Table 2.

Summary of Objectives, Approaches, and Results for Analyses M1–M5

	M1: Comparison of Mobile and Fixed-Site Monitoring in Oakland	M2: Spatiotemporal Analysis of UFPs
How Analysis Addresses Overall Research Aims	Aim 1: Validate mobile monitoring against fixed-site measurements Aim 2: Compare insights from mobile monitoring with those from fixed measurements Aim 5: Evaluate relative strengths of mobile and fixed-site sampling approaches	Aim 1: Evaluate whether mobile monitoring corroborates an observation from regulatory data Aim 2: Compare insights on UFP dynamics from mobile monitoring and regulatory monitoring Aim 3: Evaluate how mobile monitoring offers complementary information on seasonal patterns of UFP concentrations
Key Measurements	BC Mobile measurements: PAX on Google Street View cars Fixed-site measurements: Custom low-cost ABCD on buildings and utility poles	UFPs, NO, NO2 + supplementary species Mobile measurements: NO, NO2, UFPs + BC on Google Street View cars Regulatory measurements: NO, NO2 UFPs + BC, CO at 4 fixed monitoring stations ,
Period of Measurements	May 19, 2017–Aug. 27, 2017	Mobile measurements: May 2015–Dec. 2017 Regulatory measurements: Full year, 2015
Geographic Coverage of Measurements	West Oakland, California (Figure 1) Mobile measurements: ~ 170 km of road network, 10 km2 Fixed-site measurements: 100 fixed locations within the neighborhood	Mobile measurements: West Oakland and Downtown Oakland (Figure 1) Regulatory measurements at 4 sites: Sebastopol (rural), Livermore (suburban), Redwood City (urban), and Laney College (near road)
Populations Covered	West Oakland, CA (~28,000 people)	Mobile measurements: West Oakland and Downtown Oakland (~50,000 people) Regulatory domain: Entire San Francisco Bay Area (~7 million)
Statistical Analysis and Modeling Approaches	Fixed-site: Time-averaged concentrations at each site Mobile data: Time averages of repeated drive passes within a spatial buffer distance of each fixed site Assessment of the concordance between mobile and fixed-site averages using R2, MAE, and other metrics of agreement	Mobile measurements: Seasonal weekday, daytime spatial patterns determined by computing medians of repeated drive-pass–mean concentrations along 30-m road segments Regulatory measurements: Seasonal diurnal profiles of hourly regulatory data
Key Results	Repeated mobile monitoring can reproduce time-averaged, fine-scale spatial patterns of BC with good fidelity, precision, and accuracy relative to a fixed-site sensor network	Data from mobile monitoring corroborates a surprising insight from regulatory data: Patterns of UFPs and NOx are coupled in the winter months (indicative of a common primary traffic source), but sharply decoupled in the winter. UFPs in the Bay Area appear to be substantially driven by secondary formation during the summer months
	M3: Assessment of local- and regional-scale air pollution disparities in the San Francisco Bay Area using mobile monitoring	M4: Scaling hyperlocal air quality mapping through mobile monitoring and LUR
How Analysis Addresses Overall Research Aims	Aim 2: Compare insights between mobile monitoring and LUR models for assessing population exposure and disparities for NO2 Aim 5: Evaluate utility of mobile monitoring for assessing population exposure distributions and racial and ethnic exposure disparities at large scale	Aim 3: Explore whether and how statistical LUR-K models can make mobile monitoring more scalable by replacing labor-intensive measurements with statistical predictions trained on a more limited set of observations Aim 5: Evaluate the utility of dense “data-only” mobile monitoring approach that covers every city block vis-à-vis an alternative approach where mobile monitoring data are used only for training an LUR-K model
Key Measurements	NO, NO2, BC, UFPs	NO, BC
Period of Measurements	May 2015–Dec. 2017	May 2015–May 2017
Geographic Coverage of Measurements	The 13 communities across the San Francisco Bay Area (93 km2) that are mapped in Figure 1	West Oakland, Downtown Oakland, East Oakland (Figure 1) ~ 490 km of road network, 30 km2
Populations Covered	~ 450,000 people; in this analysis data were explicitly aggregated to census-block geographies to permit assessment of the demographic factors and social disparities associated with air pollution gradients	~ 103,000 people in these three neighborhoods
Statistical Analysis and Modeling Approaches	Aggregation of repeated drive-by on-road measurements to estimate median long-term weekday, daytime median concentrations for surrounding U.S. Census blocks Computation of cumulative population- weighted exposure distributions for full population and by race and ethnicity Partitioning of total spatial variation in population exposure into within- and between-neighborhood components Assessment of relative racial and ethnic disparities at different moments of the cumulative exposure distribution	Computation of long-term weekday, daytime road segment median concentrations for repeated drive-pass mean concentrations at the 30-m road segment scale Development of LUR-K models to evaluate ability to make out-of-sample spatial predictions at unmonitored locations Monte Carlo simulations of spatial and temporal coverage in mobile mapping to assess the trade-off between the amount of data collected and fidelity of LUR-K models
Key Results	Repeated mobile monitoring can represent exposure heterogeneity across a large urban region Across the entire Bay Area region, with- in-neighborhood gradients account for a large (~30% for UFPs and NO2) to dominant (>50% for BC and NO) fraction of the overall heterogeneity in the population-concentration distribution Mobile monitoring captures a much wider range of variation in the NO2 exposure distribution than does a common nationwide NO2 LUR model Substantial racial and ethnic disparities are driven mostly by intra-neighborhood segregation	The best-performing LUR-K models we developed are limited in their ability to capture full spatial heterogeneity we measured with data-only maps (max R2 ~ 0.65) An advantage of LUR-K modeling is that there is very little penalty in model performance that arises from using a simulated mobile monitoring campaign with 10–50 times less data. It is possible to drive only a fraction of roads a few times and develop models that are nearly as good as the best models we trained. Data-only maps from repeated driving are superior to LUR-K models in terms of detecting idiosyncratic or unexpected spatial features and hotspots
	M5: Mobile monitoring in Bangalore, India
How Analysis Addresses Overall Research Aims	Aim 4: Test the mobile monitoring approach in an Indian city Aim 5: Evaluate the utility of mobile monitoring in the Indian context
Key Measurements	BC, UFPs, CO2
Period of Measurements	July 2019–March 2020
Geographic Coverage of Measurements	Residential neighborhood in Bangalore (Malleshwaram) and supplemental transects in the Central Business District and between urban core and rural periphery
Populations Covered	~ 100,000 people live in the middle-income neighborhood of Malleshwaram
Statistical Analysis and Modeling Approaches	Computation of long-term weekday, daytime road segment median concentrations for repeated drive-pass mean concentrations at the 30-m road segment scale Monte Carlo simulations to assess the trade-offs between the number of repeated mobile monitoring visits and the fidelity of the resulting spatial concentration maps
Key Results	Mobile monitoring resolves time-stable spatial patterns with high fidelity in Malleshwaram and elsewhere in our domain Localized pollution gradients are sharp and reach very high concentrations in the near-road environment Observed a convergence to time-stable spatial patterns with fewer than 20 repeated mobile sampling runs over 1 year Some questions about the degree to which on-road concentrations are representative of population exposures away from roadways, especially given the persistent traffic congestion in parts of the Bangalore road network Slow traffic speeds in Bangalore present logistical challenges for mobile monitoring

ABCD = aerosol black carbon detector; BC = black carbon; LUR = land use regression; MAE = mean absolute error; PAX = photoacoustic extinctiometer; UFPs = ultrafine particles.