outer circle, compared to ground-based measurements, inner circle, over Baltimore, US a. MODIS Dark Target 10-kilometer product b. MODIS Dark Target 3-kilometer product AOD at 0.55 µm -0.05 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 Source: Munchak and others 2013. Note that these results used an older version of the MODIS surface reflectance scheme, and more recent versions (for example, Gupta and others 2016) show much better performance over urban areas. Note: AOD = aerosol optical depth; MODIS = Moderate-Resolution Imaging Spectroradiometer. mountainous regions. However, since the MODIS instruments are aging and new VIIRS instruments continue to be launched, the VIIRS products are more likely to be available in the coming decade. There are also several research products, such as extensions of the MODIS Dark Target and Deep Blue algorithms to VIIRS and the MODIS MAIAC algorithm (one-kilometer ­ resolution with daily global coverage; Emili and ­ ­ others 2011; Lyapustin and others 2011) that may become operational in the near future. However, at the time this report was prepared, these products were available only for specific test time periods and locations. For geostationary satellites (that is, satellites that continuously observe a fixed location; figure 3.3), the SMAOL (SEVIRI-MSG Aerosol Over Land) AOD product is available in NRT over Europe and Africa (three-kilometer resolution every 15 minutes; Bernard and others 2011; Mei and others 2012), and the GOES Aerosol/Smoke Product (GASP) AOD product is produced for the continental United States (four-kilometer resolution every hour; Knapp and others 2005). In  2018, the GASP product will be replaced with the GOES-16 SMAOD ­ (suspended matter, aerosol optical depth) product1,2 covering North and South America (two-kilometer resolution every 15 minutes). Himawari provides AOD retrievals over East Asia using an instrument such as GOES-16,3,4 but the Himawari product is not freely available.5,6 Seasonal limitations in satellite coverage may result in biased estimates of annual average PM 2.5 concentrations from satellites. The potential Literature Review of Approaches Used to Combine Satellite Observations with Ground-Level Monitoring in Urban Areas | 15 FIGURE 3.3 Example of good-quality aerosol optical depth 550 nm observations from a geostationary satellite from the US GOES-R SMAOD product AOD 0.00 0.50 1.00 1.50 2.00 Source: World Bank. Note: AOD = aerosol optical depth; GOES-R SMAOD = Geostationary Operational Environmental Satellite-R series–suspended matter, aerosol optical depth; nm = nanometer. application of the currently available AOD products to three example cities (Delhi, India; Ulaanbaatar, Mongolia; and Lima, Peru) was examined as part of the literature review. None of these cities is currently covered by geosta- tionary satellites with freely available AOD products, but the upcoming geo- stationary SMAOD product should cover Lima. In general, the MODIS Dark Target algorithm does not provide complete coverage over these cities because of highly reflective surfaces in urban areas, but the MODIS Deep Blue algorithm is able to provide AOD across the city. However, even with the Deep Blue algorithm, frequent clouds in Lima between May and November prevent AOD retrieval during much of this period (figure 3.4), and winter snow cover in Ulaanbaatar between December and March prevents AOD retrievals during this highly polluted season. 16 | Getting Down to Earth FIGURE 3.4 Valid MODIS Terra aerosol-optical-depth retrievals and Lima, Peru, monitoring sites in the OpenAQ database, 2016–17 Lima MOD satellite points OpenAQ monitor sites Source: World Bank, produced using Esri ArcGIS. Note: The blue diamonds = valid satellite AOD data and the green triangles = OpenAQ monitoring sites. MOD = MODIS Terra; MODIS = Moderate-Resolution Imaging Spectroradiometer; OpenAQ = openaq.org. THE TRANSLATION OF AOD TO GROUND-LEVEL PM2.5 CONCENTRATIONS Several studies have successfully converted satellite AOD to ground-level PM2.5 estimates using statistical techniques, chemical transport model (CTM)–based approaches, or hybrid approaches, generally for continental-to-global spatial scales and monthly-to-annual time scales. Statistical approaches train statistical models, such as nonlinear generalized additive models (GAMs; Sorek-Hamer and others 2013; Strawa and others 2013), on historical GLM data to predict ground-level PM2.5 using satellite AOD measurements and other meteorological and geographic data. CTM-based approaches use computer models of air qual- ity, called “chemical transport models,” to determine a time-varying relationship between ground-level PM2.5 concentrations and satellite AOD measurements. This relationship is then used to scale the CTM aerosol profile until the CTM- calculated AOD matches the satellite AOD measurement, providing a better esti- mate of the ground-level PM2.5 concentration than would be possible from the CTM alone. Hybrid methods combine statistical and CTM-based approaches by training a statistical model to correct the errors in the initial CTM-based satellite estimates of the ground-level PM2.5 concentrations. For example, recent work Literature Review of Approaches Used to Combine Satellite Observations with Ground-Level Monitoring in Urban Areas | 17 has used a hybrid approach in which CTM-based approaches are followed by a second step that uses geographically weighted regression to correct for errors in the first-step estimates (van Donkelaar, Martin, Spurr, and Burnett 2015). These corrections require long-term (multiple years), reliable GLM data over a large region, including both urban and rural sites (see table 3.2). Table 3.2 lists the strengths and weaknesses of these approaches. The statis- tical approaches (for example, Hu and others 2014; Sorek-Hamer and others 2013, 2015; Strawa and others 2013) can be more accurate if sufficient GLM data are available for the training. However, CTM-based approaches (for example, Geng and others 2015; van Donkelaar and others 2006, 2010, 2011; van Donkelaar, Martin, Brauer, and Boys 2015) are required in areas with- out GLM data or where available GLM data are of unreliable quality. Several potential sources of data are available for planetary boundary layer height and other meteorological parameters for the statistical approaches, as well as sources of CTM AOD and aerosol-profile data for the CTM-based approaches provided free of charge from global CTM run by high-income countries (see the section “Converting AOD to ground-level PM2.5” in appendix A). LMICs, which often face competing needs for limited budgetary resources, may find that the use of freely available data is preferable to the expense of running their own CTMs or paying for data. In that case, the forecasts of ground-level concentrations and vertical profiles of PM2.5 from the National Centers for ­ Environmental Prediction Global Forecast System (Lu, da Silva, and others 2016; Lu, Wei, and others 2016) can be used if NRT o ­ utput is needed, and the optimized historical ground-level concentrations and vertical profiles of PM2.5 from the National Aeronautics and Space Administration MERRA-2 reanalysis (Provençal and others 2017) can be used for historical studies. TABLE 3.2  Strengths and weaknesses of different approaches to converting satellite aerosol-optical- depth data into estimates of ground-level PM2.5 concentrations APPROACH STRENGTHS WEAKNESSES CTM-based •• Easy to implement globally •• Can have significant biases due to errors in the CTM’s simulation of the vertical profile and optical •• Does not require any GLM data properties of the aerosols •• Can remove cases with high-altitude aerosol •• Errors are likely larger for LMICs, since estimates of plumes or low PBL heights from the analysis PM2.5 emissions are more out-of-date or uncertain for these countries •• Errors are larger if horizontal resolution of the CTM is too coarse and thus a single model PM2.5 value represents an average of PM2.5 in urban and rural air Statistical •• Locally unbiased (since these methods fit to the •• Requires a long-term (multiple years), reliable GLM average for a single city) data record with concurrent satellite observations •• Estimates error in the fit automatically •• Cannot filter for high-altitude plumes •• Does not require an initial guess at the aerosol •• Determining appropriate statistical method for a profile or optical properties given site can take many attempts Hybrid •• Easy to implement globally •• Requires a long-term (multiple years), reliable GLM data record with concurrent satellite observations •• Globally unbiased (since these methods fit to the average for all cities and monitors considered) •• Requires GLM data in many different land-use types (urban and rural sites) •• Can filter the CTM results for high-altitude aerosol plumes or low PBL heights •• Can still have significant biases at specific sites Source: World Bank. Note: CTM = chemical transport model; GLM = ground-level monitoring; LMICs = low- and middle-income countries; PBL = planetary boundary layer; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. 18 | Getting Down to Earth COMBINING SATELLITE PM2.5 ESTIMATES AND GLM DATA INTO HIGH-RESOLUTION GRIDS CENTERED ON SELECTED CITIES Co-kriging or land-use regression (LUR) can be used to combine satellite PM2.5 estimates with GLM data to calculate estimates of PM2.5 concentrations at a neighborhood scale (100 to 500 meters) across urban areas. Ordinary kriging uses a weighted average of neighboring GLM measurements to predict PM2.5 concentrations across an urban area. Co-kriging is an extension of ordinary krig- ing that can take advantage of additional data sets or variables, using both the correlations between the monitors and the cross-correlations between the mon- itor data and the additional data sets (that is, the satellite-derived PM2.5 esti- mates) to make better predictions (for example, Millar and others 2010; Pearce and others 2009; Wu, Winer, and Delfino 2006). LUR spatially links GLM mea- surements of PM2.5 with other associated variables such as elevation, distance from roads, population density, and land-use type to develop fine-scale estimates of long-term PM2.5 concentrations within an urban area that better represent gradients near highways and other pollution sources than is possible with krig- ing (for example, Lee, Chatfield, and Strawa 2016; Millar and others 2010; Vienneau and others 2013). LUR can also be used to predict the bias in CTM- based estimates of ground-level PM2.5 concentrations as part of a hybrid approach to using satellites to estimate PM2.5 exposures. Co-kriging and LUR tend to work better as the satellite horizontal resolution is increased and as the GLM network covers a wider variety of sites (for example, urban and rural). Many GLM networks in LMICs are almost entirely in cities,  which makes LUR more difficult because of the lack of variation in land-use parameters (for example, urban land fraction, distance of coasts, and distance from highways) across the network. Co-kriging could be used with sat- ellite PM2.5 estimates to reduce the number of GLM stations required to ade- quately cover an urban area (see discussion in chapter 5) and can be used to obtain either daily average or annual average estimates of PM2.5. However, this study did not find any example in the literature where co-kriging had previously been used to combine satellite AOD and GLM PM2.5 data. LUR has been successfully used by several groups to provide high horizontal resolution (100- to 500-meter), annual average estimates on PM2.5 across regions and within cities using both satellite and GLM data (for example, Lee, Chatfield, and Strawa 2016; Vienneau and others 2013). Daily average estimates of PM2.5 are also possible if meteorological data that vary daily and other data are used in the LUR. The needed data for LUR at 500-meter resolution is freely available for the entire globe, and most LUR studies have used the ArcGIS software package to perform the regression with geographical information system data sets. NOTES 1. https://www.goes-r.gov/products/baseline-aerosol-opt-depth.html. 2. https://www.goes-r.gov/downloads/users/conferencesAndEvents/2014/GOES-R_Series​ _Program/04-Laszlo_abstract.pdf. 3. https://www.eorc.jaxa.jp/ptree/documents/Himawari_Monitor_Aerosol_Product_v5.pdf. 4. http://www.data.jma.go.jp/mscweb/technotes/msctechrep61-6.pdf. 5. http://www.jmbsc.or.jp/en/meteo-data.html. 6. http://www.jmbsc.or.jp/en/Data/Himawari-8-JMBSC-HP(2017.02.20).pdf. Literature Review of Approaches Used to Combine Satellite Observations with Ground-Level Monitoring in Urban Areas | 19 REFERENCES Bernard, E., C. Moulin, D. Ramon, D. Jolivet, J. Riedi, and J.-M. Nicolas. 2011. “Description and Validation of an AOT Product over Land at the 0.6 μm Channel of the SEVIRI Sensor Onboard MSG.” Atmospheric Measurement Techniques 4: 2543–65. Emili, E., A. Lyapustin, Y. Wang, C. Popp, S. Korkin, M. Zebisch, S. Wunderle, and M. Petitta. 2011. “High Spatial Resolution Aerosol Retrieval with MAIAC: Application to Mountain Regions.” Journal of Geophysical Research 116: D23211. Geng, G., Q. Zhang, R. V. Martin, A. van Donkelaar, H. Huo, H. Che, J. Lin, and K. He. 2015. “Estimating Long-Term PM2.5 Concentrations in China Using Satellite-Based Aerosol Optical Depth and a Chemical Transport Model.” Remote Sensing of Environment 166: 262–70. Gupta, P., R. C. Levy, S. Mattoo, L. A. Remer, and L. A. Munchak. 2016. “A Surface Reflectance Scheme for Retrieving Aerosol Optical Depth over Urban Surfaces in MODIS Dark Target Retrieval Algorithm.” Atmospheric Measurement Techniques 9: 3293–308. Hsu, N. C., M. J. Jeong, C. Bettenhausen, A. M. Sayer, R. Hansell, C. S. Seftor, J. Huang, and S. C. Tsay. 2013. “Enhanced Deep Blue Aerosol Retrieval Algorithm: The Second Generation.” Journal of Geophysical Research: Atmospheres 118 (16): 9296–315. Hu, X., L. A. Waller, A. Lyapustin, Y. Wang, M. Z. Al-Hamdan, W. L. Crosson, M. G. Estes, S. M. Estes, D. A. Quattrochi, S. J. Puttaswamy, and Y. Liu. 2014. “Estimating Ground-Level PM2.5 Concentrations in the Southeastern United States Using MAIAC AOD Retrievals and a Two-Stage Model.” Remote Sensing of Environment 140: 220–32. Jackson, J. M., H. Liu, I. Laszlo, S. Kondragunta, L. A. Remer, J. Huang, and H.-C. Huang. 2013. “Suomi-NPP VIIRS Aerosol Algorithms and Data Products.” Journal of Geophysical Research: Atmospheres 118: 12673–89. Knapp, K. R., R. Frouin, S. Kondragunta, and A. Prados. 2005. “Toward Aerosol Optical Depth Retrievals over Land from GOES Visible Radiances: Determining Surface Reflectance.” International Journal of Remote Sensing 26 (18): 4097–116. Lee, H. J., R. B. Chatfield, and A. W. Strawa. 2016. “Enhancing the Applicability of Satellite Remote Sensing for PM2.5 Estimation Using MODIS Deep Blue AOD and Land Use Regression in California, United States.” Environmental Science & Technology 50 (12): 6546–55. Levy, R. C., S. Mattoo, L. A. Munchak, L. A. Remer, A. M. Sayer, F. Patadia, and N. C. Hsu. 2013. “The Collection 6 MODIS Aerosol Products over Land and Ocean.” Atmospheric Measurement Techniques 6: 2989–3034. Lu, C.-H., A. da Silva, J. Wang, S. Moorthi, M. Chin, P. Colarco, Y. Tang, P. S. Bhattacharjee, S.-P. Chen, H.-Y. Chuang, H.-M. H. Juang, J. McQueen, and M. Iredell. 2016. “The Implementation of NEMS GFS Aerosol Component (NGAC) Version 1.0 for Global Dust Forecasting at NOAA/NCEP.” Geoscientific Model Development 9: 1905–19. Lu, C.-H., S.-W. Wei, S. Kondragunta, Q. Zhao, J. McQueen, J. Wang, and P. Bhattacharjee. 2016. “NCEP Aerosol Data Assimilation Update: Improving NCEP Global Aerosol Forecasts Using JPSS-NPP VIIRS Aerosol Products.” Paper presented at the “8th International Cooperative for Aerosol Prediction.” http://icap.atmos.und.edu/ICAP8/Day3/Lu_NCEP_ThursdayAM​.pdf. Lyapustin, A., Y. Wang, I. Laszlo, R. Kahn, S. Korkin, L. Remer, R. Levy, and J. S. Reid. 2011. “Multiangle Implementation of Atmospheric Correction (MAIAC): 2. Aerosol Algorithm.” Journal of Geophysical Research 116: D03211. Mei, L., Y. Xue, G. de Leeuw, T. Holzer-Popp, J. Guang, Y. Li, L. Yang, H. Xu, X. Xu, C. Li, Y. Wang, C. Wu, T. Hou, X. He, J. Liu, J. Dong, and Z. Chen. 2012. “Retrieval of Aerosol Optical Depth over Land Based on a Time Series Technique Using MSG/SEVIRI Data.” Atmospheric Chemistry and Physics 12: 9167–85. Millar, G., T. Abel, J. Allen, P. Barn, M. Noullett, J. Spagnol, and P. L. Jackson. 2010. “Evaluating Human Exposure to Fine Particulate Matter Part II: Modeling.” Geography Compass 4 (7): 731–49. Munchak, L. A., R. C. Levy, S. Mattoo, L. A. Remer, B. N. Holben, J. S. Schafer, C. A. Hostetler, and R. A. Ferrare. 2013. “MODIS 3 km Aerosol Product: Applications over Land in an Urban/ Suburban Region.” Atmospheric Measurement Techniques 6: 1747–59. 20 | Getting Down to Earth Pearce, J. L., S. L. Rathbun, M. Aguilar-Villalobos, and L. P. Naeher. 2009. “Characterizing the Spatiotemporal Variability of PM2.5 in Cusco, Peru Using Kriging with External Drift. Atmospheric Environment 43 (12): 2060–69. Provençal, S., V. Buchard, A. M. da Silva, R. Leduc, N. Barrette, E. Elhacham, and S. H. Wang. 2017. “Evaluation of PM2.5 Surface Concentrations Simulated by Version 1 of NASA’s MERRA Aerosol Reanalysis over Israel and Taiwan.” Aerosol and Air Quality Research 17: 253–61. Sorek-Hamer, M., I. Kloog, P. Koutrakis, A. W. Strawa, R. Chatfield, A. Cohen, W. L. Ridgway, and D. M. Broday. 2015. “Assessment of PM2.5 Concentrations over Bright Surfaces Using MODIS Satellite Observations.” Remote Sensing of Environment 163: 180–85. Sorek-Hamer, M., A. W. Strawa, R. B. Chatfield, R. Esswein, A. Cohen, and D. M. Broday. 2013. “Improved Retrieval of PM2.5 from Satellite Data Products Using Non-Linear Methods.” Environmental Pollution 182: 417–23. Strawa, A. W., R. B. Chatfield, M. Legg, B. Scarnato, and R. Esswein. 2013. “Improving Retrievals of Regional Fine Particulate Matter Concentrations from Moderate Resolution Imaging Spectroradiometer (MODIS) and Ozone Monitoring Instrument (OMI) Multisatellite Observations.” Journal of the Air & Waste Management Association 63 (12): 1434–46. van Donkelaar, A., R. V. Martin, M. Brauer, and B. L. Boys. 2015. “Use of Satellite Observations for Long-Term Exposure Assessment of Global Concentrations of Fine Particulate Matter.” Environmental Health Perspectives 123 (2): 135–43. van Donkelaar, A., R. V. Martin, M. Brauer, R. Kahn, R. Levy, C. Verduzco, and P. J. Villeneuve. 2010. Global Estimates of Ambient Fine Particulate Matter Concentrations from Satellite- Based Aerosol Optical Depth: Development and Application.” Environmental Health Perspectives 118 (6): 847–55. van Donkelaar, A., R. V. Martin, R. C. Levy, A. M. da Silva, M. Krzyzanowski, N. E. Chubarova, E. Semutnikova, and A. J. Cohen. 2011. “Satellite-Based Estimates of Ground-Level Fine Particulate Matter during Extreme Events: A Case Study of the Moscow Fires in 2010.” Atmospheric Environment 45 (34): 6225–32. van Donkelaar, A., R. V. Martin, and R. J. Park. 2006. “Estimating Ground-Level PM2.5 Using Aerosol Optical Depth Determined from Satellite Remote Sensing.” Journal of Geophysical Research: Atmospheres 111: D21201. van Donkelaar, A., R. V. Martin, R. J. Spurr, and R. T. Burnett. 2015. “High-Resolution Satellite- Derived PM2.5 from Optimal Estimation and Geographically Weighted Regression over North America.” Environmental Science & Technology 49 (17): 10482–91. Vienneau, D., K. de Hoogh, M. J. Bechle, R. Beelen, A. van Donkelaar, R. V. Martin, D. B. Millet, G. Hoek, and J. D. Marshall. 2013. “Western European Land Use Regression Incorporating Satellite- and Ground-Based Measurements of NO2 and PM10.” Environmental Science & Technology 47 (23): 13555–64. Wu, J., A. M. Winer, and R. J. Delfino. 2006. “Exposure Assessment of Particulate Matter Air Pollution before, during, and after the 2003 Southern California Wildfires.” Atmospheric Environment 40 (18): 3333–48. 4 Quality-Assurance Procedures in Low- and Middle-Income Countries To better estimate the exposure of their citizens to particulate matter with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5 ) and the associ- ated health effects, and to design policies to minimize these effects, low- and middle-income countries (LMICs) are working to begin, expand, and improve their air-quality monitoring efforts. However, some LMICs (Type I) currently have no air-quality data. Other LMICs (Type II) have air-quality data, but, because instrument maintenance and other “quality-assurance” procedures have not been fully followed, their data are potentially inaccurate. Still other LMICs (Type III) have accurate data but with insufficient spatial coverage. Satellites have been suggested as a way to reduce or eliminate the need for a ground-level-monitoring (GLM) network. However, accurate GLM data are needed to determine the performance of the satellite approaches to air-quality monitoring and to use the statistical and hybrid satellite approaches. Thus, to evaluate the ability of satellites to assist with air-quality monitoring in LMICs (chapter 5), this chapter discusses the issues affecting the availability and accuracy of GLM data in LMICs in the first main section, the potential of automated methods to flag potentially incorrect GLM data in the second, and the need to examine the metadata, warning files, and operator logs for the GLM instruments to ensure the data are of sufficient accuracy to be used to develop and evaluate satellite approaches to PM2.5 monitoring. ISSUES AFFECTING THE AVAILABILITY AND ACCURACY OF GLM DATA IN LMICs Recent high-air-pollution events in many LMICs have highlighted the need for improved dissemination of PM2.5 concentrations obtained via GLM. Currently, one of the most promising dissemination mechanisms for real-time air-quality information in LMICs is the open-access web platform OpenAQ (see box 4.1). In the research underlying this report, we used data from OpenAQ for several LMICs to evaluate the ability of satellites to estimate daily average ground-level PM2.5 concentrations in cities within LMICs. OpenAQ, and similar accessible  21 22 | Getting Down to Earth BOX 4.1 Open-access web platform for air quality data OpenAQ (openaq.org) is an open-source platform and in the world to freely access station-level, physical par- community that makes air-quality data accessible ticulate matter (PM2.5, PM10, BC) and gaseous (O3, NO2, worldwide. The platform, launched in 2015, automati- SO2, CO) air-quality data in aggregate, in near real- cally aggregates, at 10-minute intervals, ground-level, time or historically, that transparently traces back to near real-time air-quality data, shared publicly in dis- their originating sources (for example, a government parate formats by government entities across the air-quality agency website, ftp server, or API). The world. The data are synthesized into a universal format open-source system is built and maintained by a com- and made freely available via an application program- munity of software developers, scientists, open-data ming interface (API). To date, OpenAQ is the only site enthusiasts, and others across the world. Note: BC = black carbon; CO = carbon monoxide; NO2 = nitrogen dioxide; O3 = ozone; PM2.5/PM10 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns/10 microns, respectively; SO2 = sulfur dioxide. air-quality web platforms, can provide LMIC governments and citizens with important and timely access to local PM2.5 concentrations. However, simply rereporting PM2.5 levels reported by LMICs does not guarantee that those data are of sufficient accuracy and precision for air-quality monitoring, or that those data can be usefully combined with satellite measure- ments to determine human exposures to PM2.5. Many different instruments and techniques provide a measure of PM2.5 with significantly different estimates of their accuracy (that is, how close they get to the true value) and precision (that is, how close two measurements of the same true concentration will be). In addition, poor instrument operation, lack of calibrations, and incomplete quality assurance and quality controls can reduce the accuracy of PM 2.5 measurements. Given the seriousness of the health effects of PM2.5, inaccurate measurements have the potential to mislead the public into a false sense of secu- rity (if the monitors are underreporting mass concentration) or incite alarm (if they are overreporting). These errors would also have economic impacts, such as reduced labor productivity due to people staying home during false alarms or getting ill by going out in missed poor air-quality conditions. Thus, it is important that GLM in LMICs provide reliable PM2.5 measurements with well-understood uncertainties. The instruments commonly used to measure PM2.5 and report results to OpenAQ are approved by the US Environmental Protection Agency (EPA)1 as federal equivalent methods (FEMs) for PM2.5 characterization and thus are expected to provide daily average PM2.5 concentrations within 10 percent of the true value when properly operated and calibrated. However, other “low-cost” instruments can have much higher errors. Common instruments for measuring PM2.5 in use around world include (1) Met One Beta Attenuation Mass Monitor model BAM-1020; (2) Thermo Fisher Scientific model 5014i; (3) Thermo Fisher Scientific Tapered Element Oscillating Microbalance (TEOM, models 1400a, 1400b, and 1405); and (4) Grimm Aerosol Environmental Dust Monitor (EDM model 180). All these instruments have been approved by the EPA as FEMs for PM2.5 characterization. To qualify as a “FEM,” the PM2.5 measurement technique must demonstrate differences of less than 10 percent relative to the federal ref- erence method (FRM). In such head-to-head comparisons, the 10 percent threshold is applied to the intercept and slope of a correlation scatterplot of the Quality-Assurance Procedures in Low- and Middle-Income Countries | 23 test method versus the FRM. The EPA defines the FRM as gravitational (weight-based) PM2.5 measurements, integrated or averaged over a 24-hour sam- ple time. Figure 4.1 shows an example of the allowed uncertainties in the correla- tion (red hexagon outline) for the test instrument to qualify as a FEM. It is important to note the 24-hour averaging interval in the FEM designation. Given the dynamic and spatially heterogeneous nature of ambient PM2.5 air pol- lution, it is often necessary to characterize PM2.5 concentrations with faster time resolution (hourly or even one-minute data). Under these faster time-response conditions, the FEM status of a given PM2.5 instrument may not hold. Each of the common PM2.5 measurement methods (BAMM, TEOM) rely on different techniques for measuring PM2.5 (see appendix D) and thus have differ- ent precisions and accuracies. Although they output PM2.5 mass concentrations, their measurement uncertainties vary for each device. Table 4.1 provides the manufacturer’s stated uncertainties for each PM2.5 instrument. Manufacturers provide precision uncertainties (µg/m3) for a 1- and/or 24-hour average(s), stat- ing that any single measurement over that averaging interval should be that close to the true value 99.7 percent of the time. Measurement uncertainties (for a well-operated system under the EPA requirements) would be the greater of these uncertainties or the 10 percent FEM error estimate for a given measure- ment. More details about the BAMM and TEOM methods are given in the sec- tions “Operating methodologies and uncertainties associated with the BAMM technique” and “Operating methodologies and uncertainties associated with the TEOM technique,” respectively, in appendix D. Although it is important to know and understand the manufacturer’s stated uncertainties, the true uncertainties of PM2.5 measurements will be significantly larger, due to uncontrolled and unknown factors including site-specific FIGURE 4.1 Example of US federal equivalent method measurement and uncertainties: Data set slope, intercept, and limits 8 6 4 Intercept (µg/m3) 2 0 –2 –4 –6 –8 0.8 0.9 1.0 1.1 1.2 Slope Perfect agreement between methods Boundaries of acceptable agreement between methods Source: Modified from US Code of Federal Regulations (Figure C-2 to Subpart C of Part 53, Title 40, 72 CFR 32204, June 12, 2007). Note: μg/m3 = micrograms per cubic meter. 24 | Getting Down to Earth TABLE 4.1  PM2.5 instrument manufacturers’ stated uncertainties INSTRUMENT/ PRECISION ACCURACY MANUFACTURER MODEL (1 σ IN 1 HR) UNITS (1 σ IN 1 HR) UNITS COMMENTS Grimm EDM — — — — Optical Particle Counter; provides 180 only “reproducibility” value of ± 3% MetOne BAMS 2.4 μg/m3 — — No accuracy provided 1020 Thermo Fisher TEOM 2 μg/m3 0.75 % None Scientific 1400 Thermo Fisher BAMS 2 μg/m3 5 % None Scientific 5014i Source: World Bank. Note: PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter; — = not available; σ = standard deviation. sampling considerations, volatile PM losses, operator errors, and others. These considerations can lead to inaccurate PM2.5 data even when FEM instruments are used, and thus additional steps must be taken to ensure the accuracy of the GLM data. These steps can include the following: utomated processing of the PM2.5 data to flag potentially erroneous data • A values (see the next section) and • Examining instrument metadata, warning or error files, and operation logs to ensure that proper calibration, maintenance, and other quality-assurance procedures are being followed (see the section “Examining metadata, warn- ing files, and operator logs” in chapter 4). USING AUTOMATED SCRIPTS TO FLAG POTENTIALLY INCORRECT GLM DATA To provide more information about the uncertainties and reliability of dis- seminated air-quality data (that is, in addition to manufacturers’ stated pre- cision uncertainty), OpenAQ created a data-quality wrapper to establish a baseline approach toward flagging potentially erroneous data values within the OpenAQ database (see also the section “Co-kriging and kriging with sat- ellite and OpenAQ data” in appendix C). Guidance points for the wrapper development are outlined below: (1) Develop flags on a per-measurement (rather than per-station) basis—Aim to justify “trustworthiness” on the basis of what is known or expected from a specific technique and avoid the negative political consequences of blacklist- ing specific sites altogether. Repeat data flags—When consecutive data points in the OpenAQ database (2)   from one PM2.5 instrument repeat, generate a flag to examine the data more closely, since it is unlikely that identical data values are measured over time. Include the capacity for the user to input an adjustable parameter to define the number of repeat indexes necessary to trigger the flag. Zero points—If reported values are in fact 0.00000, generate a zero flag. (3)   Understand the extent to which the instrument is reporting zero due to an on-site zero calibration, instrument electronic glitch, or encountering PM2.5 concentrations below the mass detection limits of the device. Quality-Assurance Procedures in Low- and Middle-Income Countries | 25 (4)  Negative values—Typically, instrument manufacturers and/or instrument operators will generate −999 or −9999 error flags within the output data stream to indicate a broader context on what was happening with the instru- ment at the time of the reported error code. This could correspond with on-site calibration (zero and span) or be an indicator that the system was off-line. Encourage OpenAQ users to flag and store these negative values within their initial QA/QC steps. Missing data—Within the OpenAQ database, one often finds periods with (5)   missing data (data not found at a given interval). The fact that the data are simply removed leaves some open questions, such as the following: How (if at all) do the missing data affect the resultant “average” values reported? What is the start and stop time for each reported average, and what fraction of this time was removed? In this context, “missing” data should be flagged, but not necessarily as −999. Tracking the extent of missing data for a given instrument may also be a useful indicator of that instrument’s overall reliability. EXAMINING METADATA, WARNING FILES, AND OPERATOR LOGS To accurately evaluate satellite approaches and to combine satellite PM2.5 esti- mates with GLM data, the GLM PM2.5 measurements need to be provided in parallel with information on the instrument or technique type, estimates of mea- surement uncertainties, relevant metadata, and operational history. If the only publicly available information is the manufacturer and instrument type or model, one can identify rational lower-limit estimates of the expected uncertain- ties (assuming standard operational protocols for instrument upkeep are main- tained). However, as noted above, the true uncertainties of PM2.5 measurements can be significantly larger than the manufacturer’s stated uncertainties. Without the necessary metadata to assess the quality of the PM2.5 measurements, the sta- tistical and hybrid approaches to satellite estimates of PM2.5 cannot be used, and combining the GLM data with the satellite estimates may result in inaccurate estimates of PM2.5 exposure and health effects. Thus, it is important that each PM2.5 measurement not only identifies the type of instrument used for the measurement, but also tracks the instrument’s relevant history of use and calibration. This includes all “meta” data from instru- ment data files, including “housekeeping” files. Housekeeping files are often autogenerated by manufacturers to help identify deviation from normal operation. Such files would include error codes for alarm states, loss of sample flow, pressure changes, or periods where the instrument was out-of-spec (for example, temperature exceedances). In addition, any recorded results from tests of standard operating procedures and quality-assurance protocols should be collected and disseminated along with the PM2.5 measurements to establish an instrument track record over time providing a transparent, data-centric measure of instrument reliability and ­ performance. This is typically done by following established protocols, such as defined by the EPA. As determined during this project, US Diplomatic Posts are equipped to ­measure PM2.5 with Met One beta attenuation monitor systems and follow the current EPA guidelines for PM2.5 monitoring. However, their sampling location may not be representative of the typical air-quality conditions within a given city. 26 | Getting Down to Earth These EPA guidelines provide a well-established protocol for conducting PM2.5 measurements with a high level of quality assurance and quality control. The US embassies typically provide these PM2.5 data to host governments, and these data are now being collected on OpenAQ. Although it is important that these measurements get disseminated, it is important to stress that each measurement should be backed up with open access to the quality assurance and control (QA/QC) protocol metadata used to conduct the measurements and to QA/QC the measurements. Currently, US embassies do not publish their full QA/QC protocol metadata along with the PM2.5 levels, thereby reducing the use of these measurements as well-established and controlled reference mea- surements. It is recommended that the US Diplomatic Posts publish their full QA/QC protocol and relevant metadata along with the PM2.5 levels to make their data more useful. NOTE 1. Other national air quality agencies do their own certification of methods. Most of the instruments listed are certified in multiple countries. 5 Using Satellite Data for Daily, City-Scale PM2.5 Monitoring As noted in chapter 2, many low- and middle-income countries (LMICs) do not have ground-level-monitoring (GLM) data of sufficient reliability for monitor- ing concentrations of particulate matter with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5 ). This leads to uncertainties related to the reliabil- ity of GLM data for use in designing policies to protect public health. It has been suggested that satellites could be used instead of GLM in these countries to mon- itor PM2.5 or that satellites could be used in combination with GLM networks to improve the estimates of human exposure to PM2.5. However, most previous studies of the potential of satellites to help with air-quality monitoring have been performed for countries with long-established, well-maintained GLM networks and have focused on global, annual average estimates of PM2.5 concentrations and exposures (chapter 3). Thus, there is a need to provide guidance to LMICs on how satellite data can best be used in city-scale, daily air-quality monitoring; what the uncertainties (that is, expected errors) and biases of the satellite esti- mates of ground-level PM2.5 concentrations are; and what local conditions (altitude, distance from coast, and so forth) lead to larger errors in the satellite estimates of ground-level PM2.5 concentrations. To address these needs, the results of the literature review (chapter 3) were used to identify promising new pathways for LMICs to use satellite observations in their city-scale, daily air-quality monitoring and forecasting. One of those novel approaches is statistical, and another is based on the chemical transport model (CTM); both approaches were tested. The statistical approach used the generalized additive model approach of Sorek-Hamer and others (2013), modi- fied to use the ratio of the aerosol optical depth (AOD) to the planetary boundary layer (PBL) as the primary predictor of ground-level PM2.5 concentrations, as in Chatfield and others (2017). To our knowledge, this particular statistical approach has not been tested before. The CTM-based approach used AODs and aerosol vertical profiles from the MERRA-2 reanalysis to derive a ground-level PM2.5 estimate following the methods of van Donkelaar and others (2010). The use of standard CTM output freely provided by high-income countries for the CTM-based approach is novel, since most previous studies have invested signif- icant effort in developing customized CTM simulations, and thus the  27 28 | Getting Down to Earth performance of these approaches when using the model output available to LMICs has not been previously tested. Subsequently, the GLM network data included in the OpenAQ database (chapter 4) was used to evaluate (1) the ability of satellites to predict daily aver- age ground-level PM2.5 concentrations and (2) their variations within cities, using these promising new approaches. The next section describes the initial testing of the satellite approaches (using polar-orbiting satellite data) using 2017 GLM data sets for three cities: Delhi, India; Ulaanbaatar, Mongolia; and Lima, Peru. These cities were chosen because they have near real-time PM2.5 observations at multiple locations in each city from local government air-quality agencies and US Diplomatic Post PM2.5 data. A city in Africa (Accra, Ghana) was added to explore the ability of geostationary satellites to predict ground-level PM2.5 concentrations as compared with polar-orbiting satellites. This work was performed to answer the following questions: • Are the GLM data from the local government consistent with the data from the US Diplomatic Posts? • How do the CTM-based and statistical-satellite estimates of ground-level PM2.5 concentrations compare with the GLM data and with each other? Can either satellite approach correctly represent the day-to-day and site-to-site variation of PM2.5 concentrations within a city? • Do the satellite estimates of surface PM2.5 improve when higher-resolution polar satellite or geostationary satellite data are used instead of the MODIS observations? • Does a hybrid technique (using a statistical model to correct the biases in the CTM-based approach) perform better than a statistical or CTM-based approach alone? • Can satellites reduce the number of required monitoring sites in a GLM network? The last section describes the results of evaluating the satellite approaches for nine cities in LMICs using GLM data from 2016–17 where available. This work was performed to answer the following questions: • How well do the satellite estimates of ground-level PM2.5 concentrations compare with available GLM data in different cities? • Can we identify patterns in the errors in the satellite estimates of ground- level PM2.5 concentrations in the different cities based on their location (for example, distance from coasts, distance from deserts, and altitude)? INITIAL TESTING OF SATELLITE PRODUCTS IN SELECTED CITIES A full description of the methodology and results for this study is included in appendix B. The selected satellite data sets were from the MODIS (Moderate- Resolution Imaging Spectroradiometer) combined Deep Blue and Dark Target product (10-kilometer resolution, with data from the Aqua and Terra satellites tested separately), the standard VIIRS (Visible Infrared Imaging Radiometer Suite) AOD product (6-kilometer resolution), and the SEVIRI (Spinning Enhanced Visible and Infrared Imager) geostationary AOD product Using Satellite Data for Daily, City-Scale PM2.5 Monitoring | 29 (3-kilometer resolution, used as an example of the capability of next-generation geostationary satellites). Each satellite AOD data set also provides information about the quality of the AOD retrieval through a set of variables referred to as “quality flags” and provides recommendations for the minimum acceptable val- ues for these flags. This is analogous to the quality filter for GLM data discussed in chapter 4. However, note that the quality flags only remove bad data; they do not provide information on the relative uncertainties in the remaining AOD measurements. All recommended quality flags were applied to filter the AOD data before analysis. One statistical approach and one CTM-based approach were tested for each city for 2017 (except in Accra and Lima, where only a CTM approach could be used due to limited 2017 GLM data for these cities). Geographically weighted regression (GWR) was then performed to relate the errors between the GLM PM2.5 observations and the two satellite-based estimates of ground-level PM2.5 concentrations, as in van Donkelaar, Martin, Spurr, and Burnett (2015). Consistency of local government and US Diplomatic Post data In Delhi and Ulaanbaatar, the local GLM network data appear to be consistent with the US Diplomatic Post data, and thus the two data sets can be used together in the evaluation of the satellite methods. A test was performed in Delhi to deter- mine if the US Diplomatic Post data were roughly consistent with the local GLM network data by using ordinary kriging of the local GLM network data to predict the annual average ground-level PM2.5 concentration at the US Diplomatic Post. Kriging the local GLM network data captures the measured value at the US Diplomatic Post well, suggesting the two data sets are consistent. Further analy- sis (see the section “Interpreting ground-level PM2.5 data from LMICs with lim- ited metadata” in appendix D) suggested that the US Diplomatic Post data were consistent with local government monitors in Delhi but showed differences in Lima, where the local government monitors reported concentrations 30–50 ­ percent lower than the US embassy data. Further investigation would be required to identify the sources of this difference. Performance of different approaches for monitoring PM2.5 with satellites All cities had significant limitations in the availability of satellite AOD data that resulted in biased satellite estimates of annual average ground-level PM2.5 con- centrations. This is due to persistent clouds in Lima, wintertime snow cover in Ulaanbaatar, wet-season clouds in Delhi, and mixed water and land and bright surfaces in Accra. Satellite coverage was poorest in the coastal city of Lima and thus did not allow for a meaningful evaluation of a CTM-based approach for converting satellite AOD to ground-level PM2.5 concentrations. In Ulaanbaatar, no satellite measurements are available for the high PM2.5 winter months of December to mid-March, which would underestimate the true annual average PM2.5 concentration for this city by 50 percent, even if satellites were able to perfectly predict ground-level PM2.5 concentrations at other times. Delhi has sat- ellite measurements in all months but substantially fewer observations in the peaks of the wet and dry seasons (December, January, July, and August), which would also result in a slight (about 10 percent) underestimate of the true annual average PM2.5 concentration. 30 | Getting Down to Earth The CTM-based approach provided an accurate estimate (within about 10 percent) of the citywide annual average PM2.5 concentration (necessary for assessing chronic health effect impacts) in Delhi, but the CTM-based approach underestimated the citywide annual average PM2.5 concentration in Ulaanbaatar by a factor of 10. This is likely due to Ulaanbaatar’s location in a river valley sur- rounded by mountains and mostly rural land. Thus, the coarse resolution of global CTMs (combined with likely inaccurate emission inventories for Ulaanbaatar) means that CTMs are not able to correctly represent the aerosol profile within the city, instead using an average profile that is more representa- tive of the surrounding rural area. These same issues are present in the Global Burden of Disease (GBD) data set before the GWR correction with global GLM data is applied.1 Statistical approaches were generally able to correctly represent the annual average PM2.5 concentrations if the AOD data were available for the entire year. The estimates of the ground-level daily average PM2.5 concentrations from the CTM-based and statistical approaches generally had a low correlation (R < 0.3; tables 5.1 and 5.2) with the true daily average PM2.5 concentrations within a city. Consequently, using satellite data alone would likely result in incor- rect estimates of acute PM2.5 exposure and health effects. The techniques also showed little efficacy in representing the variability in annual average PM2.5 between surface sites, and so these techniques generally will not be able to iden- tify areas of persistent high air pollution (that is, “hot spots”) within a city. Site-specific annual average values were not represented well in Delhi, lim- iting the usefulness of satellites to study variations in chronic health effects within the city (table 5.3). In addition, the lack of satellite AOD measurements TABLE 5.1  Statistics for the Delhi, India, satellite ground-level PM2.5 products tested in this work MODIS TERRA MODIS AQUA VIIRS S-NPP STAT. CTM STAT. CTM STAT. CTM MB (micrograms per cubic meter) 0.045 −0.822 0.113 −29.275 −0.014 −58.102 MNB (%) 30.8 21.7 24.4 −7.5 32.3 −27.3 MNGE (%) 51.5 54.5 44.1 46.8 54.2 52.3 Correlation coefficient (R) 0.18 0.17 0.09 0.04 0.11 0.10 Source: Original calculations for this publication. Note: CTM = chemical transport model–based method; MB = mean bias; MNB = mean normalized bias; MNGE = mean normalized gross error; MODIS = Moderate-Resolution Imaging Spectroradiometer; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; S-NPP = Suomi National Polar-Orbiting Partnership; STAT. = statistical method; VIIRS = Visible Infrared Imaging Radiometer Suite. TABLE 5.2  Statistics for the Ulaanbaatar, Mongolia, satellite ground-level PM2.5 products tested in this work MODIS TERRA MODIS AQUA STAT. CTM STAT. CTM MB (micrograms per cubic meter) 0.082 −31.174 0.443 −33.139 MNB (%) 54.0 −72.2 54.8 −73.1 MNGE (%) 78.5 78.2 78.3 78.8 Correlation coefficient (R) 0.17 0.00 0.30 0.01 Source: Original calculations for this publication. Note: CTM = chemical transport model–based method; MB = mean bias; MNB = mean normalized bias; MNGE = mean normalized gross error; MODIS = Moderate-Resolution Imaging Spectroradiometer; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; STAT. = statistical method. Using Satellite Data for Daily, City-Scale PM2.5 Monitoring | 31 TABLE 5.3  Annual average PM2.5 surface concentrations for Delhi, India, 2017 Micrograms per cubic meter GROUND AVERAGE CTM-BASED AVERAGE MERRA-2 AVERAGE Anand Vihar 153.5 125.8 62.7 Delhi Technological University 123.9 113.5 58.5 Institute of Human Behavior and Allied Sciences 98.9 124.9 63.0 Income Tax Office 119.0 116.9 58.0 Mandir Marg 97.3 111.6 57.4 Netaji Subhas Institute of Technology Dwarka 140.2 112.9 62.4 Punjabi Bagh 103.6 117.0 55.8 Ramakrishna Puram 143.2 122.5 67.9 Shadipur 128.7 117.1 58.1 US Diplomatic Post 113.8 116.3 62.0 Urban average 122.2 117.9 60.6 Source: Original calculations for this publication. Note: As determined by the ground-level-monitoring (GLM) data, the chemical transport model (CTM)–based satellite approach using MODIS (Moderate-Resolution Imaging Spectroradiometer) Terra aerosol optical depth (AOD), and the original MERRA-2 (Modern Era Retrospective-analysis for Research and Applications) output. Note these averages include only those days with both a valid AOD and a valid GLM daily average PM2.5 (particulate matter with an aerodynamic diameter less than or equal to 2.5 microns) value. in several months means that the satellites did a poor job of representing the observed seasonal cycle of PM2.5 in Delhi. Results of using alternative AOD products to estimate PM2.5 concentrations Using VIIRS instead of MODIS did not appreciably change the ability of the sat- ellite approaches studied here to simulate the measured ground-level PM2.5 con- centrations in Delhi (figure 5.1). However, for cities where the morning MODIS Terra product performed better than the afternoon MODIS Aqua product, using the afternoon VIIRS product will also likely result in poorer predictions than using the lower-resolution MODIS Terra product, but this was tested directly only for Delhi. The SEVIRI AOD product does provide a higher horizontal resolution than the MODIS product, but it appears to have lower coverage for coastal cities such as Accra, because of an AOD data quality filter that removes data near coastlines (to avoid mixed water and land surfaces; see figure 5.2) Thus, the SEVIRI AOD product is not suitable for use in coastal cities. Using a hybrid technique to bias-correct CTM-based estimates Land-use variables (percentage of urban land cover, elevation, and population density) had little ability to predict the bias in the CTM-based ground-level PM2.5 estimates in Delhi or Ulaanbaatar (R < 0.01). This is likely because there are 10 or fewer GLM sites in each city within about 30 kilometers of each other, as opposed to the 1,440 sites used in the GWR bias correction of van Donkelaar, Martin, Spurr, and Burnett (2015) over the United States. Further analysis suggests that the limitations of the satellite techniques in Delhi and Ulaanbaatar cannot be easily addressed using post hoc bias correc- tions. Applying a simple linear bias-correction model to the CTM-based PM2.5 32 | Getting Down to Earth FIGURE 5.1 Comparison of the chemical-transport-model–based estimates for daily average ground-level PM2.5 concentrations in micrograms per cubic meter over Delhi, India, using VIIRS and MODIS Terra, November 1, 2017 a. Using VIIRS Delhi MOD VIIRS Data November 1, 2017 Delhi VIIRS CTM PM2.5 79.241139–110.780485 110.780486–142.319832 142.319833–173.859178 173.859179–205.398525 205.398526–236.937871 NAN b. Using MODIS Terra Delhi MOD CTM Data November 1, 2017 Delhi MOD CTM PM2.5 79.241139–110.780485 110.780486–142.319832 142.319833–173.859178 173.859179–205.398525 205.398526–236.937871 NAN Source: World Bank, produced with Esri ArcGIS. Note: CTM = chemical transport model-based method; MOD = MODIS Terra; MODIS = Moderate-Resolution Imaging Spectroradiometer; NAN = not a number; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; VIIRS = Visible Infrared Imaging Radiometer Suite. Using Satellite Data for Daily, City-Scale PM2.5 Monitoring | 33 FIGURE 5.2 Comparison of the chemical-transport-model–based estimates for daily average ground-level PM2.5 concentrations in micrograms per cubic meter over Accra, Ghana, using SEVIRI and MODIS Aqua, December 26, 2017 a. Using SEVIRI Accra SEVIRI CTM Data December 26, 2017 Cities SEVIRI CTM PM2.5 11.674649–34.925684 34.925685–58.176718 58.176719–81.437752 81.437753–104.678786 104.678787–127.929820 NAN b. Using MODIS Aqua Accra MYD CTM Data December 26, 2017 Cities MYD CTM PM2.5 11.674649–34.925684 34.925685–58.176718 58.176719–81.437752 81.437753–104.678786 104.678787–127.929820 NAN Source: World Bank, produced with Esri ArcGIS. Note: CTM = chemical transport model-based method; MODIS = Moderate-Resolution Imaging Spectroradiometer; MYD = MODIS Aqua; NAN = not a number; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; SEVIRI = Spinning Enhanced Visible and Infrared Imager. 34 | Getting Down to Earth estimates did improve the mean bias in Delhi and Ulaanbaatar but did not greatly improve the ability of the land-use regression approach to predict the daily and site-to-site variation of PM2.5 within an urban area. In Delhi, the Stage 1 CTM- based estimate was not a significant predictor of the daily average of PM2.5 in the bias-correction model. In fact, a statistical model that included only the MERRA-2 speciated aerosol variable did as good a job at p ­ redicting the daily average of PM2.5 in Delhi as a model using those MERRA-2 variables and the satellite Stage 1 CTM-based estimate. Can satellites reduce the number of required monitoring sites in a GLM network? Another goal of this work was to determine if satellite data could be combined with GLM data in such a way as to reduce the number of GLM sites required to quantify exposure in an urban population (see the section “Co-kriging and krig- ing with satellite and OpenAQ data” in appendix C). To test this, ordinary kriging was done for each day using the monitor station daily averages of PM2.5 in Delhi and Ulaanbaatar excluding the US Diplomatic Post. The concentration at the US Diplomatic Post for each day was then predicted with ordinary kriging. This tested how well the other GLM sites alone could characterize the PM2.5 concen- trations at an unmeasured site. The CTM-based satellite PM2.5 estimates were then combined with the GLM data (minus the US Diplomatic Post) via co-kriging, and we again estimated the concentrations at the US Diplomatic Post to test if the satellite data could compensate for the loss of one GLM station. The Delhi results suggest that adding in the satellite data would not help to reduce the number of GLM monitoring sites there, since the root-mean-square error (a measure of prediction accuracy calculated as the square root of the aver- age of the squared differences between the predicted and observed PM2.5 con- centrations at the US Diplomatic Post) increased above that for ordinary kriging when the CTM-based satellite data are added via co-kriging. This finding sug- gests that including the CTM-based satellite data resulted in poorer interpola- tion predictions than using GLM data alone. However, the Ulaanbaatar results suggest that the Diplomatic Post could be eliminated by adding satellite data to the rest of the GLM network in this city, but again only for the eight months each a year that satellites actually provide data. Overall, these results suggest that the use of satellite data may allow LMICs to reduce the number of GLM sites needed in an area, but that this ability varies from city to city and could result in large errors in annual average estimates for cities with a seasonal (Ulaanbaatar) or persistent (Lima) lack of satellite AOD coverage. IDENTIFYING THE TYPE OF CITIES WHERE SATELLITES WORK BEST We applied the CTM-based and statistical approaches discussed above to a set of nine cities in LMICs that represent a diversity of regimes (coastal versus inland, high versus low altitude, tropical versus temperate, and so forth), as shown in table 5.4. These cities also have GLM data needed for validation, through either US Diplomatic Posts, local monitoring networks, or both, although the amount of available GLM data varies significantly between cities. Using Satellite Data for Daily, City-Scale PM2.5 Monitoring | 35 TABLE 5.4  Cities included in this work CITY COUNTRY LOCATION INCOME GROUP NUMBER OF GLM SITES Accra Ghana Coastal, low altitude Lower middle 4 Addis Ababa Ethiopia Inland, high altitude Low 2a Dakar Senegal Coastal, low altitude Lower middle 4b Delhi India Inland, low altitude Lower middle 10 Hanoi Vietnam Inland, low altitude Lower middle 1a Kampala Uganda Inland but near lake, high Low 1a altitude Kathmandu Nepal Inland, high altitude Lower middle 4 Lima Peru Coastal, low altitude Upper middle 10 Ulaanbaatar Mongolia Inland, high altitude Lower middle 8 Source: World Bank. Note: GLM = ground-level monitoring. a. Only US Diplomatic Post data are available. b. Dakar data are available only as an average of the four reporting sites. Income groups correspond to World Bank Country Classifications by Income Level: 2021–22. Income classifications are affected by several factors and are subject to change in time. For the nine cities, the satellite-based estimates of ground-level PM2.5 concen- trations used the MODIS Combined (Deep Blue plus Dark Target) Product, 2 since this provides the most coverage for cities, is easy to access, has been exten- sively validated, and has a long historical record. The tests above explored sev- eral other publicly available satellite sources of AOD but did not find any that provided significant advantages over this product. Both the CTM-based and sta- tistical approaches discussed in the first main section of this chapter were applied to each city in table 5.4, and the mean bias and error in the satellite esti- mates for each city were calculated. The evaluation results are discussed in detail in appendix C and summarized below. Evaluating satellite methods in multiple cities in LMICs Table 5.5 and figures 5.3 to 5.5 show the statistics for the evaluations of the sta- tistical and CTM-based methods for whichever satellite (MODIS Terra or MODIS Aqua) gave the best performance for a given city. The statistical methods generally performed better than the CTM-based estimates, but this may be due to the low numbers of points available for model training. The cities with the highest R also had the lowest number of evaluation points—for example, Kathmandu had an R of 66 percent, mainly because there were only 50 days in the two-year evaluation period (2016–17) that had both valid daily PM2.5 averages from the GLM data and valid AOD retrievals from the satellite. This is an exam- statistical model “overfitting” where, due to the small number of data ple of ­ points relative to the degrees of freedom of the model, the model corresponds too closely to the data with which the model was created. Consequently, the model may fail to fit additional data or predict future observations reliably. Thus, the statistical approach in these cases is likely to have a much poorer perfor- mance on future AOD data. The satellite-derived PM2.5 estimates from the CTM-based method tended to have low correlation (R < 0.5) with the observed daily average PM2.5 concentra- tions. The highest R value obtained using a CTM-based method was in Delhi, with a value of 44 percent, but other cities had much lower R values. This low 36 | Getting Down to Earth TABLE 5.5  Correlation coefficient, mean normalized bias, and mean normalized gross error for each city MOST- MEAN MEAN CORRELATED CORRELATION NORMALIZED NORMALIZED CITY COUNTRY LOCATION SATELLITE COEFFICIENT (R) BIAS GROSS ERROR Accra Ghana Coastal, low altitude Aqua —b — — 0.40 −23% 56% Addis Ababa Ethiopia Inland, high altitude Terra 0.63c 7% 24% 0.40 36% 65% Dakar Senegal Coastal, low altitude Aqua —b — — 0.52 68% 85% Delhi India Inland, low altitude Terra 0.46 37% 60% 0.44 27% 60% Hanoi Vietnam Inland, low altitude Terra 0.63 21% 43% 0.30 22% 53% Kampala Uganda Inland near lake, high Aqua 0.51c 6.5% 21% altitude 0.14 −70% 70% Kathmandu Nepal Inland, high altitude Aqua 0.66c 8% 24% 0.13 −19% 48% Lima Peru Coastal, low altitude — a — — — Ulaanbaatar Mongolia Inland, high altitude Aqua 0.44 54% 77% 0.15 −70% 78% Source: Original calculations for this publication. Note: The terms correlation coefficient, mean normalized bias, and mean normalized gross error are described in the glossary of technical terms. Only results for the satellite with the highest correlation with the GLM data are presented. Statistical method results are shown in the top row of each city entry in regular type. CTM-based estimates are in the bottom row in italics. — = not available. AOD = aerosol optical depth; CTM = chemical transport model; GLM = ground-level monitoring. a. Not available; the satellite AOD coverage was too poor to perform an evaluation. b. Not enough GLM data were available to attempt the statistical method. c. High R for the statistical method may be an artifact of the low number of points with both valid AOD and GLM data. value of the correlation means that the CTM-based method was not able to pre- dict more than 44 percent of the variation in daily average PM2.5 concentrations between sites in a city or between days. This implies that current satellite meth- ods may not be able to estimate exposures for studies of the acute health effects of PM2.5 (which require knowledge of day-to-day variations) or of the changes in chronic exposure to PM2.5 within a city (which requires knowledge of site-to-site variation). Satellite AOD coverage for some cities was too sparse to allow for its use in PM2.5 monitoring. Consistent with the results in this chapter, Lima had little valid AOD data coverage for the city. Consequently, satellite methods could not be used for this location. Ulaanbaatar had seasonal gaps in coverage, with no satellite AOD available in the winter months when PM2.5 concentrations are the highest. The uncertainty in all satellite-based estimates of the daily average PM2.5 con- centration at a given location in a city (as estimated with the mean normalized gross error) tended to be very large (21–77 percent for the statistical methods, and 48–85 percent for the CTM-based methods). This is consistent with the results of van Donkelaar, Martin, Brauer, and Boys (2015), who found an error in monthly average PM2.5 concentrations over East Asia of about 50 percent. This error is much larger than usually allowed for “equivalent methods” for Using Satellite Data for Daily, City-Scale PM2.5 Monitoring | 37 FIGURE 5.3 Correlation coefficient R for the statistical and chemical-transport-model–based methods for the different low- and middle-income country cities 0.7 0.6 0.5 0.4 R 0.3 0.2 0.1 0 na , op a, di , na i, da , pa , ga , lia , ha a In lhi an ala Ne ndu ne kar go tar et no G ccr hi b ia a m l l e Et Aba on a Ug mp Se Da Vi Ha D M nba a A m Ka a th aa di Ka Ul Ad Statistical method Contaminant transport model-based method Source: Original calculations for this publication. FIGURE 5.4 Mean normalized bias for the statistical and chemical-transport-model–based methods for the different low- and middle-income country cities 80 60 40 Mean normalized bias 20 0 –20 –40 –60 –80 na , op a, ga , di , na i, da , pa , lia , ha a ne kar In lhi an ala Ne ndu go tar et no G ccr hi b ia l a m l e Et Aba on a Ug mp Se Da Vi Ha D M nba a A m Ka a th aa di Ka Ul Ad Statistical method Contaminant transport model-based method Source: Original calculations for this publication. 38 | Getting Down to Earth FIGURE 5.5 Mean normalized gross error for the statistical and chemical-transport-model–based methods for the different low- and middle-income country cities 90 80 70 Mean normalized gross error 60 50 40 30 20 10 0 na , op a, ga , di , na i, da , pa , lia , ha a ne kar In lhi an ala Ne ndu go tar et no G ccr hi b ia l a m l e Et Aba on a Ug mp Se Da Vi Ha D M nba a A m Ka s th aa di Ka Ul Ad Statistical method Contaminant transport model-based method Source: Original calculations for this publication. measuring PM2.5 by the US Environmental Protection Agency (chapter 4) and ­ ethods are not a replacement for GLM data in thus suggests that satellite m LMICs. Patterns in satellite-method performance A goal of the research presented in this report was to determine if there were patterns in the performance of satellite methods in predicting ground-level PM2.5 concentrations based on factors such as the distance of the city from a coast, the altitude of the city, and so forth. The results in table 5.5 and appendix C do suggest some patterns may be present in the performance of satellite meth- ods in these cities. CTM-based methods tend to underestimate PM2.5 concentrations in cities at high altitudes or in mountain valleys such as Ulaanbaatar (altitude 1.7 kilome- ters), Kathmandu (altitude 1.4 kilometers), and Kampala (altitude 1.2 kilome- ters). This is likely due to these cities being in small mountain valleys surrounded by rural land, and the aerosol profile from the CTM is more appropriate for rural land, which leads to an underestimate of PM2.5 in the city (see above). CTM-based methods may overestimate PM2.5 concentrations in cities where desert dust is a large fraction of the PM2.5. Although Addis Ababa is a high-­ altitude city (altitude 2.4 kilometers), the CTM-based method tends to overesti- mate in this city. This may be due to this high-altitude city being near the Sahara Using Satellite Data for Daily, City-Scale PM2.5 Monitoring | 39 Desert, and the dust from this area could be leading to the CTM overestimate in this city. Dust makes up more than half of the CTM estimate of PM2.5 in Addis Ababa and Dakar, and both show significant overestimates when the CTM-based method is used. However, Accra also has a high percentage of PM2.5 from dust, but for this coastal city, the CTM-based method underestimates PM2.5. Coastal cities (including cities near large lakes) appear to have poor satellite coverage, due to either persistent clouds or the mixture of land and water sur- faces in the satellite retrieval of AOD. Accra, Dakar, and Lima all had too few matching GLM and satellite AOD points to perform the statistical method, and Kampala (near Lake Victoria) had very few points as well. Satellite-based methods appear to work best for low-altitude, inland cities such as Delhi and Hanoi, but still have significant errors (43–60 percent) in pre- dictions of daily average PM2.5 concentrations at sites within the city. The performance of the satellite approaches in estimating ground-level PM2.5 did not appear to be affected by the type of instrument used to measure GLM data. Some cities had data only from GLM instruments at US Diplomatic Posts, which all used the beta attenuation mass monitor (BAMM) method. Accra used GLM instruments based on gravimetric methods, but the performance of satel- lite methods in this city was not clearly better or worse than the BAMM method. Ulaanbaatar and Delhi also included the Grimm environmental dust monitor and gravimetric filters, respectively, but this did not have a noticeable impact on the performance of satellite methods. Recommended next steps for each city Table 5.6 summarizes the lessons learned for each city and the recommended next steps to improve the monitoring of PM2.5 in each to better understand and reduce the health impacts of PM2.5. The GLM networks in each of these cities should be expanded to better char- acterize the variations of PM2.5 in the cities and surrounding suburban and rural areas, which will also allow for the derivation of better statistical models for translating satellite AOD to ground-level PM2.5 estimates. Only Delhi and Ulaanbaatar had sufficient GLM data to derive robust statistical models, although in Ulaanbaatar this was possible for only eight months of the year due to snow cover preventing the measurement of AOD by the satellites. Lima and Ulaanbaatar have sufficiently poor AOD coverage that preclude the use of satellite observa- tions for annual PM2.5 monitoring, although satellites may still be useful in Ulaanbaatar for the eight months each year when AOD measurements can be taken. Similarly, the problems with poor AOD coverage near coastal regions make it unlikely that satellite methods would be successful in Accra, Dakar, or Kampala. For inland cities without persistent seasonal snow cover, satellite observa- tions may provide some benefits to GLM monitoring, but the current CTM- based approaches tend to underestimate significantly for the high-altitude cities. Thus, for inland cities without persistent seasonal snow (Addis Ababa, Delhi, Hanoi, and Kathmandu), a SPARTAN network site (see the section “Chemical transport model–based approaches” in appendix A) should be added to the local GLM network to directly measure the relationship between PM2.5 and AOD using ground-based instruments, which will allow this relationship to be used instead of the CTM-based relationships in these cities. 40 | Getting Down to Earth TABLE 5.6  Lessons learned and recommended next steps to improve monitoring of PM2.5 in each city to better u­ nderstand and reduce the health impacts of PM2.5 CITY COUNTRY LOCATION LESSONS LEARNED RECOMMENDED NEXT STEPS Accra Ghana Coastal, • Insufficient GLM and AOD data • Expand current GLM network and improve QA/QC low for statistical approaches to be practices, including public access to instrument metadata altitude used Addis Ethiopia Inland, • Low amount of GLM and AOD • Expand current GLM network and improve QA/QC Ababa high data may be resulting in an practices, including public access to instrument metadata altitude over-fitted statistical approach • Add SPARTAN network site • CTM-based approach biased • After the GLM network is expanded and multiple years high, apparently due to high dust of data are available, test combining the GLM data with impacts the satellite data via co-kriging to see if this approach is useful for this city Dakar Senegal Coastal, • Insufficient GLM and AOD data • Expand current GLM network and improve QA/QC low for statistical approaches to be practices, including public access to instrument metadata altitude used • Satellite approaches unlikely to be useful even with • CTM-based approach biased very increased GLM data due to coastal location high Delhi India Inland, • Satellites can provide a citywide • Add rural sites to current GLM network to allow use of low annual average estimate, but do hybrid satellite approaches and land-use regression altitude not represent day-to-day or • Add SPARTAN network site site-to-site variability • Statistical and CTM-based approaches perform similarly Hanoi Vietnam Inland, • Statistical and CTM-based • Expand current GLM network and improve QA/QC low approaches perform similarly practices, including public access to instrument metadata altitude • Add SPARTAN network site • After the GLM network is expanded and multiple years of data are available, test combining the GLM data with the satellite data via co-kriging to see if this approach is useful for this city Kampala Uganda Inland • Low amount of GLM and AOD • Expand current GLM network and improve QA/QC near lake, data may be resulting in an practices, including public access to instrument metadata high over-fitted statistical approach • Satellite approaches unlikely to be useful even with altitude • CTM approach biased very low increased GLM data due to coastal location Kathmandu Nepal Inland, • Low amounts of GLM and AOD • Expand current GLM network and improve QA/QC high data may be resulting in an practices, including public access to instrument metadata altitude over-fitted statistical approach • Add SPARTAN network site • After the GLM network is expanded and multiple years of data are available, test combining the GLM data with the satellite data via co-kriging to see if this approach is useful for this city Lima Peru Coastal, • Satellite AOD coverage is too • Expand current GLM network and improve QA/QC low poor, and thus satellite approach- practices, including public access to instrument metadata altitude es cannot be used for PM2.5 moni- • Investigate inconsistency between local network and US toring in this city Diplomatic Post data Ulaanbaatar Mongolia Inland, • Poor satellite coverage in winter • Expand current GLM network and improve QA/QC high means satellite approaches practices, including public access to instrument altitude cannot be used for annual PM2.5 metadata monitoring in this city but may • Add SPARTAN network site be useful for the eight months a • Combine satellite data during the eight-month year when satellite AOD snow-free period with GLM data via co-kriging to measurements exist provide more-accurate neighborhood-scale exposure estimates Source: World Bank. Note: AOD = aerosol optical depth; CTM = chemical transport model; GLM = ground-level monitoring; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; QA/QC = quality assurance/quality control; SPARTAN = Surface Particulate Matter Network. Using Satellite Data for Daily, City-Scale PM2.5 Monitoring | 41 NOTES 1. Applying the GWR correction does improve the GBD estimate for Ulaanbaatar, because this correction uses GLM data in other urban areas to increase the satellite-derived PM2.5 estimate for this city. However, the corrected value still underestimates the true annual average by a factor of two. 2. Note that here the “combined” product refers to the combination of the Dark Target and Deep Blue algorithms for a given version of the MODIS instrument (either MODIS Terra or MODIS Aqua). Thus, there are separate “combined” products from MODIS Terra and MODIS Aqua. REFERENCES Chatfield, R., M. Sorek-Hamer, A. Lyapustin, and Y. Liu. 2017. “Daily Kilometer-Scale MODIS Satellite Maps of PM2.5 Describe Wintertime Episodes.” Paper 263193 presented at the “Air and Waste Management Association 2017 Annual Conference and Exhibition,” Pittsburgh, PA, June 5–9. Sorek-Hamer, M., A. W. Strawa, R. B. Chatfield, R. Esswein, A. Cohen, and D. M. Broday. 2013. “Improved Retrieval of PM2.5 from Satellite Data Products Using Non-Linear Methods.” Environmental Pollution 182: 417–23. van Donkelaar, A., R. V. Martin, M. Brauer, and B. L. Boys. 2015. “Use of Satellite Observations for Long-Term Exposure Assessment of Global Concentrations of Fine Particulate Matter.” Environmental Health Perspectives 123 (2): 135–43. van Donkelaar, A., R. V. Martin, M. Brauer, R. Kahn, R. Levy, C. Verduzco, and P. J. Villeneuve. 2010. Global Estimates of Ambient Fine Particulate Matter Concentrations from Satellite- Based Aerosol Optical Depth: Development and Application.” Environmental Health Perspectives 118 (6): 847–55. van Donkelaar, A., R. V. Martin, R. J. Spurr, and R. T. Burnett. 2015. “High-Resolution Satellite- Derived PM2.5 from Optimal Estimation and Geographically Weighted Regression over North America.” Environmental Science & Technology 49 (17): 10482–91. 6 Conclusions and Recommendations To accomplish the objectives of this report, a literature review of the approaches used to combine satellite observations with ground-level monitoring (GLM) measurements of particulate matter with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5) was conducted. The GLM data provided by US Diplomatic Posts and other sources contained in the OpenAQ database were investigated to determine the level of quality assurance and likely uncertainties in the GLM data. Several methods for using satellite aerosol optical depth (AOD) from the publicly available sources to estimate ground-level daily average PM2.5 concentrations within cities were then tested in nine cities in low- and middle-income countries (LMICs) to look for patterns in the performance of the satellite methods with the altitude and location of the city. Based on the results of this work, GLM and satellite data are best thought of as complements to each other. Many GLM networks could be improved by con- sidering satellite data, but all approaches using satellite data improve as the number of high-quality GLM sites is increased. Thus, it is important that LMICs continue to support the establishment of GLM networks to measure air pollut- ants that cause mortality, notably fine particulate matter (PM2.5), in regions such as Sub-Saharan Africa and other regions with many LMICs before they can take full advantage of satellite data. The GLM data must have adequate quality assur- ance and quality control and follow standard operating procedures to ensure the data are of sufficient quality to be used to estimate PM2.5 exposures for health studies and to be combined with the satellite estimates. Satellite data may be useful for estimating air quality for countries or large areas based on average estimates. However, for the purpose of protecting human health, LMICs need to prioritize establishment or strengthening of GLM net- works where they are lacking or weak. These networks can measure air quality at the level where human activity is typically carried out and where people are exposed to air pollutants, notably PM2.5, that are harmful to health and can cause death. The next section summarizes the conclusions of this report, and the final section provides recommendations for the use of satellite data and the improve- ment of GLM in LMICs.  43 44 | Getting Down to Earth CONCLUSIONS Literature review The literature review covered the approaches used (and their strengths and lim- itations) in each of the three key steps entailed in combining satellite observa- tions with GLM measurements of PM2.5: (1) the retrieval of AOD from satellite observations of reflected solar radiation, (2) the translation of the vertically inte- grated AOD observation into an estimate of the ground-level PM2.5 concentra- tion, including any bias corrections, and (3) simultaneously interpolating these satellite-derived PM2.5 estimates and the available GLM data. The major findings of the review follow: • Although many satellite AOD products exist, only a small number of daily products are currently freely available shortly after the observation is made at a horizontal resolution of 10 kilometers or finer. ▪▪ Some satellite AOD products that have been used in previous studies in the literature (for example, MAIAC [Multi-Angle Implementation of Atmospheric Correction] AOD) were not operational products as of the start of this work and thus are not available for all regions in near real- time. Other satellite data (for example, the Himawari-8 and -9 data) must be purchased by the user and thus may be outside the budget for most LMICs. ▪▪ Polar-orbiting satellite instruments (for example, MODIS [Moderate- Resolution Imaging Spectroradiometer] and VIIRS [Visible Infrared Imaging Radiometer Suite]) have the advantage of daily global coverage but tend to have relatively coarse horizontal resolution (6 to 10 kilometers). Geostationary satellites (for example, GOES and MSG-SEVIRI) make several observations each day at a higher horizontal resolution (2 to 4  kilometers) but observe only one side of the globe, and their AOD retrievals tend to exclude coastal areas. ▪▪ Accessing the VIIRS and GOES products is difficult, since the download process cannot be easily automated. ▪▪ The MODIS combined Deep Blue and Dark Target AOD retrieval is free, is available in near real-time, is easy to access, has been extensively validated, and has good coverage over urban and coastal areas. • Several studies have successfully converted satellite AOD to monthly average and annual average ground-level PM2.5 estimates using statistical techniques, chemical transport model (CTM)–based approaches, or hybrid approaches with typical errors of 30–50 percent, but very few have tried to predict daily PM2.5 concentrations or their variation within a city. ▪▪ CTM-based approaches are easy to implement globally and do not require any GLM data. Consequently, CTM-based approaches can be imple- mented for all LMICs. However, these approaches can have significant biases due to errors in the CTM’s simulation of the vertical profile of the aerosols, as well as in the mismatch between the horizontal resolution of the CTM and the distribution of population within a city. ▪▪ Statistical approaches are locally unbiased (since these methods train a statistical model to minimize the mean bias) and do not depend on an initial estimate of the aerosol vertical profile. However, statistical approaches require a multiyear GLM data record with concurrent satel- lite observations to train the statistical model. Only a few LMICs Conclusions and Recommendations | 45 (for example, India or Mongolia) have sufficient GLM data to use these approaches. ▪▪ Hybrid approaches that correct CTM-based approaches with statistical models trained with GLM data tend to give the best performance. However, they also require a multiyear GLM data record with concurrent satellite observations over many different land-use types (for example, urban and rural sites or coastal and inland sites) before the statistical model can be trained. • Co-kriging and land-use regression (LUR) can be used to combine satellite-derived PM2.5 estimates with GLM data to provide estimates of PM2.5 concentrations at a neighborhood scale across urban areas. ▪▪ Co-kriging can be used to estimate variations in daily average PM2.5 con- centrations, whereas LUR can be used for annual or longer-term averages. ▪▪ Co-kriging and LUR tend to work better as the satellite horizontal resolu- tion is increased and as the GLM network covers a wider variety of sites (for example, urban and rural). Quality assurance of ground-level monitoring data in LMICs All the cities evaluated in this project would benefit from expanding their GLM networks and enhancing their quality-assurance procedures, and this should be a key next step in monitoring PM2.5 in these cities. Missing or unreliable GLM data prevent the use of statistical or hybrid approaches to incorporating satellites observations into PM2.5 monitoring. When GLM data are absent, only the CTM-based approaches can be applied. When GLM data are present but of poor quality (that is, unreliable), a bias correction can be calculated for the CTM-based approaches, but there is no assurance that the “corrected” satellite value is more accurate than the original estimate. Finally, poor maintenance of GLM instruments can lead to large data gaps, which either prevent the use of statistical methods or bias the statistical models trained with the data. Most PM2.5 instruments in use around the world have been approved by the US Environmental Protection Agency (EPA) as federal equivalent methods (FEMs) for PM2.5 characterization and thus can provide reliable measurements of ground-level PM2.5 if they are properly maintained and calibrated. To qualify as a FEM, the PM2.5 measurement technique must demonstrate uncertainties less than 10 percent relative to the federal reference method for 24-hour average PM2.5 concentrations. However, the actual accuracy and precision of the data will depend on whether proper maintenance and procedures are followed, and if the instruments are monitored for errors. US Diplomatic Posts are equipped to measure PM2.5 with Met One beta attenuation monitor (BAM) systems and abide by the current EPA guidelines for PM2.5 monitoring. Thus, these posts are sources of high-quality GLM data for the locations where they operate. However, these data may not be represen- tative of average air quality in the city where these posts are located. Whereas the Delhi, India, local government GLM network appears to be consistent with the US Diplomatic Post data, the local network in Lima, Peru, is inconsistent. More details on the maintenance and quality-assurance procedures in the GLM network in Lima would be needed to determine the causes of the observed differences. 46 | Getting Down to Earth Application of satellite approaches to cities in LMICs The results of the application of satellite approaches to cities in LMICs suggest that satellites cannot be a replacement for a high-quality GLM network. This is because several limitations to the use of satellite AOD are found in the cities tested: • Coastal cities (Accra, Dakar, Kampala, and Lima) and cities with persistent snow cover (Ulaanbaatar) have significant limitations in satellite AOD coverage. ▪▪ Satellite coverage was poorest in the coastal city of Lima, and thus an evaluation of CTM-based approaches with GLM network data was not possible. The other coastal cities (including cities near large lakes) examined in this study had poor satellite coverage, because of either persistent clouds or the mixture of land and water surfaces in the satellite retrieval of AOD. ▪▪ In Ulaanbaatar, no satellite AOD products are available for the high PM2.5 winter months of December to mid-March. • The ground-level PM2.5 estimates from the CTM-based and statistical approaches for converting satellite AOD generally had a low correlation with the true daily average PM2.5 values within a city over the year. • The uncertainty (that is, the average error that is expected for a single esti- mate) in all satellite-based estimates of the daily average PM2.5 concentration at a given location in a city tended to be very large (21–77 percent for the sta- tistical methods, and 48–85 percent for the CTM-based methods), and thus these methods cannot be considered “equivalent methods” when compared with United States– or European Union–approved GLM techniques.1 • CTM-based satellite approaches tend to underestimate PM2.5 in high-altitude cities. For example, annual average PM2.5 concentrations in Ulaanbaatar were underestimated by a factor of 10. • Estimating variations in annual average PM2.5 within a city (to estimate chronic health effects on a neighborhood scale) is unlikely to be possible with satellite AOD data using the approaches tested here. This will require GLM- network data and LUR, with the satellite AOD product or the satellite-based PM2.5 estimate used as a variable in the LUR. • Both the CTM-based and statistical approaches tested here showed little ability to represent the day-to-day variability in PM2.5 concentrations, with average absolute errors of ±50 percent for the best approaches within each city. Thus, studies of acute health effects will likely require GLM data. However, a few cases and conditions are found where the use of satellite data may be appropriate as a complement (but not a replacement) for GLM networks. For example, satellite-based methods appear to work best for low-altitude, inland cities such as Delhi and Hanoi but still have significant errors (43–60 percent) in predictions of daily average PM2.5 concentrations at sites within these cities. In addition, the results suggest that under some conditions, adding satellite data to GLM network data via co-kriging may reduce the num- ber of GLM sites needed to characterize PM2.5 concentrations and exposure within a city, although this ability varies from city to city and could lead to large errors in annual estimates for cities with a seasonal or persistent lack of satellite AOD coverage. Conclusions and Recommendations | 47 RECOMMENDATIONS Use of satellite data in PM2.5 monitoring Table 6.1 summarizes typical problems associated with the use of satellite data for PM2.5 monitoring in LMICs, the consequence of the problem, and the recom- mended action to address each problem. Inland cities without persistent seasonal snow cover may be able to benefit from the use of satellite observations in their PM2.5 monitoring, and so a key next step should be establishing a SPARTAN network site within these cities to directly measure the relationship between AOD and ground-level PM2.5 concen- trations, thus reducing the biases that come from using CTM-based estimates of this relationship. The research conducted for this report also suggests the following recom- mendations for the use of satellite observations to supplement GLM data in LMICs based on the typology proposed in chapter 1 (see also table 6.2): • For Type I countries, since no GLM data exist, the only possible approach to convert AOD to ground-level PM2.5 concentrations for these countries is a CTM-based approach. Thus, the raw CTM-based estimates of ground-level PM2.5 concentrations at the native resolution of the satellite AOD product should be used. However, the derived ground-level PM2.5 values should be TABLE 6.1  Summary of typical problems in using satellites in PM2.5 monitoring, the consequences, and the recommended actions to further the use of satellite data in PM2.5 monitoring to better understand and reduce the health impacts of PM2.5 in low- and middle-income countries PROBLEM CONSEQUENCE RECOMMENDATION Lack of GLM data Only CTM-based approaches can be Install additional GLM sites and follow used to monitor PM2.5 with satellites, rigorous quality-assurance and mainte- (Type I LMICs). and accuracy cannot be verified. nance procedures. GLM data exist but are of poor Only CTM-based approaches can be Establish and follow rigorous quality (Type II LMICs). used to monitor PM2.5 with satellites, quality-assurance and maintenance and accuracy cannot be verified. procedures for GLM data and report GLM metadata with actual measurements. City is surrounded by rural land within CTM-based aerosol profile is incorrect, Establish a SPARTAN network site to direct- typical resolution of global chemical biasing satellite results. ly measure AOD and PM2.5 relationship. transport models (that is, cities in mountain valleys). City has seasonal persistent snow or Satellite estimates will be unavailable Satellite approaches should not be used cloud cover. during key seasons, biasing annual for annual averages but may be used in averages. certain seasons. City is near a major body of water. Many AOD products will have poor Satellite coverage for city should be coverage, since they filter out coastal assessed on a case-by-case basis. regions. City has only urban GLM sites. Land-use-regression and hybrid Establish a few nearby rural GLM sites approaches are not useful because of to provide data on urban-rural variation the lack of data on suburban and rural of PM2.5. PM2.5 concentrations. Source: World Bank. Note: AOD = aerosol optical depth; CTM = chemical transport model; GLM = ground-level monitoring; LMIC = low- and middle-income country; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns, SPARTAN = Surface Particulate Matter Network. 48 | Getting Down to Earth TABLE 6.2  Recommendations for the use of satellite observations to supplement ground-level monitoring data in low- and middle-income countries COUNTRY TYPE RECOMMENDATIONS Type I: No air-quality data • The raw CTM-based estimates of ground-level PM2.5 concentrations at the native resolution of the satellite AOD product should be used. • The satellite estimates should be assigned a high uncertainty (at least 50 percent). Type II: Countries with some • Both the raw and bias-corrected CTM-based air-quality data but of variable estimates of ground-level PM2.5 concentrations quality without rigorous QA at the native resolution of the satellite AOD procedures product should be reported and stored. • The satellite estimates should be assigned a high uncertainty (at least 50 percent). • As quality-assurance procedures improve (Type III), the satellite bias corrections should be recal- culated. Type III: Countries that possess • Both statistical and bias-corrected CTM-based reliable air-quality information approaches for converting AOD to PM2.5 should but with poor spatial or tempo- be tested, and the most reliable approach for a ral coverage given city selected. • Test combining these satellite estimates with GLM data via LUR and co-kriging. Type IV: Good, reliable air-quality • Both statistical and bias-corrected CTM-based monitoring underway or being approaches for converting AOD to PM2.5 should established be tested, and the most reliable approach for a given city selected. • Test combining these satellite estimates with GLM data via LUR and co-kriging. Type V: Routine, long-term • Both statistical and bias-corrected CTM-based air-quality monitoring approaches for converting AOD to PM2.5 should be tested, and the most reliable approach for a given city selected. • Test combining these satellite estimates with GLM data via LUR and co-kriging. Source: World Bank. Note: AOD = aerosol optical depth; CTM = chemical transport model; GLM = ground-level monitoring; LUR = land-use regression; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; QA = quality assurance. assigned a high uncertainty that reflects not only the uncertainty in the AOD but also the estimated uncertainty in the CTM-derived AOD to PM2.5 relation- ship (for example, about 50 percent based on van Donkelaar and others 2015 or the 22–85 percent errors found by the research conducted for this report). • For Type II countries, the small amount of GLM data (with variable quality) that exists will provide at least some ability to derive a bias estimate for the raw CTM-based estimates of ground-level PM2.5 concentrations discussed above. However, without improvement in the quality-assurance procedures in these countries, it will not be possible to be certain that the bias-­corrected estimate is truly more accurate than the raw estimate. Thus, it is advisable to report and store both values. This will allow for reprocessing of the sat- ellite estimates when more-rigorous quality-assurance procedures are developed. Conclusions and Recommendations | 49 • The GLM data in Type III countries potentially allow both statistical and bias-corrected CTM-based approaches for converting AOD to PM2.5 to be considered. CTM-based approaches likely would be the best approach for these countries, but statistical approaches should also be tested before mak- ing a final decision. LUR can be combined with either technique to provide a finer-scale estimate of annual average ground-level PM2.5 concentrations (at about 0.5-kilometer resolution), and co-kriging can be used to derive high-resolution daily estimates across the city. • Type IV and V countries will be able to take advantage of the same approaches outlined above for Type III countries to fill the gaps in the existing GLM net- work to help cover more of the country’s population. The more extensive GLM networks in these countries will allow for more-accurate estimates of the geographical and seasonal variation in the AOD to PM2.5 relationship. Thus, at this point, purely statistical approaches may begin to outperform CTM-based estimates, especially for regions where the model’s emission inventories are out-of-date or otherwise inaccurate. Presented below are this report’s recommendations on data sources and other methodological concerns for LMICs interested in using satellite data in their PM2.5 monitoring and exposure assessment: • Since the uncertainties of the different standard satellite AOD products are similar, all countries should use the satellite product that offers the best bal- ance of satellite AOD coverage and fine horizontal resolution for their coun- try. The highest horizontal resolution possible for satellite AOD will likely be between 2 and 4 kilometers (for geostationary observations) or 1 and 10 kilo- meters (for polar observations). Consequently, finer-scale predictions will be possible only for areas where extensive, high-quality GLM data exist. • The satellite AOD coverage, both spatially and seasonally, should be assessed for the country before any attempt to integrate satellite data is performed. The physical limits of satellite AOD coverage will make their use in PM2.5 monitoring difficult in many cities. ▪▪ In cities with persistent wintertime snow cover, such as Ulaanbaatar, or seasonal persistent cloud cover, such as in Lima, all approaches using sat- ellite AOD will likely not give accurate annual averages from current satellites. ▪▪ In coastal cities, the mix of water and land surfaces within a satellite foot- print and persistent clouds may also mean that using satellite AOD is not an option. • For countries that do not have the capacity to perform their own meteorolog- ical or chemical transport modeling, the freely available global model data sets provided by organizations in the United States for estimates of planetary boundary layer height, aerosol vertical profiles, and other parameters can be used in a CTM-based approach to estimating ground-level PM2.5 from satellite measurements. For long-term studies, the National Aeronautics and Space Administration’s MERRA-2 reanalysis can be used, because this will represent the best estimate of the historical atmospheric state. However, the reanalysis takes two months to produce, and thus for short-term forecasts and advisories, the National Centers for Environmental Prediction’s Global Forecast System (NCEP GFS) output (including the NEMS GFS Aerosol Component aerosol forecasts) can be used. The initial tests suggest that these data sets perform like the custom model data sets used in the GBD 2016 database for annual averages. 50 | Getting Down to Earth • Purchasing modeling data sets with higher resolution than the MERRA-2 and NCEP GFS products could address the errors seen in CTM-based methods for cities in appreciably different air-quality environments than their sur- roundings, such as cities in mountain valleys surrounded by rural land such as Ulaanbaatar. Quality-assurance procedures for ground-level data Based on the research conducted in preparing this report, it is important that LMICs continue to support the establishment of GLM networks in regions such as Sub-Saharan Africa and other regions where GLM networks are weak or non- existent. These GLM data must have adequate quality assurance and quality control and follow standard operating procedures to ensure the data are of suffi- cient quality. In addition, the results of the research conducted for this report suggest the following recommendations for LMICs and the US State Department for the quality-assurance procedures and the reporting of metadata for their GLM measurements: • PM2.5 measurements need to be provided along with information on the instrument/technique type, estimates of measurement uncertainties, relevant metadata, and operational history. It is important that each PM2.5 measure- ment not only identifies the type of instrument used for the measurement, but also tracks the instrument’s relevant history of use and calibration. This includes all “meta” data from instrument data files, including “housekeeping” files. Any recorded results from standard operating procedure and quality-assur- • ance procedures tests should be collected and disseminated along with the PM2.5 measurements to establish an instrument track record over time pro- viding a transparent, data-centric measure of instrument reliability and performance. • Currently, US embassies do not publish their full quality-assurance protocol metadata along with the PM2.5 levels, thereby reducing the use of these measurements as well-established and controlled reference measurements. Thus, it is recommended that US embassies publish their full quality-assurance protocol and relevant metadata along with the PM2.5 levels. • Near real-time GLM data should be assessed using automated scripts such as the OpenAQ quality wrapper developed in this project to identify potential problems quickly so they can be addressed. This will help focus the time of air-quality staff in LMICs, allowing for higher-quality data while keeping the required number of trained technicians small. NOTE 1. These percentages are the mean normalized gross error (average of the absolute value of the differences between the satellite estimate and the observations divided by the observa- tions) for satellite-based predictions of daily average PM2.5 concentrations at each GLM site within a city and thus are an estimate of the error that could be expected when using the satellite methods to predict daily average PM2.5 concentrations at a location within the city that does not include a GLM site. Conclusions and Recommendations | 51 REFERENCES GBD 2016 Risk Factors Collaborators. 2017. “Global, Regional, and National Comparative Risk Assessment of 84 Behavioural, Environmental and Occupational, and Metabolic Risks or Clusters of Risks, 1990–2016: A Systematic Analysis for the Global Burden of Disease Study 2016.” Lancet 390: 1345–422. van Donkelaar, A., R. V. Martin, M. Brauer, and B. L. Boys. 2015. “Use of Satellite Observations for Long-Term Exposure Assessment of Global Concentrations of Fine Particulate Matter.” Environmental Health Perspectives 123 (2): 135–43. APPENDIX A Literature Review Current approaches to converting total column aerosol optical depth (AOD) retrieved from satellites into estimates of ground-level PM2.5 (particulate matter with an aerodynamic diameter less than or equal to 2.5 microns) concentrations have various limitations. First, retrieved total column AODs from satellites have errors associated with some issues. Such issues include excessive spatial averag- ing that mixes urban and rural areas and thus underestimates AOD in the more polluted urban areas (for example, Hu, Waller, Lyapustin, Wang, Al-Hamdan, and others 2014; Hu, Waller, Lyapustin, Wang, and Liu 2014; Lyapustin and oth- ers 2011); reflective surfaces, such as desert dust, that make it more difficult to separate the reflection of sunlight by aerosols from the surface reflection (for example, Sorek-Hamer and others 2015); “patchwork” surfaces in urban areas, where several different land-use types, each with different albedos, may all be present within a single satellite pixel (for example, Oo and others 2010); and the presence of undetected thin clouds, which can also create a positive bias in AOD retrievals (for example, Sun and others 2011). These limitations in satellite AOD products are consistent with some of the errors identified in using satellite observations to determine ground-level PM2.5 concentrations. For example, the high estimates in Egypt, Qatar, and Saudi Arabia are consistent with overesti- mates of AOD in regions with highly reflective surfaces. These problems can be mitigated by careful analysis and identification of the satellite observations and retrieval algorithms that give the most accurate and precise estimates of AOD for a given region, rather than trying to find one satellite product that works suffi- ciently well for all conditions around the globe. A deep understanding of the strengths and weaknesses of the different satellite AOD products is thus essen- tial to determining which product will be most useful for a given set of geo- graphic and meteorological conditions. Most previous validation work on the use of satellite observations to estimate ground-level PM2.5 has been performed in developed countries with extensive, well-calibrated, long-term ground-level-monitoring (GLM) observations of PM2.5 (Type V countries, using the proposed typology in chapter 1). However, the performance of the statistical models and chemical transport models (CTMs) used to derive these relationships is not as well characterized in many low- and middle-income countries (LMICs), where GLM is infrequent or absent and few  53 54 | Getting Down to Earth field campaigns have been performed to make PM2.5 observations with which to train and test the models. In addition, any GLM observations that are available in LMICs may be difficult to access or may require significant effort to ­quality-assure the data. Being able to access these data sets and assess the quality of the measurements is thus a critical part of this project. The lack of GLM in many LMICs also introduces difficulties in trying to relate the spatially averaged AOD observations to variations in PM2.5 within a city (due to distance from road- ways and other pollution sources), which suggests the need to combine satellite observations with GLM observations and data on pollution sources within a city using interpolation techniques (for example, de Hoogh and others 2016; Vienneau and others 2013). This appendix contains a literature review that discusses the approaches used (and their strengths and limitations) in each of the three key steps in com- bining satellite observations with GLM measurements of PM2.5: • The retrieval of AOD from satellite observations of reflected solar radiation (second section) • The translation of the vertically integrated AOD observation into an estimate of the ground-level PM2.5 concentration, including any bias corrections (third section) • Simultaneously interpolating these satellite-derived PM2.5 estimates and the available GLM data (fourth section). Based on this review, recommendations are made for LMICs on how best to incorporate satellite data into their monitoring plans (fifth section). The recom- mendations use the following proposed typology of countries based on their levels of engagement with air-quality monitoring: ­ • Type I: Countries with no existing measurements and no history of routine measurements of atmospheric composition of any kind. Some anecdotal measurements or one-time sampling may have taken place. ­ • Type II: Countries with some information on atmospheric composition avail- able (perhaps PM10 or total suspended particulates [TSPs]) but of variable quality without rigorous quality assurance (QA) procedures. • Type III: Countries that possess reliable information but with poor spatial or temporal coverage. For example, monitoring may exist in only one city or rou- tine monitoring may exist for a period of a year or two but is no longer being collected because of equipment malfunction or lack of repairs. • Type IV: Good, reliable AQ monitoring underway or being established. • Type V: Routine, long-term AQ monitoring. AVAILABLE AEROSOL-OPTICAL-DEPTH PRODUCTS Monitoring the distribution and properties of atmospheric aerosols from satel- lites is a crucial component of establishing the radiative forcing of aerosols and their impact on climate, and it is of growing interest for regional and urban air-quality monitoring. The degree to which aerosol particles suspended in the atmosphere interact with light depends on the size and shape of the particles. Smaller particles that scatter light at visible (VIS) and shorter ultraviolet (UV) wavelengths have negligible effect at longer near-infrared (NIR) wavelengths where (in the absence of clouds) there is a relatively clear view to the surface. Literature Review | 55 These dependencies on spectral properties can be exploited through satellite observations in the VIS through NIR wavelengths to infer properties of the aero- sols classified according to a limited set of models. The ability to measure aero- sols is dependent on there being reasonably good contrast between the light scattered from the particles and that reflected from the surface at VIS wave- lengths. For this reason, aerosol retrievals are often limited to darker surfaces (for example, ocean or vegetation). Integral to this approach is the need to infer the surface reflectance at VIS wavelengths in the absence of aerosols. This is typically achieved either by using observations from previous times (when aero- sols were not present) or by relying on relationships with the signals at longer wavelengths unaffected by the aerosols. These methods tend to also work best for darker surfaces. The monitoring of aerosols based on reflected light is also restricted to daytime conditions. Radiative-transfer (RT) models enable accurate simulations of reflectance at the top of the atmosphere (TOA) in the presence of aerosol layers with a variety of optical properties. By properly accounting for surface reflectance from land and ocean backgrounds, it is possible to retrieve several important aerosol prop- erties by comparing satellite-observed TOA reflectances to those calculated from an RT model. Typically, to optimize processing time, TOA reflectances are precalculated by the RT model for several combinations of aerosols and are stored in look-up tables (LUTs) accessed by a retrieval algorithm. The RT calcu- lations are performed for a range of aerosol optical thicknesses, so each stored reflectance value in the LUT corresponds to an aerosol optical thickness (AOT). Because the surface type plays such an important role in aerosol retrievals, and because concentrations of aerosols often correspond to land or ocean prov- enance regions, somewhat different approaches are implemented over land and ocean. For land retrievals, specification of land surface reflectance is more diffi- cult because of the greater heterogeneity of global land cover. For this reason, direct observations of land-surface reflectance are preferred over using simu- lated land-surface reflectances. Threshold tests are typically applied to all land pixels to ensure that only dark vegetated pixels are processed. The total TOA reflectance is the sum of the observed land surface and atmospheric contribu- tions. The matching aerosol model is one with a TOA reflectance that produces the smallest residual with respect to observed TOA reflectances in bands sensi- tive to the presence of aerosols. The AOD and other properties associated with the selected model are reported. Variations on this retrieval approach have been used to generate aerosol products from satellite observations. The World Meteorological Organization (WMO) maintains an online record of satellite technology and products related to weather, water, and climate called the Observing Systems Capability Analysis and Review Tool (OSCAR)1 that is a very good reference of all past, current, and future aerosol products. The operational and science remote sensing capabilities collected there provide different levels of information about the aerosol constit- uents. Multispectral imaging radiometers (for example, Advanced Baseline Imager [ABI], Moderate-Resolution Imaging Spectroradiometer [MODIS], and Visible Infrared Imaging Radiometer Suite [VIIRS]) are capable of column AOD and aerosol particle size determinations. Multiangle radiometers (for example, Multi-angle Imaging Spectroradiometer [MISR]) and polarization imagers (for example, Multi-viewing, Multi-channel, Multi-Polarization Imaging mission [3MI] or Polarization and Directionality of the Earth’s Reflectances [POLDER]) 56 | Getting Down to Earth can provide more detail on aerosol size or type information, and space-based lidars (light detection and ranging devices) (for example, Cloud-Aerosol Lidar with Orthogonal Polarization [CALIOP]) are used for aerosol and cloud profiling. Aerosol products derived from instruments aboard polar satellites in low earth orbit (LEO) (for example, Advanced Very High Resolution Radiometer [AVHRR], MODIS, and VIIRS) provide global coverage roughly once per day, with spatial resolution on the order of 1 to 10 kilometers, and that depends on the orbit path. Products derived from instruments on geo- stationary (GEO) satellites (for example, ABI, Advanced Himawari Imager [AHI], Geostationary Operational Environmental Satellite [GOES-Imager], and Spinning Enhanced Visible and Infrared Imager [SEVIRI]) provide cov- erage over the full field of view (that is, Full Disk) at less than 30-minute frequency with a spatial resolution of roughly 2 to 4 kilometers, and that varies with the satellite zenith angle. As a resource for air-quality monitoring at specific locations around the world, it is necessary that remotely sensed aerosol-optical-depth products used to infer ground-level PM2.5 concentra- tions are available in near real-time (NRT) and that the products provide adequate coverage for regions of interest. Figure A.1 illustrates the coverage provided by LEO and GEO satellites in the northern hemisphere. A summary of the aerosol products produced by the current series of LEO and GEO sat- provided in table A.1. ellites is ­ Satellite-based AOD products typically undergo a very rigorous validation process before they are officially released. This process includes extensive com- parisons to ground-based AOD measurements (for example, Aerosol Robotic Network [AERONET]; Holben and others 2001) as well as intercomparisons with other established AOD products and with other science missions that can provide more detailed aerosol information (for example, aerosol profiles via CALIOP). As a result, the errors are fairly well documented and are typically expressed as a bias from truth (that is, accuracy) and the scatter relative to this bias (that is, the precision), though the metrics used can vary from program to program. In addition, products are always accompanied by data quality flags that identify locations of good and degraded performance with strong recommenda- tions to refer to these ratings when used in any application. For PM2.5 air-quality applications, these error specifications should be used to propagate uncertain- ties in the satellite retrievals of AOD to the estimates of ground-level PM 2.5 concentrations. The next two subsections provide an overview of satellite-based AOD retriev- als from polar and geostationary platforms. One significant finding of this review is that alternative products have sometimes been developed by different organi- zations, but not all AOD products described in the published literature are cur- rently made available to users. The National Oceanic and Atmospheric Administration (NOAA) MODIS products are currently the most accessible, but the long-term status of MODIS is uncertain, with the Joint Polar Satellite System (JPSS) VIIRS instrument expected to provide continuing coverage into the future. The AOD products from GEO satellites can provide greater temporal coverage of aerosol events, but currently the only reliable coverage is from the the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) Meteosat Second Generation (MSG) satellite providing coverage of Europe and Africa. The Geostationary Operational Environmental Satellite-R Literature Review | 57 FIGURE A.1 Coverage provided by available polar and geostationary weather satellites a. Current coverage provided by available polar b. Future coverage after planned updates of and geostationary satellites geostationary satellites Source: World Bank. Note: The three bands in each panel illustrate three consecutive swaths from the Moderate-Resolution Imaging Spectroradiometer in sun-synchronous orbit providing once per day coverage. Green, red, turquoise, magenta, and blue points represent coverage from the geostationary satellites. The left panel shows current coverage; the right panel shows future coverage with roughly two times higher spatial resolution. TABLE A.1  Summary of operational aerosol products AGENCY SATELLITE ORBIT REFRESH NEAR REAL-TIME [DEVELOPER] [INSTRUMENT] [TYPE] PRODUCT(S) [COVERAGE] AVAILABILITY EUMETSAT MSG 0E SMAOL 15 minutes Yes [CNRS-ICARE] [SEVIRI] [GEO] [Europe, Africa] EUMETSAT Metop-A 9:30 dsc Polar Multi-Sensor Aerosol Daily Yes Metop-B Product [EUMETSAT] 9:30 dsc [Global] [AVHRR, HSR 5 × 40 10 × 40 kilometer GOME-2, IASI] [LEO] swath 960 × 1920 kilometer JMA Himawari-8/9 141. E AOD [East Asia, Indonesia, TBD Australia] [JMA-MSC] [AHI] [GEO] NASA Terra 10:30 dsc 10 km Dark Target AOD Daily Yes [NASA-GSFC] Aqua 13:30 asc 10 km Deep Blue AOD [Global] [MODIS] [LEO] 10 km combined AOD 3 km Dark Target AOD NASA Terra 10:30 dsc MAIAC AOD (1 kilometer) Daily TBD gridded Amazon data set [NASA-GSFC] Aqua 13:30 asc [Global] 2000–12 [MODIS] [LEO] continued 58 | Getting Down to Earth TABLE A.1, continued AGENCY SATELLITE ORBIT REFRESH NEAR REAL-TIME [DEVELOPER] [INSTRUMENT] [TYPE] PRODUCT(S) [COVERAGE] AVAILABILITY NOAA NOAA-20 13:30 asc 10 kilometer Dark Target AOD Daily TBD [NASA-GSFC] [VIIRS] [LEO] 10 kilometer Deep Blue AOD [Global] 10 kilometer Combined AOD 3 kilometer Dark Target AOD NASA Terra 10:30 dsc AOD (MIL2ASAE) 9 days Yes [JPL] [MISR] [LEO] [Global] NOAA NOAA-19 15:20 asc AOT Daily Yes [NOAA-NCEI] NOAA-15 18:15 asc [Global; Ocean only] NOAA-18 19:02 asc [AVHRR] [LEO] NOAA GOES-13 75. W GASP-East AOD 1 hour Yes [NOAA-SPSD] GOES-15 135. W GASP-West AOD [CONUS] [IMAGER] [GEO] NOAA NOAA-20 13:30 asc AOT Daily Yes [NOAA-STAR] [VIIRS] [LEO] [Global] NOAA GOES-16 75. W SMAOD 5, 10, or 15 minutes Yes [NOAA-STAR] GOES-17 137. W [North and South America] [ABI] [GEO] Source: World Bank. Note: In column 3, Low Earth orbits (LEOs) are all sun-synchronous and identified in terms of the daytime equator crossing time and direction (that is, ascending [asc] or descending [dsc]); Geostationary (GEO) orbits are identified by their satellite longitude. ABI = Advanced Baseline Imager; AOD = aerosol optical depth; AVHRR = Advanced Very High Resolution Radiometer; CNRS-ICARE = Centre National de la Recherché Scientific–Institut de Combustion, Réactivité et Environnement; CONUS = continental United States; EUMETSAT = European Organisation for the Exploitation of Meteorological Satellites; GASP = GOES Aerosol/Smoke Product; GOES = Geostationary Operational Environmental Satellite; GOME-2 = Global Ozone Monitoring Experiment-2; GSFC = Goddard Space Flight Center; IASI = Infrared Atmospheric Sounding Interferometer; IMAGER = GOES Imager; JMA = Japan Meteorology Agency; JPL = Jet Propulsion Laboratory; MAIAC = Multi-Angle Implementation of Atmospheric Correction; MIL2ASAE = MISR level 2 aerosol parameters; MISR = Multi-angle Imaging Spectroradiometer; MODIS = Moderate-Resolution Imaging Spectroradiometer; MSC = Meteorological Satellite Center; MSG = Meteosat Second Generation; NASA = National Aeronautics and Space Administration; NOAA = National Oceanic and Atmospheric Administration; SEVIRI = Spinning Enhanced Visible and Infrared Imager; SMAOD = suspended matter, aerosol optical depth; SMAOL = SEVIRI-MSG Aerosol Over Land; STAR = Center for Satellite Applications and Research; TBD = to be determined; VIIRS = Visible Infrared Imaging Radiometer Suite. series (GOES-R) products over North and South America were made available in 2018. However, the availability of Japan Meteorology Agency (JMA) Himawari data with coverage over East Asia and Indonesia is uncertain, and no current geostationary satellite mission is providing AOD products in the region of India and western Asia. Furthermore, the availability of satellite coverage does not ensure that AOD products will be produced, because the retrievals are limited by environmental conditions (for example, must be cloud-free) and surface condi- tions (for example, must be snow-free). The third subsection looks at three urban locations to review how satellite-derived AOD coverage might be ­provided to support air-quality missions. Polar-orbiting satellites MODIS products Aerosol products derived from observations by the MODIS aboard the National Aeronautics and Space Administration’s (NASA’s) Terra and Aqua satellites are among the most widely studied products in the aerosol remote sensing Literature Review | 59 ­ ommunity. This pair of instruments (MODIS-Terra and MODIS-Aqua) has c created a climate record extending over 15 years. Although MODIS has already exceeded its design lifetime, calibration and validation efforts continue to maintain the calibration accuracy of the instruments to ensure the integrity of the ongoing observations. Studies based on MODIS data provide some of the best examples of the application of space-based AOD to the understanding of global climate and regional air-quality conditions (see the third section below). Two primary AOD products are produced from the MODIS observations based on the Dark Target2 and Deep Blue3 aerosol-retrieval algorithms. These two products are made readily available to users. A third research product from MODIS, Multi-Angle Implementation of Atmospheric Correction (MAIAC), has been used in many studies but is not yet processed operationally. Dark Target: The NASA Dark Target algorithm produces AOD over ocean and other dark land (for example, vegetated) surfaces using two separate algo- rithms for the different regimes. The standard product for climate studies is AOD aggregated at 10-kilometer resolution (Levy and others 2013). Over land, the retrieval produces AOD at 550 nanometers, the AOD model/weight, and sur- face reflectance at 2,130 nanometers based on 500-meter resolution observa- tions at 470, 650, and 2,130 nanometers and using a LUT of RT-model–computed reflectances as a function of aerosol and surface properties. Dark targets are identified as pixels where the 2,130-nanometer reflectance is between 0.01 and 0.25. For these locations, the quality flag is set based on the goodness of the fit between the observed and modeled reflectance. Targets with reflectance up to 0.4 are processed through a separate algorithm path but flagged as reduced qual- ity. To meet the growing interest in AOD for regional studies, the product over land is now also produced at a 3-­ kilometer resolution (Remer and others 2013). The error estimate of the 10-kilometer AOD based only on good-quality retriev- als has been assessed against AERONET measurements to be  ±(0.05 + 0.15 AODAERONET),4 whereas the error in the 3-kilometer p ­ roduct is ±(0.05 + 0.20 AODAERONET). This product is available through the NASA Level-1 and Atmospheric Archive and Distribution System (LAADS) Distributed Active Archive Center (DAAC)5,6 or through the NASA WORLDVIEW website.7 Investigations of the use of the 3-kilometer Dark Target product to provide better definition of local aerosol gradients (Munchak and others 2013) suggest that the 3-kilometer product does provide better spatial coverage (see figure A.2). However, this product tends to overestimate the aerosol loading ­ and can be susceptible to noise problems in urban areas, likely because of inad- equate characterization of surface features. For such applications, some cau- tion is advised. Deep Blue: The Deep Blue aerosol algorithm explicitly models surface reflec- tance, rather than estimating it from observations at 2,130 nanometers as in the Dark Target algorithm. This explicit modeling allows the Deep Blue algorithm to be applied to bright surfaces such as desert, semiarid, and urban regions (Hsu and others 2013). To enable retrievals over brighter surfaces, the algorithm uses the 412-nanometer or “deep blue” band on MODIS (where the surface reflec- tance over land is much lower than in low-frequency visible bands), thus provid- ing better contrast with the aerosol signal. This product is computed over land at 1-kilometer resolution and aggregated to 10 kilometers. Surface reflectance is determined by one of several methods depending on the surface type, including use of a database that takes account of seasonal changes and the variability of urban landscapes based on the normalized difference vegetation index (NDVI). 60 | Getting Down to Earth FIGURE A.2 MODIS Dark Target 10-kilometer and 3-kilometer aerosol-optical-depth products retrieved for clear land and ocean fields of view and the local 5-kilometer average derived from the products, outer circle, compared to ground-based measurements, inner circle, over Baltimore, US a. MODIS Dark Target 10-kilometer product b. MODIS Dark Target 3-kilometer product AOD at 0.55 µm -0.05 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 Source: Munchak and others 2013. Note that these results used an older version of the MODIS surface reflectance scheme, and more recent versions (for example, Gupta and others 2016) show much better performance over urban areas. Note: AOD = aerosol optical depth; MODIS = Moderate-Resolution Imaging Spectroradiometer. The (prognostic) uncertainty derived by Deep Blue is comparable to the error estimate of the Dark Target product for typical aerosol levels in the range from 0.1 to 0.5, ± 0.03 + 0.2τDB. The combined Deep Blue and Dark Target aerosol product was developed to provide users with the alternative of a single merged MODIS product (Sayer and others 2014). It applies NDVI criteria based on a monthly composite to select either the Dark Target or Deep Blue AOD for a given 10-kilometer pixel location: • NDVI ≤ 0.2 Use Deep Blue • NDVI ≥ 0.3 Use Dark Target • 0.2 < NDVI < 0.3 Use product with highest quality or report mean value. As a result, Deep Blue is reported over desert regions, Dark Target is selected over permanent vegetation, and in other areas the selection of Deep Blue or Dark Target varies with the season. Comparison of the two products to AERONET indicates that neither Deep Blue nor Dark Target consistently outperforms the other (see map A.1). Dark Target tends to have slightly smaller overall errors compared to AERONET, but Deep Blue provides additional coverage and tends to perform better for low-AOD conditions. Both the Deep Blue and the com- bined product are available through the NASA LAADS DAAC or through the NASA WORLDVIEW website. MAIAC: The Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm8 is an atmospheric correction algorithm designed for Literature Review | 61 MAP A.1 Fraction of good-quality attempted retrievals from Deep Blue and Dark Target algorithms showing differences in coverage over desert regions, and showing differences in coverage due to the quality checks applied in each algorithm a. Deep Blue fraction assigned good QA 80° 40° Latitude 0° –40° –80° –180° –90° 0° 90° 180° Longitude Fraction of retrievals that passed QA checks (range 0–1) 0 0.25 0.5 0.75 1.0 b. Dark Target / ocean fraction assigned good QA 80° 40° Latitude 0° –40° –80° –180° –90° 0° 90° 180° Longitude Source: Sayer and others 2014. Note: QA = quality assurance. MODIS that performs simultaneous retrievals of atmospheric aerosols and bidi- rectional surface reflectance (Emili and others 2011; Lyapustin and others 2011). Unlike the Dark Target and Deep Blue algorithms, the MAIAC algorithm does not rely on any predetermined relationships with the shortwave information to constrain the surface retrieval but instead invokes a multiangle, multitemporal retrieval of the surface bidirectional reflectance distribution function and aero- sol loading, based on assumptions that the surface reflectance does not change over a 16-day period and that AOD changes little over 25-kilometer distances. The resulting AOD product is generated on a one-kilometer s ­ inusoidal grid. Over vegetated regions (forest, cropland, grassland, and savanna), more than 66 percent of retrievals are within the expected error ±(0.05 + 0.05t) with a cor- relation coefficient to AERONET better than 0.86 (Martins and others 2017). 62 | Getting Down to Earth Results over other bright backgrounds were less precise, especially for low aero- sol signals where the retrieval problem is not as well constrained (for example, 55 percent of retrievals within the expected error for urban regions). The MAIAC aerosol product is nominally distributed through the NASA LAADS DAAC9,10 but appears to be currently unavailable except for a limited data set over the Amazon basin in South America from 2000 to 2012.11,12 Online information indicates support of the product is ongoing13,14,15 with a potential extension to process VIIRS data in the works.16 VIIRS products The Visible Infrared Imaging Radiometer Suite (VIIRS) sensor is a scanning radiometer on Suomi-National Polar-orbiting Partnership (Suomi-NPP) and the Joint Polar Satellite System (JPSS) satellites operated by NOAA and NASA. Over land, the NOAA VIIRS AOD product17 is generated with a LUT- based retrieval of AOD, based on five aerosol models using observed reflec- tance in 750-meter moderate-resolution bands at 412, 445, 488, 672, and 2,250 nanometers (Jackson and others 2013). The surface reflectance in the red (672 nanometers) and blue (488 nanometers) bands is inferred based on an empirical relationship with the shortwave infrared (2,250 nanometers) band that holds mainly for dark surfaces. The AOD is then determined from observed blue-to-red reflectance ratios by matching the observed signal to computed values for a given aerosol model. AOD retrievals are not performed for clouds, snow, fire, glint, and other bright surfaces identified, based on a shortwave infrared NDVI threshold (NDVISWIR < 0.05 with a 2,250-­nanometer reflectance > 0.3). NDVISWIR > 0.2, consistent with vegetation backgrounds, is required for high-quality retrievals. The AOD output is available as an intermediate product at 750-meter resolu- tion and as a 6-kilometer (eight-by-eight aggregated) product, thus providing some advantage over the MODIS 3- and 10-kilometer products for analysis of the localized aerosol distribution. While studies of regional applications of VIIRS AOD products have been presented as conference papers, no compre- hensive review of the product has been published. The overall performance of the VIIRS product has been validated against requirements with uncertainties similar to that of MODIS Dark Target (Huang and others 2016). However, com- parisons of the 1-kilometer MAIAC AOD product to the high-resolution VIIRS product (see figure A.3) found that the VIIRS product was biased to higher AOD in urban and mountainous regions and was susceptible to errors in regions of snowmelt (Superczynski, Kondragunta, and Lyapustin 2017). The official NOAA VIIRS products are distributed through the Comprehensive Large Array-Data Stewardship System (CLASS)18 and with the most recent 90 days available via file transport protocol (FTP).19 Alternatives to the operational NOAA VIIRS AOD product are in develop- ment. For example, the NASA Dark Target algorithm was applied to VIIRS data, and results were compared with the operation NOAA algorithm and to MODIS (Levy and others 2015). The NASA Deep Blue algorithm20 also has been applied to VIIRS data. In addition, the NOAA Center for Satellite Applications and Research (STAR) has published work describing an enhanced AOD algo- rithm capable of retrieving aerosols over brighter surfaces and comparable to the Deep Blue product (Zhang and others 2016). Finally, VIIRS products based on the MAIAC algorithm have been referenced in online presentations, but the status of such products is unclear. Once available, these alternative products Literature Review | 63 FIGURE A.3 Comparison of VIIRS 750-meter and MAIAC one-kilometer aerosol- optical-depth products during western US fires, 2013 a. VIIRS 750-meter AOD product b. MAIAC one-kilometer AOD product AOD 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Source: Superczynski, Kondragunta, and Lyapustin 2017. Note: AOD = aerosol optical depth; MAIAC = Multi-Angle Implementation of Atmospheric Correction; VIIRS = Visible Infrared Imaging Radiometer Suite. may provide improved coverage, in particular for urban areas, potentially increasing their utility for regional air-quality applications. MISR products The Multi-angle Imaging Spectroradiometer (MISR) instrument uses multian- gle techniques in the VIS and NIR to acquire stereoscopically resolved observa- tions of AOD under sunlit conditions (Diner and others 1998). All 36 channels (nine cameras and four spectral bands) are used when performing aerosol retrievals over land (Martonchik and others 2004). MISR measurements (between 70.5° forward and 70.5° aft) span a wide range of scattering angles and 64 | Getting Down to Earth air mass factors providing information on aerosol microphysical properties and yielding sensitivity to optically thin aerosols. The MISR retrieval approach relies on describing the change in surface contrast with view (camera) angle over a region 17.6 kilometers by 17.6 kilometers in size, composed of 16 by 16 (256) sub- kilometer by 1.1 kilometer. This subregion size is the nominal regions, each 1.1 ­ spatial resolution of MISR when observing in Global Mode, and 17.6 kilometers is the spatial resolution of MISR’s retrieved aerosol product. Aerosol retrievals are performed on only a regional and not a subregional basis, because the entire region is used to determine the surface contribution to the reflectance. In addi- tion to having a lower horizontal resolution than the MODIS and VIIRS prod- ucts, the MISR instrument has a smaller swath than MODIS (400 kilometers versus 2,330 kilometers) and thus only observes a given location once every nine days, unlike once a day for MODIS and VIIRS. The performance of AOD retrievals from MISR is comparable to that of MODIS. Martonchik and others (2004) compared MISR AOD with AERONET observations and found that at 17.6-kilometer spatial resolution, the estimated uncertainty is about 0.08. When the spatial resolution was degraded to 52.8 kilo- meters, the estimated MISR AOD uncertainty decreased to about 0.05. Khan and others (2005) found that 66 percent of the measurements agreed with AERONET within ±(0.05 + 0.2τ). Kahn and others (2009) compared MISR with MODIS AOD and found that where coincident AOD retrievals are obtained over the ocean, the MISR-MODIS correlation coefficient is about 0.9 with a slope of 0.75; over land, the correlation coefficient is about 0.7 with a slope of 0.60. MISR AOD data are available from the Jet Propulsion Laboratory.21 AVHRR products The Advanced Very High Resolution Radiometer (AVHRR/3) instrument pro- vides visible-through-IR remote sensing observations on the LEO satellites flown by NOAA (for example, NOAA-15) and by the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) (for example, Metop-A). However, the operational NOAA AOD product22 is produced only over oceans, though the application for AOD retrievals over land has also been investigated (Li and others 2013). The NASA Deep Blue algorithm has been applied to the NOAA data to produce an aggregated 8.8-kilometer product over land and ocean with an expected error over land of ±(0.05 + 0.25τ) (Sayer and others 2017). This application of Deep Blue uses a database of surface reflectance for retrievals over bright surfaces, but because AVHRR lacks a ­ 412-­nanometer band, the red band at 620 nanometers is used instead. Because the surface at 630 nanometers is not as dark as 412 nanometers, the sensitivity of the aerosol retrieval is limited compared to MODIS and VIIRS. AVHRR also lacks a shortwave band that is used in the MODIS/VIIRS algorithms to derive the surface reflectance for dark targets. For AVHRR, the surface is modeled as an empirical function of NDVI. However, this product was available online only through 2011.23 EUMETSAT produces an operational multisensor aerosol product based on observations from AVHRR, the Global Ozone Monitoring Experiment (GOME; a medium-resolution double UV-VIS spectrometer), and the infrared atmospheric sounding interferometer (IASI)–Fourier transform spectrome- ter.24,25 However, the primary input for the AOD retrieval is derived from GOME, and the resolution of the product (5 kilometers by 40 kilometers for Literature Review | 65 Metop-A or 10 kilometers by 40 kilometers for Metop-B) is dictated by this instrument. Geostationary satellites GOES-NOP imager products The GOES-N series geostationary satellites (GOES-NOP; for example, GOES- 13 East at 75° W and GOES-15 West at 135° W) operated by NOAA are able to provide Full Disk imagery of the Earth covering North and South America in a single VIS band and four IR bands with a refresh rate of 30 minutes. For over 15 years, NOAA has supported the generation of an operational AOD product based on the GOES observations. Being limited to a single (uncalibrated) VIS band, the accuracy of this product is somewhat less than other, more advanced products, but the GOES product is still of significant value owing to its unique temporal information. The GOES Aerosol/Smoke Product (GASP)26 is derived using a surface reflectance based on a 28-day composite (converted to albedo using an RT model) and AOD retrieved using an RT-model–computed LUT and comparing observations to model reflectances in the VIS band (Knapp others 2005). The retrieval is limited to a single continental aerosol model and ­ and restricted to regions with dark vegetation. GASP is produced at four-­ kilometer resolution (at nadir) with a quoted precision of ±0.13 and a correla- tion of R of 0.72 with AERONET. The current product is limited to the continental United States, and products are available in binary format in NRT from the NOAA Satellite Products and Services Division website.27 This prod- uct was discontinued in 2018 from GOES-13 (East) when the GOES-16 prod- ucts became operational. SEVIRI-MSG products The SEVIRI-MSG Aerosol Over Land (SMAOL) product28 (based on observa- tions from the Spinning Enhanced Visible and Infrared Imager [SEVIRI] on Meteosat Second Generation [MSG]) was developed by HYGEOS for the Centre National de la Recherché Scientific (CNRS). SMAOL is distributed in NRT though the ICARE Data and Services Center (Bernard and others 2011; Mei and others 2012). The algorithm approach is similar to that of the GOES GASP product but with greater sensitivity. SEVIRI observations at 630, 810, and 1640 nanometers (corrected for gas absorption and molecular scattering) are used to derive AOD at 550 nanometers chosen from between five aerosol models using a LUT approach to minimize the differences at the three wavelengths. The product is generated at the three-kilometer (at nadir) resolution and produced every 15 minutes during daytime/cloud-free conditions and subject to some viewing and scattering angle restrictions. With the MSG satellite at 0° longi- tude, the product provides coverage of Europe and Africa. The surface reflec- tance used in the retrieval is based on a fit to the minimum reflectance over a 14-day period. This methodology is valid only for dark targets; therefore, desert regions (such as the Sahara, Sahel, and Namib) are excluded. The product has been validated with comparisons to AERONET and to the MODIS Dark Target AOD product and found to provide good estimates of both diurnal and daily variations in AOD (see figure A.4). Error sources include possible subpixel cloud contamination, errors in surface reflectance estimation, and possible temporal noise in model selection. 66 | Getting Down to Earth FIGURE A.4 Detection of temporal variations in aerosol optical depth with the SEVIRI-MSG aerosol over land product compared to AERONET measurements at Palaiseau, France, and to MODIS, July 14, 2006 0.6 Aerosol optical thickness (0.6 micrometers) 0.5 0.4 0.3 0.2 0.1 0 0 5 7.0 7.5 0 5 0 5 .0 .5 .0 .5 .0 .5 .0 .5 .0 .5 .0 .5 .0 .5 .0 .5 .0 .5 6. 6. 8. 8. 9. 9. 17 17 11 11 13 13 15 15 10 10 12 12 16 16 18 18 14 14 Time (UT) Aeronet 675nm SEVIRI 635nm MODIS Aqua 660nm (1 pix) MODIS Terra 660nm (1 pix) Source: ©Bernard and others 2011; CC BY 3.0. Note: AERONET = Aerosol Robotic Network; AOT = aerosol optical thickness; MODIS = Moderate-Resolution Imaging Spectroradiometer; MSG = Meteosat Second Generation; nm = nanometer; SEVIRI = Spinning Enhanced Visible and Infrared Imager; UT = Universal Time. GOES-R ABI products GOES-R is the first in a series of next-generation geostationary environ- mental satellites covering the western hemisphere. 29 It was launched as GOES-16 in November 2016. For earth remote sensing, observations are col- lected by the Advanced Baseline Imager (ABI), which offers much-improved spatial, temporal, and spectral information over the preceding GOES-NOP Imager series. For the past year, the ABI instrument products have under- gone an intensive validation process. In December 2017, the satellite was moved to the GOES East position, with products scheduled to become oper- ational in 2018. The GOES-R AOD and aerosol-particle–sized products30,31 are derived from ABI reflectance measurements through physical retrievals that utilize a LUT of TOA reflectance that is precalculated using an RT model. Retrievals are per- formed separately over land (for dark surfaces) and ocean. The baseline AOD product is generated at two-kilometer resolution (at nadir) for Full Disk and continental United States (CONUS) geographic coverage areas. The Full Disk AOD output product is generated at 15-minute intervals (see figure A.5), whereas the CONUS product is generated at five-minute intervals. The GOES-R product performance specification is a function of AOD value. For intermediate AOD loading between 0.04 and 0.8, the accuracy over land is ± 0.04 and the precision Literature Review | 67 FIGURE A.5 Full disk coverage at 550 nanometers of GOES-R SMAOD product over ocean and land AOD 0.00 0.50 1.00 1.50 2.00 Source: World Bank. Note: AOD = aerosol optical depth; GOES = Geostationary Operational Environmental Satellite R series; SMAOD = suspended matter, aerosol optical depth. is ± 0.25, which is similar to the equivalent polar products. Once approved, the GOES-R AOD product will be distributed through the NOAA’s Comprehensive Large Array-Data Stewardship System.32,33 Although the GOES-R program product is the official operational AOD product, alternative algorithms have been developed to also measure aerosol properties from GOES-R data. For example, the Enterprise Processing System (EPS) AOD algorithm34 is an algorithm from NOAA STAR intended to support the generation of an AOD product from GOES-R ABI and JPSS VIIRS with a common methodology.35 The status of the NOAA EPS product is uncertain at this time. Himawari-8/9 AHI products Himawari-8/9 is a new series of geostationary weather satellites operated by the JMA. The imagery collected by Himawari-8/9 is produced by the Advanced Himawari Imager (AHI), which is very similar in design to the ABI instrument flown on GOES-16. Although the MODIS Deep Blue, GOES-R, and EPS 68 | Getting Down to Earth algorithms have been applied and tested on AHI data, these products are not currently available. The JMA advertises their own environmental products,36,37 but these are also not freely available.38,39 Case studies In table A.2 three urban locations with poor air quality are reviewed to evaluate the available satellite AOD products and to assess the utility and limitations of these products for the regions. Each location includes a link to the NASA WORLDVIEW website with visualizations of the Dark Target and Deep Blue aerosol products from MODIS-Terra and MODIS-Aqua. In Delhi, India, the air quality in late fall is affected by widespread smoke from rural crop fires combined with the localized urban smog trapped in the TABLE A.2  Case study of satellite aerosol-optical-depth applicability for three representative low- and middle-income country urban areas ­ LOCATION COORDINATES LAND TYPE/CLIMATE COVERAGE Delhi, India 28°36′36″N, Land type: Semivegetated, desert, urban MODIS 77°13′48″E Topography: Yamuna River basin VIIRS Climate: Rain Jun to Oct MISR Air quality: Widespread smoke from rural crop fires WORLDVIEWa mixed with urban smog Lima, Peru 12°2′36″S, Land type: Coastal desert, urban, semivegetated MODIS 77°1′42″W Topography: Andes foothills VIIRS Climate: Persistent clouds/fog May to Nov MISR Air quality: Urban (localized) smog ABI WORLDVIEWb Ulaanbaatar, Mongolia 47°55′N, Land type: Grasslands (steppes) MODIS 106°55′E Topography: Tuul River valley VIIRS Climate: Seasonal snow cover MISR Air quality: Desert dust, urban smog, coal/wood fires AHI (winter) WORLDVIEWc Source: World Bank, produced with Esri ArcGIS. Note: AHI = Advanced Himawari Imager; MISR = Multi-angle Imaging Spectroradiometer; MODIS = Moderate-Resolution Imaging Spectroradiometer; VIIRS = Visible Infrared Imaging Radiometer Suite.  ttps://worldview.earthdata.nasa.gov/?p=geographic&l=MODIS_Aqua_SurfaceReflectance_Bands143(hidden),MODIS_Terra_SurfaceReflectance​ a. h _Bands143(hidden),MODIS_Aqua_CorrectedReflectance_TrueColor(hidden),MODIS_Terra_CorrectedReflectance_TrueColor,MODIS_Terra​ _­NDVI_8Day(hidden),MODIS_Terra_AOD_Deep_Blue_Combined(hidden),MODIS_Terra_AOD_Deep_Blue_Land(hidden),MODIS_Terra​ _­Aerosol(hidden),MODIS_Terra_Aerosol_Optical_Depth_3km(hidden),MODIS_Aqua_Aerosol(hidden),MODIS_Aqua_AOD_Deep_Blue​ _Combined(hidden),MODIS_Aqua_Aerosol_Optical_Depth_3km(hidden),MODIS_Aqua_AOD_Deep_Blue_Land(hidden),Reference_Labels,Reference​ _Features(hidden),Coastlines&t=2017-03-04&z=3&v=75.58614905641733,27.64425081916523,78.66232093141733,29.24825472541523&ab=off& as=2017-03-04&ae=2017-03-1 1&av=3&al=true.  ttps://worldview.earthdata.nasa.gov/?p=geographic&l=MODIS_Aqua_SurfaceReflectance_Bands143(hidden),MODIS_Terra_SurfaceReflectance​ b. h _Bands143(hidden),MODIS_Aqua_CorrectedReflectance_TrueColor(hidden),MODIS_Terra_CorrectedReflectance_TrueColor,MODIS_Terra​ _­NDVI_8Day(hidden),MODIS_Terra_AOD_Deep_Blue_Combined(hidden),MODIS_Terra_AOD_Deep_Blue_Land(hidden),MODIS_Terra​ _­Aerosol(hidden),MODIS_Terra_Aerosol_Optical_Depth_3km(hidden),MODIS_Aqua_Aerosol(hidden),MODIS_Aqua_AOD_Deep_Blue​ _­Combined(hidden),MODIS_Aqua_Aerosol_Optical_Depth_3km,MODIS_Aqua_AOD_Deep_Blue_Land(hidden),Reference_Labels,Reference_Features​ (hidden),Coastlines&t=2017-05-25&z=3&v=-77.91445867492484,-12.577473819870836,-76.37857000304984,- 11.608479679245836&ab=off&as=2017-03-04&ae=2017-03-1 1&av=3&al=true.  ttps://worldview.earthdata.nasa.gov/?p=geographic&l=MODIS_Aqua_SurfaceReflectance_Bands143(hidden),MODIS_Terra_SurfaceReflectance​ c. h _Bands143(hidden),MODIS_Aqua_CorrectedReflectance_TrueColor(hidden),MODIS_Terra_CorrectedReflectance_TrueColor,MODIS_Terra​ _­NDVI_8Day(hidden),MODIS_Terra_AOD_Deep_Blue_Combined(hidden),MODIS_Terra_AOD_Deep_Blue_Land(hidden),MODIS_Terra​ _­Aerosol(hidden),MODIS_Terra_Aerosol_Optical_Depth_3km(hidden),MODIS_Aqua_Aerosol(hidden),MODIS_Aqua_AOD_Deep_Blue​ _­Combined(hidden),MODIS_Aqua_Aerosol_Optical_Depth_3km(hidden),MODIS_Aqua_AOD_Deep_Blue_Land(hidden),Reference_Labels,Reference​ _­Features(hidden),Coastlines&t=2017-04-25&z=3&v=106.0165808187288,47.38272533863788,107.5524694906038,48.35171947926288&ab=off& as=2017-03-04&ae=2017-03-1 1&av=3&al=true. Literature Review | 69 Yamuna River basin around the city. When conditions are clear, the MODIS Deep Blue product typically provides complete coverage across the region, but the Dark Target product is excluded from parts of the urban area where the background signal is high. Also, the coverage from the Dark Target algorithm varies with the season depending on the stage of vegetation growth in the sur- rounding rural area. On very bad days the smoke can be so widespread and thick that localized variations in air quality cannot be determined. During these peri- ods, the Dark Target product is sometimes not produced at all for the region. The pollution in Lima, Peru, is largely due to localized urban smog. Lima is located on the west coast of South America with weather patterns greatly influ- enced by the Humboldt Current and its location just west of the foothills of the Andes mountains. For significant portions of the year (May to November), Lima is often engulfed in fog, and during these periods the remote sensing of AOD is not possible. In addition, the background signal from the city and the surrounding des- ert is relatively bright, such that the coverage from the Dark Target product is min- imal. The Deep Blue product does provide coverage over the urban environment but with frequent gaps due to clouds. In addition, as Lima is situated on the coast, AOD retrievals are not produced for any pixels with a mix of land and water, fur- ther reducing the coverage of the Deep Blue product for this city. Ulaanbaatar, Mongolia, is in the Tuul River valley at the foot of the heavily forested Bogd Kahn Uul mountains and is surrounded by a steppe ecoregion where the largely grassland vegetation varies with the seasons. In winters, the entire region is covered with snow. The air-quality problems in Ulaanbaatar are highest in the winter months when use of coal and wood for heat gives rise to significant smoke emissions that couple with the local smog from cars and other sources. Because satellite-based AOD retrievals are not capable of distinguishing aerosol from the bright background of snow, neither the Dark Target nor Deep Blue products are useful under these conditions. At other times of year on clear days, the Deep Blue product does provide fairly good coverage over the region, but the coverage from the Dark Target algorithm is spotty, often limited only to the nearby vegetated mountainous areas and not providing information in the urban and steppe environments. CONVERTING AOD TO GROUND-LEVEL PM2.5 Many studies have attempted to convert satellite AOD to ground-level PM2.5 con- centration estimates using either statistical techniques (the first subsection below), approaches based on the chemical transport model (CTM) (discussed in the second subsection), or hybrid approaches that mix the two (discussed in the third subsection). Statistical approaches Several studies have used a purely statistical approach, where linear mixed effects models (for example, Hu, Waller, Lyapustin, Wang, Al-Hamdan, and oth- ers 2014; Sorek-Hamer and others 2015) or nonlinear generalized additive mod- els (GAMs; Sorek-Hamer and others 2013; Strawa and others 2013) have been trained on historical GLM data to predict ground-level PM2.5 using AOD and other meteorological and geographic variables as input variables. The advan- tages of statistical approaches are that the statistical models can be trained for 70 | Getting Down to Earth specific areas and are easier to run than CTMs. However, unlike CTM approaches, statistical models require a substantial amount of GLM data for the model train- ing and can be used only in the region in which they were trained. Statistical approaches also provide an estimate of the uncertainty in their pre- dictions. In addition, the uncertainty in AOD, planetary boundary layer (PBL) height, and other variables can be propagated to the ground-level PM2.5 esti- mates by running the statistical model for the high and low error bounds of the variable, and then combining that uncertainty with the estimated uncertainty of the statistical fit itself. Linear mixed-effects models Linear mixed-effects models assume a linear relationship between the predic- tors and the modeled variable but allow the intercepts and slopes to vary with the day by fitting fixed (that is, constant) and random (that is, daily varying) val- ues for those parameters, with the limitation that the random values cannot also vary with location. Routines for training these models are included in the open source R statistical program. Hu, Waller, Lyapustin, Wang, Al-Hamdan, and others (2014) used a two-stage linear mixed-effects model to estimate ground-level PM2.5 concentrations at one-kilometer resolution from MAIAC AOD data over the Southeast United States. In the first stage the predictors included MAIAC AOD, wind speed, and elevation, as well as the length of major roads, percent forest cover, and point emissions of PM2.5 within one kilometer of the site. The second stage used geo- graphically weighted regression (GWR) to predict the residuals from the first- stage model using the MAIAC AOD. GWR is an extension of least-squares regression that allows predictor coefficients to vary spatially by weighting the estimate-observation pairs according to the inverse-squared distance from indi- vidual observation sites, resulting in a spatially continuous prediction of PM2.5 over an area at one-kilometer resolution. The GLM data used in the training had 166 monitors over a domain of 800 by 1,200 square kilometers. Hu, Waller, Lyapustin, Wang, Al-Hamdan, and others (2014) found that all of the first-stage predictors were significant at an α = 0.05 level. The first-stage model explained 64 percent of the variability with a mean error of 2.8 micrograms per cubic meter and a root-mean-square error (RMSE) of 3.9 micrograms per cubic meter. The second-stage GWR model increased the variability explained to 67 percent and reduced the mean error to 2.5 micro- grams per cubic meter but had little impact on the RMSE. Sorek-Hamer and others (2015) used a linear mixed-effects model to predict ground-level PM2.5 at 10-kilometer resolution over Israel, using MODIS Deep Blue AOD data to provide better retrievals over the desert surfaces in the area. The MODIS AOD was the only predictor variable used in the model. Over Israel, this model explained 45 percent of the variability in PM2.5 and 69 percent of the variability in PM10, with RMSE of 12.1 and 27.9 micrograms per cubic meter, respectively. Lv and others (2016) developed a method to account for the spatial and tem- poral variations in PM2.5 and missing AOD observations before applying a Bayesian hierarchical mixed-effects model. This gap-filling approach uses the observed seasonal mean AOD/PM2.5 ratio at a given site and the measured PM2.5 concentration to produce an estimated AOD for the site even on days when clouds or other effects prevent AOD retrieval. Ordinary kriging (see the section “Co-kriging” in this appendix) is then applied to the retrieved and estimate AOD Literature Review | 71 values to provide a daily AOD field over the entire region. Temperature, relative humidity, and planetary boundary layer height (PBLH) were taken from NCEP forecast data, and MODIS land cover and USGS elevation data were also used as spatial predictors. The final model explained 78 percent of the variability in PM2.5 in north China and provided ground-level PM2.5 concentration fields at 12-kilometer resolution with complete spatial coverage. Generalized additive models Generalized additive models (GAMs) are a generalization of linear regression models that are able to account for the potentially nonlinear dependence of the modeled variable on the values of the predictors. The functional dependence of each predictor is determined during the fit as a linear combination of basis func- tions, with a penalty applied for the number of degrees of freedom included in each functional form. Routines for training GAMs are included in the open source R statistical program. Strawa and others (2013) used a weighted GAM to predict daily PM2.5 concen- trations at sites in the San Joaquin Valley in California. They found that a weighted GAM including MODIS dark-target AOD, ozone-monitoring instru- ment (OMI) AOD, OMI tropospheric nitrogen dioxide (NO2) columns, and a day- of-year variable explained 74 percent of the variability in PM2.5, as compared to 17 percent from a linear model. Sorek-Hamer and others (2013) also used MODIS and OMI data to predict daily PM2.5 concentrations in the San Joaquin Valley. Their GAM used MODIS Dark Target AOD, MODIS Deep Blue AOD, OMI tropospheric NO2 columns, and a day-of-year variable. This explained 61 percent of the variability in PM2.5 with an RMSE of 13.0 micrograms per cubic meter. Sources of planetary boundary layer height data A key property in determining the relationship between total column AOD and ground-level PM2.5 is the height of the planetary boundary layer (PBL; for exam- ple, Alexeeff and others 2015; Chatfield and others 2017; Lee, Chatfield, and Strawa 2016), which varies significantly with time of day, geography, season, and meteorological conditions. For example, PBLH can vary on an urban scale with distance from the city center, distance from coastlines, and other factors, as shown in figure A.6. Including predicted PBLH from meteorological model sim- ulations (especially as an AOD/PBLH ratio) has been shown to significantly increase the performance of the statistical approaches (for example, Chatfield and others 2017). When there is a well-mixed convective PBL and the aerosols in the PBL dom- inate the total AOD, there should be a nearly linear relationship between ground- level PM2.5 and the AOD/PBLH ratio, and thus for PBL variability to account for much of the variability in the relationship of AOD to PM2.5. However, for areas with very shallow PBLs, or when a smoke or dust plume is present above the PBL, the total column AOD may have no relationship to the ground-level PM2.5 at all. Thus, in those circumstances, the PBLH along with other parameters can be used to filter suspect estimates of ground-level PM2.5 concentrations. Public sources of PBLH data include in situ radiosonde and aircraft vertical profiles, ground-based and satellite remote sensing retrievals, and numerical weather prediction model data. These sources are discussed further below. Radiosonde and aircraft profiles: Various methods are available for calcu- lating the PBLH from radiosonde profiles of temperature, relative humidity, and 72 | Getting Down to Earth FIGURE A.6 Planetary boundary layer height in areas surrounding Washington, DC, US, on July 14, 2011, at 21:00 UTC 78°W 77°W 76°W 75°W 40°N 75°W 40°N 79°W 39°N 39°N 38°N 79°W 39°N 78°W 77°W 76°W Height of the planetary boundary layer (meters) 1,000 1,200 1,400 1,600 1,800 2,000 Source: World Bank, using Python software. Note: Planetary-boundary-layer (PBL) height was calculated using the Weather Research and Forecasting (WRF) model. UTC = Coordinated Universal Time. wind (for example, Seidel, Ao, and Li 2010). Though considered the gold stan- dard for observing the vertical structure of the atmosphere, radiosondes in global operational networks are generally launched only twice daily and often not at ideal times to estimate the peak daytime PBLH. Furthermore, as shown in map A.2, the radiosonde launch locations are concentrated in higher-income countries in Europe, North America, and Southeast Asia, and there is notably poorer coverage in Africa, Central Asia, and Central and South America. Therefore, the measured profiles and PBLH may not be representative of the cities in LMICs. The Aircraft Meteorological Data Relay (AMDAR), which includes data from US aircraft from the Meteorological Data Collection and Reporting System (MDCRS), is a global data set produced by commercial aircraft equipped with instruments to measure meteorological data during flights (for example, Drüe and others 2008; Fleming 1996; Zhu and others 2015). Meteorological data Literature Review | 73 MAP A.2 Radiosonde launch locations for 00:00 UTC, December 8, 2017 90° 60° 30° Latitude 0° –30° –60° –90° 0° 60° 120° 180° –120° –60° 0° Longitutde Source: The plot was generated from an application on the Naval Research Laboratory (NRL) website: http://www​ .usgodae​.org/cgi-bin/cvrg_con.cgi. Note: UTC = Coordinated Universal Time. measured at different heights during takeoff and landing have been used to cal- culate PBLH around the globe (McGrath-Spangler and Denning 2012, 2013). The AMDAR coverage is more variable than that of radiosondes but also tends to be concentrated in the higher income regions (map A.3). Ground-based remote sensing: Ground-based remote sensing systems, including lidar, ceilometer, sonic detection and ranging (SODAR), and Doppler wind profilers (DWPs), are capable of providing observations of the PBLH for field campaigns or as part of operational networks. Of these, currently only lidar networks routinely produce PBLH products available for public use. There are three networks contributing to WMO’s Global Atmospheric Watch (GAW) Aerosol Lidar Observations Network (GALION)40 that provide PBLH data: • The European Aerosol Research Lidar Network (EARLINET)41 consists of 28 stations distributed over Europe. • The Asian Dust and Aerosol Lidar Observation Network (AD-Net) includes 20 stations in Asia including Ulaanbaatar, Mongolia, and Phimai, Thailand, with NRT coverage (map A.3). • The NASA Micro-Pulse Lidar Network (MPLNET) is a federated network of currently 23 active Micro-Pulse Lidar (MPL) systems located around the globe, including stations at Kanpur, India; Omkoi, Thailand; and Windpoort, Namibia (map A.3). An improved PBLH retrieval algorithm has recently been incorporated into the latest version (version 3) of the operational product that is less susceptible to contamination by clouds and residual layers that can result in errors (Lewis and others 2013). Retrievals of PBLH with this new algorithm have been validated with PBLH calculated from ozonesonde 74 | Getting Down to Earth MAP A.3 Aircraft observation coverage from AMDAR and MDCRS, December 8, 2017 a. AMDAR 90° 60° 30° Latitude 0° –30° –60° –90° 0° 60° 120° 180° –120° –60° 0° Longitude b. MDCRS 90° 60° 30° Latitude 0° –30° –60° –90° 0° 60° 120° 180° –120° –60° 0° Longitude Source: Generated from an application on the Naval Research Laboratory (NRL) website: http://www.usgodae. org/cgi-bin​/­cvrg_con.cgi. Note: AMDAR = Aircraft Meteorological Data Relay; MDCRS = Meteorological Data Collection and Reporting System. Red lines = Canadian AMDAR data; green lines = other AMDAR data; blue lines = MDCRS data. Literature Review | 75 soundings produced during the DISCOVER-AQ field campaign and c ­ ompared to high-resolution Weather Research and Forecasting (WRF) model simula- tions (Hegarty and others 2018; Lewis and others 2013). The data are ­available in NRT (approximately one-hour processing delay). Satellite data: Only a few studies have examined PBLH using satellite data (Martins and others 2010). Radio occultation data from global positioning sys- tem (GPS) satellites have been used to determine PBLH but because of the long tangential path of the GPS signal through the atmosphere the horizontal resolu- tion is coarse ranging from tens to hundreds of kilometers (Ao and others 2012; Guo and others 2011). PBLH retrievals from the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP; Winker, Hunt, and McGill 2007; Winker and others 2009) onboard the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) satellite have been used to evaluate numerical weather prediction model reanalysis data (Jordan, Hoff, and Bacmeister 2010) during 2006. More recently, a global data set of CALIPSO PBLH retrievals was gener- ated for June 2006 to December 2012, evaluated with PBLHs calculated with AMDAR meteorological data, and used to examine global seasonal variability in the midday PBLH (McGrath-Spangler and Denning 2012, 2013). In addition, CALIPSO PBLHs were compared favorably with radiosondes over China during 2011 to 2014 (Zhang and others 2016). The CALIPSO cloud and aerosol products used to derive the PBLHs for these studies are available through the NASA Langley Data Archive Center (DARC).42 Unfortunately, the PBLHs were pro- duced independently for only these limited research projects and currently are not yet produced regularly for public dissemination. However, no continuous data sets of satellite-derived PBLHs are publicly available. NWP model data: Numerical weather prediction (NWP) models assimi- late all types of meteorological and environmental observations to produce three-­dimensional meteorological fields that are as accurate a representation of the true atmospheric state as possible at a given time. There are two types of NWP data: operational and reanalysis. Operational forecast centers run NWP models several times per day to provide numerical guidance to human weather forecasters and inputs to air-quality forecast models. Operational NWP data have the advantage of being available in near real-time, often within three to four hours of the beginning of a forecast cycle, whereas reanalysis NWP data have a latency period of several months. However, operational NWP models are continuously being updated to correct bugs and address problems in physical parameterization schemes, complicating long-term his- torical analysis. On the other hand, reanalysis NWP data are generated with models similar to those used in operational centers but frozen in development. When enough changes in the state of NWP modeling accumulates, a new ver- sion of the reanalysis, usually with the same starting date as the previous ver- sion, is generated with an updated model. Thus, reanalysis data are more appropriate for looking at temporal trends and are also subject to more thor- ough evaluations and analysis. Both types of NWP data could be used depend- ing on the time requirements on data availability of a particular research task. The observational data described above when available will be used to com- plement the NWP data, particularly for situations in which the observational data provide a better representation of the local conditions than the grid-aver- aged NWP data. 76 | Getting Down to Earth Operational NWP: The National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) is a global model with four daily run cycles beginning at 0000, 0600, 1200, and 1800 UTC (Coordinated Universal Time). For each cycle the GFS model is run for 384 hours, and the initial analysis and forecast fields are available on global latitude-longitude grids of 0.25°, 0.5°, and 1.0° resolution in grib2 format.43 Forecast output is available at each hour for the first 180 forecast hours, every 3 hours from 180 to 240 hours, and then every 12 hours until 384 hours. The analysis and forecast fields include PBLH and tem- perature, horizontal winds, vertical velocity, specific and relative humidity, geo- potential height, and cloud water mixing ratios at the surface and 30 pressure levels with 25 hecto Pascals (about 200 meters) spacing up to 900 hecto Pascals. The model output is available from the NOAA NCEP ftp server generally within 5 hours after the beginning of the forecast cycle. The Canadian Meteorological Center (CMC) produces a global NWP forecast twice daily at 0000 and 1200 UTC for 240 hours called the Global Deterministic Prediction System (GDPS).44 The CMC GDPS initial analysis and forecast data are available on a 0.24° latitude-longitude grid in grib2 format.45 The data are available every 3 hours out to 140 hours and every 6 hours to 240 hours. The variables include temperature, winds, and relative humidity at the surface and 23 pressure levels. A PBLH diagnostic is not included but could be calculated from vertical model profiles; however, the vertical resolution is coarse ranging from 15 hecto Pascals (about 120 meters) from the surface to 50 hecto Pascals (more than 450 meters) above 900 hecto Pascals, and this would affect the accuracy of the PBLH calculation. The latency time between the beginning of the model cycle and model data availability is approximately 4 to 5 hours. The JMA runs a Global Spectral Model (GSM) four times a day. The runs at 0000, 0600, and 1800 UTC are for 84 hours, and the 1200 UTC run is for 264 hours. Output data are available in grib2 format on a 0.5° latitude-longitude grid at 3-hour intervals up to 84 hours and then 6-hour intervals up to 264 hours from the WMO Global Information System Centres (GISCs) website. The PBLH diag- nostic is not included in the available output fields but could be calculated from temperature, wind, and moisture variables. The data latency time and specific access procedures were not clearly ­ documented on the GISC website. The JMA also produces a mesoscale model forecast with five-kilometer resolution for a domain centered over Japan but including eastern Mongolia. However, only graphical outputs seem to be available in the public domain. The Fleet Numerical Oceanographic Center produces a global NWP forecast with the Navy Global Environmental Model (Hogan and others 2014) four times per day with a horizontal resolution of about 37 kilometers. However, the grid- ded output fields are not readily accessible and may be restricted to organiza- tions within the United States. The European Center for Medium Range Weather Forecasts (ECMWF) and United Kingdom Meteorological Office (UKMET) also produce global NWP forecasts,46 but the output data are only provided for a licensing fee based on the amount and type of data being requested. Operational models such as the NCEP GFS are continuously being updated with new physical parameterization schemes (for example, Han and Pan 2011) and new sources of observational data to be assimilated through observing sys- tem experiments and observing system simulation experiments (for example, Atlaskin and Vihma 2012). The improvements are often evaluated at the Literature Review | 77 continental to hemispheric scale using geopotential height anomaly correlations and at regional and local scales against temperature, winds, and moisture obser- vations (for example, Cucurull and Derber 2008) but not PBLH. A Google Scholar search for evaluations of operational model PBLH outputs over land produced no results, perhaps because of the general lack of continuous PBLH observations in most locations. Nevertheless, data from operational models are used in satellite algorithms; for example, the MODIS cloud height algorithm uses GFS temperature profile data (for example, Holtz and others 2008). Thus, given the general lack of observations, it seems to be a reasonable procedure to use GFS PBLHs as an estimate of the PBLH in LMICs to determine PM2.5 from column AOD measurements. Furthermore, in 2016 NOAA announced the devel- opment of a new model to replace the current GFS. The model will still be called the GFS and but will feature a new dynamic core, the Finite Volume on a Cubed Sphere (FV3), that is expected to increase the model’s accuracy and numerical efficiency.47 The new GFS model is projected to be operational in 2019 (Schneider 2016). MERRA reanalysis: The Modern Era Retrospective-analysis for Research and Applications (Bosilovich 2008; MERRA) is a NASA reanalysis from 1979 to the present day produced using the Goddard Earth Observing System (GEOS) Data Assimilation System (DAS) Version 5 (GEOS-5 DAS; Rienecker and others 2008). The reanalysis data are on the GEOS-5 native 576 by 361 grid with 0.625° by 0.5° resolution at the surface and include PBLH and temperature, winds, rel- ative humidity at the surface, and either 42 pressure levels at three hourly inter- vals or 72 pressure levels at six-hour intervals. The 42-level product has 8 levels below 800 hecto Pascals, while the 72-level product has 12. An updated version of the reanalysis (Bosilovich and others 2015; Gelaro and others 2017; MERRA- 2) extends from 1980 to the present with a latency time of approximately one to two months. MERRA PBLH data have been evaluated using PBLH retrievals from the CALIPSO satellite (Winker, Hunt, and McGill 2007; Winker and others 2009) over a western hemisphere domain from 60° S to 60° N for August and December 2006 and Africa in August 2006 (Jordan, Hoff, and Bacmeister 2010). The evaluation indicated a better MERRA-CALIPSO correlation in August in both domains (R of 0.73) than in December (R of 0.47). Over the Sahara Desert the MERRA PBLHs were clustered around 1 to 3 kilometers above ground level and generally lower than those of CALIPSO that had sev- eral clusters around 1 to 2 kilometers, 3.5 to 4.5 kilometers, and 5 to 6 kilome- ters above ground level. The PBLHs from the atmospheric modeling component of GEOS-5 (Atmospheric General Circulation Model) used to derive MERRA have also been evaluated with micro-MPL retrievals at the NASA Goddard Space Flight Center (GSFC) in Maryland for the period of 2001 to 2008 (Lewis and others 2013). The model and MPL diurnal cycles agreed well but the model underestimated the maximum daily PBLH compared to the MPL retrievals by about 0.4 kilometers. There is inherent uncertainty in any source of PBLH data. This uncer- tainty can be attributed to three sources: (1) the methods used to calculate it from in situ observations (for example, Hegarty and others 2018; Seidel, Ao, and Li 2010), (2) the retrieval method (for example, Lewis and others 2013; Winker, Hunt, and McGill 2007), or (3) error in the NWP model (for example Jordan, Hoff, and Bacmeister 2010; Lewis and others 2013; McGrath-Spangler 78 | Getting Down to Earth and Denning 2012, 2013). This uncertainty can be near to 50 percent (for example, Lewis and others 2013; Seidel and others 2012). The uncertainty can also be dependent on the synoptic weather conditions. Hegarty and oth- ers (2018) found poor agreement between PBLHs calculated from ozone- sondes with different methods, MPL retrievals, and mesoscale model outputs on days with southerly and southwesterly flow in the Baltimore–Washington, DC, area, but good agreement between PBLHs calculated with all methods, MPL retrievals, and model data on days with northerly winds. This suggests that, whenever possible, data from both observations and models should be compared to quantify the uncertainty of the PBLH inputs to the algorithms used to determine the ground-level PM2.5 concentrations for satellite AOD retrievals. Chemical transport model–based approaches Several studies (for example, van Donkelaar, Martin, and Park 2006) have used CTMs to determine a time-varying relationship between ground-level PM2.5 concentrations and satellite AOD observations. This relationship is then used to scale the CTM aerosol profile to match the satellite-observed AOD, providing an improved estimate of the ground-level PM2.5 concentration than would be possi- ble from the CTM alone. The advantages of this approach include that it can be generalized to apply to any region of the globe, can provide daily estimates of the relationship between AOD and PM2.5 (similar to mixed effects models) (see the section “Statistical approaches” in this appendix), and can account for the impact of elevated PM2.5 layers on this relationship, thereby identifying periods where the AOD is likely dominated by aerosols above the PBL and thus do not provide good data on ground-level PM2.5 concentrations. The disadvantages are that it requires the running of a CTM, which can be a complicated and labor-intensive process, and that it is highly dependent on the ability of the CTM to correctly predict aerosol vertical profiles, even when the CTM predictions of AOD are off. However, the use of publicly available CTM output (see the section “Chemical transport model–based approaches” in this appendix) could reduce the labor issues for LMICs. Errors in AOD can be propagated through the CTM methods by repeating the process for the high and low error bounds on the AOD. However, quantifying the uncertainty in the CTM estimate of the relationship between AOD and surface concentrations is not straightforward, since this relationship is sensitive to errors in multiple aerosol components at multiple vertical levels, as well as errors in relative humidity. Thus, the method used here to account for this error is to assume a constant relative error based on post hoc validation of the CTM PM2.5 estimates (for example, ±47 percent, based on the “rest of world” results of van Donkelaar, Martin, Brauer, and Boys 2015). Van Donkelaar, Martin, and Park (2006) used MODIS Dark Target AOD and MISR AOD separately to predict annual mean ground-level PM2.5 concentrations over North America using the GEOS-Chem model (v7-02-01, 1° by 1° nested reso- lution over North America, driven with GEOS-3 meteorology) to predict the rela- tionship between AOD and PM2.5. Using MISR AOD explained 34 percent of the variability, and the MISR-predicted values had a mean bias of 3.1 micrograms per cubic meter and a slope of 0.57. Using MODIS Dark Target explained 48 percent of the variability with a mean bias of 5.1 micrograms per cubic meter and a slope of 0.82. Literature Review | 79 Van Donkelaar and others (2010) used MODIS Dark Target AOD and MISR AOD together to provide long-term average (2001–06) global estimates of ground-level PM2.5 concentrations at about 10 kilometers (0.1° by 0.1°) resolution using the GEOS-Chem model (v8-01-04, 2° by 2.5° resolution, driven with GEOS-4 meteorology) to predict the relationship between AOD and PM2.5. In the present study, the MODIS and MISR data were filtered to remove observations that had an anticipated bias greater than the larger of ±0.1 or ±20 percent. Remaining MODIS and MISR AOD retrievals were averaged to produce a single value for a grid cell. Over North America, the model explained 59 percent of the observed variability with a slope of 1.07 and a mean bias of −1.75 micrograms per cubic meter, both substantial improvements over the estimates of van Donkelaar, Martin, and Park (2006). Van Donkelaar and others (2011) used MODIS Dark Target AOD and the GEOS-Chem model (v8-03-01, 2° by 2.5° resolution, driven with GEOS-5 meteorology) to estimate daily PM2.5 concentrations during a major biomass ­ burning event around Moscow in the summer of 2010. During this event, the standard MODIS retrieval incorrectly identified some of the aerosol as cloud due to the large AOD values. Relaxing the cloud screening increased MODIS ­coverage by 21 percent with no evidence of false aerosol detection. GLM PM2.5 data were estimated from PM10 observations for several nearby sites, because only two sites had PM2.5 observations. The satellite product explained 85 percent of the variability in these estimated PM2.5 observations with a slope of 1.06. ­ Geng and others (2015) used the combined MODIS Dark Target and MISR AOD product of van Donkelaar and others (2010) to determine long-term aver- age (2006–12) PM2.5 concentrations over China at about 10 km (0.1° by 0.1°) res- olution using the GEOS-Chem model (version 9-01-02, 0.5° by 0.667° nested resolution over China, driven with GEOS-5 meteorology) to predict the relation- ship between AOD and PM2.5. Comparison with ground-level PM2.5 concentra- tions observations showed the satellite-based product explained 55 percent of the variability in PM2.5 with a slope of 0.77. Van Donkelaar and others (2015) used satellite AOD observations and the GEOS-Chem model to produce annual average PM2.5 estimates (1998–2012) over the globe. The MODIS radiance observations were used in an optimal esti- mation framework (van Donkelaar and others 2013) to derive PM2.5 estimates that were consistent with the GEOS-Chem aerosol scheme for 2004–10, which were then combined with the product of van Donkelaar and others (2010) for 2001–03 to produce a global, decadal average PM2.5 estimate at about 10-kilome- resolution. The work of Boys and others (2014), which used AOD from the ter ­ SeaWIFS and MISR satellites and the GEOS-Chem model to estimate the tem- poral variation in PM2.5, was then applied to this decadal average to produce a 15-year global estimate of ground-level PM2.5. Comparisons with decadal mean GLM data over North America gave an R of 58 percent with a slope of 0.96 and a 1−σ error of −1 microgram per cubic meter + 16 percent. Over Europe, the com- parison gave an R of 53 percent with a slope of 0.78 and a 1−σ error of 1 micro- gram per cubic meter +21 percent, whereas over the rest of the world the R was 0.66, the slope was 0.68, and the 1−σ error was −1 microgram per cubic meter + 47 percent. Sources of AOD-PM2.5 relationships from observations Until recently, few data were available from colocated observations of AOD and ground-level PM2.5. The Surface Particulate Matter Network (SPARTAN) was established to address this need (Snider and others 2015). The network includes 80 | Getting Down to Earth a global federation of ground-level monitors of hourly PM2.5, primarily in highly populated regions in proximity to existing ground-based sun photometers (for example, AERONET sites) that measure AOD. Together, these instruments pro- vide an empirical measure of the AOD/PM2.5 ratio that is used to relate satellite AOD retrievals to ground-level PM2.5. The current SPARTAN network is fairly sparse, but adding an AERONET and SPARTAN site to a given city would provide valuable data on the local variation of the AOD/PM2.5 ratio at a relatively low cost. However, in the absence of these data, this ratio can be estimated from CTM output. Sources of aerosol profile data from CTMs For LMICs to use the CTM approach without having to run CTMs themselves, they need publicly available CTM data sets that can be used to match aerosol profiles with AOD observations. Several potential sources of these data for LMICs are discussed below. MERRA-2 Reanalysis: The NASA MERRA-2 reanalysis (Provençal and oth- ers 2017) provides hourly average AOD at 550 nm both by total and by aerosol component, as well as the surface mass concentration for each component, at a horizontal resolution of 0.5° by 0.625°. However, aerosol vertical profiles are only provided every three hours as instantaneous (snapshot) profiles. MERRA-2 uses the NASA GEOS-5 atmospheric general circulation model (Colarco and others 2014). The aerosol component of MERRA-2 is based on the Goddard Chemistry Aerosol Radiation and Transport Model (GOCART; Chin and others 2000; Colarco and others 2010, 2014; Kim and others 2013). Funded mainly by NASA Earth Science programs, the GOCART model was developed to simulate atmospheric aerosols (including sulfate, black carbon, organic carbon, dust, and sea salt), carbon monoxide, and sulfur gases. The reanalysis has about a two- month latency. All of the data are freely available for download from NASA via MDISC,48 managed by the NASA Goddard Earth Sciences (GES) Data and Information Services Center (DISC). NCEP NGAC forecasts: NEMS GFS Aerosol Component (NGAC; Lu, da Silva, and others 2016) is a global inline aerosol forecast system. The forecast model component of the NGAC is the GFS based on the NOAA Environmental Modeling System (NEMS; Black and others 2007, 2009) which, in turn, is based on the common modeling framework using Earth System Modeling Framework (ESMF). The aerosol component of the NGAC is GOCART (Chin and others 2000; Kim and others 2013). The initial production implementation of NGAC with global-dust-only forecast was implemented in September 2012 (Lu, da Silva, and others 2016). The current NGAC operational forecast produces 120-hour global multispe- cies forecasts including dust, sea salt, sulfate, and carbonaceous aerosols. The system runs twice daily within the NCEP Production Job Suite at the 0000Z cycle and 1200Z cycle. Although no satellite data assimilation is cur- rently performed, research is underway to add the assimilation of the NOAA VIIRS AOD product to NGAC (Lu, Wei, and others 2016). Output is posted to a 1° by 1° longitude/latitude grid with a 3-hour forecast interval to 120 hours. The files (in grib2 format) include three-dimensional profiles of individual aerosol components (dust, sea salt, sulfate, organic carbon, and black car- bon), temperature, and relative humidity, as well as two-dimensional fields including the model-calculated total AOD and the AOD from each aerosol component. Literature Review | 81 NCAR MOZART-4/GEOS-5 forecasts: NCAR provides output from their MOZART-4/GEOS-5 global air-quality simulations (Emmons and others 2010) for public use, both in NRT and in a historical archive going back to 2007.49 The representation of tropospheric aerosols in MOZART-4 is based on the work of Tie and others (2001, 2005) and includes the calculation of sulfate, black carbon, primary organic, secondary organic aerosols (SOAs), ammonium nitrate, and sea salt (Lamarque and others 2005). Because only the bulk mass is calculated, a lognormal number distribution is assumed for all aerosols to calculate the sur- face area (and thus AOD), using a different geometric mean radius and standard deviation for each type of aerosol (based on Chin and others 2002). Sea salt aero- sols are included in the model with four size bins (0.1–0.5, 0.5–1.5, 1.5–5, and 5–10 micrometers), and emissions are calculated online (Mahowald, Lamarque, and others 2006). However, the distributions of four sizes of dust (0.05–0.5, 0.5–1.25, 1.25–2.5, and 2.5–5.0 micrometers) are set from monthly mean distributions taken from online calculations in the Community Atmosphere Model (CAM) (Mahowald, Muhs, and others 2006), and thus the model does not include prog- nostic dust. Hygroscopic growth of the aerosols is determined from the ambient relative humidity, with different rates for each type of aerosol (Chin and others 2002). Washout of all aerosols, except hydrophobic black carbon and organic carbon, is set to 20 percent of the washout rate of nitric acid (HNO3) (Horowitz 2006; Tie and others 2005). Comparisons of calculated AOD over the ocean to AOD retrievals from the MODIS satellite instrument indicate this is a reasonable washout rate. The NCAR MOZART-4/GEOS-5 aerosol profiles are provided at a horizontal resolution of 1.875° by 2.5° every six hours. However, the output files include only the bulk mass for each size bin, and so the model AOD has to be calculated from these data by the user. UK Met Office research ensemble forecast: The UK Met Office provides an operational global forecast at 17-kilometer resolution that includes prognostic and interactive dust for a four-member hybrid ensemble. An additional aerosol forecasting global model, including carbonaceous aerosols and sulfate aerosols, is run as a research product using a copy of the operational model, with aerosol chemistry using the CLASSIC scheme (Bellouin and others 2011; Mulcahy and others 2014). However, the additional aerosols do not feed back to the dynamics. This research product is available only via the principal investigator (Malcolm Brooks, malcolm.e.brooks@metoffice.gov.uk). ECMWF Copernicus Atmospheric Monitoring Service: The ECMWF Copernicus Atmospheric Monitoring Service (CAMS) produces twice-daily 120-hour global forecasts of total aerosols, aerosol components (dust, sea salt, biomass burning, sulfate), and AOD at about 40-kilometer spatial resolution (0.5625° by 0.5625°) at a three-hour temporal resolution. CAMS also produces the Monitoring Atmospheric Chemistry and Climate (MACC) reanalysis (Toll and others 2015) of AOD and component AOD at about 80-kilometer resolution (1.125° by 1.125°). However, both products currently require purchasing a license from ECMWF and thus may not be the correct solution for many LMICs. CAMS currently uses a simple bin scheme (IFS-LMD [Integrated Forecast System–Laboratoire de Météorologie Dynamique]) for its near real-time fore- casts and the MACC reanalysis. IFS-LMD mainly follows the aerosol treatment in the Laboratoire d’Optique Atmosphérique / LMD-Z Laboratoire de Météorologie Dynamique–Zoom (LOA/LMD-Z) model (Boucher and others 2003; Reddy and others 2005). Five types of tropospheric aerosols are 82 | Getting Down to Earth considered: sea salt, dust, organic carbon, black carbon, and sulfate aerosols. Prognostic aerosols of natural origin, such as mineral dust and sea salt, are described using three size bins. Emissions of dust depend on the 10-meter wind, soil moisture, the UV-VIS component of the surface albedo, and the fraction of land covered by vegetation when the surface is snow-free. A correction to the 10-meter wind to account for gustiness is also included (Morcrette and others 2009). Sea-salt emissions are diagnosed using a source function based on work by Guelle and others (2001) and Schulz, de Leeuw, and Balkanski (2004). Sources for the other aerosol types are taken from the Speciated Particulate Emission Wizard (SPEW) and the Emission Database for Global Atmospheric Research (EDGAR) annual or monthly mean climatologies (Dentener and others 2006). Emissions of organic matter, black carbon, and SO2 linked to fire emissions are obtained using the Global Fire Assimilation System based on MODIS satellite observations of fire radiative power, as described in Kaiser and others (2012). In addition, the ECMWF forecasts assimilate MODIS AOD data at 550 nano- meters in a 4D-Var framework that has been extended to include the aerosol total mixing ratio as an extra control variable (Benedetti and others 2009). Hybrid approaches Hybrid approaches are those in which both the output from CTMs and statisti- cal models are used to convert satellite AOD retrievals into predictions of ground-level PM2.5. For example, van Donkelaar, Martin, Spurr, and Burnett (2015) used geographically weighted regression (GWR) to relate the errors between the GLM PM2.5 observations and their estimate of ground-level PM2.5 based on MODIS AOD observations and GEOS-Chem simulations (van Donkelaar and others 2013). They used the Matlab function glmfit, but open- source options in Python are available (for example, the pyglmnet library). The spatial predictors used were the percentage of urban land cover (at one-kilometer resolution) and the subgrid elevation difference (that is, the dif- ference between the mean elevation for a 0.5° by 0.67° GEOS-Chem grid box, the local elevation from high-resolution maps). In addition, the GEOS-Chem pre- dictions for the relative contributions of nitrate, primary carbonaceous, and sec- ondary organic aerosol to the total PM2.5 were used as additional predictors. GWR was then used to predict the bias between monthly mean ground-level PM2.5 estimates and the satellite-derived estimate of ground-level PM2.5. This gave spatially varying, but constant with time, bias-correction estimates at a fine resolution (one kilometer) over North America. However, although the bias-cor- rection is constant with time, the uncorrected values vary with time (daily for areas with MODIS coverage) and thus provide daily estimates of ground-level PM2.5. The bias-corrected satellite product showed a much higher correlation with surface observation and a better regression slope but had similar RMSE with respect to surface monitors as the uncorrected product. COMBINING GLM AND SATELLITE DATA The two most common methods that have been used to simultaneously interpo- late satellite-derived PM2.5 estimates and GLM data to obtain city-scale PM2.5 esti- mates are co-kriging and simultaneous land-use regression (for example, Alexeeff and others 2015; de Hoogh and others 2016; Lee, Chatfield, and Strawa 2016). Literature Review | 83 Co-kriging Like most spatial interpolation schemes, ordinary kriging uses a weighted aver- age of neighboring samples (that is, from a spatially distributed set of ground- level monitors of PM2.5) to predict PM2.5 concentrations across an urban area. However, in ordinary kriging the weights depend not only on the distance between the prediction location and the measured points, but also on the spatial autocorrelation in the data (Millar and others 2010). Kriging also provides error estimates for its predictions. Co-kriging is an extension of ordinary kriging that can take advantage of additional data sets or variables, using both the correlations between the monitors and the cross-correlations between the monitor data and the addi- tional data sets to make better predictions. In the application of interest here, the second data set would be the satellite-derived ground-level PM2.5 esti- mates produced using one of the methods outlined earlier in this appendix, but additional data sets, such as meteorological data, can be incorporated via co-kriging (Millar and others 2010; Pearce and others 2009; Wu, Winer, and Delfino 2006). Kriging is generally done using licensed software such as ArcGIS, but open source options (for example, the pyKrige toolkit in Python) also are available. For LMICs, the advantage of combining satellite data with GLM data via co-kriging is that it potentially reduces the number of GLM stations required to adequately cover an urban area via ordinary kriging. Kriging is most suitable for cases where the GLM network (1) is spatially dense, (2) captures the general spatial variation of PM2.5, and (3) adequately covers the edges of the study area. Applying kriging to sparse data will either result in over-smoothed surfaces or over-fitted predictions. Using the more spatially dense satellite estimates of sur- face level PM2.5 in co-kriging could thus address these limitations of kriging for urban areas with sparse GLM networks. Co-kriging has other potential advantages. It does not require any data beyond the GLM network data and the satellite-derived ground-level PM2.5 esti- mates. In contrast, land-use regression requires substantial additional data (see the sections “Co-kriging” and “Land-use regression” in this appendix). Furthermore, co-kriging can be applied independently at each time point, so that the temporal resolution of the maps produced can be the same as the initial GLM and satellite data, allowing for short-term average (that is, daily) predictions of PM2.5 concentrations. In fact, Alexeeff and others (2015) showed that to estimate chronic (annual) exposure to PM2.5, kriging applied to annual average data was insufficient to capture the exposure accurately, and thus kriging had to be applied to the daily data and then averaged to provide annual estimates. However, the spatial resolution of the predictions is limited by the spatial density and res- olution of the input data sets, making predictions on a one-kilometer scale or finer difficult. An example of the use of co-kriging to supplement a sparse GLM network was provided by Singh and others (2011), who demonstrated how CTM predic- tions of ground-level PM2.5 concentrations could be used as a secondary vari- able in a co-kriging interpolation of GLM data to predict PM10 concentration over northern Italy. They used 14 PM10 monitors and five-kilometer resolution CTM predictions over a 60-kilometer by 60-kilometer domain. Their co-krig- ing models were able to explain 88–90 percent of the variability in PM10 with an RMSE of 16 micrograms per cubic meter, which was lower than the kriging 84 | Getting Down to Earth value (20 micrograms per cubic meter). In addition, they found that the pre- dicted maps of PM10 concentration were substantially different between krig- ing and co-kriging, with the co-kriging map capturing more of the expected special variation of PM10 in the domain. In addition, Singh, Venkatachalam, and Gautam (2017) used co-kriging to fill in missing MISR AOD data using observations from the MODIS instruments. However, there is no example of co-kriging with satellite-based PM2.5 products in the literature. Land-use regression Land-use regression (LUR) spatially links GLM network data of PM2.5 with other associated variables (such as elevation, distance from roads, population density, and land-use type) to develop finer-scale and more accurate estimates of long- term PM2.5 concentrations within an urban area than is possible with kriging (Millar and others 2010). The most predictive variables are identified through least-squares linear regression (Briggs and others 1997). Other studies (for exam- ple, Vienneau and others 2013) have incorporated data on meteorology (for example, wind direction; Arain and others 2007; Ryan and others 2008) and sat- ellite observations of aerosols, NO2, and other pollutants (for example, Lee, Chatfield, and Strawa 2016) into the regression as well. The development of pre- dictor variables from geographic information system (GIS) files is generally per- formed using licensed software such as ArcGIS, but the regression can be performed using open-source software such as R. LUR is an attractive option for estimating PM2.5 concentrations at a neighbor- hood scale within urban areas with sparse measurement networks, since the regression can be used to predict concentration at the scale of the land-use data in unmonitored areas. However, the method assesses only spatial variation and thus can predict only long-term average concentrations (for example, monthly or annual). In addition, the method generally requires a significant amount of GIS skill to apply and requires much more input data than co-kriging (see the section “Co-kriging” in this appendix). Several studies have used satellite estimates of ground-level PM2.5 within LUR models. For example, Vienneau and others (2013) used MODIS-derived ground- level PM2.5 estimates (at 10-kilometer resolution) as a predictor in a land-use regression model to estimate annual average PM10 concentrations over Europe at a resolution of 100 meters. The LUR models using the satellite data had consis- tently higher correlations and generally lower mean bias and mean absolute errors (MAEs) than the LUR models that did not use satellite data. The models using satellite data to predict annual average PM10 generally explained between 35 and 50 percent of the variability in PM10 (mean absolute error of 4.6 to 6.0 micrograms per cubic meter). Lee, Chatfield, and Strawa (2016) adopted a different approach in which LUR is included in the linear mixed-effects statistical model used to relate MODIS Deep Blue AOD to ground-level PM2.5 estimates for California (see, for example, the section “Statistical approaches” in this appendix). Separate LUR models were built for each season and trained using three years of data from 142 GLM sites. The land-use predictors included distance from major highways, popula- tion density, estimated pollutant emissions from the US EPA National Emission Inventory, elevation, percentage of developed land, distance from the coast, and the air basin in which each measurement site was located. In addition, daily val- ues for AOD, wind speed, relative humidity, temperature, and PBLH from the Literature Review | 85 weather stations nearest to the site were also used as predictors. This addition of temporally varying data with land-use data allows for predictions of daily and annual average PM2.5 at the fine spatial scales possible with land-use regression. Lee, Chatfield, and Strawa (2016) found that the model was able to explain 66 percent of the variability of PM2.5 over California (MAE of 3.7 micrograms per cubic meter, RMSE of 5.7 micrograms per cubic meter) and that the modeled relationship between AOD and PM2.5 varied significantly by day. De Hoogh and others (2016) used the 10-kilometer resolution satellite-­ derived estimated of ground-level PM2.5 from van Donkelaar, Martin, Spurr, and Burnett (2015) as a predictor in a land-use regression model of PM2.5 over west- ern Europe, predicting annual average PM2.5 concentrations at a resolution of 100 meters. The model also included satellite estimates of tropospheric NO2 col- umns at a 10-kilometer resolution, as well as independent CTM predictions of ground-level PM2.5 and NO2 from the MACC-II ENSEMBLE model, which was the median value of seven individual regional CTMs (CHIMERE, EMEP, EURAD, LOTOS-EUROS, MATCH, MOCAGE, and SILAM). Additional predic- tors included high-resolution data on land cover, roads, and altitude. They found that the PM2.5 LUR models that included both the satellite and CTM estimates of ground-level PM2.5 explained about 60 percent of the spatial variation in mea- sured (annual average) PM2.5 concentrations, substantially more than was explained by the LUR model without satellite or CTM data (33–38 percent). Data sets for land-use regression If an LMIC has detailed GIS data on land use, population density, roads, pollut- ant emissions, and other fields needed for LUR, those should be used, because they would reflect the best local knowledge regarding the conditions likely to lead to high levels of PM2.5. However, since some LMICs may not have these data freely available, high-resolution, global data sets for the predictors used in the LUR studies discussed above were identified in this review: • Land-cover type can be determined from the 0.5-kilometer MODIS-based Global Land Cover Climatology provided by USGS.50 This map is based on based on 10 years (2001–10) of Collection 5.1 MCD12Q1 land-cover–type data and used 17 land-cover categories. The map is generated by choosing, for each pixel, the land cover classification with the highest overall confidence from 2001 to 2010 (Broxton and others 2014). These data can also be used to calcu- late parameters such as the percentage of developed land within a given radius of the site. • Distance from coasts can be determined using the World Water Bodies GIS layer.51 • GIS data on population density52 and major roads53 are available through the NASA Socioeconomic Data and Applications Center (SEDAC) hosted at Columbia University. • The SEDAC Gridded Population of the World, version 4 (GPWv4) Population Density product (CIESIN 2016) consists of estimates of human population density, based on counts consistent with national censuses and population registers, for the years 2000, 2005, 2010, 2015, and 2020. A proportional allo- cation gridding algorithm, utilizing approximately 12.5 million national and subnational administrative units, is used to assign population values to 30-arc-second (about one-kilometer) grid cells. 86 | Getting Down to Earth • The SEDAC Global Roads Open Access Data Set (CIESIN and ITOS 2013; gROADS) combines the best available roads data by country into a global roads data set, using the UN Spatial Data Infrastructure Transport (UNSDI-T) version 2 as a common data model. Because the data are compiled from mul- tiple sources, the date range for road network representations ranges from the 1980s to 2010 depending on the country (most countries have no con- firmed date), and spatial accuracy varies. • In addition, OpenStreetMap54 is an open source, community project that maintains data about roads, trails, railway stations, and other landmarks, all over the world. OpenStreetMap emphasizes local knowledge, and contribu- tors use aerial imagery, GPS devices, and low-tech field maps to verify that OpenStreetMap is accurate and up to date. The inspection of these data sug- gests that the data are more detailed and occasionally more accurate than the gROADS database, but the data sources are also less well documented. • Topography data can be obtained from the Global Multi-resolution Terrain Elevation Data 2010 (GMTED2010)55 created by the USGS and the National Geospatial-Intelligence Agency (NGA). This data set has multiple spatial res- olutions, 30, 15, and 7.5 arcseconds, and is a replacement for the previous GTOPO30 data. • Spatially resolved global air pollutant emissions at a 0.1° by 0.1° resolution (about 10 kilometers) are available for the years 2008 and 2010 from the EDGAR-HTAP_V2 emission inventory for several pollutants (CH4, CO, SO2, NOx, nonmethane volatile organic compounds, NH3, PM10, PM2.5, black car- bon, and organic carbon). RECOMMENDATIONS FOR LMICS These recommendations for the use of satellite observations to supplement GLM data in LMICs (using the typology proposed in chapter 1 and reproduced at the beginning of this appendix) are based on the literature review above: • For Type I countries, since no GLM data are available, the only possible approach to convert AOD to ground-level PM2.5 for these countries is a CTM- based approach. Bias correction of the CTM-based estimates is also not pos- sible for these countries, because there are no data with which to characterize the model bias. Thus, the raw CTM-based estimates of ground-level PM2.5 at the native resolution of the satellite AOD product should be used, but the derived ground-level PM2.5 values should be assigned a high uncertainty that reflects not only the uncertainty in the AOD but also the estimated uncer- tainty in the CTM-derived AOD to PM2.5 relationship (for example, about 50 percent based on van Donkelaar, Martin, Brauer, and Boys 2015). It is likely in this case that only city-average estimates will be possible, and that the annual estimates will be more reliable than the daily estimates. • For Type II countries, the small amount of available GLM data (with variable quality) will provide at least some ability to derive a bias estimate for the raw CTM-based estimates of ground-level PM2.5 discussed above. However, with- out improvement in the QA procedures in these countries, it will not be pos- sible to be certain that the bias-corrected estimate is truly more accurate than the raw estimate, and thus both values should be reported and stored. This will allow for reprocessing of the satellite estimates when more rigorous QA Literature Review | 87 procedures are developed. Similar to Type I countries, it is likely that only city-average estimates will be possible, and that the annual estimates will be more reliable than the daily estimates. • The GLM data in Type III countries allow both statistical and bias-­ corrected CTM-based approaches for converting AOD to PM2.5 to be con- sidered. CTM-based approaches may still be the best approach for these countries, but statistical approaches should also be tested before making a final decision. If GLM data from several sites exist, but only for a short period, those data can be used to develop seasonally and geographically varying statistical models and/or seasonally and geographically varying bias corrections for CTM-based approaches. If only a single site exists in the urban area, a seasonally varying statistical model and/or bias correc- tion is still possible, but both the model and the bias correction would have to be applied uniformly across the urban area. Land-use regression (LUR; see the section “Land-use regression” in this appendix) can be combined with either technique to provide a finer-scale estimate of annual average ground-level PM2.5 (at about 0.5-kilometer resolution) using the data sets described in the section “Data sets for land-use regres- sion” in this appendix, and co-kriging (see the section “Co-kriging” in this appendix) can be used to derive daily estimates across the city. • Type IV countries will be able to take advantage of the same approaches out- lined above for Type III countries. However, the more extensive measure- ment networks will allow for more accurate estimates of the geographical and seasonal variation in the AOD to PM2.5 relationship. Thus, at this point, purely statistical approaches may begin to outperform CTM-based estimates, espe- cially for regions where the model’s emission inventories are out of date or otherwise inaccurate. • In Type V countries, the main use of satellite observations is to help fill in the gaps of the existing monitoring network to cover more of the country’s popu- lation. These countries also tend to have access to higher-resolution, regional CTM simulations that are customized for their region. Thus, both statistical and CTM-based approaches can be used. In addition, the following recommendations on data sources and other meth- odological concerns are made for all countries interested in using satellite data to supplement GLM data: • All countries should use the satellite product that offers the best balance of accurate AOD and spatial resolution for their country. For example, regions near bright urban or desert surfaces should prefer the MODIS Deep Blue product (see the section “Polar-orbiting satellites” in this appendix), since this has been demonstrated to give better performance in these regions. The extension of this algorithm to VIIRS will help maintain this capability over the next two decades. However, the MAIAC algorithm or the application of the Deep Blue algorithm to geostationary observations may provide better AOD observations in the near future. • The highest-possible spatial resolution for satellite AOD will likely be between 2 and 4 kilometers (for geostationary observations) or 1 and 10 kilometers (for polar observations). Consequently, finer predictions will be ­ possible only for cases where GLM data exist (via LUR; see the section “­Land-use regression” in this appendix). 88 | Getting Down to Earth • Some cities may have poor satellite coverage due to clouds or snow-covered surfaces regardless of the choice of the AOD product. This will reduce the number of days for which a daily estimate is possible and will need to be accounted for in the annual averages. • If local data on PBLH are available in a city (for example, Ulaanbaatar, Mongolia), the measured PBLH should be used in statistical approaches. For countries without the capacity to perform their own meteorological or • chemical transport modeling, the freely available global model data sets pro- vided by organizations in the United States for estimates of PBL height (see the section “Statistical approaches” in this appendix), aerosol vertical profiles (see the section “Chemical transport model–based approaches” in this appen- dix), and other parameters should be used. For long-term studies, the NASA MERRA-2 reanalysis should be used, because this will represent the best esti- mate of the historical atmospheric state. However, the reanalysis takes two months to produce, and thus for short-term forecasts and advisories the NCEP GFS output (including the NGAC aerosol forecasts) should be used. • In all cases, the errors from the AOD retrieval, the CTM estimate of the aero- sol profiles, and the statistical models used (either directly or for bias correc- tion) should be quantified and used to quantify the error in the derived PM2.5 estimate. • In addition, cases where the NCEP NGAC forecasts and/or the MERRA-2 reanalyses suggest that the AOD is dominated (more than 70 percent) by aerosols above the PBL height should be flagged as potentially poor-quality satellite data, and thus poor-quality ground-level PM2.5 estimates. NOTES  1. https://www.wmo-sat.info/oscar/gapanalyses?variable=6.  2. https://darktarget.gsfc.nasa.gov/.  3. https://deepblue.gsfc.nasa.gov/.  4. This is a formula for the total error, which depends on the AERONET AOD value. At low AOD values, the error will be at least 0.05 whereas at high AOD values the error will be at 15 percent.  5. https://ladsweb.modaps.eosdis.nasa.gov/tools-and-services/#stage_MODIS.pl.  6. https://ladsweb.modaps.eosdis.nasa.gov/search/order/2/MOD04_L2--6,MYD04_L2​ --6,MOD04_3K--6,MYD04_3K--6.  7. https://worldview.earthdata.nasa.gov.  8. https:// ladsweb.modaps.eosdis.nasa.gov/missions-and-measurements/modis​ /­MAIAC_ATBD_v1.pdf.  9. https://ladsweb.modaps.eosdis.nasa.gov/missions-and-measurements/science-domain​ /­maiac/. 10. https:// ladsweb.modaps.eosdis.nasa.gov/missions-and-measurements/products​ /­MCD19A2/. 11. https://ladsweb.modaps.eosdis.nasa.gov/archive/allData/6/MCD19A2/. 12. https://ladsweb.modaps.eosdis.nasa.gov/archive/allData/6/MCD19A2/. 13. https://landweb.modaps.eosdis.nasa.gov/cgi-bin/QA_WWW/newPage.cgi. 14. https://landweb.modaps.eosdis.nasa.gov/cgi-bin/QA_WWW/newPage.cgi?fileName​ =sciTestMenu_C6. 15. https://landweb.modaps.eosdis.nasa.gov/QA_WWW/forPage/C6_test_MOD19.html. 16. https://modis.gsfc.nasa.gov/sci_team/meetings/201606/presentations/land/devadiga.pdf. 17. https://www.star.nesdis.noaa.gov/smcd/emb/viirs_aerosol/index.php/. 18. https://www.avl.class.noaa.gov/saa/products/search?datatype_family=VIIRS_EDR. 19. https://www.avl.class.noaa.gov/saa/products/search?datatype_family=VIIRS_EDR. Literature Review | 89 20. https://deepblue.gsfc.nasa.gov/data. 21. https://misr.jpl.nasa.gov/getData/accessData/. 22. https://www.ncdc.noaa.gov/cdr/atmospheric/avhrr-aerosol-optical-thickness. 23. https://portal.nccs.nasa.gov/datashare/AVHRRDeepBlue/. 24. https://www.eumetsat.int/what-we-monitor/atmosphere. 25. https://www.eumetsat.int/what-we-monitor/atmosphere. 26. https://www.star.nesdis.noaa.gov/smcd/emb/aerosols/products_geo.php. 27. http://www.ssd.noaa.gov/PS/FIRE/GASP/gasp.html. 28. http://www.icare.univ-lille1.fr/projects/seviri-aerosols. 29. https://www.goes-r.gov. 30. https://www.goes-r.gov/products/baseline-aerosol-opt-depth.html. 31. https://www.goes-r.gov/downloads/users/conferencesAndEvents/2014/GOES-R_Series​ _Program/04-Laszlo_pres.pdf. 32. https://www.avl.class.noaa.gov/saa/products/welcome. 33. https://www.ncdc.noaa.gov/goes-r-series-satellites/goes-r-series-frequently​ -asked-questions. 34. https://www.star.nesdis.noaa.gov/jpss/documents/ATBD/ATBD_EPS_Aerosol​ _AOD_v3.0.1.pdf. 35. https://www.star.nesdis.noaa.gov/star/documents/meetings/2015JPSSAnnual​ /­d ayFive/06_Session9_Wolf_STAREnterpriseAlgorithmsPlanJPSSAnnualMeeting​ 082815.pdf. 36. https://www.jma.go.jp/jma/jma-eng/satellite/index.html. 37. http://www.data.jma.go.jp/mscweb/technotes/msctechrep61-6.pdf. 38. http://www.jmbsc.or.jp/en/meteo-data.html. 39. http://www.jmbsc.or.jp/en/Data/Himawari-8-JMBSC-HP(2017.02.20).pdf. 40. https://www.gaw-wdca.org/. 41. https://www.earlinet.org/index.php?id=earlinet_homepage. 42. http://www-calipso.larc.nasa.gov. 43. http://www.nco.ncep.noaa.gov/pmb/products/gfs. 44. https://weather.gc.ca. 45. https://weather.gc.ca/grib/grib2_glb_25km_e.html. 46. https://www.metoffice.gov.uk/services/data. 47. http://www.noaa.gov/media-release/noaa-to-develop-new-global-weather-model. 48. https://disc.sci.gsfc.nasa.gov/datasets?page=1&keywords=MERRA-2. 49. https://www.acom.ucar.edu/wrf-chem/mozart.shtml. 50. https://lpdaac.usgs.gov/products/mcd12q1v006/. 51. https://www.arcgis.com/home/item.html?id=e750071279bf450cbd510454a80f2e63. 52. http://sedac.ciesin.columbia.edu/data/set/gpw-v4-population-density/data-download. 53. http://sedac.ciesin.columbia.edu/data/set/groads-global-roads-open-access-v1​ /­data-download. 54. http://www.openstreetmap.org. 55. https://www.usgs.gov/core-science-systems/eros/coastal-changes-and-impacts​ /­gmted2010?qt-science_support_page_related_con=0#qt-science_support_page_related​ _con. REFERENCES Alexeeff, S. E., J. Schwartz, I. Kloog, A. Chudnovsky, P. Koutrakis, and B. A. Coull. 2015. “Consequences of Kriging and Land Use Regression for PM2.5 Predictions in Epidemiologic Analyses: Insights into Spatial Variability Using High-Resolution Satellite Data.” Journal of Exposure Science & Environmental Epidemiology 25 (2): 138–44. Ao, C. O., D. E. Waliser, S. K. Chan, J. L. Li, B. Tian, F. Xie, and A. J. Mannucci. 2012. “Planetary Boundary Layer Heights from GPS Radio Occultation Refractivity and Humidity Profiles.” Journal of Geophysical Research 117: D16117. Arain, M. A., R. Blair, N. Finkelstein, J. R. Brook, T. Sahsuvaroglu, B. Beckerman, L. Zhang, and M. Jerrett. 2007. “The Use of Wind Fields in a Land Use Regression Model to Predict Air Pollution Concentrations for Health Exposure Studies.” Atmospheric Environment 41 (16): 3453–64. 90 | Getting Down to Earth Atlaskin, E., and T. Vihma. 2012. “Evaluation of NWP Results for Wintertime Nocturnal Boundary-Layer Temperatures over Europe and Finland.” Quarterly Journal of the Royal Meteorological Society 138: 1440–51. Bellouin, N., J. Rae, J. A. Jones, C. Johnson, J. Haywood, and O. Boucher. 2011. “Aerosol Forcing in the Climate Model Intercomparison Project (CMIP5) Simulations by HadGEM2-ES and the Role of Ammonium Nitrate.” Journal of Geophysical Research 116: D20206. Benedetti, A., J. J. Morcrette, O. Boucher, A. Dethof, R. Engelen, M. Fisher, H. Flentje, N. Huneeus, L. Jones, J. Kaiser, S. Kinne, A. Mangold, M. Razinger, A. J. Simmons, and M. Suttie. 2009. “Aerosol Analysis and Forecast in the European Centre for Medium-Range Weather Forecasts Integrated Forecast System: 2. Data Assimilation.” Journal of Geophysical Research: Atmospheres (1984–2012) 114: D13205. Bernard, E., C. Moulin, D. Ramon, D. Jolivet, J. Riedi, and J.-M. Nicolas. 2011. “Description and Validation of an AOT Product over Land at the 0.6 μm Channel of the SEVIRI Sensor Onboard MSG.” Atmospheric Measurement Techniques 4: 2543–65. Black, T., H. M. H. Juang, and M. Iredell. 2009. “The NOAA Environmental Modeling System at NCEP.” Preprint for the American Meteorological Society, Joint Session 2A.6, “23rd Conference on Weather Analysis and Forecasting/19th Conference on Numerical Weather Prediction,” Omaha, NE, June 1–5, 2009. Black, T., H. M. H. Juang, W. Y. Yang, and M. Iredell. 2007. “An ESMF Framework for NCEP Operational Models.” Paper presented at the American Meteorological Society, Joint Session J3.1, “22nd Conference on Weather Analysis and Forecasting/18th Conference on Numerical Weather Prediction,” Park City, UT, June 25–29, 2007. Bosilovich, M. G. 2008. “NASA’s Modern Era Retrospective-Analysis for Research and Applications: Integrating Earth Observations.” Earthzine (September 26, 2008). https:// earthzine.org/nasas-modern-era-retrospective-analysis/. Bosilovich, M. G., S. Akella, L. Coy, R. Cullather, C. Draper, R. Gelaro, R. Kovach, Q. Liu, A. Molod, P. Norris, K. Wargan, W. Chao, R. Reichle, L. Takacs, Y. Vikhliaev, S. Bloom, A. Collow, S. Firth, G. Labow, G. Partyka, S. Pawson, O. Reale, S. D. Schubert, and M. Suarez. 2015. “MERRA-2: Initial Evaluation of the Climate.” Technical Report Series on Global Modeling and Data Assimilation no. 43. Greenbelt, MD: NASA Goddard Space Flight Center. Boucher, O., C. Moulin, S. Belviso, O. Aumont, L. Bopp, E. Cosme, R. von Kuhlmann, M. G. Lawrence, M. Pham, M. S. Reddy, J. Sciare, and C. Venkataraman. 2003. “DMS Atmospheric Concentrations and Sulphate Aerosol Indirect Radiative Forcing: A Sensitivity Study to the DMS Source Representation and Oxidation.” Atmospheric Chemistry and Physics 3: 49–65. Boys, B. L., R. V. Martin, A. van Donkelaar, R. J. MacDonell, N. C. Hsu, M. J. Cooper, R. M. Yantosca, Z. Lu, D. G. Streets, Q. Zhang, and S. W. Wang. 2014. “Fifteen-Year Global Time Series of Satellite-Derived Fine Particulate Matter.” Environmental Science & Technology 48 (19): 11109–18. Briggs, D. J., S. Collins, P. Elliott, P. Fischer, S. Kingham, E. Lebret, K. Pryl, H. van Reeuwijk, K.  Smallbone, and A. van Der Veen. 1997. “Mapping Urban Air Pollution Using GIS: A Regression-Based Approach.” International Journal of Geographical Information Science 11 (7): 699–718. Broxton, P. D., X. Zeng, D. Sulla-Menashe, P. A. Troch. 2014. “A Global Land Cover Climatology Using MODIS Data.” Journal of Applied Meteorology and Climatology 53: 1593–605. Chatfield, R., M. Sorek-Hamer, A. Lyapustin, and Y. Liu. 2017. “Daily Kilometer-Scale MODIS Satellite Maps of PM2.5 Describe Wintertime Episodes.” Paper 263193 presented at the “Air and Waste Management Association 2017 Annual Conference and Exhibition,” Pittsburgh, PA, June 5–9. Chin, M., P. Ginoux, S. Kinne, O. Tores, B. N. Holben, B. N. Duncan, R. V. Martin, J. A. Logan, A. Higurashi, and T. Nakajima. 2002. “Tropospheric Aerosol Optical Thickness from the GOCART Model and Comparisons with Satellite and Sun Photometer Measurements.” Journal of Atmospheric Science 59: 461–83. Chin, M., R. B. Rood, S. J. Lin, J. F. Müller, and A. M. Thompson. 2000. “Atmospheric Sulfur Cycle Simulated in the Global Model GOCART: Model Description and Global Properties.” Journal of Geophysical Research: Atmospheres 105 (D20): 24671–87. Literature Review | 91 CIESIN (Center for International Earth Science Information Network, Columbia University). 2016. “Gridded Population of the World, Version 4 (GPWv4): Population Density.” Palisades, NY: NASA Socioeconomic Data and Applications Center (SEDAC). https://doi.org//10.7927​ /H4NP22DQ. CIESIN (Center for International Earth Science Information Network, Columbia University) and ITOS (Information Technology Outreach Services, University of Georgia). 2013. “Global Roads Open Access Data Set, Version 1 (gROADSv1).” Palisades, NY: NASA Socioeconomic Data and Applications Center (SEDAC). https://doi.org/10.7927/H4VD6WCT. Colarco, P., A. da Silva, M. Chin, and T. Diehl. 2010. “Online Simulations of Global Aerosol Distributions in the NASA GEOS-4 Model and Comparisons to Satellite and Ground-Based Aerosol Optical Depth.” Journal of Geophysical Research: Atmospheres 115: D14207. Colarco, P. R., E. P. Nowottnick, C. A. Randles, B. Yi, P. Yang, K. M. Kim, J. A. Smith, and C. G. Bardeen. 2014. “Impact of Radiatively Interactive Dust Aerosols in the NASA GEOS-­5 Climate Model: Sensitivity to Dust Particle Shape and Refractive Index.” Journal of Geophysical Research: Atmospheres 119 (2): 753–86. Cucurull, L., and J. C. Derber. 2008. “Operational Implementation of COSMIC Observations into NCEP’s Global Data Assimilation System.” Weather and Forecasting 23: 702–11. de Hoogh, K., J. Gulliver, A. van Donkelaar, R. V. Martin, J. D. Marshall, M. J. Bechle, G. Cesaroni, M. C. Pradas, A. Dedele, M. Eeftens, and B. Forsberg. 2016. “Development of West-European PM2.5 and NO2 Land Use Regression Models Incorporating Satellite-Derived and Chemical Transport Modelling Data.” Environmental Research 151: 1–10. Dentener, F., S. Kinne, T. Bond, O. Boucher, J. Cofala, S. Generoso, P. Ginoux, S. Gong, J. J. Hoelzemann, and A. Ito, and L. Marelli. 2006. “Emissions of Primary Aerosol and Precursor Gases in the Years 2000 and 1750 Prescribed Data-Sets for AeroCom.” Atmospheric Chemistry and Physics 6 (12): 4321–44. Diner, D. J., J. C. Beckert, T. H. Reilly, C. J. Bruegge, J. E. Conel, R. A. Kahn, J. V. Martonchik, T. P. Ackerman, R. Davies, S. A. Gerstl, and H. R. Gordon. 1998. “Multi-Angle Imaging SpectroRadiometer (MISR) Instrument Description and Experiment Overview.” IEEE Transactions on Geoscience and Remote Sensing 36 (4): 1072–87. Drüe, C., W. Frey, A. Hoff, and T. Hauf. 2008. “Aircraft-Type Specific Errors in AMDAR Weather Reports from Commercial Aircraft.” Quarterly Journal of the Royal Meteorological Society 134: 229–39. Emili, E., A. Lyapustin, Y. Wang, C. Popp, S. Korkin, M. Zebisch, S. Wunderle, and M. Petitta. 2011. “High Spatial Resolution Aerosol Retrieval with MAIAC: Application to Mountain Regions.” Journal of Geophysical Research 116: D23211. Emmons, L. K., S. Walters, P. G. Hess, J.-F. Lamarque, G. G. Pfister, D. Fillmore, C. Granier, A. Guenther, D. Kinnison, T. Laepple, J. Orlando, X. Tie, G. Tyndall, C. Wiedinmyer, S. L. Baughcum, and S. Kloster. 2010. “Description and Evaluation of the Model for Ozone and Related Chemical Tracers, Version 4 (MOZART-4).” Geoscientific Model Development 3: 43–67. Fleming, R. J. 1996. “The Use of Commercial Aircraft as Platforms for Environmental Measurements.” Bulletin of the American Meteorological Society 77: 2229–42. Gelaro, R., W. McCarty, M. J. Suárez, R. Todling, A. Molod, L. Takacs, C. A. Randles, A. Darmenov, M. G. Bosilovich, R. Reichle, K. Wargan, L. Coy, R. Cullather, C. Draper, S. Akella, V. Buchard, A. Conaty, A. M. da Silva, W. Gu, G.-K. Kim, R. Koster, R. Lucchesi, D. Merkova, J. E. Nielsen, G. Partyka, S. Pawson, W. Putman, M. Rienecker, S. D. Schubert, M. Sienkiewicz, and B. Zhao. 2017. “The Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2).” Journal of Climate 30 (14): 5419–54. Geng, G., Q. Zhang, R. V. Martin, A. van Donkelaar, H. Huo, H. Che, J. Lin, and K. He. 2015. “Estimating Long-Term PM2.5 Concentrations in China Using Satellite-Based Aerosol Optical Depth and a Chemical Transport Model.” Remote Sensing of Environment 166: 262–70. Guelle, W., M. Schulz, Y. Balkanski, and F. Dentener. 2001. “Influence of the Source Formulation on Modeling the Atmospheric Global Distribution of Sea Salt Aerosol.” Journal of Geophysical Research: Atmospheres 106 (D21): 27509–24. Guo, P., Y.-H. Kuo, S. V. Sokolovskiy, and D. H. Lenschow. 2011. “Estimating Atmospheric Boundary Layer Depth Using COSMIC Radio Occultation Data.” Journal of the Atmospheric Sciences 68 (8): 1703–13. 92 | Getting Down to Earth Gupta, P., R. C. Levy, S. Mattoo, L. A. Remer, and L. A. Munchak. 2016. “A Surface Reflectance Scheme for Retrieving Aerosol Optical Depth over Urban Surfaces in MODIS Dark Target Retrieval Algorithm.” Atmospheric Measurement Techniques 9: 3293–308. Han, J., and H.-L. Pan. 2011. “Revision of Convection and Vertical Diffusion Schemes in the NCEP Global Forecast System.” Weather and Forecasting 26: 520–33. Hegarty, J. D., J. Lewis, L. McGrath-Spangler, J. Henderson, A. J. Scarino, R. Adams-Selin, M. Hicks, R. Ferrare, P. DeCola, and E. Welton. 2018. “Analysis of the Planetary Boundary Layer Height during DISCOVER-AQ Baltimore–Washington, DC with Lidar and High- Resolution WRF Modeling.” 2018. Journal of Applied Meteorology and Climatology 57: 2679–96. Hogan, T. F., M. Liu, J. A. Ridout, M. S. Peng, T. R. Whitcomb, B. C. Ruston, C. A. Reynolds, S. D. Eckermann, J. R. Moskaitis, N. L. Baker, J. P. McCormack, K. C. Viner, J. G. McLay, M. K. Flatau, L. Xu, C. Chen, and S. W. Chang. 2014. “The Navy Global Environmental Model.” Oceanography 27 (3): 116–25. Holben, B. N., D. Tanre, A. Smirnov, T. F. Eck, I. Slutsker, N. Abuhassan, W. W. Newcomb, J. S. Schafer, B. Chatenet, F. Lavenu, and Y. J. Kaufman. 2001. “An Emerging Ground-Based Aerosol Climatology: Aerosol Optical Depth from AERONET.” Journal of Geophysical Research: Atmospheres 106 (D11): 12067–97. Holtz, R. E., S. A. Ackerman, F. W. Nagle, R. Frey, S. Dutcher, R. E. Kuehn, M. A. Vaughan, and B. Baum. 2008. “Moderate Resolution Imaging Spectroradiometer (MODIS) Cloud Detection and Height Evaluation Using CALIOP.” Journal of Geophysical Research 113: D00A19. Horowitz, L. W. 2006. “Past, Present, and Future Concentrations of Tropospheric Ozone and Aerosols: Methodology, Ozone Evaluation, and Sensitivity to Aerosol Wet Removal.” Journal of Geophysical Research 111: D22211. Hsu, N. C., M. J. Jeong, C. Bettenhausen, A. M. Sayer, R. Hansell, C. S. Seftor, J. Huang, and S. C. Tsay. 2013. “Enhanced Deep Blue Aerosol Retrieval Algorithm: The Second Generation.” Journal of Geophysical Research: Atmospheres 118 (16): 9296–315. Hu, X., L. A. Waller, A. Lyapustin, Y. Wang, M. Z. Al-Hamdan, W. L. Crosson, M. G. Estes, S. M. Estes, D. A. Quattrochi, S. J. Puttaswamy, and Y. Liu. 2014. “Estimating Ground-Level PM2.5 Concentrations in the Southeastern United States Using MAIAC AOD Retrievals and a Two-Stage Model.” Remote Sensing of Environment 140: 220–32. Hu, X., L. A. Waller, A. Lyapustin, Y. Wang, and Y. Liu. 2014. “10-year Spatial and Temporal Trends of PM2.5 Concentrations in the Southeastern US Estimated Using High-Resolution Satellite Data.” Atmospheric Chemistry and Physics 14 (12): 6301–14. Huang, J., S. Kondragunta, I. Laszlo, H. Liu, L. A. Remer, H. Zhang, S. Superczynski, P. Ciren, B. N. Holben, and M. Petrenko. 2016. “Validation and Expected Error Estimation of Suomi- NPP VIIRS Aerosol Optical Thickness and Ångström Exponent with AERONET.” Journal of Geophysical Research: Atmospheres 121: 7139–60. Jackson, J. M., H. Liu, I. Laszlo, S. Kondragunta, L. A. Remer, J. Huang, and H.-C. Huang. 2013. “Suomi-NPP VIIRS Aerosol Algorithms and Data Products.” Journal of Geophysical Research: Atmospheres 118: 12673–89. Jordan, N. S., R. M. Hoff, and J. T. Bacmeister. 2010. “Validation of Goddard Earth Observing System-Version 5 MERRA Planetary Boundary Layer Heights Using CALIPSO.” Journal of Geophysical Research: Atmospheres 115 (D24): D24218. Kahn, R. A., B. J. Gaitley, J. V. Martonchik, D. J. Diner, K. A. Crean, and B. Holben. 2005. “Multiangle Imaging Spectroradiometer (MISR) Global Aerosol Optical Depth Validation Based on 2 Years of Coincident Aerosol Robotic Network (AERONET) Observations.” Journal of Geophysical Research: Atmospheres 110: D10S04. Kahn, R. A., D. L. Nelson, M. J. Garay, R. C. Levy, M. A. Bull, D. J. Diner, J. V. Martonchik, S. R. Paradise, E. G. Hansen, and L. A. Remer. 2009. “MISR Aerosol Product Attributes and Statistical Comparisons with MODIS.” IEEE Transactions on Geoscience and Remote Sensing 47 (12): 4095–114. Kaiser, J., A. Heil, M. Andreae, A. Benedetti, N. Chubarova, L. Jones, J.-J. Morcrette, M. Razinger, M. Schultz, M. Suttie, and G. R. van der Werf. 2012. “Biomass Burning Emissions Estimated with a Global Fire Assimilation System Based on Observed Fire Radiative Power.” Biogeosciences 9 (1): 527–54. Literature Review | 93 Kim, D., M. Chin, H. Bian, Q. Tan, M. E. Brown, T. Zheng, R. You, T. Diehl, P. Ginoux, and T. Kucsera. 2013. “The Effect of the Dynamic Surface Bareness on Dust Source Function, Emission, and Distribution. Journal of Geophysical Research: Atmospheres 118 (2): 871–86. Knapp, K. R., R. Frouin, S. Kondragunta, and A. Prados. 2005. “Toward Aerosol Optical Depth Retrievals over Land from GOES Visible Radiances: Determining Surface Reflectance.” International Journal of Remote Sensing 26 (18): 4097–116. Lamarque, J.-F., J. T. Kiehl, P. G. Hess, W. D. Collins, L. K. Emmons, P. Ginoux, C. Luo, and X. Tie. 2005. “Response of a Coupled Chemistry-Climate Model to Changes in Aerosol Emissions: Global Impact on the Hydrological Cycle and the Tropospheric Burdens of OH, Ozone and NO.” Geophysical Research Letters 32: L16809. Lee, H. J., R. B. Chatfield, and A. W. Strawa. 2016. “Enhancing the Applicability of Satellite Remote Sensing for PM2.5 Estimation Using MODIS Deep Blue AOD and Land Use Regression in California, United States.” Environmental Science & Technology 50 (12): 6546–55. Levy, R. C., S. Mattoo, L. A. Munchak, L. A. Remer, A. M. Sayer, F. Patadia, and N. C. Hsu. 2013. “The Collection 6 MODIS Aerosol Products over Land and Ocean.” Atmospheric Measurement Techniques 6: 2989–3034. Levy, R. C., L. A. Munchak, S. Mattoo, F. Patadia, L. A. Remer, and R. E. Holz. 2015. “Towards a Long-Term Global Aerosol Optical Depth Record: Applying a Consistent Aerosol Retrieval Algorithm to MODIS and VIIRS-Observed Reflectance.” Atmospheric Measurement Techniques 8: 4083–110. Lewis, J. R., E. J. Welton, A. M. Molod, and E. Joseph 2013. “Improved Boundary Layer Depth Retrievals from MPLNET.” Journal of Geophysical Research: Atmospheres 118: 9870–79. Li, Y., Y. Xue, G. de Leeuw, C. Li, L. Yang, T. Hou, and F. Marir. 2013. “Retrieval of Aerosol Optical Depth and Surface Reflectance over Land from NOAA AVHRR Data.” Remote Sensing of Environment 133: 1–20. Lu, C.-H., A. da Silva, J. Wang, S. Moorthi, M. Chin, P. Colarco, Y. Tang, P. S. Bhattacharjee, S.-P. Chen, H.-Y. Chuang, H.-M. H. Juang, J. McQueen, and M. Iredell. 2016. “The Implementation of NEMS GFS Aerosol Component (NGAC) Version 1.0 for Global Dust Forecasting at NOAA/NCEP.” Geoscientific Model Development 9: 1905–19. Lu, C.-H., S.-W. Wei, S. Kondragunta, Q. Zhao, J. McQueen, J. Wang, and P. Bhattacharjee. 2016. “NCEP Aerosol Data Assimilation Update: Improving NCEP Global Aerosol Forecasts Using JPSS-NPP VIIRS Aerosol Products.” Paper presented at the “8th International Cooperative for Aerosol Prediction.” http://icap.atmos.und.edu/ICAP8​ /­Day3/Lu_NCEP_ThursdayAM.pdf. Lv, B., Y. Hu, H. H. Chang, A. G. Russell, and Y. Bai. 2016. “Improving the Accuracy of Daily PM2.5 Distributions Derived from the Fusion of Ground-Level Measurements with Aerosol Optical Depth Observations, a Case Study in North China.” Environmental Science & Technology 50  (9): 4752–59. Lyapustin, A., Y. Wang, I. Laszlo, R. Kahn, S. Korkin, L. Remer, R. Levy, and J. S. Reid. 2011. “Multiangle Implementation of Atmospheric Correction (MAIAC): 2. Aerosol Algorithm.” Journal of Geophysical Research 116: D03211. Mahowald, N., J.-F. Lamarque, X. Tie, and E. Wolff. 2006. “Sea Salt Aerosol Response to Climate Change: Last Glacial Maximum, Pre-Industrial and Doubled Carbon Dioxide Climates.” Journal of Geophysical Research 111: D05303. Mahowald, N. M., D. R. Muhs, S. Levis, P. J. Rasch, M. Yoshioka, C. S. Zender, and C. Luo. 2006. “Change in Atmospheric Mineral Aerosols in Response to Climate: Last Glacial Period, Preindustrial, Modern, and Doubled Carbon Dioxide Climates.” Journal of Geophysical Research 111: D10202. Martins, V. S., A. Lyapustin, L. A. S. de Carvalho, C. C. F. Barbosa, and E. M. L. M. Novo. 2017. “Validation of High-Resolution MAIAC Aerosol Product over South America.” Journal of Geophysical Research: Atmospheres 122: 7537–59. Martins, J. P. A., J. Teixeira, P. M. M. Soares, P. M. A. Miranda, B. H. Kahn, V. T. Dang, F. W. Irion, E. J. Fetzer, and E. Fishbein. 2010. “Infrared Sounding of the Trade-Wind Boundary Layer: AIRS and the RICO Experiment.” Geophysical Research Letters 37 (24): L24806. 94 | Getting Down to Earth Martonchik, J. V., D. J. Diner, R. Kahn, B. Gaitley, and B. N. Holben. 2004. “Comparison of MISR and AERONET Aerosol Optical Depths over Desert Sites.” Geophysical Research Letters 31 (16): L16102. McGrath-Spangler, E. L., and A. S. Denning. 2012. “Estimates of North American Summertime Planetary Boundary Layer Depths Derived from Space-Borne Lidar.” Journal of Geophysical Research: Atmospheres 117: D15101. McGrath-Spangler, E. L., and A. S. Denning. 2013. “Global Seasonal Variations of Midday Planetary Boundary Layer Depth from CALIPSO Space-Borne LIDAR.” Journal of Geophysical Research: Atmospheres 118: 1226–33. Mei, L., Y. Xue, G. de Leeuw, T. Holzer-Popp, J. Guang, Y. Li, L. Yang, H. Xu, X. Xu, C. Li, Y. Wang, C. Wu, T. Hou, X. He, J. Liu, J. Dong, and Z. Chen. 2012. “Retrieval of Aerosol Optical Depth over Land Based on a Time Series Technique Using MSG/SEVIRI Data.” Atmospheric Chemistry and Physics 12: 9167–85. Millar, G., T. Abel, J. Allen, P. Barn, M. Noullett, J. Spagnol, and P. L. Jackson. 2010. “Evaluating Human Exposure to Fine Particulate Matter Part II: Modeling.” Geography Compass 4 (7): 731–49. Morcrette, J.-J., O. Boucher, L. Jones, D. Salmond, P. Bechtold, A. Beljaars, A. Benedetti, A. Bonet, J. Kaiser, M. Razinger, M. Schulz, S. Serrar, A. J. Simmons, M. Sofiev, M. Suttie, A. M. Tompkins, and A. Untch. 2009. “Aerosol Analysis and Forecast in the European Centre for Medium-Range Weather Forecasts Integrated Forecast System: Forward Modeling.” Journal of Geophysical Research: Atmospheres (1984–2012) 114 (D6): D06206. Mulcahy, J. P., D. N. Walters, N. Bellouin, and S. F. Milton. 2014. “Impacts of Increasing the Aerosol Complexity in the Met Office Global Numerical Weather Prediction Model.” Atmospheric Chemistry and Physics 14: 4749–78. Munchak, L. A., R. C. Levy, S. Mattoo, L. A. Remer, B. N. Holben, J. S. Schafer, C. A. Hostetler, and R. A. Ferrare. 2013. “MODIS 3 km Aerosol Product: Applications over Land in an Urban​ /Suburban Region.” Atmospheric Measurement Techniques 6: 1747–59. Oo, M. M., M. Jerg, E. Hernandez, A. Picon, B. M. Gross, F. Moshary, and S. A. Ahmed. 2010. “Improved MODIS Aerosol Retrieval Using Modified VIS/SWIR Surface Albedo Ratio over Urban Scenes.” IEEE Transactions on Geoscience and Remote Sensing 48 (3): 983–1000. Pearce, J. L., S. L. Rathbun, M. Aguilar-Villalobos, and L. P. Naeher. 2009. “Characterizing the Spatiotemporal Variability of PM2.5 in Cusco, Peru Using Kriging with External Drift. Atmospheric Environment 43 (12): 2060–69. Provençal, S., V. Buchard, A. M. da Silva, R. Leduc, N. Barrette, E. Elhacham, and S. H. Wang. 2017. “Evaluation of PM2.5 Surface Concentrations Simulated by Version 1 of NASA’s MERRA Aerosol Reanalysis over Israel and Taiwan.” Aerosol and Air Quality Research 17: 253–61. Reddy, M. S., O. Boucher, N. Bellouin, M. Schulz, Y. Balkanski, J.-L. Dufresne, and M. Pham. 2005. “Estimates of Global Multicomponent Aerosol Optical Depth and Direct Radiative Perturbation in the Laboratoire de Météorologie Dynamique General Circulation Model.” Journal of Geophysical Research: Atmospheres (1984–2012) 110 (D10): D10S16. Remer, L. A., S. Mattoo, R. C. Levy, and L. A. Munchak. 2013. “MODIS 3 km Aerosol Product: Algorithm and Global Perspective.” Atmospheric Measurement Techniques 6: 1829–44. Rienecker, M. M., M. J. Suarez, R. Todling, J. Bacmeister, L. Takacs, H.-C. Liu, W. Gu, M. Sienkiewicz, R. D. Koster, R. Gelaro, I. Stajner, and J. E. Nielsen. 2008. “The GEOS-5 Data Assimilation System—Documentation of Versions 5.0.1, 5.1.0, and 5.2.0.” NASA GSFC Technical Report Series on Global Modeling and Data Assimilation vol. 27, NASA​ /­TM-2007-104606. Ryan, P. H., G. K. LeMasters, L. Levin, J. Burkle, P. Biswas, S. Hu, S. Grinshpun, and T. Reponen. 2008. “A Land-Use Regression Model for Estimating Microenvironmental Diesel Exposure Given Multiple Addresses from Birth through Childhood.” Science of the Total Environment 404 (1): 139-47. Sayer, A. M., N. C. Hsu, J. Lee, N. Carletta, S.-H. Chen, and A. Smirnov. 2017. “Evaluation of NASA Deep Blue/SOAR Aerosol Retrieval Algorithms Applied to AVHRR Measurements.” Journal of Geophysical Research: Atmospheres 122: 9945–67. Literature Review | 95 Sayer, A. M., L. A. Munchak, N. C. Hsu, R. C. Levy, C. Bettenhausen, and M.-J. Jeong. 2014. “MODIS Collection 6 Aerosol Products: Comparison between Aqua’s e-Deep Blue, Dark Target, and “Merged” Data Sets, and Usage Recommendations.” Journal of Geophysical Research: Atmospheres 119: 13965–89. https://doi.org/10.1002/2014JD022453. Schneider, T. 2016. “Next Generation Global Prediction System (NGGPS) Phase 2 Atmospheric Dynamic Core Evaluation.” Presentation for UMAC, June 22, 2016. https://www.weather​ .gov/media/sti/nggps/Phase%202%20Dycore%20Evaluation%20Briefing%2022%20 June%202016%20UMAC%20Final%20for%20posting%2006272016.pdf. Schulz, M., G. de Leeuw, and Y. Balkanski. 2004. “Sea-Salt Aerosol Source Functions and Emissions.” In Emissions of Atmospheric Trace Compounds, edited by C. Granier, P. Artaxo, and C. E. Reeves, 333–59. Dordrecht, the Netherlands: Kluwer. Seidel, D. J., C. O. Ao, and K. Li. 2010. “Estimating Climatological Planetary Boundary Layer Heights from Radiosonde Observations: Comparison of Methods and Uncertainty Analysis.” Journal of Geophysical Research 115 (D16): D16113. Seidel, D. J., Y. Zhang, A. C. M. Beljaars, J.-C. Golaz, A. R. Jacobson, and B. Medeiros. 2012. “Climatology of the Planetary Boundary Layer over the Continental United States and Europe.” Journal of Geophysical Research 117: D17106. Singh, V., C. Carnevale, G. Finzi, E. Pisoni, and M. Volta. 2011. “A Cokriging Based Approach to Reconstruct Air Pollution Maps, Processing Measurement Station Concentrations and Deterministic Model Simulations.” Environmental Modelling & Software 26 (6): 778–86. Singh, M. K., P. Venkatachalam, and R. Gautam. 2017. “Geostatistical Methods for Filling Gaps in Level-3 Monthly-Mean Aerosol Optical Depth Data from Multi-Angle Imaging Spectroradiometer.” Aerosol and Air Quality Research 17: 1963–74. Snider, G., C. L. Weagle, R. V. Martin, A. van Donkelaar, K. Conrad, D. Cunningham, C. Gordon, M. Zwicker, C. Akoshile, P. Artaxo, and N. X. Anh. 2015. “SPARTAN: A Global Network to Evaluate and Enhance Satellite-Based Estimates of Ground-Level Particulate Matter for Global Health Applications.” Atmospheric Measurement Techniques 8: 505–21. Sorek-Hamer, M., I. Kloog, P. Koutrakis, A. W. Strawa, R. Chatfield, A. Cohen, W. L. Ridgway, and D. M. Broday. 2015. “Assessment of PM2.5 Concentrations over Bright Surfaces Using MODIS Satellite Observations.” Remote Sensing of Environment 163: 180–85. Sorek-Hamer, M., A. W. Strawa, R. B. Chatfield, R. Esswein, A. Cohen, and D. M. Broday. 2013. “Improved Retrieval of PM2.5 from Satellite Data Products Using Non-Linear Methods.” Environmental Pollution 182: 417–23. Strawa, A. W., R. B. Chatfield, M. Legg, B. Scarnato, and R. Esswein. 2013. “Improving Retrievals of Regional Fine Particulate Matter Concentrations from Moderate Resolution Imaging Spectroradiometer (MODIS) and Ozone Monitoring Instrument (OMI) Multisatellite Observations.” Journal of the Air & Waste Management Association 63 (12): 1434–46. Sun, W., G. Videen, S. Kato, B. Lin, C. Lukashin, and Y. Hu. 2011. “A Study of Subvisual Clouds and Their Radiation Effect with a Synergy of CERES, MODIS, CALIPSO, and AIRS Data.” Journal of Geophysical Research: Atmospheres 116: D22207. Superczynski, S. D., S. Kondragunta, and A. I. Lyapustin. 2017. “Evaluation of the Multi-Angle Implementation of Atmospheric Correction (MAIAC) Aerosol Algorithm through Intercomparison with VIIRS Aerosol Products and AERONET.” Journal of Geophysical Research: Atmospheres 122: 3005–22. https://doi.org/10.1002/2016JD025720.  Tie, X., G. Brasseur, L. Emmons, L. Horowitz, and D. Kinnision. 2001. “Effects of Aerosols on Tropospheric Oxidants: A Global Model Study.” Journal of Geophysical Research 106: 2931–64. Tie, X., S. Madronich, S. Walters, D. Edwards, P. Ginoux, N. Mahowald, R. Zhang, C. Luo, and G. Brasseur. 2005. “Assessment of the Global Impact of Aerosols on Tropospheric Oxidants.” Journal of Geophysical Research 110: D03204. Toll, V., K. Reis, R. Ots, M. Kaasik, A. Männik, M. Prank, and M. Sofiev. 2015. “SILAM and MACC Reanalysis Aerosol Data Used for Simulating the Aerosol Direct Radiative Effect with the NWP Model HARMONIE for Summer 2010 Wildfire Case in Russia.” Atmospheric Environment 121: 75–85. 96 | Getting Down to Earth van Donkelaar, A., R. V. Martin, M. Brauer, and B. L. Boys. 2015. “Use of Satellite Observations for Long-Term Exposure Assessment of Global Concentrations of Fine Particulate Matter.” Environmental Health Perspectives 123 (2): 135–43. van Donkelaar, A., R. V. Martin, M. Brauer, R. Kahn, R. Levy, C. Verduzco, and P. J. Villeneuve. 2010. Global Estimates of Ambient Fine Particulate Matter Concentrations from Satellite- Based Aerosol Optical Depth: Development and Application.” Environmental Health Perspectives 118 (6): 847–55. van Donkelaar, A., R. V. Martin, and R. J. Park. 2006. “Estimating Ground-Level PM2.5 Using Aerosol Optical Depth Determined from Satellite Remote Sensing.” Journal of Geophysical Research: Atmospheres 111: D21201. van Donkelaar, A., R. V. Martin, R. J. Spurr, and R. T. Burnett. 2015. “High-Resolution Satellite- Derived PM2.5 from Optimal Estimation and Geographically Weighted Regression over North America.” Environmental Science & Technology 49 (17): 10482–91. van Donkelaar, A., R. V. Martin, R. J. Spurr, E. Drury, L. A. Remer, R. C. Levy, and J. Wang. 2013. “Optimal Estimation for Global Ground-Level Fine Particulate Matter Concentrations.” Journal of Geophysical Research: Atmospheres 118 (11): 5621–36. Vienneau, D., K. de Hoogh, M. J. Bechle, R. Beelen, A. van Donkelaar, R. V. Martin, D. B. Millet, G. Hoek, and J. D. Marshall. 2013. “Western European Land Use Regression Incorporating Satellite- and Ground-Based Measurements of NO2 and PM10.” Environmental Science & Technology 47 (23): 13555–64. Winker, D. M., W. H. Hunt, and M. J. McGill. 2007. “Initial Performance Assessment of CALIOP.” Geophysical Research Letters 34 (19): L19803. Winker, D. M., M. A. Vaughan, A. Omar, Y. Hu, K. A. Powell, Z. Liu, W. H. Hunt, and S. A. Young. 2009. “Overview of the CALIPSO Mission and CALIOP Data Processing Algorithms.” Journal of Atmospheric and Oceanic Technology 26 (11): 2310–23. Wu, J., A. M. Winer, and R. J. Delfino. 2006. “Exposure Assessment of Particulate Matter Air Pollution before, during, and after the 2003 Southern California Wildfires.” Atmospheric Environment 40 (18): 3333–48. Zhang, H., S. Kondragunta, I. Laszlo, H. Liu, L. A. Remer, J. Huang, S. Superczynski, and P. Ciren. 2016. “An Enhanced VIIRS Aerosol Optical Thickness (AOT) Retrieval Algorithm over Land Using a Global Surface Reflectance Ratio Database.” Journal of Geophysical Research: Atmospheres 121: 10717–38. Zhu, Y., J. C. Derber, R. J. Purser, B. A. Ballish, and J. Whiting. 2015. “Variational Correction of Aircraft Temperature Bias in the NCEP’s GFI Analysis System.” Monthly Weather Review 143: 3774–803. APPENDIX B Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 In this task of the research underlying this report, data from publicly available ground-level-monitoring (GLM) networks (as included in the OpenAQ d ­ atabase) were used to identify promising new pathways for low- and ­ middle-income countries (LMICs) to use satellite observations in their air-quality monitoring and forecasting. As called for in the Terms of Reference for this report, this work “(i) compare[d] and contrast[ed] GLMs in selected LMICs; and (ii) examine[d] novel methods for improving the conversion of aerosol ­ optical depth (AOD) and other remotely sensed data to concentration estimates of surface-level particu- late matter with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5) in environments typical of LMIC conditions.” The GLM data sets for testing the satellite approaches were data for three cities currently included in the OpenAQ database, which include near real-time (NRT) PM2.5 observations from local government air-quality agencies and US embassy and consulate PM2.5 data (table B.1). Kriging of the local government data was used to estimate PM2.5 values for comparison with the US embassy and consulate PM2.5 values, and discrepancy between these estimates was used as a quality check on the local government data. The use of various satellite data sets was examined for these three cities. The literature review (appendix A) suggested that the Moderate-Resolution Imaging Spectroradiometer (MODIS) combined Deep Blue–Dark Target product was likely the best in terms of the coverage in urban areas, the public availability of the data set, and the long-term global record for training statistical models. The Aqua and Terra data sets were tested separately because they have different overhead crossing times, and the one that gave the best fit with the GLM net- work data was determined. Two other satellite data sets were also tested. In Delhi, India, the Visible Infrared Imaging Radiometer Suite (VIIRS) six-­ kilometer resolution AOD retrievals was tested as a potential alternative to the MODIS data. Both MODIS instruments are well past their design lifetimes, and new VIIRS instruments will be launched over the next decade as part of the US JPSS program, so the VIIRS product is more likely to be available in the future than MODIS. The Spinning Enhanced Visible and Infrared Imager (SEVIRI) data set was also used as an example of the capability of next-generation geosta- tionary satellites. Because SEVIRI did not cover the three initially selected test cities, Accra, Ghana, was added as an additional test city for this satellite.  97 98 | Getting Down to Earth TABLE B.1  Three focus cities where OpenAQ aggregates real-time local government particulate matter data overlap with US Diplomatic Post measurements NUMBER OF CITY COUNTRY STATIONS ORIGINATING DATA SOURCE Delhi India 11 Central Pollution Control Board http://www.cpcb.gov.in/CAAQM/mapPage/frmindiamap.aspx Lima Peru 10 Servicio Nacional de Meteorología e Hidrología del Perú http://peruclima.pe/?p=calidad-de-aire Ulaanbaatar Mongolia 7 Agaar.mn/National Agency of Meteorology and Environment Monitoring http://agaar.mn/ Source: World Bank. However, the GLM data for Accra are not yet available in OpenAQ, and thus this work merely compared the SEVIRI results with those using the MODIS prod- ucts for this city. For each satellite data source, all recommended quality flags were applied to filter the AOD data, accepting only the highest quality data from each source. One statistical approach and one chemical-transport-model (CTM) –based approach were tested for each city (except Accra and Lima, where only the CTM-based approaches were possible). The statistical approach used the gener- alized additive model (GAM) approach of Sorek-Hamer and others (2013), mod- ified to use the ratio of the AOD to planetary boundary layer (PBL) as the primary predictor, as in Chatfield and others (2017). The CTM approach used AODs and aerosol vertical profiles from the Modern Era Retrospective-analysis for Research and Applications (MERRA-2) reanalysis to derive a ground-level PM2.5 estimate following the methods of van Donkelaar and others (2010). Geographically weighted regression (GWR) was then attempted to relate the errors between the GLM PM2.5 observations and the two satellite-based esti- mates of ground-level PM2.5, as in van Donkelaar and others (2015). METHODOLOGY Satellite data MODIS and VIIRS data MODIS and VIIRS products were collected for a two-year period (2016 and 2017) for Delhi, India; Lima, Peru; and Ulaanbaatar, Mongolia. The MODIS data are available through the National Aeronautics and Space Administration’s (NASA’s) Distributed Active Archive Center (DAAC),1 where all the MODIS product files for each year and for the Aqua and Terra platforms are available for download. These files are simple to download and inspect, check the spatial range covered in the granule, and keep the granules that cover the region(s) of interest. This is very time consuming but can be automated, as was done for this work. For VIIRS, the Comprehensive Large Array-Data Stewardship System (CLASS)2 web-based system was used for ordering data. Here the user has to work through the CLASS graphical user interface and wait for the order to be available; the process cannot be automated further. The process is the following: Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 99 assign a date range and draw a box around the region of interest (note the box configuration can be saved to be reloaded for future use to ensure the same region is being searched for each time range). Note that for VIIRS to specifically request that the Earth location data be included (that is, the latitudes and longi- tudes are not included with the aerosol product), the user must request a final product file that is a merger of the VIIRS geolocation product (GAERO) and the aerosol product (VAOOO). The AOD data files for the MODIS product contain several different “prod- ucts” based on the Dark Target and Deep Blue algorithms. The combined Dark Target–Deep Blue product was chosen for this project. This seems to satisfy the requirements and needs of this project the best by providing the greatest possi- ble coverage over urban areas. For VIIRS there is only a single AOD product. For both VIIRS and MODIS, a region around the focus sites is extracted and a subset database is created for further study. The region extracted is defined for a 2° by 2° region surrounding the center of each city. Table B.2 lists the variables included in the postprocessed MODIS and VIIRS files. Table B.3 discusses how these variables were mapped from the variables in the MODIS and VIIRS files from the NASA DAAC and CLASS, respectively. Note the different definitions for the MODIS and VIIRS products for some of the data products in the focus site files. An uncertainty estimate was also added to the subset files based on the analysis of Remer and others (2013) following the equation AOTUNC = AOTBIAS + AOTRMS × AOT, (B.1) where AOTBIAS = 0.03 over ocean and 0.05 over land, and AOTRMS = 0.05 over ocean and 0.15 over land, and BIAS = bias, RMS = root mean square, and UNC = uncertainty. Figure B.1 illustrates the geographic coverage for an approximately 1° by 1° box around each of the focus regions, which was the area used for the analysis. TABLE B.2  Contents of the subsampled aerosol-optical-depth product files PRODUCT NAME DESCRIPTION Algorithm_Flag MODIS: Defines the algorithm used for generating the product, 0: Dark Target, 1: Deep Blue, 2: Mixed VIIRS: 0: Dark Target AOD_550 MODIS: AOD at 0.55 µm for combined Dark Target–Deep Blue products VIIRS: Dark Target product Error_Estimate The expected error for each product Land_Sea_Flag [0,1,2]: Land/sea/coast flag, 0: ocean, 1: land, 2: coastal Latitude Pixel latitude degree Longitude Pixel longitude degree nPix Number of pixels in file ObsTimeInSec Observation time in seconds from start of day QA_Flag MODIS: Combined Dark Target–Deep Blue confidence flag (0: no confidence, 1: marginal, 2: good, 3: very good) VIIRS: Dark Target aerosol confidence–quality flag (0: not produced, 1: low, 2: medium, 3: high) Scan_Start_Time Start time of scan: the pixel is in seconds since January 1, 1993 Source: World Bank. Note: Note the different product meaning for VIIRS (Visible Infrared Imaging Radiometer Suite) and MODIS (Moderate-Resolution Imaging Spectroradiometer) products. AOD = aerosol optical depth; QA = quality assurance. 100 | Getting Down to Earth TABLE B.3  Mapping between the parsed data file variable names and the aerosol-optical-depth products in the L2 files PRODUCT NAME INSTRUMENT IN INSTRUMENT FILE Algorithm_Flag MODIS AOD_550_Dark_Target_Deep_Blue_Combined_Algorithm_Flag VIIRS Algorithm_Flag_Land AOD_550 MODIS AOD_550_Dark_Target_Deep_Blue_Combined VIIRS AerosolOpticalDepth_at_550 nm QA_Flag MODIS AOD_550_Dark_Target_Deep_Blue_Combined_QA_Flag VIIRS QF1_VIIRSAEROEDR: first 2 bits, AOT Quality High: Number of good-quality pixel AOT retrievals more than 16 (1/4 the total number of pixels in aggregated horizontal cell) Medium: Number of good-quality retrievals less than or equal to 16 and the number of good/degraded quality retrievals greater than or equal to 16 Low: Number of good/degraded-quality retrievals less than 16 Not produced: No good/degraded-quality retrievals, neither land nor seawater dominant (number of land or ocean pixels less than half the number of good/ degraded pixels in the horizontal cell), ellipsoid fill in the geolocation, night scan, has a pixel with a solar zenith angle between 80° and 85° but no pixel with a solar zenith angle between 65° and 80° Source: World Bank. Note: AOD = aerosol optical depth; AOT = aerosol optical thickness; L2 = level 2; MODIS = Moderate-Resolution Imaging Spectroradiometer; QA = quality assurance; VIIRS = Visible Infrared Imaging Radiometer Suite. FIGURE B.1 Illustration of the regions covered by a 1° by 1° box around each focus region a. Delhi, India b. Lima, Peru c. Ulanbaatar, Mongolia Latitude Latitude Latitude Longitude Longitude Longitude Source: World Bank, produced using Google Maps and Esri ArcGIS. The image presented for each city covers approximately the circa 2° by 2° box in the subset files. SEVIRI data The SEVIRI–Meteosat Second Generation (MSG) Aerosol Over Land (SMAOL) product data were provided through ICARE.3 Using a Python script to access the ICARE FTP server, the HDF4-formatted, near real-time SMAOL v.1.3.6 product files were collected and data values within ±2° in latitude and longitude of Accra  were extracted. The extracted information (72 by 69 pixels at Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 101 three-kilometer resolution) was saved in netCDF4 files (see table B.4). The lati- tude and longitude for the SEVIRI grid was obtained from the file “MSG+0000.3km.hdf.” The SMAOL product files “SMAOL-AOT-NRT_V1.3.6_*. hdf” include the AOD retrievals for each of six aerosol models (1: Continental WMO, 2: Moderately Absorbing, 3: Urban Industrial, 4: Smoke, 5: Spheroidal Dust, and 6: Maritime) in addition to the best model AOD and TABLE B.4  Region of interest data fields in the SEVIRI netCDF output files STANDARD FIELDS DESCRIPTION/SOURCE AOT_DQF AOT quality flag: 0 = Clean AOT, 1 = Best AOT, 2 = No AOT (derived from AOT550_clean from SMAOL-AOT-NRT_V1.3.6_*.hdf) AOT_LND AOT over land from AOT550_Best_Model_unfiltered from SMAOL-AOT-NRT_V1.3.6_*.hdf AOT_UNC AOT uncertainty estimated based on published performance CloudMask Four-Level Cloud Mask from SMAOL quality flags (extracted from Quality_Flags from SMAOL-AOT-NRT_V1.3.6_*.hdf) Landmask Landmask: 0 = Land, 1 = Ocean, from SMAOL quality flags (extracted from Quality_Flags from SMAOL-AOT-NRT_V1.3.6_*.hdf) Latitude Latitude from MSG+0000.3km.hdf Longitude Longitude from MSG+0000.3km.hdf SMAOL DIAGNOSTIC FIELDS DESCRIPTION/SOURCE _AOT_LND_Aerosol_Model Aerosol_Model from SMAOL-AOT-NRT_V1.3.6_*.hdf _AOT_LND_Model1 AOT550_model_1_unfiltered from SMAOL-AOT-NRT_V1.3.6_*.hdf _AOT_LND_Model2 AOT550_model_2_unfiltered from SMAOL-AOT-NRT_V1.3.6_*.hdf _AOT_LND_Model3 AOT550_model_3_unfiltered from SMAOL-AOT-NRT_V1.3.6_*.hdf _AOT_LND_Model4 AOT550_model_4_unfiltered from SMAOL-AOT-NRT_V1.3.6_*.hdf _AOT_LND_Model5 AOT550_model_5_unfiltered from SMAOL-AOT-NRT_V1.3.6_*.hdf _AOT_LND_Model6 AOT550_model_6_unfiltered from SMAOL-AOT-NRT_V1.3.6_*.hdf _AOT_LND_Processing_Summary Processing_Summary from SMAOL-AOT-NRT_V1.3.6_*.hdf _AOT_LND_Quality_Flags Quality_Flags from SMAOL-AOT-NRT_V1.3.6_*.hdf _AOT_LND_Surf_Refl_VIS06 Surf_Refl_VIS06 from SMAOL-AOT-NRT_V1.3.6_*.hdf OPTIONAL OCEAN DIAGNOSTIC FIELDS DESCRIPTION/SOURCE _AOT_OCN Aerosol_Optical_Depth from SEV_AER-OC-L2_*_V1-04.hdf _AOT_OCN_Angstrom_Exponent Angstrom_Exponent from SEV_AER-OC-L2_*_V1-04.hdf _AOT_OCN_DQX DQX from SEV_AER-OC-L2_*_V1-04.hdf _AOT_OCN_Model_Number Model_Number from SEV_AER-OC-L2_*_V1-04.hdf Source: World Bank. Note: AOT = aerosol optical thickness; MSG = Meteosat Second Generation; NRT = near real-time; SEVIRI = Spinning Enhanced Visible and Infrared Imager; SMAOL = SEVIRI-MSG Aerosol Over Land. 102 | Getting Down to Earth model identification and the clean AOD product that reports AOD only for high-quality pixels. Also included in the product is the near-clear-sky reflec- tance at 630 nanometers, quality flags (including land and water and cloud mask identification), and processing information. Each of these fields was extracted for the region of interest. All information was preserved for the region as poten- tial diagnostic information. An additional field representing an estimate of the nominal AOD uncertainty was generated based on AOTUNC = AOTBIAS + AOTRMS × SMAOL (B.2) with values of AOTBIAS = 0.017 and AOTRMS = 0.63 estimated based on published performance. This formulation does not consider errors due to undetected cloud and other degenerate conditions that might lead to larger errors. A separate AOD product is available through ICARE over ocean: “SEV_AER- OC-L2_*_V1-04.hdf.” Because many regions of interest are located on the coast, the AOD values from the ocean product were optionally included in the extraction process to be used as potential diagnostic information (that is, for nearby coastal pixels). This product includes the AOD, Angstrom Exponent, aerosol model, and quality flags. The SEVIRI AOD product(s) are reported with a 15-minute refresh for day- time conditions. For the Accra region, SMAOL products were collected between 08Z and 17Z for the period from January 2017 to December 2017. The product was not available through ICARE for only a few days during this period. The ICARE Data and Services Center provided access to the data used in this study. MERRA CTM data The MERRA-2 data were downloaded through the Goddard Earth Sciences (GES) Data and Information Services Center (DISC).4 The MERRA-2 reanalysis product is organized into 26 file collections of instantaneous or time-averaged fields on three-dimensional native, pressure, or edge-level vertical grids or two-dimensional grids of related meteorological, ground surface, chemical, or aerosol variables. To provide inputs to the CTM and statistical approaches, sub- sets of five data file collections were downloaded in netCDF4 format. The file collections were the inst3_2d_gas_Nx for instantaneous three-hourly aerosol optical depth, tavg1_2d_flx_Nx for time-averaged hourly planetary boundary layer height (PBLH), surface temperature, and surface horizontal wind compo- nents (U and V), inst3_3d_aer_Nv for instantaneous three-hourly vertical profile on the 72 model native levels of speciated aerosol, tavg1_2d_aer_Nx for time-av- eraged hourly speciated surface aerosol and column mass density, and inst3_3d_ asm_Nv for surface geopotential and mid-layer height of the 71 model layers. The data file collections and extracted variables from each are summarized in table B.5. The data from all these file collections were processed and combined with satellite optical depths in a Python script and then written to a comma-­ separated values file for input to the CTM and statistical approach R scripts. Ground-level monitoring data and land-use variables 2017 GLM data for each city were extracted from the OpenAQ database. Several surface sites had slightly different names in the database because of capitaliza- tion and spacing differences, even though they represented data from the same site. Data from these sites were combined before further processing. Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 103 TABLE B.5  MERRA-2 data file collections and extracted variables for input into the chemical transport model and statistical approach scripts COLLECTION VARIABLES TIME GRID inst3_2d_gas_Nx Aerosol optical depth Instantaneous, three-hourly 2D inst3_3d_aer_Nv Air density Instantaneous, three-hourly 3D Mixing ratio of: Hydrophilic black carbon Hydrophobic black carbon Hydrophilic organic carbon Hydrophobic organic carbon Dust mixing ratio in five size bins Sea salt in five size bins Sulfate aerosol Inst3_3d_asm_Nv Surface geopotential Instantaneous, three-hourly 3D Midlayer height tavg1_2d_aer_Nx Surface mass concentration and Time-averaged, one-hourly 2D column mass density of: Black carbon Organic carbon Sulfate Dust Dust PM2.5 Sea Salt Sea Salt PM2.5 tavg1_2d_flx_Nx PBLH Time-averaged, one-hourly 2D Surface temperature Surface U and V Source: World Bank. Note: 2D = two-dimensional; 3D = three-dimensional; MERRA = Modern Era Retrospective–Analysis for Research and Applications; PBLH = planetary boundary layer height; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. The GLM data generally included 1-hour average PM2.5 data, or 15-minute average data for Delhi. These data were used to calculate daily average PM2.5 concentrations. Negative values were presumed to be errors and were removed from the data set. Only days with at least 19 hours of valid data at a given site were used to calculate daily averages. Land-use variables (including percentage of land that was classified as urban, grassland/wetland/natural, water, barren, or trees; distance to coast; major and minor road length; elevation; and average population density) were calculated for different buffer distances around each GLM network site (100, 200, 500, 750, 100, 1,500, 2,000, 2,500, 3,000, and 3,500 meters) using the data sets for land-use regression in appendix A. Each variable and buffer distance were then used one at a time in a simple linear regression to predict the ground- level PM2.5 concentrations. The buffer distance that gave the best overall fit for each variable ­(measured with the R correlation statistic) was used for all further tests of land-use variables. 104 | Getting Down to Earth CTM-based approach The CTM-based approach used the instantaneous three-hourly MERRA-2 AODs closest to the satellite overpass time and the daily average surface (dry) PM2.5 values. This approach assumes that there is a conversion factor of 1.4 to go from organic carbon to organic matter. Ground-level PM2.5 values were cal- culated using the MERRA-2 output for the PM2.5 contribution of dust and sea salt aerosols and assuming that all other aerosol components were in the PM2.5 fraction. Ground-level PM10 was also calculated by including all modeled aero- sol species. Other variables, such as the PBL height, the percentage contribu- tion of different aerosol species to the ground-level PM2.5, and the percentage of aerosol mass present above the PBL were also calculated based on the MERRA-2 output. The CTM results included for this report did not use the percentage of aerosol mass present above the PBL metric due to computational time constraints, but this screening metric will be included in the final deliv- ered products. Statistical approach The statistical approach fitted a GAM using the R software package to predict the ground-level PM2.5 concentrations at all the available GLM network sites within a given city. The initial fit used day of week and the satellite AOD divided by the PBL height (AOD/PBLH) as the primary predictors and a log link function. RESULTS Local government versus US Diplomatic Post data A test was performed to determine if the US Diplomatic Post data were roughly consistent with the local GLM network data. The approach used kriging of the local GLM network data to predict the annual average ground-level PM2.5 concentration ­ at the US Diplomatic Post and to compare this with the measured annual average at the post. An example of this analysis is shown in the top panel of figure B.2 for Ulaanbaatar. The gridded field produced via kriging the local GLM network data captures the measured value at the US Diplomatic Post well and somewhat captures the large gradient between the post and the Bukhiin Urguu monitoring site. The bottom panel of figure B.2 shows the kriging esti- mate when the US Diplomatic Post is included. The predictions near the post do not change much, but the combined kriging estimate suggests a more general region of low PM2.5 concentrations between Bukhiin Urguu and Amgalan. Based on this analysis, the local GLM network data can be used in combination with the US Diplomatic Post data in the training and evaluation of the satellite PM2.5 approaches. CTM-based approaches versus statistical approaches The CTM-based and statistical approaches were applied to the MODIS com- bined product (10-km resolution) to try to predict the daily average PM2.5 con- centration in Delhi, Lima, and Ulaanbaatar. Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 105 FIGURE B.2 Local Ulaanbaatar, Mongolia, ground-level-monitoring network and US Diplomatic Post annual average PM2.5 concentrations as well as a gridded estimate produced via ordinary kriging Micrograms per cubic meter a. Results without US Diplomatic Post Ulaanbaatar Kriging MNB US Diplomatic Post Tolgoit Bukhiin urguu Amgalan Baruun 4 zam Nisekh Annual average Kriging US NOT included 0.000000–37.238725 0.000000–37.238725 37.238726–57.079439 37.238726–57.079439 57.079440–67.650529 57.079440–67.650529 67.650530–87.491243 67.650530–87.491243 87.491244–124.729968 87.491244–124.729968 b. Results with US Diplomatic Post Ulaanbaatar Kriging MNB US Diplomatic Post Tolgoit Bukhiin urguu Amgalan Baruun 4 zam Nisekh Annual average Kriging US included 0.000000–37.238725 0.000000–37.238725 37.238726–57.079439 37.238726–57.079439 57.079440–67.650529 57.079440–67.650529 67.650530–87.491243 67.650530–87.491243 87.491244–124.729968 87.491244–124.729968 Source: World Bank, produced using Esri ArcGIS. Note: Spacing between kriging points is 0.01° latitude-longitude (about one kilometer). In panel a, the kriging estimate was produced using local ground-level-monitoring (GLM) network data only. In panel b, the kriging estimate was produced using local GLM data and US Diplomatic Post data. MNB = mean normalized bias; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. 106 | Getting Down to Earth After filtering for valid AOD values over surface sites from OpenAQ with valid daily average PM2.5 measurements, there were no matching observations for the city of Lima, Peru (figure B.3). This is because of the presence of persistent clouds over the region, the fact that this is a coastal city and thus many satellite footprints will have a mix of land and water surfaces, and the relative lack of GLM network data currently in OpenAQ for this city. Thus, Lima could not be evaluated any further in this initial work, but CTM-based ground-level PM2.5 estimates were produced for it. Future work will determine if this coverage issue is likely to be a large problem for all coastal cities. Similarly, in Ulaanbaatar, there were no valid MODIS AOD values for the entire winter (December, January, February, and most of March) because of per- sistent snow cover. This limits the ability of satellites to represent the true annual average ground-level PM2.5 concentrations in this city, as discussed further below. Statistical approach In Delhi, the statistical method AOD/PBLH was a statistically significant (at the α = 0.001 level), but fairly poor, predictor of daily average ground-level PM2.5 values, with R of 0.09 (for MODIS Aqua, 13:30 local standard time overpass) and 0.18 (for MODIS Terra, 13:30 local standard time overpass). Sensitivity studies fitting AOD and PBLH separately slightly increased the correlation (R of 0.21), but in general there was little correlation of MODIS combined product (Deep Blue plus Dark Target) AOD with ground-level PM2.5 estimates. AOD/PBLH was a slightly better predictor in Ulaanbaatar, with R values of 0.30 and 0.18 for MODIS Aqua and Terra, respectively. However, the overall cor- relation was still fairly poor. In addition, note that there was not a consistent pattern in which the MODIS satellite gave the better correlation with ground- level data, with the afternoon Aqua values performing better in Ulaanbaatar and the morning Terra values performing better in Delhi. Assessing the GAM fits by examining the residuals showed significant issues. The residuals deviated significantly from a normal distribution and showed a strong dependence on the predicted value—both indications of a poor statistical fit. An example for Ulaanbaatar using MODIS Aqua is shown in figure B.4, which was the best performing statistical fit. Performing simple linear fits of daily aver- age PM2.5 and AOD also gave poor correlations, and thus the poor fit is likely not due to the chosen statistical approach (that is, choice of GAM link function), but rather is representative of the low correlation between daily MODIS AOD values and daily average ground-level PM2.5 values. Figure B.5 shows a scatterplot of the observed and GAM-predicted daily aver- age PM2.5 values for the Income Tax Office at Delhi (left) and the US Diplomatic Post at Ulaanbaatar (right) for both Aqua (top) and Terra (bottom). The US Diplomatic Post at Delhi was not plotted because a couple of extreme values make it difficult to evaluate the plots. Also note that as only days with a valid AOD value were used, the days plotted for Aqua and Terra were not the same subset of days. The GAM-predicted values tend to have a smaller range of vari- ability than the observed values and exhibit a high level of scatter about the 1:1 line. Tables B.6 and B.7 give the 2017 mean bias (MB), mean normalized bias (MNB), mean normalized gross error (MNGE), and R statistics for all satellite AOD methods tested for Delhi and Ulaanbaatar, respectively. As expected, the statistical approach tends to have a low MB, because the GAM fitting procedure attempts to minimize the bias. However, the MNGE can still be very large Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 107 FIGURE B.3 Example of satellite coverage showing the CTM–based estimates for daily average ground-level PM2.5 concentrations for all valid MODIS Terra and Aqua aerosol-optical-depth retrievals, Lima, Peru, May 12, 2017 Micrograms per cubic meter a. MODIS Terra Cities MOD CTM PM2.5 2.201392–5.856531 5.856832–9.512269 9.512270–13.167708 13.167709–16.823146 16.823147–20.478584 NAN b. MODIS Aqua Cities MYD CTM PM2.5 2.201392–5.856531 5.856832–9.512269 9.512270–13.167708 13.167709–16.823146 16.823147–20.478584 NAN Source: World Bank, produced using Esri ArcGIS. Note: CTM = chemical transport model; MOD = MODIS Terra; MODIS = Moderate-Resolution Imaging Spectroradiometer; MYD = MODIS Aqua; NAN = not a number; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. 108 | Getting Down to Earth FIGURE B.4 Residual evaluation plots for Ulaanbaatar, Mongolia, using the MODIS Aqua aerosol- optical-depth product a. Residuals versus theoretical quantiles b. Residuals versus linear predictions 100 100 Residuals (µg/m3) Residuals (µg/m3) 50 50 -50 -50 -100 -100 -50 0 50 3.0 3.5 4.0 4.5 Theoretical quantiles Linear predictors c. Histogram of residuals d. Responses versus fitted values 120 200 100 Responses (µg/m3) 150 80 Frequency 60 100 40 50 20 0 0 -50 0 50 100 20 40 60 80 100 120 140 Residuals (µg/m3) Fitted values (µg/m3) Source: World Bank. Note: Panel a shows the residual percentiles (y) plotted against the predicted quantiles if the residuals were following a normal distribution (x). Panel b shows the residuals (y) plotted against the value of the linear predictor. The cone shape of the plot suggests a poor statistical fit. Panel c shows the histogram of the residuals. Panel d shows the plot of the generalized additive model response (y) versus the measurements (fitted values). MODIS = Moderate-Resolution Imaging Spectroradiometer; μg/m3 = micrograms per cubic meter. TABLE B.6  Statistics for the Delhi, India, satellite ground-level PM2.5 products tested in this work MODIS TERRA MODIS AQUA VIIRS SUOMI-NPP STATISTIC STAT. CTM STAT. CTM STAT. CTM MB (micrograms per cubic meter) 0.045 −0.822 0.113 −29.275 −0.014 −58.102 MNB (%) 30.8 21.7 24.4 −7.5 32.3 −27.3 MNGE (%) 51.5 54.5 44.1 46.8 54.2 52.3 Correlation coefficient (R) 0.18 0.17 0.09 0.04 0.11 0.10 Source: World Bank. Note: “STAT.” is the statistical method, and “CTM” is the chemical transport model-based method. MB = mean bias; MNB = mean normalized bias; MNGE = mean normalized gross error; MODIS = Moderate-Resolution Imaging Spectroradiometer; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; Suomi-NPP = Suomi-National Polar-Orbiting Partnership; VIIRS = Visible Infrared Imaging Radiometer Suite. Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 109 FIGURE B.5 Observed and GAM–predicted daily average PM2.5 values for the US Diplomatic Posts in Delhi, India, and Ulaanbaatar, Mongolia, for both MODIS Aqua and Terra Micrograms per cubic meter a. MODIS Aqua, Income Tax Office, b. MODIS Aqua, Diplomatic Post, Delhi, India Ulaanbaatar, Mongolia 300 100 GAM-predicted average (µg/m3) GAM-predicted average (µg/m3) 250 80 200 60 150 40 100 50 20 0 0 0 50 100 150 200 250 300 0 20 40 60 80 100 OpenAQ average (µg/m3) OpenAQ average (µg/m3) c. MODIS Terra, Income Tax Office, d. MODIS Terra, Diplomatic Post, Delhi, India Ulaanbaatar, Mongolia 500 100 GAM-predicted average (µg/m3) GAM-predicted average (µg/m3) 400 80 300 60 200 40 100 20 0 0 0 100 150 200 250 300 0 20 40 60 80 100 OpenAQ average (µg/m3) OpenAQ average (µg/m3) Source: World Bank. Note: GAM = generalized additive model; MODIS = Moderate-Resolution Imaging Spectroradiometer; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. (44–78 percent), again showing that the statistical models have little skill in capturing daily variability or the variability between sites in these cities. ­ CTM-based approach Figure B.6 shows a scatterplot of the observed and CTM-based estimates of daily average PM2.5 values for the Income Tax Office at Delhi (left) and the US Diplomatic Post at Ulaanbaatar (right) for both the MODIS Aqua (top) and Terra (bottom). The CTM-based estimate for Delhi does a reasonable job representing the mean and variability of the observed values, but individual daily predictions can have significant errors. In contrast, in Ulaanbaatar, the CTM-based estimate 110 | Getting Down to Earth TABLE B.7  Statistics for the Ulaanbaatar, Mongolia, satellite ground-level PM2.5 products tested in this work MODIS TERRA MODIS AQUA STATISTIC STAT. CTM STAT. CTM MB (micrograms per cubic meter) 0.082 −31.174 0.443 −33.139 MNB (%) 54.0 −72.2 54.8 −73.1 MNGE (%) 78.5 78.2 78.3 78.8 Correlation coefficient (R) 0.17 0.0 0.30 0.01 Source: World Bank. Note: “STAT.” is the statistical method, and “CTM” is the chemical transport model-based method. MB = mean bias; MNB = mean normalized bias; MNGE = mean normalized gross error; MODIS = Moderate-Resolution Imaging Spectroradiometer; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. FIGURE B.6 Observed and CTM–based estimates daily average PM2.5 values for the US Diplomatic Posts in Delhi, India, and Ulaanbaatar, Mongolia, for both MODIS Aqua and Terra Micrograms per cubic meter a. MODIS Aqua, Income Tax Office, b. MODIS Aqua, Diplomatic Post, Delhi, India Ulaanbaatar, Mongolia 300 100 250 80 CTM average (µg/m3) CTM average (µg/m3) 200 60 150 40 100 20 50 0 0 0 50 100 150 200 250 300 0 20 40 60 80 100 OpenAQ average (µg/m3) OpenAQ average (µg/m3) c. MODIS Terra, Income Tax Office, d. MODIS Terra, Diplomatic Post, Delhi, India Ulaanbaatar, Mongolia 500 100 400 80 CTM average (µg/m3) CTM average (µg/m3) 300 60 200 40 100 20 0 0 0 100 200 300 400 500 0 20 40 60 80 100 OpenAQ average (µg/m3) OpenAQ average (µg/m3) Source: World Bank. Note: CTM = chemical transport model; MODIS = Moderate-Resolution Imaging Spectroradiometer; OpenAQ = openaq.org; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 111 almost always underestimates the observed values and appears to be doing a poorer job than the statistical approach. This is consistent with the statistics shown in tables B.6 and B.7. The R sta- tistics for the CTM-based approaches using MODIS data tend to be very low (0.01–0.17), which combined with the high MNGE values suggests that the CTM-based method also has little skill in capturing daily variability or the variability between sites in these cities. However, in contrast to the statistical models, MODIS Terra AOD seems to perform better for Ulaanbaatar in the CTM-based approach. The CTM-based approach has difficulty reproducing daily average values. However, further examination of the CTM-based results for Delhi shows that the scaling of the MERRA-2 daily average PM2.5 surface concentrations by the MODIS Terra AOD did result in reasonable values for the annual averages at each surface site. Table B.8 shows the comparison between the observed annual average at each surface site (averaged only over those days with both a valid AOD and a valid daily average PM2.5 value, and with the outliers—greater than 600 micrograms per cubic meter—filtered for the US Diplomatic Post), the CTM- based estimate of this annual average, and the annual average predicted by the MERRA-2 model. Using the satellite AOD tends to increase the MERRA-2 ground-level PM2.5 concentration by about a factor of two, bringing it much closer to the observed annual averages at each surface site for those days where the surface sites had a valid satellite AOD. However, the low spatial resolution of MODIS AOD product and the MERRA-2 data mean that the CTM-based prod- uct has little skill in predicting the variations in the annual average PM2.5 values at each site. TABLE B.8  Annual average PM2.5 surface concentrations for Delhi, India, 2017 Micrograms per cubic meter LOCATION GROUND AVERAGE CTM-BASED AVERAGE MERRA-2 AVERAGE Anand Vihar 153.5 125.8 62.7 Delhi Technological University 123.9 113.5 58.5 Institute of Human Behavior and Allied Sciences 98.9 124.9 63.0 Income Tax Office 119.0 116.9 58.0 Mandir Marg 97.3 111.6 57.4 Netaji Subhas Institute of Technology Dwarka 140.2 112.9 62.4 Punjabi Bagh 103.6 117.0 55.8 Ramakrishna Puram 143.2 122.5 67.9 Shadipur 128.7 117.1 58.1 US Diplomatic Post 113.8 116.3 62.0 Urban average 122.2 117.9 60.6 Source: World Bank. Note: Annual average surface concentrations are as determined by the GLM data from OpenAQ, the CTM-based satellite approach using MODIS (Moderate-Resolution Imaging Spectroradiometer) Terra AOD, and the original MERRA-2 (Modern Era Retrospective-analysis for Research and Applications) output. These averages include only those days with both a valid AOD and a valid GLM daily average PM2.5 value. AOD = aerosol optical depth; CTM = chemical transport model; GLM = ground-level monitoring; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. 112 | Getting Down to Earth In addition, the uneven satellite AOD coverage over the year could potentially bias the urban annual averages, since some months would be weighted more heavily than others. Thus, the annual average for all GLM network sites in Delhi was calculated as well, not just those with valid satellite AODs. For Delhi, this true GLM average (133.8 micrograms per cubic meter) is 9.5 percent higher than that calculated using only those sites with valid AOD observations, likely because of the poor satellite coverage during the dry season when PM2.5 concentrations are expected to peak. This suggests that, even if the conversion of satellite AOD to ground-level PM2.5 were perfected for this city, the uneven coverage would bias the satellite-based annual average about 10 percent lower than the true average. The research conducted for this report tested whether the CTM-based approach in Delhi could be used to estimate monthly averages at a given site. The US Diplomatic Post site was used for this comparison. The monthly average con- centrations are shown in figure B.7 for the GLM data, the CTM-based estimate, and the monthly average predicted by the MERRA-2 model. As before, the scaling with the satellite AOD improves upon the MERRA-2 prediction, but neither the MERRA-2 output nor the CTM-based estimate captures the observed seasonal cycle of PM2.5 at this location, tending to overestimate during the wet season and underestimate during the dry. Several of these months (December, January, July, FIGURE B.7 Monthly average PM2.5 surface concentration, 2017 Micrograms per cubic meter 800 700 600 PM2.5 (µg/m3) 500 400 300 200 100 0 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month OpenAQ (AOD) CTM GAM predict OpenAQ (All) Source: World Bank. Note: Monthly average PM2.5 surface concentration is as determined by the GLM data, the CTM-based satellite approach using MODIS Terra AOD, and the original MERRA-2 output. These averages include only those days with both a valid AOD and a valid GLM daily average PM2.5 value. AOD = aerosol optical depth; CTM = chemical transport model; GLM = ground-level monitoring; MERRA = Modern Era Retrospective–analysis for Research and Applications; OpenAQ = openaq.org; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 113 and August) also had fewer than 10 valid matches with the MODIS Terra AOD, which may contribute to the errors. Thus, although the CTM-based approach appears to do a reasonable job for a citywide annual average in Delhi, it does not show any skill in determining the spatial or seasonal variation of PM2.5. The situation is worse for Ulaanbaatar. The use of satellite AOD to scale the MERRA-2 predictions of ground-level PM2.5 does improve upon the raw MERRA-2 output for citywide annual averages. However, the initial MERRA-2 estimate is so low (6.1 micrograms per cubic meter) that the scaled value (7.1 micrograms per cubic meter) is still a dramatic underprediction of the observed value (40.3 micro- grams per cubic meter) for days when there is a valid satellite AOD. This suggests that in Ulaanbaatar, the relationship between ground-level PM2.5 and the AOD is so different from the MERRA-2 estimate that the CTM-based approach does not appear to add any value to the existing GLM data set. Thus, using the CTM-based approach in Ulaanbaatar would require running a CTM that gives a better initial estimate of the vertical profile and optical properties of the aerosol in this region. Further progress in using satellite observations in Ulaanbaatar would likely bene- fit from adding a SPARTAN network site to the city. However, due to the lack of satellite AOD during the winter months, the true annual average PM2.5 concentration (88.1 micrograms per cubic meter) is much higher than that calculated when only the ground sites with valid AOD observa- tions are used (40.3 micrograms per cubic meter). This suggests that, even if the conversion of satellite AOD to ground-level PM2.5 were perfected for Ulaanbaatar, the lack of satellite AOD values in winter would bias the satellite-based annual average about 50 percent lower than the true average. Comparison with GBD 2016 estimates In Delhi, the Global Burden of Disease (GBD) 2016 data set5 gives 2016 annual averages of about 135 micrograms per cubic meter at both 0.1° by 0.1° resolution and 0.01° by 0.01° resolution after GWR bias correction. However, the GBD 2016 estimate before GWR bias correction is 105 micrograms per cubic meter, consis- tent with the 2017 CTM-based estimate for Delhi of 112 micrograms per cubic meter, and with the explanation of how the relative lack of valid satellite AOD data during December–January could bias the satellite-derived annual average low for these cities. In Ulaanbaatar, the GBD 2016 data set interestingly gives 2016 annual averages of 30 micrograms per cubic meter at 0.1° by 0.1° resolution and 42 micrograms per cubic meter at 0.01° by 0.01° resolution (after GWR bias correction). These results are pretty close to the “best possible” satellite estimate of 40.3 micrograms per cubic meter, given the lack of wintertime AOD data. However, the GBD 2016 estimate before GWR bias correction is 7 micrograms per cubic meter, consistent with the 2017 CTM-based estimate for this city. Thus, it appears that (1) the GBD 2016 data set also has difficulty representing the annual average in this city, (2) most of the improvement in Ulaanbaatar in GBD 2016 is coming from the global GWR bias correction, and (3) even with this correction, the annual aver- age is underestimated by about a factor of two. Thus, both the GBD 2016 results and our results suggest that GLM is required to get an accurate annual average PM2.5 concentration in Ulaanbaatar. 114 | Getting Down to Earth Comparing satellite AOD data sets VIIRS versus MODIS in Delhi, India Using the VIIRS AOD retrievals instead of the MODIS products in both the CTM-based and statistical approaches was attempted for Delhi. In the statistical approach, the quality of fit using the VIIRS AOD was not appreciably better than that using MODIS Aqua (which has a similar afternoon equator crossing time). With VIIRS the R was 0.1, as opposed to 0.08 for MODIS Aqua, and there was similar evidence of a poor statistical fit. The scatterplot of the GAM predictions versus the observed daily average values at the Income Tax Office site (panel a of figure B.8) also does not indicate any improved performance relative to using MODIS Aqua or Terra data. The MNB and MNGE are slightly worse when VIIRS is used instead of MODIS Aqua (table B.6). Using the CTM-based approach with VIIRS AOD data does result in greater spatial coverage over a city (figure B.9), but it results in poorer performance rel- ative to using MODIS (panel b of figure B.8). When the VIIRS data are used, the CTM approach tends to underestimate the ground-level PM2.5 concentration, leading to an underestimate of about a factor of two of the annual average ground-level PM2.5 concentration (that is, VIIRS-based prediction of 74.8 micro- grams per cubic meter versus an observed value of 140.3 micrograms per cubic meter). Thus, unlike the case when the MODIS Terra AOD was used, using the afternoon VIIRS AOD does not significantly improve the MERRA-2 predicted annual average (62.9 micrograms per cubic meter). This is reflected in the large mean bias for this approach in table B.6. MODIS versus SEVIRI in Accra, Ghana The predictions of the CTM-based method using the geostationary SEVIRI instrument were compared with those from the polar-orbiting MODIS FIGURE B.8 Observed values and VIIRS-based estimates of daily average PM2.5 concentrations for the Income Tax Office in Delhi, India, using GAM and CTM–based approaches Micrograms per cubic meter a. GAM results b. CTM results 600 600 GAM-predicted average (µg/m3) 500 500 CTM average (µg/m3) 400 400 300 300 200 200 100 100 0 0 0 100 200 300 400 500 600 0 100 200 300 400 500 600 OpenAQ average (µg/m ) 3 OpenAQ average (µg/m ) 3 Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter; VIIRS = Visible Infrared Imaging Radiometer Suite. Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 115 FIGURE B.9 A comparison of the CTM–based estimates for daily-average ground- level PM2.5 concentrations over Delhi, India, using VIIRS and MODIS Terra, November 1, 2017 Micrograms per cubic meter a. Using VIIRS Delhi MOD VIIRS Data November 1, 2017 Delhi VIIRS CTM PM2.5 79.241139–110.780485 110.780486–142.319832 142.319833–173.859178 173.859179–205.398525 205.398526–236.937871 NAN b. Using MODIS Terra Delhi MOD CTM Data November 1, 2017 Delhi MOD CTM PM2.5 79.241139–110.780485 110.780486–142.319832 142.319833–173.859178 173.859179–205.398525 205.398526–236.937871 NAN Source: World Bank, produced using Esri ArcGIS. Note: CTM = chemical transport model; MOD = MODIS Terra; MODIS = Moderate-Resolution Imaging Spectroradiometer; NAN = not a number; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; VIIRS = Visible Infrared Imaging Radiometer Suite. 116 | Getting Down to Earth FIGURE B.10 A comparison of the CTM–based estimates for daily-average ground-level PM2.5 concentrations over Accra, Ghana, using SEVIRI and MODIS Aqua, December 26, 2017 Micrograms per cubic meter a. SEVIRI data Cities SEVIRI CTM PM2.5 11.674649–34.925684 34.925685–58.176718 58.176719–81.427752 81.427753–104.678786 104.678787–127.929820 NAN b. MODIS Aqua data Cities MYD CTM PM2.5 11.674649–34.925684 34.925685–58.176718 58.176719–81.427752 81.427753–104.678786 104.678787–127.929820 NAN Source: World Bank, produced using Esri ArcGIS. Note: CTM = chemical transport model; MODIS = Moderate-Resolution Imaging Spectroradiometer; MYD = MODIS Aqua; NAN = not a number; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; SEVIRI = Spinning Enhanced Visible and Infrared Imager; VIIRS = Visible Infrared Imaging Radiometer Suite. Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 117 instrument, using the coastal city of Accra as the test case. The results are shown in figure B.10 for a single day in December. The SEVIRI product has a much higher spatial resolution, as expected, but also has less coverage near the coast. Thus, the SEVIRI product has limited coverage near the city of Accra itself, although it does have reasonable coverage for the area further inland. In con- trast, the MODIS combined (Deep Blue and Dark Target) AOD product has much better coverage for this period, although the over-ocean CTM-based val- ues appear higher than expected. The low coastal coverage of the SEVIRI prod- uct suggests that it likely would not improve the ability of satellites to monitor PM2.5 in Accra over using the MODIS product. Bias corrections of the CTM-based estimates A GWR correction to the Stage 1 CTM-based ground-level PM2.5 estimates in Delhi was attempted with the spgwr software package in the R statistical soft- ware platform using only the in-city data. A GWR is like a simple linear regres- sion except the coefficients are allowed to vary with latitude and longitude. The first attempt tried to fit the daily bias in the Stage 1 estimates to land-use vari- ables (percentage of urban land cover, elevation, and population density) using a simple linear regression model. This showed that these variables have little abil- ity to predict the bias in the Stage 1 estimates (R less than 0.01). However, the MERRA-2 predicted values of the concentration (micrograms per cubic meter) of ground-level PM2.5 due to organic carbon, sulfate, dust, and sea salt had some skill in predicting the bias, with an R value of 0.37 for a simple linear model fit, but implementing this regression equation as a GWR model instead does not appreciably improve the model fit (R of 0.37). This is likely because Delhi has only 10 GLM sites, as opposed to the 1,440 sites used in van Donkelaar and others (2015), and thus allowing the coefficients to vary with location does not provide much of a benefit. Allowing the regression coefficients to vary with location likely will not provide much of a benefit on a city scale, but this approach could be useful when using high-quality, harmonized GLM measurement networks on country or continent scales, such as exist in North America (van Donkelaar and others 2015) and western Europe (Vienneau and others 2013). Using the simple linear model to predict the bias and applying this bias correc- tion did improve the correlation of the daily average CTM-based satellite predic- tions (R of 0.23 compared to 0.15 before bias correction). However, this simple bias correction results in several data points predicted to have negative values, and thus the bias-corrected value produced a much lower citywide annual average (85.4 micrograms per cubic meter) than the initial CTM-based estimate. Training a simple linear model to predict the actual PM2.5 daily average con- centration using the above MERRA-2 variables and the Stage 1 CTM-based estimate was then attempted. This did remove the model mean bias for the urban-scale annual average and gave an R value of 0.33, an MNB of 20 percent, and an MNGE of 38 percent. However, the Stage 1 CTM-based estimate was not  a significant predictor in this model, and a similar fit can be found by removing the satellite AOD-derived estimate entirely and just fitting the MERRA-2–speciated aerosol variables. Similarly, applying a bias correction to the Ulaanbaatar data does fix the mean bias in the citywide annual average, but the CTM-based estimate was still unable to predict daily and site-to-site variability (R of 0.07). Thus, the statistical bias corrections of the CTM-based estimates attempted here do not significantly 118 | Getting Down to Earth improve the ability of the satellite data to represent the daily and site-to-site variability of PM2.5 in either Delhi or Ulaanbaatar. CONCLUSIONS AND RECOMMENDATIONS Several novel methods for using satellite AOD from publicly available sources to predict observed ground-level PM2.5 daily averages in Lima, Delhi, Ulaanbaatar, and Accra, for 2017 were tested. The research conducted for this report found several limitations in the use of satellite AOD in these cities. For example, all cities had significant limitations in satellite AOD coverage due to persistent clouds, wintertime snow cover, or mixed water-land and other bright surfaces. Satellite coverage was poorest in the coastal city of Lima, and thus a meaningful evaluation of a CTM-based method or the creation of a statistical method was not possible for this city. In Ulaanbaatar, no satellite observations are available for the high PM2.5 winter months of December to mid-March, and thus even a perfect method for converting satellite AOD to ground-level PM2.5 would under- estimate the true annual average PM2.5 concentration for this city by 50 percent. Delhi has observations in all months, but substantially fewer observations in the peaks of the wet and dry seasons (December, January, July, and August), which would also result in a slight (about 10 percent) underestimate of the true annual average. A CTM-based approach could provide a reasonable citywide annual average in Delhi, but the CTM-based approach leads to dramatic underestimates of the citywide annual average in Ulaanbaatar. Similar results were seen in the GBD 2016 data set before GWR bias correction. This is likely due to Ulaanbaatar being in a river valley surrounded by mountains and mostly rural land. Thus, the coarse resolution of global CTMs (combined with likely inaccurate emission invento- ries for Ulaanbaatar) means that CTMs are not able to correctly represent the aerosol profile within the city, instead using an average profile more representa- tive of the surrounding rural area. However, neither the CTM-based approach nor the statistical approach tested here were able to represent more than 30 percent of the variability in the daily average PM2.5 values within a city over the year, and many techniques had little correlation with the corresponding daily averages. Site-specific monthly average values were more reasonable in Delhi, but the lack of data during several months means that the satellites did a poor job of representing the observed sea- sonal cycle of PM2.5. Using VIIRS instead of MODIS did not appreciably change the overall evalu- ation results in Delhi, but the VIIRS product does provide a slightly higher spa- tial resolution than MODIS (6-kilometer versus 10-kilometer). Although the SEVIRI AOD product does provide a higher spatial resolution, it appears to have lower coverage for coastal cities such as Accra. The following recommendations are made for LMICs considering the use of satellite AOD in their PM2.5 monitoring: • The physical limits of satellite AOD coverage will make their use in PM2.5 monitoring difficult in many cities. ■ In cities with persistent wintertime snow cover, like Ulaanbaatar, it is likely not possible to get accurate annual averages from current satellites. Converting Satellite Aerosol Optical Depth to Ground-Level PM2.5 | 119 ■ In coastal cities, the mix of water and land surfaces within a satellite foot- print and persistent clouds may also mean that using satellite AOD is not an option. • CTM-based methods are also likely to fail for cities that are in apprecia- bly different air-quality environments than their surroundings, such as cities in mountain valleys surrounded by rural land. This means that such cities in Tier I and II countries (that lack the GLM data for statisti- cal approaches) will likely not be able to use satellites in their PM2.5 monitoring. • Estimating variations in annual average PM2.5 within a city (to estimate chronic health effects) is unlikely to be possible with satellite AOD data using the approaches tested here. This will require GLM network data and land-use regression, with the satellite AOD product or the satellite-based PM2.5 esti- mate used as a predictor variable in the LUR. • Both the CTM-based and statistical approaches tested here showed little ability to represent the day-to-day variability in PM2.5 concentrations, with average errors (MNGE) of ±50 percent for the best approaches within each city. Thus, studies of acute health effects will likely require GLM data. • For cities interested in using satellite AOD in their PM2.5 monitoring, estab- lishing at least one SPARTAN network site within the city to directly measure the relationship between AOD and ground-level PM2.5 will likely help to reduce the errors in the approaches tested here. NOTES 1. https://ladsweb.modaps.eosdis.nasa.gov/archive/allData/6/MOD04_L2. 2. https://www.avl.class.noaa.gov/saa/products/welcome. 3. http://www.icare.univ-lille1.fr. 4. https://disc.gsfc.nasa.gov/. 5. Downloaded from http://fizz.phys.dal.ca/~atmos/martin/?page_id=140. REFERENCES Chatfield, R., M. Sorek-Hamer, A. Lyapustin, and Y. Liu. 2017. “Daily Kilometer-Scale MODIS Satellite Maps of PM2.5 Describe Wintertime Episodes.” Paper 263193 presented at the “Air and Waste Management Association 2017 Annual Conference and Exhibition,” Pittsburgh, PA, June 5–9. GBD (Global Burden of Disease) 2016 Risk Factors Collaborators. 2017. “Global, Regional, and National Comparative Risk Assessment of 84 Behavioural, Environmental and Occupational, and Metabolic Risks or Clusters of Risks, 1990–2016: A Systematic Analysis for the Global Burden of Disease Study 2016.” Lancet 390: 1345–422. Remer, L. A., S. Mattoo, R. C. Levy, and L. A. Munchak. 2013. “MODIS 3 km Aerosol Product: Algorithm and Global Perspective.” Atmospheric Measurement Techniques 6: 1829–44. Sorek-Hamer, M., A. W. Strawa, R. B. Chatfield, R. Esswein, A. Cohen, and D. M. Broday. 2013. “Improved Retrieval of PM2.5 from Satellite Data Products Using Non-Linear Methods.” Environmental Pollution 182: 417–23. van Donkelaar, A., R. V. Martin, M. Brauer, R. Kahn, R. Levy, C. Verduzco, and P. J. Villeneuve. 2010. Global Estimates of Ambient Fine Particulate Matter Concentrations from Satellite- Based Aerosol Optical Depth: Development and Application.” Environmental Health Perspectives 118 (6): 847–55. 120 | Getting Down to Earth van Donkelaar, A., R. V. Martin, R. J. Spurr, and R. T. Burnett. 2015. “High-Resolution Satellite- Derived PM2.5 from Optimal Estimation and Geographically Weighted Regression over North America.” Environmental Science & Technology 49 (17): 10482–91. Vienneau, D., K. de Hoogh, M. J. Bechle, R. Beelen, A. van Donkelaar, R. V. Martin, D. B. Millet, G. Hoek, and J. D. Marshall. 2013. “Western European Land Use Regression Incorporating Satellite- and Ground-Based Measurements of NO2 and PM10.” Environmental Science & Technology 47 (23): 13555–64. APPENDIX C Evaluation of Satellite Approaches The methods (discussed in appendix B) for using satellite observations to determine ground-level particulate matter with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5) were applied to nine selected low- and middle-income country (LMIC) cities (table C.1) that cover a mixture of regimes (coastal versus inland, high versus low altitude, tropical versus temperate, and so forth). These cities also have available the ground-level-­ ­ monitoring (GLM) data needed for validation, through either US Diplomatic Posts, local monitoring networks, or both. This appendix presents the results of the evaluation of the satellite approaches for each city in table C.1. A 1° by 1° latitude-longitude swath of Moderate-Resolution Imaging Spectroradiometer (MODIS) afternoon Aqua (MYD) and morning Terra (MOD) data was downloaded for each city, centered on the coordinates listed in table C.2. Modern Era Retrospective-analysis for Research and Applications (MERRA) data were also downloaded for each site. For PM2.5 monitor data, OpenAQ data from 2016 and 2017 were used and processed for the cities. In addition, individ- ual PM2.5 data sets for Accra, Ghana, and Dakar, Senegal, were provided respec- tively by the Environmental Protection Agency of the Ghana Ministry of Environment, Science, Technology and Innovation (EPA Ghana) and the Center for Air Quality Management, a part of the Senegal Ministry of the Environment and Sustainable Development. The analyses of the data sets for Accra and Dakar focused on the years 2015 and 2011, respectively, since these years had the most-recent fairly continuous PM2.5 data for each city. As noted in table C.2, not all cities had GLM sites, and some had only a few. For each city, both the statistical and chemical-transport-model (CTM)– based approaches described in appendix B were attempted to determine ground- level PM2.5 concentrations. Some cities had too few GLM observations to allow a statistical approach to be performed; for these, only the CTM-based approach was evaluated. The statistical and CTM-based PM2.5 estimates were then com- pared with the GLM values for the cases where a valid satellite aerosol optical depth (AOD) existed for the GLM site, and a statistical and graphical evaluation of the match between the satellite estimates and the GLM observations was performed.  121 122 | Getting Down to Earth TABLE C.1  Cities included in this study CITY COUNTRY LOCATION INCOME GROUP Accra Ghana Coastal, low altitude Lower middle Addis Ababa Ethiopia Inland, high altitude Low Dakar Senegal Coastal, low altitude Low Delhi India Inland, low altitude Lower middle Hanoi Vietnam Inland, low altitude Lower middle Kampala Uganda Inland but near lake, high altitude Low Kathmandu Nepal Inland, high altitude Low Lima Peru Coastal, low altitude Upper middle Ulaanbaatar Mongolia Inland, high altitude Lower middle Source: World Bank. TABLE C.2  Coordinates and number of ground-level monitoring sites, by city CITY COUNTRY LATITUDE (°) LONGITUDE (°) SITES Accra Ghana 5.55° −0.2° 1 Addis Ababa Ethiopia 9.03 38.74 2 Dakar Senegal 14.6928 −17.4467 4a Delhi India 28.61 77.23 10 Hanoi Vietnam 21.0283 105.854 1 Kampala Uganda 0.313611 32.5811 1 Kathmandu Nepal 27.7114 85.3086 4 Lima Peru −12.0433 −77.0283 10 Ulaanbaatar Mongolia 47.9167 106.917 8 Source: World Bank. a. Dakar data are available only as an average of the four reporting sites. In addition, for cities with hourly PM2.5 data from the GLM monitoring per- formed at US Diplomatic Posts, the average ground-level PM2.5 concentrations during the two-hour window around the satellite overpass time (about 10:30 and 13:30 local solar time at the equator for Terra and Aqua, respectively) were com- pared to the daily average concentrations. This was done to estimate how well a “perfect” polar-orbiting satellite, which is able to observe a location only once per day, would be able to determine the daily average PM2.5 concentration in a city. This evaluation, which requires only GLM data, provides an upper limit on the ability of polar-orbiting satellites to determine ground-level PM2.5 ­concentrations in each city and can be used by LMICs to assess whether satellite data will be appropriate for their monitoring plans. The second section contains the evaluations for each city. In addition, the third section has data on the changes that can be introduced by applying the OpenAQ-developed Quality Filter to the OpenAQ GLM data. The last section describes an experiment to see if the use of satellite data could allow for a reduc- tion in the number of GLM sites needed to characterize ground-level PM2.5 con- centrations and exposure within a city, using Delhi, India, and Ulaanbaatar, Mongolia, as test cases. Evaluation of Satellite Approaches | 123 EVALUATION FOR EACH CITY Delhi, India Delhi, India, was one of the few cities that had enough data to train a statistical (generalized additive model or GAM; see appendix B) model with both Aqua and Terra satellite data. Figure C.1 shows the evaluation plots for the residuals of the GAM trained on the Terra data. The response residuals do not follow a normal distribution, which indicates that the statistical fit is fairly poor. Figure C.2 shows that the statistical estimate of ground-level PM2.5 increases monotonically with the ratio of the satellite aerosol optical depth (AOD) to the planetary boundary layer (PBL) height, as expected, but the slope of the fit appears to be dominated by two high values of AOD / boundary layer height (PBLH). Figure C.3 shows a scatterplot of the ground-level PM2.5 estimate from the CTM-based method and the statistical method, respectively, against the OpenAQ data for the US Diplomatic Post monitor site. The other OpenAQ sites have sim- ilar scatterplots (not shown). Table C.3 shows the evaluation statistics from both FIGURE C.1 Residual evaluation plots for the generalized additive model trained on the Terra satellite data, Delhi, India a. Residuals versus theoretical quantiles b. Residuals versus linear predictions 1,500 1,500 Residuals (µg/m3) Residuals (µg/m3) 1,000 1,000 500 500 0 0 -300 -200 -100 0 100 200 300 4.5 5.0 5.5 6.0 Theoretical quantiles Linear predictors c. Histogram of residuals d. Responses versus fitted values 800 1,500 Responses (µg/m3) 600 Frequency 1,000 400 200 500 0 0 0 500 1,000 1,500 100 200 300 400 500 600 Residuals (µg/m )3 Fittted values (µg/m ) 3 Source: World Bank. Note: μg/m3 = micrograms per cubic meter. 124 | Getting Down to Earth FIGURE C.2 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Terra aerosol optical depth and the MERRA planetary-boundary-layer height, Delhi, India 4 2 log of PM2.5 concentration 0 -2 1 0 0 0.0005 0.0010 0.0015 0.0020 0.0025 0.0030 Aerosol Optical Depth divided by planetary boundary layer height Source: World Bank. Note: Green dots = individual observations; red bar = uncertainty of the fit; MERRA = Modern Era Retrospective-analysis for Research and Applications; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. FIGURE C.3 Scatterplots of the Terra satellite estimate of PM2.5 from the generalized additive model and the chemical transport model versus the OpenAQ data, US Diplomatic Post in Delhi, India a. GAM results b. CTM results GAM-predicted average (µg/m3) CTM average (µg/m3) 1,500 1,500 1,000 1,000 500 500 0 0 0 500 1000 1500 0 500 1000 1500 OpenAQ average (µg/m3) OpenAQ average (µg/m3) Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. Evaluation of Satellite Approaches | 125 TABLE C.3  Delhi, India, evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods TERRA AQUA STATISTIC GAM CTM GAM CTM Correlation coefficient (R) 0.456 0.437 0.328 0.273 Mean bias (micrograme percubic meter) −0.475 −0.884 0.079 −30.267 Mean normalized bias (%) 36.9 26.9 29.4 −6.1 Mean normalized gross error (%) 59.8 60.4 50.1 49.3 Root-mean-square error (micrograms per cubic meter) 89.4 98.3 83.5 95.5 Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model. the statistical and CTM-based methods with respect to the OpenAQ data. The correlation coefficients are fairly low (maximum R of 0.46) but the statistical and CTM-based methods using the Terra data have low mean biases (less than 1 microgram per cubic meter). Comparing these correlation coefficients with the correlation coefficients between the GLM PM2.5 concentrations averaged around the satellite over- pass time (two-hour window) and the daily average PM2.5 concentrations provides an upper limit on the correlation between the once-daily polar-or- biting satellite observations and the GLM data. For Delhi, these c ­ orrelations are 55 percent at the Terra overpass time and 64 percent at the Aqua overpass time. Interestingly, these are the opposite of the results in table C.3, which imply that Terra AOD has a better correlation than Aqua AOD. The current performance of the Terra AOD (R of 44–46 percent for the daily average PM2.5) is fairly close to the “upper limit” of 55 percent expected for this sat- ellite based on the GLM data correlations. Figure C.4 shows the monthly averages calculated using the CTM-based and statistical methods and compares them to the values calculated using the OpenAQ data, either using all the available data (OpenAQ All) or only the data on days with a valid AOD retrieval (OpenAQ AOD). During the summer months (wet season), both the CTM-based and statistical methods overestimate the PM2.5 values, and they tend to underestimate in the winter months (dry season). Figure C.5 shows the annual average PM2.5 estimated for each method by mon- itoring site. As noted in appendix B, the satellite methods generally do a poor job of representing the site-to-site variations of the annual average PM2.5 within the city. Ulaanbaatar, Mongolia Ulaanbaatar was another city with several monitor sites, and so generalized additive models (GAMs) were fitted for both the Terra and Aqua data, both gen- erally indicating a poor fit. Figure C.6 shows the evaluation plots for the residu- als of the GAM trained on the Aqua data. The response residuals do not follow a normal distribution, which indicates that the statistical fit is fairly poor. Figure C.7 shows that the statistical estimate of ground-level PM2.5 increases with the ratio of the satellite AOD to the PBL height for low values of AOD/ PBLH, as expected, but the fit is highly unconstrained at larger values, showing what are likely false maxima and minima. 126 | Getting Down to Earth FIGURE C.4 Monthly average PM2.5 for each method, Delhi, India 300 250 200 PM2.5 (µg/m3) 150 100 50 0 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Month OpenAQ (AOD) OpenAQ (All) CTM GAM prediction Source: World Bank. Note: AOD = aerosol optical depth; CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. FIGURE C.5 Annual average PM2.5 for each method, by monitor site, Delhi, India 180 160 140 120 PM2.5 (µg/m3) 100 80 60 40 20 0 r sit l S ce g ka gh m r el , er ca D st ha u BA ar ra ip w Po ar ffi y Ba hi iv gi Vi rM IH Pu ad w O Un lo Ne tic i d D ab no Sh di x na an Ta a IT an nj m ch ish An NS Pu e M lo Te m r ip ak co hi D m In el US Ra D Monitor site OpenAQ (AOD) OpenAQ (ALL) CTM GAM prediction Source: World Bank. Note: AOD = aerosol optical depth; CTM = chemical transport model; GAM = generalized additive model; NSIT = Netaji Subhas Institute of Technology; OpenAQ = openaq.org; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. Table C.4 shows the evaluation statistics from both the statistical and CTM-based methods with respect to the OpenAQ data. It is important to note that no satellite data were available for the winter months in Ulaanbaatar because of persistent snow cover. This means no data were available in Evaluation of Satellite Approaches | 127 FIGURE C.6 Residual evaluation plots for the generalized additive model trained on the Aqua satellite data, Ulaanbaatar, Mongolia a. Residuals versus theoretical quantiles b. Residuals versus linear predictions 150 150 100 Residuals (µg/m3) 100 Residuals (µg/m3) 50 50 0 0 -50 -50 -50 0 50 3.0 3.5 4.0 4.5 Theoretical quantiles Linear predictors c. Histogram of residuals d. Responses versus fitted values 150 200 Responses (µg/m3) 150 100 Frequency 100 50 50 0 0 -50 0 50 100 150 20 40 60 80 100 120 Residuals (µg/m3) Fittted values (µg/m3) Source: World Bank. Note: μg/m3 = micrograms per cubic meter. December–February for the GAM fits or CTM methods to use in the evalua- tion. The correlation coefficients are fairly low (maximum R of 0.44), and although the mean bias is fairly low for the statistical method as expected, the CTM-based method is consistently biased low, consistent with the results from appendix B. For Ulaanbaatar, the correlations (R) between the GLM PM2.5 concentra- tions averaged around the satellite overpass time (two-hour window) and the daily average PM2.5 concentrations are 66 percent at the Terra overpass time and only 31 percent at the Aqua overpass time. This correlation at the Aqua overpass time is lower than the R for the Aqua GAM in table C.4, which sug- gests that the statistical approach may be overfitting the data for this case. The high ­correlation with the Terra overpass time suggests that a better satellite instrument or analysis approach may be possible for this city, but it is unlikely that any new approach or instrument will overcome the loss of data during the snow-covered winter months. 128 | Getting Down to Earth FIGURE C.7 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Aqua aerosol optical depth and the MERRA planetary-boundary-layer height, Ulaanbaatar, Mongolia 4 2 log of PM2.5 concentration 0 -2 -4 0.0000 0.0005 0.0010 0.0015 0.0020 0.0025 0.0030 0.0035 Aerosol optical depth divided by planetary boundary layer height Source: World Bank. Note: Green dots = individual observations; red bar = uncertainty of the fit; MERRA = Modern Era Retrospective-analysis for Research and Applications; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. TABLE C.4  Evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods, Ulaanbaatar, Mongolia TERRA AQUA STATISTIC GAM CTM GAM CTM Correlation coefficient (R) 0.379 0.12 0.442 0.15 Mean bias (micrograms per cubic meter) 0.088 −28.473 0.232 −30.521 Mean normalized bias (%) 52.7 −66.5 53.9 −69.7 Mean normalized gross error (%) 77.2 78.6 78.5 78.1 Root-mean-square error (micrograms per cubic meter) 33.2 48.7 28.9 46.7 Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model. Figure C.8 shows a scatterplot of the GAM prediction method and the CTM method against the OpenAQ data, respectively, for the US Diplomatic Post mon- itor site. The CTM-based method is substantially overestimating the concentra- tion. For the statistical method, the bias is lower, but it still shows a very poor fit Evaluation of Satellite Approaches | 129 FIGURE C.8 Scatterplots of the generalized-additive-model– and chemical-transport-model–based concentrations using the Aqua aerosol-optical-depth data, versus OpenAQ data, US Diplomatic Post in Ulaanbaatar, Mongolia a. GAM results b. CTM results 100 100 GAM-predicted average (µg/m3) 80 80 CTM average (µg/m3) 60 60 40 40 20 20 0 0 0 20 40 60 80 100 0 20 40 60 80 100 OpenAQ average (µg/m ) 3 OpenAQ average (µg/m ) 3 Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; μg/m3 = micrograms per cubic meter. to the data, also reflected in the large mean normalized bias and mean normal- ized gross error in table C.4. Figure C.9 shows the citywide monthly averages calculated using each method. This plot shows only the months March through November as stated earlier because of the lack of satellite data during the winter months. Except for July, the CTM-based method greatly underestimates PM2.5. The statistical method does a better job of tracking the seasonal cycle of the citywide monthly average, except for November, likely because of low satellite coverage in this month. Figure C.10 shows the annual average PM2.5 for each method by monitor- ing site. It is reinforced here how the CTM-based method grossly underesti- mates PM2.5 for each site. The statistical method does well except for the sites with the highest values: Bayankhoshuu, Mongolia National Broadcaster, and Tolgoit. This may be due to the influence of local sources of PM 2.5 at these sites. Lima, Peru Lima, Peru, had 10 monitoring sites in the OpenAQ database. However, because of persistent cloud cover over Lima and the exclusion of mixed water and land pixels in the satellite retrieval process, only about 10 valid data points were available over OpenAQ monitors for Aqua and 12 for Terra in the years 2016–17. Figure C.11 shows an aerial view of the Lima area with the OpenAQ monitoring sites in green and the valid satellite AOD data for the two years in blue. Little to no data are available along the coast or over most of Lima, and thus satellite data cannot be used to monitor PM2.5 in this city. 130 | Getting Down to Earth FIGURE C.9 Monthly average PM2.5 for each method, Ulaanbaatar, Mongolia 160 140 120 PM2.5 (µg/m3) 100 80 60 40 20 0 Mar Apr May Jun Jul Aug Sep Oct Nov Month OpenAQ (AOD) CTM GAM prediction OpenAQ (All) Source: World Bank. Note: AOD = aerosol optical depth; CTM = chemical transport mode; GAM = generalized additive model; OpenAQ = openaq.org; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. FIGURE C.10 Annual average (March–November) PM2.5 for each method by monitor site, Ulaanbaatar, Mongolia 80 70 60 PM2.5 (µg/m3) 50 40 30 20 10 0 n m uu u te l kh t ta , as na aa st oi gu la za nb Po se sh lg r r ga dc io ur To Ni o oa at 4 Am aa ic kh iin Br ia N un Ul at n m kh ru ya ol lo Ba Bu Ba g ip on D M US Monitor site OpenAQ (AOD) CTM GAM prediction OpenAQ (All) Source: World Bank. Note: AOD = aerosol optical depth; CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; PM = particulate matter. PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. Evaluation of Satellite Approaches | 131 FIGURE C.11 Valid Terra aerosol-optical-depth retrievals for monitoring sites in the OpenAQ database, Lima, Peru, 2016–17 Lima Valid satellite AOD data OpenAQ Monitor Sites Source: World Bank, produced using Esri ArcGIS. Note: OpenAQ = openaq.org. Accra, Ghana Accra had no monitoring sites in the OpenAQ database, but data for PM2.5 con- centrations at several sites in 2015 were provided by the Environmental Protection Agency of the Ghana Ministry of Environment, Science, Technology and Innovation and were compared to the CTM-based estimates at these locations. Table C.5 shows the evaluation statistics for the CTM-based Terra estimates at each site. The correlation coefficients are fairly low (maximum R of 0.4, but for a site where the results are negatively correlated), with significant underesti- mates of PM2.5 at each site and large mean absolute errors. In addition, as shown in the scatterplot in figure C.12, there are very few data points with both valid AOD retrievals and GLM daily averages, and thus attempting a statistical method is unlikely to give much benefit for this city. 132 | Getting Down to Earth TABLE C.5  Evaluation statistics for the chemical-transport-model–based method using Terra data, Accra, Ghana TETTEH QUARSHIE WEIJA INTERCHANGE MALAM JUNCTION GRAPHIC STATISTIC TERRA AQUA TERRA AQUA TERRA AQUA TERRA AQUA Correlation coefficient (R) 0.04 0.40 0.05 0.04 0.12 0.06 0.06 0.04 Mean bias (micrograms per cubic −44.5 −38.5 −6.7 −16.9 −50.7 −56.4 −25.4 −37.4 meter) Mean normalized bias (%) −33.3 −22.6 −40.8 −2.9 −46.9 −18.4 −28.5 −15.8 Mean normalized gross error (%) 79.5 55.5 92.3 68.0 51.2 82.5 55.0 71.1 Root-mean-square error 76.1 71.1 63.3 50.0 72.5 84.6 49.3 75.0 (micrograms per cubic meter) Source: World Bank. FIGURE C.12 Scatterplot of the chemical-transport-model–based PM2.5 concentrations using the Terra aerosol-optical-depth data, versus ground-level-monitoring PM2.5 data, Malam Junction site in Accra, Ghana 200 180 160 140 CTM PM2.5 (µg/m3) 120 100 80 60 40 20 0 0 50 100 150 200 Malam Junction PM2.5 (µg/m3) Source: World Bank. Note: CTM = chemical transport model; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. Kathmandu, Nepal Kathmandu, Nepal, had two OpenAQ monitoring sites, both of which are US Diplomatic Posts, and these sites had data only for 2017. These sites had enough data for the statistical method to be used, but there were only a total of 50 matching satellite-GLM data points, and so monthly and annual averages could not be investigated. Figure C.13 shows the evaluation plots for the residuals of the GAM trained on the Aqua data. The response residuals do follow a normal distribution, which indicates that this statistical fit may be more realistic than those for Delhi and Ulaanbaatar. However, figure C.14 shows that the statistical estimate of ground-level PM2.5 does not strongly depend on the ratio of the sat- ellite AOD to the PBL height, with likely unphysical maxima and minima around the relatively low response. This suggests that the statistical fit may be Evaluation of Satellite Approaches | 133 FIGURE C.13 Residual evaluation plots for the generalized additive model trained on the Aqua satellite data, Kathmandu, Nepal a. Residuals versus theoretical quantiles b. Residuals versus linear predictions 40 40 Residuals (µg/m3) Residuals (µg/m3) 20 20 0 0 –20 –20 –40 –20 –20 20 40 3.6 3.8 4.0 4.2 4.4 4.6 Theoretical quantiles Linear predictors c. Histogram of residuals d. Responses versus fitted values 12 100 Responses (µg/m3) 10 80 Frequency 8 6 60 4 40 2 0 20 –40 –20 0 40 40 40 50 60 70 80 90 100 Residuals (µg/m3) Fitted values (µg/m3) Source: World Bank. Note: μg/m3 = micrograms per cubic meter. primarily fitting the errors in the comparison between the AOD/PBLH and the ground-level PM2.5 concentrations. Figure C.15 shows a scatterplot of the statistical (GAM) method and the CTM-based method against the OpenAQ data, respectively, for one of the US Diplomatic Post monitor sites. The statistical method seems to overestimate the monitored values, while the CTM-based method seems to be distributed better but tends to underestimate the values. Table C.6 shows the evaluation statistics from both the statistical and CTM-based methods with respect to the OpenAQ data. The correlation coef- ficient for the statistical model trained on the Aqua data is fairly high (R of 0.66), but this is likely an artifact of the relatively low number of matching satellite and GLM data points. This is supported by the fairly low correlation between the GLM data measured at the Aqua overpass time and the GLM daily average PM 2.5 ( R of 32  percent; value of Terra overpass time is 43 ­percent). The CTM-based method is biased low, similar to the results for  Ulaanbaatar, another high-altitude urban site in a mountain valley surrounded by relatively rural land. ­ 134 | Getting Down to Earth FIGURE C.14 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Aqua aerosol optical depth and the MERRA planetary-boundary-layer height, Kathmandu, Nepal 4 2 log of PM2.5 concentration 0 –2 –4 0.0000 0.0001 0.0002 0.0003 0.0004 0.0005 0.0006 Aerosol optical depth divided by planetary boundary layer height Source: World Bank. Note: Green dots = individual observations; red bar = uncertainty of the fit; MERRA = Modern Era Retrospective-analysis for Research and Applications; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. FIGURE C.15 Scatterplots of the generalized-additive-model– and chemical-transport-model–based concentrations using the Aqua aerosol-optical-depth data, versus OpenAQ data, US Diplomatic Post in Kathmandu, Nepal a. GAM results b. CTM results 100 100 GAM-predicted average (µg/m3) 80 80 CTM average (µg/m3) 60 60 40 40 20 20 0 0 0 20 40 60 80 100 0 20 40 60 80 100 OpenAQ average (µg/m3) OpenAQ average (µg/m3) Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; μg/m3 = micrograms per cubic meter. Evaluation of Satellite Approaches | 135 TABLE C.6  Evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods, Kathmandu, Nepal TERRA AQUA STATISTIC GAM CTM GAM CTM Correlation coefficient (R) 0.479 0.256 0.656 0.130 Mean bias (micrograms per cubic meter) 0.179 −19.768 0.255 −15.801 Mean normalized bias (%) 19.3 −30.1 8.4 −18.5 Mean normalized gross error (%) 39.1 52.1 24.4 47.8 Root-mean-square error (micrograms per cubic meter) 19.1 33.6 14.9 30.1 Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model. Addis Ababa, Ethiopia Addis Ababa, Ethiopia, had two sites in the OpenAQ database. Figure C.16 shows the evaluation plots for the residuals of the GAM trained on the Terra data. The response residuals do follow a normal distribution. Figure C.17 shows that the statistical estimate of ground-level PM2.5 is linear with the ratio of the satel- lite AOD to the PBL height, as expected. FIGURE C.16 Residual evaluation plots for the generalized additive model trained on the Terra satellite data, Addis Ababa, Ethiopia a. Residuals versus theoretical quantiles b. Residuals versus linear predictions 30 30 20 20 Residuals (µg/m3) Residuals (µg/m3) 10 10 0 0 –10 –10 –20 –20 –10 0 10 2.8 3.0 3.2 3.4 3.6 Theoretical quantiles Linear predictors c. Histogram of residuals d. Responses versus fitted values 70 60 15 Responses (µg/m3) 50 Frequency 10 40 30 5 20 0 10 –20 –10 0 10 20 30 15 20 25 30 35 40 Residuals (µg/m )3 Fitted values (µg/m ) 3 Source: World Bank. Note: μg/m3 = micrograms per cubic meter. 136 | Getting Down to Earth Figure C.18 shows the scatterplot of the statistical and CTM-based satellite estimates versus the OpenAQ data, and table C.7 shows the statistics from both the statistical and CTM-based methods with respect to the OpenAQ data. The CTM-based estimates tend to be biased high. The GAM trained on Terra data gives a fairly high value for the correlation coefficient, which is slightly higher than the correlation between the GLM data at the Terra overpass time and the GLM daily average PM2.5 (59 percent; 81 percent for Aqua overpass time), suggesting that the fit may be due to the low number of points with a valid AOD estimate. FIGURE C.17 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Terra aerosol optical depth and the MERRA planetary-boundary-layer height, Addis Ababa, Ethiopia 4 2 log of PM2.5 concentration 0 –2 –4 0.00005 0.00010 0.00015 0.00020 Aerosol optical depth divided by planetary boundary layer height Source: World Bank. Note: Green dots = individual observations; red bar = uncertainty of the fit; MERRA = Modern Era Retrospective-analysis for Research and Applications; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. TABLE C.7  Evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods, Addis Ababa, Ethiopia TERRA AQUA STATISTIC GAM CTM GAM CTM Correlation coefficient (R) 0.63 0.399 0.441 0.054 Mean bias (micrograms per cubic meter) −0.051 6.463 0.0003 15.786 Mean normalized bias (%) 6.9 35.5 7.4 92.2 Mean normalized gross error (%) 23.5 64.8 21.3 104.5 Root-mean-square error (micrograms per cubic meter) 7.0 17.5 6.2 28.9 Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model. Evaluation of Satellite Approaches | 137 FIGURE C.18 Scatterplots of the generalized-additive-model– and chemical-transport-model–based concentrations, using the Aqua aerosol-optical-depth data, versus OpenAQ Data, US Diplomatic Post in Addis Ababa, Ethiopia a. GAM results b. CTM results 300 300 GAM-predicted average (µg/m3) 250 250 CTM average (µg/m3) 200 200 150 150 100 100 50 50 0 0 0 50 100 150 200 250 300 0 50 100 150 200 250 300 OpenAQ average (µg/m3) OpenAQ average (µg/m3) Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; μg/m3 = micrograms per cubic meter. Dakar, Senegal Figure C.19 shows a scatterplot of the Dakar PM2.5 data for 2011 versus the CTM- based satellite estimate using the Terra data. Since the Dakar GLM PM2.5 data are the average of four sites, the CTM-based estimates were calculated for each site and then averaged before comparing them with the GLM data. Two valid AOD values (and thus valid CTM-based PM2.5 estimates) were required to consider the satellite average as a valid data point. FIGURE C.19 Scatterplot of the chemical-transport-model–based PM2.5 concentrations using the Terra aerosol-optical-depth data versus citywide average PM2.5 data, Dakar, Senegal 400 350 300 CTM PM2.5 (µg/m3) 250 200 150 100 50 0 50 100 150 200 250 300 350 200 Dakar average PM2.5 (µg/m3) Source: World Bank. Note: CTM = chemical transport model; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. 138 | Getting Down to Earth TABLE C.8  Evaluation statistics for the chemical-transport-model–based method, Dakar, Senegal STATISTIC TERRA AQUA Correlation coefficient (R) 0.35 0.52 Mean bias (micrograms per cubic meter) 24.3 23.0 Mean normalized bias (%) 86.0 67.5 Mean normalized gross error (%) 101.6 85.3 Root-mean-square error (micrograms per cubic meter) 45.3 41.4 Source: World Bank. Table C.8 shows the statistics for the comparison. The correlation coefficient is fairly high (R of 52 percent), but the CTM-based approach tends to overesti- mate the PM2.5 concentrations and has very large values for mean normalized bias and mean normalized gross error (86 percent and 101 percent), indicating fairly poor performance in this city. Kampala, Uganda Kampala, Uganda, has one OpenAQ site, but there were very few valid daily aver- ages in the OpenAQ data, which combined with the relatively low coverage of the AOD retrievals left few matching points for evaluation. There were only 14 matching points for the Terra data, and only 22 for Aqua, so only the Aqua data are discussed further here. The low number of points in the statistical fit for the Aqua data do not give us confidence in the results. The GAM evaluation plots in figure C.20 look reasonable, as does the linear dependence of the ground-level PM2.5 estimate on AOD/PBLH in figure C.21, but the low number of valid AOD points suggests that this may be coincidental. Figure C.22 shows the scatterplot of the statistical and CTM-based satellite estimates versus the OpenAQ data, and table C.9 shows the statistics from both the statistical and CTM-based methods with respect to the OpenAQ data. The GAM seems to perform well, but this is likely due to the low number of points with a valid AOD. The correlation coefficient for the GAM is larger than the cor- relations between the GLM PM2.5 data at the Aqua overpass time and the daily average PM2.5 (45 percent). In contrast, the CTM-based estimates appear to be biased low even for the small number of points available, and the correlation coefficient is very poor (14 percent). Evaluation of Satellite Approaches | 139 FIGURE C.20 Residual evaluation plots for the generalized additive model trained on the Aqua satellite data, Kampala, Uganda a. Residuals versus theoretical quantiles b. Residuals versus linear predictions 30 30 20 20 Residuals (µg/m3) Residuals (µg/m3) 10 10 0 0 –10 –10 –20 –20 –30 –20 –10 0 10 20 30 3.7 3.8 3.0 4.0 4.1 Theoretical quantiles Linear predictors c. Histogram of residuals d. Responses versus fitted values 30 7 6 20 Responses (µg/m3) 5 10 Frequency 4 0 3 –10 2 1 –20 0 –30 –20 –10 0 10 20 30 40 45 50 55 60 Residuals (µg/m3) Fitted values (µg/m3) Source: World Bank. Note: μg/m3 = micrograms per cubic meter. 140 | Getting Down to Earth FIGURE C.21 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Aqua aerosol optical depth and the MERRA planetary-boundary-layer height, Kampala, Uganda 4 log of PM2.5 concentration 2 0 -2 -4 0.0005 0.0010 0.0015 0.0020 0.0025 0.0030 Aerosol optical depth divided by planetary boundary layer height Source: World Bank. Note: Green dots = individual observations; red bar = uncertainty of the fit; MERRA = Modern Era Retrospective-analysis for Research and Applications; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. FIGURE C.22 Scatterplots of the generalized-additive-model– and chemical-transport-model–based PM2.5 concentrations, using the Aqua aerosol-optical-depth data, versus OpenAQ data, US Diplomatic Post, Kampala, Uganda a. GAM results b. CTM results 100 100 GAM-predicted average (µg/m3) 80 80 CTM average (µg/m3) 60 60 40 40 20 20 0 0 0 20 40 60 80 100 0 20 40 60 80 100 OpenAQ average (µg/m ) 3 OpenAQ average (µg/m ) 3 Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; μg/m3 = micrograms per cubic meter. Evaluation of Satellite Approaches | 141 TABLE C.9  Evaluation statistics for the generalized-additive-model– and chemical-transport-model–based methods, Kampala, Uganda AQUA STATISTIC GAM CTM Correlation coefficient (R) 0.514 0.139 Mean bias (micrograms per cubic meter) 0.018 −36.16 Mean normalized bias (%) 6.5 −69.5 Mean normalized gross error (%) 20.5 69.5 Root-mean-square error (micrograms per cubic meter) 11.3 39 Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model. Hanoi, Vietnam Hanoi, Vietnam, had one site in the OpenAQ data; it was possible to complete both Terra and Aqua GAM fits. The GAM evaluation plots in figure C.23 look reasonable, as does the near linear dependence of the ground-level PM2.5 ­estimate FIGURE C.23 Residual evaluation plots for the generalized additive model trained on the Terra satellite data, Hanoi, Vietnam a. Residuals versus theoretical quantiles b. Residuals versus linear predictions 40 30 20 20 Residuals (µg/m3) Residuals (µg/m3) 0 0 -20 –20 -40 –40 -40 -20 0 20 40 3.6 3.8 4.0 4.2 4.4 4.6 Theoretical quantiles Linear predictors c. Histogram of residuals d. Responses versus fitted values 10 120 8 100 Responses (µg/m3) 80 Frequency 6 4 60 40 2 20 0 –60 –40 –20 0 20 40 40 50 60 70 80 90 100 110 Residuals (µg/m ) 3 Fitted values (µg/m3) Source: World Bank. Note: μg/m3 = micrograms per cubic meter. 142 | Getting Down to Earth on AOD/PBLH in figure C.24, but the low number of valid matching GLM and AOD points (less than 33 for all cases) suggests that this may be coincidental. Figure C.25 shows the scatterplot of the statistical and CTM-based satellite estimates versus the OpenAQ data, and table C.10 shows the statistics from both the statistical and CTM-based methods with respect to the OpenAQ data. The GAM seems to perform well, but this may be due to the low number of points with a valid AOD. However, the correlation coefficient for the GAM is smaller than the correlations between the GLM PM2.5 data at the Terra overpass time and the daily average PM2.5 (R of 73 percent; 71 percent for Aqua). The CTM- based estimates are also relatively unbiased, although the correlation coefficient is low (30 percent). FIGURE C.24 Plot of the dependence of the ground-level PM2.5 estimate from the generalized additive model on the ratio of the Terra aerosol optical depth and the MERRA planetary-boundary-layer height, Hanoi, Vietnam 4 2 log of PM2.5 concentration 0 -2 -4 0.0005 0.0010 0.0015 Aerosol optical depth divided by planetary boundary layer height Source: World Bank. Note: Green dots = individual observations; red bar = uncertainty of the fit; MERRA = Modern Era Retrospective-analysis for Research and Applications; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. Evaluation of Satellite Approaches | 143 FIGURE C.25 Scatterplots of the generalized-additive-model– and chemical-transport-model– based concentrations using Terra aerosol-optical-depth data, versus OpenAQ data, US Diplomatic Post in Hanoi, Vietnam a. GAM results b. CTM results 100 100 GAM-predicted average (µg/m3) 80 80 CTM average (µg/m3) 60 60 40 40 20 20 0 0 0 20 40 60 80 100 0 20 40 60 80 100 OpenAQ average (µg/m3) OpenAQ average (µg/m3) Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; μg/m3 = micrograms per cubic meter. TABLE C.10  Evaluation statistics for the generalized-additive-model– and chemical- transport-model–based methods, Hanoi, Vietnam TERRA AQUA STATISTIC GAM CTM GAM CTM Correlation coefficient (R) 0.634 0.302 0.626 0.259 Mean bias (micrograms per cubic meter) 0.347 0.175 0.76 6.367 Mean normalized bias (%) 20.9 21.8 22.2 39.6 Mean normalized gross error (%) 42.5 53.4 39.1 67.8 Root-mean-square error (micrograms per cubic meter) 22.0 34.3 21.0 31.0 Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model. OPENAQ QUALITY CHECK TOOL OpenAQ worked with the organization Development Seed to create a command line tool1 that can aid in quality checking the OpenAQ data. This tool, called open- aq-quality-checks (QA [quality assurance] Tool), can flag each data point based on metrics given by the user, for example, missing values, negative values, or values greater than some user-provided threshold. This tool was applied to the Delhi data to see the change in results when additional quality measures are applied. The tool was used to remove values noted as missing (−999), values less than 1, values over a provided threshold, and data points where the same value was repeated for mul- tiple consecutive hours at the same site. The data for this city is subhourly. At 15-minute intervals, it makes sense that the concentration values may be the same. Therefore, the repeat option for this tool is not valid for this data set. For this exer- cise the tool was run to remove only values less than 1 (including missing values) and values over a calculated threshold. This threshold was derived by finding the 144 | Getting Down to Earth standard deviation of the log of hourly values for all sites for the two years and the mean of the log for all hourly values. The concentration threshold was calculated as three times this standard deviation plus the mean of the log. The QA Tool then removes any values over this threshold. For Delhi the threshold is 1,223 micro- grams per cubic meter. This does not prove that points above this threshold actu- ally represent poor data but instead merely tests the impact of applying this threshold to the analysis. Furthermore, the initial processing already removed negative and zero values, and so the main impact of the QA Tool is expected to be from applying this upper threshold. Figure C.26 shows the difference in the monthly averages of the base process- ing done to all sites (here for Delhi) and the QA Tool equivalent. It is apparent the US Diplomatic Post had unusually high values during a few months that the QA Tool removed. Table C.11 shows the 2016–17 average PM2.5 concentrations for each OpenAQ site in Delhi. Consistent with figure C.26, the only significant changes are for the US Diplomatic Post, where large values were filtered. FIGURE C.26 Difference in base processing monthly average versus monthly averages after the quality assurance tool is applied 100 80 Difference (μg/m3) 60 40 20 0 –20 –40 –60 ay g v ec ar b p r ct n n l Ap No Ju Au Fe Se Ju Ja M O M D Month Anand vihar IHBAS Mandir Marg Punjabi Bagh Shadipur Delhi Technological University IncomeTax Office NSIT Dwarka RK Puram US Diplomatic Post, New Delhi Source: World Bank. Note: IHBAS = Institute of Human Behavior and Allied Sciences; NSIT = Netaji Subhas Institute of Technology; RK Puram = Ramakrishna Puram; μg/m3 = micrograms per cubic meter. TABLE C.11  Average PM2.5 concentrations for each monitor site in Delhi, India, 2016–17 SITE QUALITY ASSURANCE TOOL BASE PROCESSING Anand Vihar 159.2 160.7 Delhi Technological University 156.3 152.2 Institute of Human Behavior and Allied Sciences 94.7 102.2 Income Tax Office 144.4 141.4 Mandir Marg 115.6 112.8 Netaji Subhas Institute of Technology Dwarka 132.0 134.6 Punjabi Bagh 129.8 128.4 Ramakrishna Puram 136.6 138.3 Shadipur 127.2 125.7 US Diplomatic Post: New Delhi 116.3 132.0 Source: World Bank. Note: PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. Evaluation of Satellite Approaches | 145 However, the investigation into the quality assurance and control (QA/QC) pro- cesses for the US Diplomatic Post data showed that they are following US EPA procedures for QA/QC and instrument maintenance, so it would seem unlikely that these points truly represent poor quality data. Using these data from the QA Tool, the same analysis scripts as before were run to merge the satellite data and train a GAM. Table C.12 shows the fit statis- tics of the GAM using the QA Tool data versus the data used in the base pro- cessing. Both runs with the QA Tool (Terra and Aqua) show a better fit than the base processing runs; however, an R value of 0.3 is still a poor fit. Figure C.27 TABLE C.12  Generalized-additive-model output statistics of quality assurance tool versus base processing, Delhi, India BASE PROCESSING QUALITY ASSURANCE TOOL STATISTIC TERRA AQUA TERRA AQUA Correlation coefficient (R) 0.204 0.102 0.3 0.196 Deviance explained 20.8% 10.7% 30.5% 20.0% Source: World Bank. FIGURE C.27 Generalized-additive-model residual evaluation plots from Terra quality assurance tool a. Residuals versus theoretical quantiles b. Residuals versus linear predictions 300 300 200 Residuals (µg/m3) 200 Residuals (µg/m3) 100 100 0 0 -100 -100 -200 -200 -200 -100 0 -100 200 4.5 5.0 5.5 6.0 Theoretical quantiles Linear predictors c. Histogram of residuals d. Responses versus fitted values 600 800 500 Responses (µg/m3) 600 400 Frequency 300 400 200 200 100 0 0 -300 -100 0 100 200 300 400 100 200 300 400 500 Residuals (µg/m3) Fitted values (µg/m3) Source: World Bank. Note: μg/m3 = micrograms per cubic meter. 146 | Getting Down to Earth shows the GAM residual evaluation plots from the Terra QA Tool run, which can be compared to figure C.1 from the base processing run. The residuals are much closer to a normal distribution after the QA filter is applied. Figure C.28 shows a scatterplot of the ground-level PM2.5 estimate from the CTM-based method and the statistical method, respectively, against the filtered OpenAQ data for the US Diplomatic Post monitor site (compare with ­ figure C.3). Table C.13 shows the evaluation statistics from both the statistical ­ and CTM-based methods with respect to the filtered OpenAQ data (compare with table C.3). The correlation coefficients have increased somewhat (maxi- mum R from 0.46 to 0.55) but that the overall bias and error statistics have not significantly changed. FIGURE C.28 Scatterplots of the statistical and chemical-transport-model–based concentrations using the Aqua aerosol-optical-depth data versus the QA filtered OpenAQ data, US Diplomatic Post in Ulaanbaatar, Mongolia a. GAM results b. CTM results GAM-predicted average (µg/m3) 800 800 600 CTM average (µg/m3) 600 400 400 200 200 0 0 0 200 400 600 800 0 200 400 600 800 OpenAQ average (µg/m3) OpenAQ average (µg/m3) Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org; QA = quality assurance; μg/m3 = micrograms per cubic meter. TABLE C.13  Evaluation statistics (using the filtered OpenAQ data) for the statistical and chemical-transport-model–based methods, Delhi, India TERRA AQUA STATISTIC GAM CTM GAM CTM Correlation coefficient (R) 0.552 0.537 0.447 0.354 Mean bias (micrograms per cubic meter) −0.349 2.642 0.078 −27.483 Mean normalized bias (%) 34.2 28.0 26.1 −6.1 Mean normalized gross error (%) 57.5 60.2 47.7 48.1 Root-mean-square error 68.2 78.8 56.5 71.8 (micrograms per cubic meter) Source: World Bank. Note: CTM = chemical transport model; GAM = generalized additive model; OpenAQ = openaq.org. Evaluation of Satellite Approaches | 147 CO-KRIGING AND KRIGING WITH SATELLITE AND OPENAQ DATA The goal of this test was to determine if satellite data, although not a replacement for GLM data, could be combined with GLM data in a way that reduces the num- ber of GLM sites required to quantify exposure in an urban population. To test this, ordinary kriging was done for each day using the monitor station daily aver- ages, excluding the US Diplomatic Post. The concentration at the US Diplomatic Post for each day was then predicted with ordinary kriging. This tested how well the other GLM sites alone could characterize the PM2.5 concentrations at an unmeasured site. The CTM-based satellite PM2.5 estimates were then combined with the GLM data (minus the US Diplomatic Post) via co-kriging and again esti- mated the concentrations at the US Diplomatic Post. The results for Delhi and Ulaanbaatar are shown in tables C.14 and C.15, respectively. The Delhi results suggest that adding in the satellite data would not reduce the number of monitoring sites there, because the root mean square is actually higher when the CTM-based satellite data are added. The Ulaanbaatar result suggests that the Diplomatic Post could be eliminated by adding satellite data to the rest of the GLM network in this city, but again only for the eight months a year that satellites actually produce data. TABLE C.14  Kriging and co-kriging statistics, Delhi, India STATISTIC CO-KRIGING KRIGING Mean bias (micrograms per cubic meter) −4.330 −5.935 Root-mean-square error (micrograms per cubic meter) 148.527 134.374 Source: World Bank. TABLE C.15  Kriging and co-kriging statistics, Ulaanbaatar, Mongolia STATISTIC CO-KRIGING KRIGING Mean bias (micrograms per cubic meter) −1.670 −12.400 Root-mean-square error (micrograms per cubic meter) 27.558 55.204 Source: World Bank. NOTE 1. https://github.com/openaq/openaq-quality-checks. APPENDIX D Quality Considerations for Ground- Level-Monitoring Data in Low- and Middle-Income Countries This appendix provides further details on the uncertainties in ground-level-­ monitoring (GLM) data in low- and middle-income countries (LMICs). There are two points of focus here: (1) a focus on the uncertainties associated with two commonly used measurement techniques; that is, the beta attenuation mass monitors (first section) and the tapered element oscillating microbalance (sec- ond section), and (2) a focus on evaluation of the consistency between the GLM data collected by the US Diplomatic Posts and the GLM data collected by local authorities in Delhi, India (third section) and Lima, Peru (final section). OPERATING METHODOLOGIES AND UNCERTAINTIES ASSOCIATED WITH THE BAMM TECHNIQUE Beta attenuation mass monitors (BAMMs), including the Met One model BAM-1020 and the Thermo Fisher Scientific BAMM model 5014i, are the most common instruments for monitoring particulate matter with an aerodynamic diameter less than or equal to 2.5 microns (PM2.5) across the globe. US Diplomatic Posts deploy these instruments for PM2.5 measurements. A PM2.5 measurement system typically includes a sample inlet, a PM cyclone, a sample conditioning unit (such as a heater to ensure modest relative humidity inside the instrument), the PM2.5 instrument, and a pump to draw a sample through the system. The beta attenuation monitor (BAM) technique samples ambient PM onto a filter and measures the attenuation of beta rays (high-energy electrons) through the PM-loaded filter in time. The beta rays emitted from a radioactive source (14C) are attenuated (absorbed) by PM collected on a filter: 1  Io  I = Io e − µx and x = ln  , (D.1) µ  I where I is the measured beta ray intensity (counts per unit of time) of the atten- uated beta ray (particle-laden filter tape), I0 is the measured beta ray intensity of the nonattenuated beta ray (clean filter tape), μ is the beta ray attenuation cross section of the PM (square meters per kilogram), and x is the mass density of the PM (kilograms per square meter) on the filter. Key to the success of the beta  149 150 | Getting Down to Earth attenuation monitor is, in part, that μ, the absorption cross section, varies little among commonly sampled PM such as C, Fe2O3, NH4NO3, NH4(SO4)2, or SiO2. This permits the device to be calibrated during the manufacturing process and permits the user to measure PM concentrations without having to know the chemical composition of the sampled PM in advance. PM2.5 mass concentrations are calculated from the measured attenuation. The ambient PM concentration of particulate matter (kilograms per cubic meter) is ( ) ( ) x kg / m2 A m2 D  Io  ( C kg / m =3 ) Q ( m / s ) × t (s ) 3 and ( ) C µg / m3 = ln µ  I  , (D.2) where C is the ambient PM concentration (kilograms per cubic meter), A is the cross-sectional area of the tape spot over which PM is being deposited (m2), Q is the rate at which ambient air is being sampled (cubic meters per second), t is the sampling time (s), and D is the ratio of A/(Q × t) and unit conversion constant (109). The BAM-1020 operates on a one-hour cycle starting at the top of each hour with filter movement between each of the following: (1) measurements starting with an 8-minute beta measurement on a clean filter spot at the beginning of each hour; (2) a sampling period of 42 minutes; (3) an 8-minute beta measure- ment of a particle loaded filter spot; and (4) an explicit cycle taking a total of 58 minutes with the other 2 minutes of the hour used for tape and nozzle move- ments during the cycle. Figure D.1 shows a detailed picture of the BAM’s sample head, filter tape, and beta source and detector. FIGURE D.1 Picture of the Met One beta attenuation monitor unit showing the beta source and the filter tape used to collect the particulate matter samples Beta source Pinch Roller Latch Nozzle with motor with latch photosensor and two nozzle photosensors Capstan Shaft with Reference Membrane with motor and two motor and photosensor photosensors Shuttle Beam (back side) connected Shuttle Beam with to other side shuttle photosensor Left Tensioner Idler Right Tensioner no photosensors Idler with tape break and tension photosensors Tape Take-Up Tape Supply Reel with motor Reel with motor Tape Transport Assembly Beta detector Source: © Antelope Valley Air Quality Management District 2021. Reproduced from https://avaqmd.ca.gov/43ffdbb. Quality Considerations for Ground-Level-Monitoring Data in Low- and Middle-Income Countries | 151 Measurement uncertainties in the BAM’s PM2.5 can be obtained from estimates of uncertainties, the equation for PM2.5 mass concentrations given above, and standard error propagation. The random (precision) uncertainties in the BAM’s PM2.5 calculation include (1) flow rate, sample time, spot size (minor, about 2 percent); (2) calibration (μ) and span values (10 percent); (3) counting statistics (I and Io) (about 2.5 micrograms per square meter); and (4) baseline and electronic drift (minor, about 2 percent). Other uncertainties and systematic bias are more difficult to assess because they are operationally dependent on relative humidity changes during sampling, site-location factors, operator error, and volatile PM components that may evaporate from the filter during sampling. The manufacturer states an uncertainty (σ) of 2.4 micrograms per cubic meter (variance of 72 one-hour averages of zero particle measure- ments). Thus, any single hourly measurement should fall within 3σ = 7.2 micro- grams per cubic meter of the expected value. OPERATING METHODOLOGIES AND UNCERTAINTIES ASSOCIATED WITH THE TAPERED ELEMENT OSCILLATING MICROBALANCE TECHNIQUE Tapered element oscillating microbalance (TEOM) measurement principle: Sample air is drawn through a filter cartridge (Teflon-coated glass fiber) mounted on the tip of a hollow glass tube (the tapered element) at 3 liters per minute (split from a total 16.67 liters per minute flow through a size-cut sampling inlet; figure D.2). The base of the tube cannot move, but the tip is free to vibrate at its natural fre- quency. Thus, the tube acts as a tuning fork or a hollow cantilever beam with an associated spring rate and mass. Any additional weight from particles that col- lect on the filter changes the frequency at which the tube oscillates. The elec- tronic circuitry senses this change and calculates the particle mass rate from the magnitude of the frequency change. The instrument then returns the vibrating glass tube to its natural frequency ready for the next measurement. The TEOM maintains a constant temperature and flow rate and electronically smooths the readings to reduce noise. Dividing the mass rate by the flow rate provides a con- tinuous output of the particle mass concentration. Measurement of mass: The frequency of a spring-mass system follows the equation 0.5  K f =  , (D.3)  M where f is the frequency of oscillation of the tapered element with the particle filter, K is the spring constant, and M is the mass of the tapered element. To measure the particulate matter (PM) collected on the particle filter, differ- ences in mass are obtained by  1 1  dm = K0  2 − 2  , (D.4)  f1 f0  where dm is the change in mass of the tapered element due to collected PM, K0 is the spring constant and unit conversion constants (calibrated using known masses), and f0 and f1 are the initial and final frequencies (hertz) for a measurement. 152 | Getting Down to Earth FIGURE D.2 Schematic diagram of a tapered element oscillating microbalance PM10 sampling inlet Tapered element oscillating (TSP, PM2.5) microbalance (TEOM) Flow splitter 13.7 L/min 3 L/min Heated sensor unit By-pass Mass transducer flowline (tapered element) Amplifier and frequency In-line counter filters Mass flow controllers Exhaust Vacuum pump Source: © The State of Queensland 2021 (CC BY 4.0). https://www.qld.gov.au/environment/pollution​ /­monitoring/air-pollution/oscillating-microbalance. Note: L/min = liters per minute; PM = particulate matter; TSP = total suspended particulate. The ambient concentration of particulate matter (micrograms per cubic meter) is then dm ( µg ) ( C µg /m3 = ) , (D.5) ( ) Q m3 / s × t ( s ) where C is the ambient PM concentration (micrograms per cubic meter), dm is the change in mass of the tapered element due to collected PM (micrograms), Quality Considerations for Ground-Level-Monitoring Data in Low- and Middle-Income Countries | 153 Q is the rate at which ambient air is being sampled (cubic meters per second), and t is the sampling time (seconds). TEOM measurement uncertainties: dm ( C µg / m3 =) (Q ×t) (D.6) Random (precision): • Flow rate, (minor, about 2 percent) • Calibration (K0) and span values (10 percent) • Baseline and electronic drift (minor, about 2 percent) • Noise due to difference measurements (significant at short time frames) Systematic (bias): • RH dependence (addressed via sample heating) • Environmental vibrations (affect stability of oscillations) • Volatile PM (specifically nitrates and semivolatile organics) • Operator error Uncertainties typically are combined via “root-sum-of-squares”: ∂C 2 ∂C ∂C σ 2 (C ) = ∑ ∂ xi σ ( x i ) + 2∑ ∑ ( ∂ xi ∂ x j ) σ xi , x j . (D.7) The manufacturer states precision uncertainty of ±2.0 micrograms per cubic meter for a one-hour average: • If this is a single σ precision, any single measurement should fall within 3σ = 6.0 micrograms per cubic meter of the expected value • The TEOM stated accuracy uncertainty of ±0.75 percent for a one-hour aver- age on the mass measurement (dm) • If this is a single σ accuracy, any single one-hour measurement should fall within 3σ = 2.25 percent of the expected value (this needs to be combined with flow rate–measurement uncertainties) • Thus, TEOM measurement uncertainty (for a well-operated system under the Environmental Protection Agency protocols and requirements) would be the greater of these uncertainties • The unknown factors include site location, volatile PM losses, and operator errors. There are several TEOM base models (1400a, 1400b, 1405) and methods (standard and FDMS). The base TEOM instrument is the same instrument detection scheme using the same tapered element oscillating microbalance. The various TEOM model numbers refer to the following: • Model 1400a: Sampling with PM2.5 cyclone • Model 1400b: Sampling with PM10 cyclone • Model 1405: Sampling with a virtual impactor that separates particles into fine (PM2.5) and coarse particle modes (PM10–PM2.5) • Models with –F or –DF: include a Filter Dynamics Measurement System (FDMS) upstream of the TEOM (see schematic in figure D.3). The TEOM 1405-DF Dichotomous Ambient Particulate Monitor with FDMS is designated as federal equivalent method EQPM-0609-182 for PM2.5. In this 154 | Getting Down to Earth FIGURE D.3 Tapered element oscillating microbalance–filter dynamics measurement system schematic a. Assembly b. Flow schematic PM10 Inlet Inlet flow 16.67 Inlet flow 16.67 Ipm Nafion 1405-DF: 1.67 Ipm Virtual dryers impactor to coarse channel Isokinetic flow Heat Flow splitter exchanger splitter 47 mm RH Fine fraction flow: TX40 filter sensors Bypass flow: 12 Ipm 3.0 Ipm to PM2.5 (purge) channel Bupass flow Purge (12.0 Ipm) Switching outlet valve Roof line Dryer Conditioning section Fine fraction Heat flow (3.0 Ipm) Coarse fraction exchanger flow (1.67 Ipm) Purge inlet Heated Vacuum main flow air tube pump Purge assembly filter %RH sensor FDMS switching control valve Bass MC path FDMS (green only) tower Purge flow to Pump Reference Dry purge flow MC flow path Sample (green + blue) collection Main Mass Inertial mass filters flow transducer transducers Bypass Main mass mass flow flow controller controller Source: Standard Operating Procedure for the Continuous Measurement of Particulate Matter Thermo Scientific TEOM® 1405-DF Dichotomous Ambient Particulate Monitor with FDMS® Federal Equivalent Method EQPM-0609-182 for PM2.5 STI-905505.03-3657-SOP, available at https://www3.epa.gov​ /­ttnamti1/files/ambient/pm25/sop_project/905505_TEOM_SOP_Draft_Final_Sept09.pdf (pages 3-3 and 3-4). Note: FDMS = Filter Dynamics Measurement System; lmp = liters per minute; MC = mass concentration; mm = millimeters; PM = particulate matter. version of the instrument, all sample flows are passed through a diffusion dryer to remove water vapor. Every six minutes, a switching valve alternates the sam- ple flows between the base- and reference-sample periods. During the base period, the PM sample is collected normally, and the differential mass is mea- sured at 30° Celsius (C). At this temperature, volatile PM will possibly evaporate and not be measured. During the reference period, the sample flow is diverted through a chilled filter (4°C to 10°C) to remove and retain both the nonvolatile and volatile PM prior to passing through the particle filter of the TEOM. Thus, the reference period is when dry, particle-free air is passed through the particle filter of the TEOM. The “PM” mass measurement during this period is a differ- ential mass measurement that is either near zero or negative, providing an approximate measure of the amount of volatile PM that evaporates from the TEOM particle filter. Ambient PM concentrations are calculated from base-reference mass measurements. Logging parameters, recommended main- tenance, and calibration details are shown below:1 Quality Considerations for Ground-Level-Monitoring Data in Low- and Middle-Income Countries | 155 • 1405-F/DF TEOM recommended logging parameters:2 ■ FEM MC ■ PM2.5 Base MC ■ PM2.5 Ref MC ■ PMc MC ■ PMc Base MC ■ PMc Ref MC ■ Ambient temperature ■ Ambient humidity ■ Ambient pressure ■ Vacuum pump pressure ■ Status ■ TEOM A filter loading ■ TEOM A dryer temperature ■ TEOM A dryer dew point ■ TEOM B filter loading ■ TEOM B dryer temperature ■ TEOM B dryer dew point ■ Three additional available for user selection (20 total) • Recommended maintenance intervals: ■ Replace TEOM filter: Filter loading nears 75 percent or every 30 days ■ Replace chilled filter: When you exchange the TEOM filter ■ Clean PM10 inlet: With every TEOM filter exchange ■ Clean virtual impactor (1405-DF): With every TEOM filter exchange ■ Replace in-line filter: Every 6 months ■ Clean coolers: Once a year ■ Clean switching valve: Once a year ■ Clean air inlet system: Once a year ■ Rebuild vacuum pump: 12 to 18 months ■ Dryer refurbishment: Once a year • Recommended interval of frequency: ■ Ambient temperature: Audit monthly/calibration yearly ■ Ambient pressure: Audit monthly/calibration yearly ■ Flow (PM2.5, coarse, bypass): Audit monthly/calibration yearly ■ Leak check: Monthly ■ Analog outputs: Once a year ■ Mass transducer: Audit once a Year INTERPRETING GROUND-LEVEL PM2.5 DATA FROM LOW- AND MIDDLE-INCOME COUNTRIES WITH LIMITED METADATA In many low- and middle-income countries (LMICs), there are organizations, researchers, scientists, and members of the general public who are keenly interested in understanding their local air quality, especially if air pollution epi- sodes occur with any regularity and in turn directly affect the way people feel. Nearly all publicly accessible information on real-time air pollution concentra- tions is just that—a single data point, often a PM2.5 mass concentration. Little to 156 | Getting Down to Earth no metadata are available with which to draw additional concrete expectations for instrument performance, accuracy, precision, or reliability. In such situa- tions, one way to evaluate the legitimacy of the reported PM2.5 levels is to com- pare measurements from different monitoring stations across a given area of interest. More populous LMICs often have multiple government-run air-quality monitoring stations spread across each city. In recent years, US Diplomatic Posts throughout these same LMICs have established a PM2.5 monitoring capacity (equipped with BAM instruments) on embassy roofs. Accessible data from local and US government sites within the same city present the opportunity to evalu- ate the consistency of reported PM2.5 concentrations in the area. This concept is demonstrated below by examining seasonal, monthly, weekly, and diurnal trends in PM2.5 data across Delhi, India, and Lima, Peru, throughout the years 2016–17. Delhi, India Map D.1 reveals a shaded area of about 50 square miles, the perimeter of which is defined with air-quality monitoring sites in Delhi. Site 1 corresponds to the location of the US Diplomatic Post, and sites 2–7 are government-run monitoring stations equipped to measure PM2.5 mass concentrations. Working with the OpenAQ online database, PM2.5 data measured in 2016 and 2017 were retrieved for each of the stations shown on the map. The series of plots displayed below show the time-series, seasonal, monthly, weekly, and diurnal profiles in PM2.5 for each monitoring location. In this metadata “blind” scenario, the US Diplomatic Post represents slightly higher reliability than the local government stations based on the fact that embassy staff employ the same standard operating proce- dure as recommended by the EPA to keep the BAM system maintained and in MAP D.1 Proximity and distribution of government-run air-quality monitoring stations around the US Diplomatic Post in Delhi, India Source: World Bank, produced using Google Maps. Note: CPCB = Central Pollution Control Board; DPCC = Delhi Pollution Control Committee. Quality Considerations for Ground-Level-Monitoring Data in Low- and Middle-Income Countries | 157 good working order. These same assurances are not necessarily present in the local government data feeds, and thus the direct head-to-head comparisons with embassy data can be a useful tool for validating the government PM2.5 trends. The top panel of the analysis frame (figure D.4) shows the PM2.5 time series, season-average PM2.5, monthly average PM2.5, and weekly average PM2.5. The lower panel in each analysis frame breaks down the PM2.5 trends further, reveal- ing diurnal (hourly) profiles for the overall data set and then each of the four seasons. One should take careful notice of not only the relative trends observed at each location but also the magnitudes of the peaks and valleys in PM2.5. In each series of graphs, the PM2.5 concentration axes (all y-axes in the graph series) are normalized to the maximum observed concentration (or error bar) in that graph. The embassy BAM data reveal the strong seasonal dependence of Delhi’s PM2.5 concentrations, with extremely high (greater than 200 micrograms per cubic meter) levels observed between November and January. There is virtually zero weekday versus weekend difference in PM2.5 loadings, and hourly diurnal profiles demonstrate the pronounced influence of boundary layer dynamics comingled with pollution source intensity and activity profiles: higher PM2.5 concentrations overnight and into the morning hours when the boundary layer is lowest with a significant decrease in pollutant concentration in the mid-to-late afternoon as the boundary layer rises, diluting the PM2.5 levels. To provide an empirical basis to assess the consistency of ground-level PM2.5 measurements throughout Delhi, the same series of graphs are plotted in figures D.5 through D.10. Each figure corresponds to a different government-run air-quality moni- toring location (marked as 2–7 on map D.1). Generally, the observed seasonal, monthly, and weekly trends in PM2.5 levels from the government stations are consistent with the PM2.5 data acquired at the embassy. There are some FIGURE D.4 Time-series, seasonal, monthly, weekly, and diurnal PM2.5 mass concentration data measured from the US Diplomatic Post in Delhi, India a. Time-series b. Seasonal c. Monthly d. Weekly 1400 300 300 250 1200 250 250 200 1000 200 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 200 800 150 150 150 600 100 100 400 100 50 50 200 50 0 0 0 0 12/1/20163/1/2017 6/1/2017 9/1/201712/1/2017 DJF MAM JJA SON n b ar M r ay n l g p ct v ec n on es ed s i t Ju Fr Ap Sa ur No Ja Ju Su Au Fe Se O M Tu D W M Th e. Diurnal overall f. Diurnal winter g. Diurnal spring h. Diurnal summer i. Diurnal autumn 300 350 200 80 300 250 300 250 150 200 60 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 200 150 200 100 40 150 100 100 100 50 20 50 50 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. 158 | Getting Down to Earth FIGURE D.5 PM2.5 trends from data acquired at the Mandir Marg, New Delhi, India–DPCC monitoring site, 1.45 miles west of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 500 250 800 200 400 200 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 600 150 150 300 400 100 100 200 200 50 50 100 0 0 0 5/1/2016 11/1/2016 5/1/2017 11/1/2017 DJF MAM JJA SON ay g Dv ec b ar p n on es ed s i t r ct n n l Fr Sa Ap ur No Ju Au Su Fe Se Ju Ja M O M Tu W M Th e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 350 200 100 250 400 300 80 150 200 250 300 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 60 200 150 100 150 200 40 100 100 50 100 50 20 50 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. DPCC = Delhi Pollution Control Committee; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. FIGURE D.6 PM2.5 trends from data acquired at the Shadipur, New Delhi, India–CPCB monitoring site, 4.75 miles northwest of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 200 300 400 250 200 250 200 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m ) 3 200 200 300 150 150 200 200 100 100 200 100 50 50 0 0 0 0 3/1/2017 6/1/2017 9/1/2017 12/1/2017 DJF MAM JJA SON n on es ed s i t Fr Sa ur ay g Dv ec b ar p r ct Su n n l Ap Tu No Ju Au Fe Se W M Ju Ja Th M O M e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 400 400 300 300 80 250 250 300 300 60 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 200 200 200 150 200 150 40 100 100 100 20 100 50 50 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. CPCB = Central Pollution Control Board; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. Quality Considerations for Ground-Level-Monitoring Data in Low- and Middle-Income Countries | 159 FIGURE D.7 PM2.5 trends from data acquired at the Punjabi Bagh, Delhi, India–DPCC monitoring site, 6.91 miles northwest of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 350 300 800 600 300 250 500 600 250 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 200 400 200 150 400 300 150 200 100 100 200 50 100 50 0 0 0 0 5/1/2016 11/1/2016 5/1/2017 11/1/2017 DJF MAM JJA SON n on es ed ur i t n b Apr M r ay n Au l g p ct Dv ec Fr Sa Ju a Su No Ja Ju Fe Se Th Tu O M W M e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 250 500 400 100 300 400 200 80 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 300 150 300 200 60 200 200 100 40 100 100 50 100 20 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. DPCC = Delhi Pollution Control Committee; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. FIGURE D.8 PM2.5 trends from data acquired at the NSIT Dwarka, Delhi, India–CPCB monitoring site, 11 miles southwest of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 300 400 250 800 250 200 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 300 PM2.5 (µg/m3) 600 200 150 150 400 200 100 100 200 100 50 50 0 0 0 0 3/1/2017 6/1/2017 9/1/2017 12/1/2017 DJF MAM JJA SON n b Apr M r ay n Au l g p No t Dv ec n on es ed ur i t Ju Fr a c Sa Ja Ju Su Fe Se O M Th Tu W M e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 350 350 400 200 400 300 300 250 150 300 250 300 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 200 200 100 200 150 200 150 100 100 100 50 100 50 50 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. CPCB = Central Pollution Control Board; NSIT = Netaji Subhas Institute of Technology; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. 160 | Getting Down to Earth FIGURE D.9 PM2.5 trends from data acquired at the Ramakrishna Puram, Delhi, India–DPCC monitoring site, 5.69 miles south-southwest of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 500 300 250 800 250 400 200 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 600 200 300 150 400 150 200 100 100 200 100 50 50 0 0 0 0 5/1/2016 11/1/2016 5/1/2017 11/1/2017 DJF MAM JJA SON n on es ed ur i t n b ar M r ay n Au l g p No t Dv ec Fr Sa Ju Ap c Su Ja Ju Fe Se Th Tu O M W M e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 250 500 350 400 120 300 200 400 100 250 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 300 150 80 300 200 150 200 60 100 200 100 40 100 50 100 50 20 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. DPCC = Delhi Pollution Control Committee; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. FIGURE D.10 PM2.5 trends from data acquired at the Anand Vihar, Delhi, India–DPCC monitoring site, 4.6 miles east-northeast of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 400 800 400 600 500 300 600 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 300 400 200 400 200 300 200 100 200 100 100 0 0 0 0 5/1/2016 11/1/2016 5/1/2017 11/1/2017 DJF MAM JJA SON n on es ed ur i t n b ar M r ay n Au l g p No t Dv ec Fr Sa Ju Ap c Su Ja Ju Fe Se Th Tu O M W M e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 600 500 120 400 500 400 300 100 PM2.5 (µg/m3) 400 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 300 80 300 200 300 200 60 200 40 200 100 100 100 20 100 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. DPCC = Delhi Pollution Control Committee; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. Quality Considerations for Ground-Level-Monitoring Data in Low- and Middle-Income Countries | 161 differences in the reported data between stations, but no egregiously disparate values are found that would be indicative of instrument malfunction or data tam- pering across this subset of monitoring locations. Lima, Peru The embassy BAM data from Lima, Peru, reveal nearly an order of magnitude lower PM2.5 concentrations than observed in Delhi, with a far less pronounced seasonal dependence. There is little to no weekday-versus-weekend difference in PM2.5 loadings, and hourly diurnal profiles indicate that PM2.5 pollution source intensity and activity profiles result in highest PM2.5 concentrations in the mid- morning hours (except for June, July, and August in the embassy data, when peak PM2.5 concentrations shift to early afternoon). To provide an empirical basis to assess the consistency of ground-level PM2.5 measurements throughout Lima, the same series of graphs are plotted in figures D.11 through D.15. Each figure corresponds to a different government-run air-quality monitoring location (marked as 2–5 on map D.2). Generally, the observed seasonal, monthly, and weekly trends in PM2.5 levels from the government stations are reasonably con- sistent with the PM2.5 data acquired at the embassy, but important differences are seen in the magnitude of the PM2.5 concentrations reported at different loca- tions. A moderately polluted area such as Lima highlights the need for accurate, reproducible measurements in a manner that is often lost in the much more heavily polluted LMICs (such as Delhi). The embassy data shown in figure D.12 also highlight the importance of having complete data sets over which the PM2.5 trends can be a ­ nalyzed. Missing data from the embassy site complicates the head-to-head intercomparisons somewhat, especially if the surrounding sta- tions had more complete data sets (which is the case for most of the government locations examined here). Nevertheless, inferences can be drawn by completing head-to-head analyses focused on the times of year where data are available. Year-round, the embassy PM2.5 concentrations are about 30 ± 10 micrograms per cubic meter. In comparison, San Borja, Peru (2.7 miles southwest of the embassy) reports concentrations on average about 30 percent lower throughout the year, and Campo de Marte, Peru, reports concentrations about 50 percent lower than the embassy. In contrast, the other two stations, Santa Anita, Peru, and Ate, Peru, report concentrations about 10–30 percent higher than the embassy. In this way, the Lima data sets begin to reveal the complexity of com- bining PM2.5 measurements from different agencies at different locations with minimal metadata. Closer inspection of the prevailing meteorology alongside the pollution source activity profiles and density within the city would further elucidate the extent to which the differences observed across the monitoring stations are derived from atmospheric conditions or instrument performance. The diurnal analyses presented here represent a first step toward building a more comprehensive approach to validating ground-level PM2.5 concentrations in LMICs. 162 | Getting Down to Earth FIGURE D.11 Time-series, seasonal, monthly, weekly, and diurnal PM2.5 mass concentration data measured from the US Diplomatic Post in Lima, Peru a. Time-series b. Seasonal c. Monthly d. Weekly 40 50 60 150 50 40 30 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 40 30 100 20 30 20 20 10 50 10 10 0 0 0 2/9/2017 5/20/2017 DJF MAM JJA SON n b ar M r ay n l g p ct v ec n on es ed s i t Ju Fr Ap Sa ur No Ja Ju Su Au Fe Se O M Tu D W M Th e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 50 80 50 50 80 40 40 40 60 60 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 30 30 30 40 40 20 20 20 20 20 10 10 10 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. FIGURE D.12 PM2.5 trends from data acquired at the San Borja, Peru, monitoring site, 2.7 miles southwest of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 35 40 50 800 30 40 20 25 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 600 30 20 15 400 15 20 10 200 10 10 5 0 0 0 0 5/1/2016 9/1/2016 1/1/2017 5/1/2017 DJF MAM JJA SON n b ar M r ay n l g p No t v ec n on es ed s i t Ju Fr Ap c Sa ur Ja Ju Su Au Fe Se O M Tu D W M Th e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 60 40 20 50 40 50 40 30 15 40 30 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 30 30 20 10 20 20 20 10 5 10 10 10 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. Quality Considerations for Ground-Level-Monitoring Data in Low- and Middle-Income Countries | 163 FIGURE D.13 PM2.5 trends from data acquired at the Campo de Marte, Peru, monitoring site, 5.4 miles northwest of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 400 40 25 30 25 30 20 300 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 20 15 200 15 20 10 10 100 10 5 5 0 0 0 0 5/1/2016 9/1/2016 1/1/2017 5/1/2017 DJF MAM JJA SON n on es ed s i t n b ar M r ay n Au l g p No t v ec Fr Sa ur Ju Su Ap c Ja Ju Fe Se Tu O M W M D Th e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 35 100 30 20 80 30 25 80 15 25 60 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 20 20 60 15 10 40 15 40 10 10 5 20 20 5 5 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. FIGURE D.14 PM2.5 trends from data acquired at the Santa Anita, Peru, monitoring site, 3.9 miles north of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 60 500 60 80 50 50 150 40 40 60 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 30 30 100 40 20 20 50 20 10 10 0 0 0 0 5/1/2016 9/1/2016 1/1/2017 5/1/2017 DJF MAM JJA SON n on es ed s i t n b ar M r ay n l g p No t v ec Fr Ju Sa Ap c ur Su Ja Ju Au Fe Se O M Tu D W M Th e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 70 80 60 80 60 60 50 60 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 60 40 40 40 30 40 40 20 20 20 20 20 10 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. 164 | Getting Down to Earth FIGURE D.15 PM2.5 trends from data acquired at the Ate, Peru, monitoring site, 6.1 miles northeast of US Diplomatic Post a. Time-series b. Seasonal c. Monthly d. Weekly 500 80 80 80 400 60 60 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 60 300 40 40 40 200 20 20 100 20 0 0 0 0 5/1/2016 9/1/2016 1/1/2017 5/1/2017 DJF MAM JJA SON ay n b n Au l g p No t v ec M r ar n on es ed s i t Ju c Ap Fr Sa ur Ja Ju Fe Se Su O M D Tu W M Th e. Diurnal overall f. Diurnal DJF g. Diurnal MAM h. Diurnal JJA i. Diurnal SON 160 100 120 100 140 80 80 120 100 80 100 60 PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) PM2.5 (µg/m3) 80 60 60 80 60 40 40 60 40 40 40 20 20 20 20 20 0 0 0 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Source: World Bank. Note: DJF = December, January, February; MAM = March, April, May; JJA = June, July, August; SON = September, October, November. PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns; μg/m3 = micrograms per cubic meter. MAP D.2 Illustration of the proximity of the US Diplomatic Post in Lima, Peru, relative to a subset of the nearest government-run assurance quality monitoring stations for which OpenAQ PM2.5 data were available Source: World Bank, produced using Google Maps. Note: Government assurance quality stations ranged between about 2.5 to 6 miles from the US Diplomatic Post, covering an area of about 20.8 square miles (blue shaded area). OpenAQ = openaq.org; PM2.5 = particulate matter with an aerodynamic diameter less than or equal to 2.5 microns. Quality Considerations for Ground-Level-Monitoring Data in Low- and Middle-Income Countries | 165 NOTES 1. https://www3.epa.gov/ttnamti1/files/ambient/pm25/sop_project/905505_TEOM_SOP​ _Draft_Final_Sept09.pdf. 2. Bold text represents additional parameters logged for the 1045-DF TEOM; other text rep- resents the parameters logged for both the 1045-F and 1045-DF TEOM. ECO-AUDIT Environmental Benefits Statement The World Bank Group is committed to reducing its environmental footprint. In support of this commitment, we leverage electronic publishing options and print- on-demand technology, which is located in regional hubs worldwide. Together, these initiatives enable print runs to be lowered and shipping distances decreased, resulting in reduced paper consumption, chemical use, greenhouse gas emissions, and waste. We follow the recommended standards for paper use set by the Green Press Initiative. The majority of our books are printed on Forest Stewardship Council (FSC)–certified paper, with nearly all containing 50–100 percent recycled content. The recycled fiber in our book paper is either unbleached or bleached using totally chlorine-free (TCF), processed chlorine–free (PCF), or enhanced elemental chlorine–free (EECF) processes. ­ More information about the Bank’s environmental philosophy can be found at http://www.worldbank.org/corporateresponsibility. O utdoor air pollution accounts for an estimated 4.2 million deaths worldwide, the majority of which are caused by exposure to fine particulate matter (or PM2.5) air pollution. Most of these deaths occur in low- and middle-income countries (LMICs). Reducing PM2.5 air pollution is thus crucial for improving public-health outcomes in those countries. Measuring and reducing the health impacts of PM2.5 is especially challenging in many LMICs because ground-level air-quality monitoring networks are typically nonexistent. Where they are in place, they are often not properly operated and maintained and are thus unreliable. Limited local expertise in air-quality measurement presents additional challenges. These shortcomings can undermine a country’s ability to design and implement effective policies to improve outdoor air quality. Satellite technology has been used successfully for measuring air quality in high-income countries where operation of ground-level air-quality monitoring networks is well established. However, Getting Down to Earth: Are Satellites Reliable for Measuring Air Pollutants That Cause Mortality in Low- and Middle-Income Countries? investigates the performance of satellites in LMICs for predicting outdoor concentrations of PM2.5, based on case studies in nine cities in different regions representing a range of environmental conditions (including mountainous, dusty, and coastal). The report finds that the satellite-derived estimates of PM2.5 in LMICs are associated with very large uncertainty, ranging from 21 percent to 85 percent depending on the model used for translating satellite measurements of aerosol optical depth—the parameter measured by satellites—to surface-level outdoor PM2.5 concentrations. This report shows that satellites are unreliable for estimating ambient concentrations of PM2.5 in LMICs. Furthermore, satellite-derived measurements cannot replace properly operated and maintained ground-level monitoring networks for measuring the concentrations of PM2.5 that human beings are typically exposed to daily. Thus, it is important that LMICs strengthen support for the establishment of ground-level monitoring networks to measure air pollutants, notably PM2.5, that cause mortality in Sub-Saharan Africa and other regions with LMICs. ISBN 978-1-4648-1727-4 SKU 211727