Estimating daily PM2.5 concentrations in New York City at the neighborhood-scale: Implications for integrating non-regulatory measurements

Respiratory and Cardiovascular Medical Emergency Calls Related to Indoor Heat Exposure through a Case-Control Study in New York City

Understanding when and where heat adversely influences health outcomes is critical for targeting interventions and adaptations. However, few studies have analyzed the role of indoor heat exposures on acute health outcomes. To address this research gap, the study partnered with the New York City Fire Department Emergency Medical Services. Paramedics carried portable sensors that passively measured indoor temperatures at 3-min intervals while responding to calls during summer, 2016. Patient care reports provided the patient's chief health complaint and sociodemographic and health status information. Propensity score matching increased comparability between groups exposed to elevated indoor temperature versus those unexposed. To assess indoor heat-health associations, we conducted independent case-control studies between indoor heat exposures and cardiovascular (n = 735) and respiratory (n = 296) emergency medical calls when compared to heat-insensitive controls (n = 1611). Patients experiencing heat exposures (indoor temperature ≥ 28 °C) were not significantly more likely (OR, 1.15; 95% CI, 0.64-2.09) to receive care for respiratory conditions. Both outdoor and indoor temperatures increased the odds of receiving care for cardiovascular versus comparison calls. Outdoor temperatures consistently elevated cardiovascular risks (OR, 1.12; 95% CI, 1.05-1.19). There was some evidence that indoor temperatures further increased the odds of cardiovascular distress (OR, 1.44; 95% CI, 0.97-2.13). Sensitivity testing suggested indoor temperatures at a lower threshold (≥ 26 °C) were unrelated to either health outcome. Along with converging lines of evidence linking extreme heat to adverse cardiovascular outcomes, we present one of the first indoor heat observational studies.

Unmasking the sky: high-resolution PM2.5 prediction in Texas using machine learning techniques

BACKGROUND

Although PM2.5 (fine particulate matter with an aerodynamic diameter less than 2.5 µm) is an air pollutant of great concern in Texas, limited regulatory monitors pose a significant challenge for decision-making and environmental studies.

OBJECTIVE

This study aimed to predict PM2.5 concentrations at a fine spatial scale on a daily basis by using novel machine learning approaches and incorporating satellite-derived Aerosol Optical Depth (AOD) and a variety of weather and land use variables.

METHODS

We compiled a comprehensive dataset in Texas from 2013 to 2017, including ground-level PM2.5 concentrations from regulatory monitors; AOD values at 1-km resolution based on images retrieved from the MODIS satellite; and weather, land-use, population density, among others. We built predictive models for each year separately to estimate PM2.5 concentrations using two machine learning approaches called gradient boosted trees and random forest. We evaluated the model prediction performance using in-sample and out-of-sample validations.

RESULTS

Our predictive models demonstrate excellent in-sample model performance, as indicated by high R2 values generated from the gradient boosting models (0.94-0.97) and random forest models (0.81-0.90). However, the out-of-sample R2 values fall within a range of 0.52-0.75 for gradient boosting models and 0.44-0.69 for random forest models. Model performance varies slightly across years. A generally decreasing trend in predicted PM2.5 concentrations over time is observed in Eastern Texas.

IMPACT STATEMENT

We utilized machine learning approaches to predict PM2.5 levels in Texas. Both gradient boosting and random forest models perform well. Gradient boosting models perform slightly better than random forest models. Our models showed excellent in-sample prediction performance (R2 > 0.9).

Exposure to particulate matter and ozone, locations of regulatory monitors, and sociodemographic disparities in the city of Rio de Janeiro: Based on local air pollution estimates generated from machine learning models☆

South America is underrepresented in research on air pollution exposure disparities by sociodemographic factors, although such disparities have been observed in other parts of the world. We investigated whether exposure to and information about air pollution differs by sociodemographic factors in the city of Rio de Janeiro, the second most populous city in Brazil with dense urban areas, for 2012-2017. We developed machine learning-based models to estimate daily levels of O3, PM10, and PM2.5 using high-dimensional datasets from satellite remote sensing, atmospheric and land variables, and land use information. Cross-validations demonstrated good agreement between the estimated levels and measurements from ground-based monitoring stations: overall R 2 of 76.8 %, 63.9 %, and 69.1 % for O3, PM2.5, and PM10, respectively. We conducted univariate regression analyses to investigate whether long-term exposure to O3, PM2.5, PM10 and distance to regulatory monitors differs by socioeconomic indicators, the percentages of residents who were children (0-17 years) or age 65+ years in 154 neighborhoods. We also examined the number of days exceeding the Brazilian National Air Quality Standard (BNAQS). Long-term exposures to O3 and PM2.5 were higher in more socially deprived neighborhoods. An interquartile range (IQR) increment of the social development index (SDI) was associated with a 3.6 μg/m3 (95 % confidence interval [CI]: 2.9, 4.4; p-value≤0.001) decrease in O3, and 0.3 μg/m3 (95 % CI: 0.2, 0.5; p-value = 0.010) decrease in PM2.5. An IQR increase in the percentage of residents who are children was associated with a 4.1 μg/m3 (95 % CI: 3.1, 5.0; p-value≤0.001) increase in O3, and 0.4 μg/m3 (95 % CI: 0.3, 0.6; p-value = 0.009) increase in PM2.5. An IQR increase in the percentage of residents age ≥65was associated with a 3.3 μg/m3 (95 % CI: 2.4, 4.3; p-value=<0.001) decrease in O3, and 0.3 μg/m3 (95 % CI: 0.1, 0.5; p-value = 0.058) decrease in PM2.5. There were no apparent associations for PM10. The association for daily O3 levels exceeding the BNAQS daily standard was 0.4 %p-0.8 %p different by the IQR of variables, indicating a 7-15 days difference in the six-year period. The association for daily PM2.5 levels exceeding the BNAQS daily standard showed a 0.7-1.5 %p difference by the IQR, meaning a 13-27 days difference in the period. We did not find statistically significant associations between the distance to monitors and neighborhood characteristics but some indication regarding SDI. We found that O3 levels were higher in neighborhoods situated farther from monitoring stations, suggesting that elevated levels of air pollution may not be routinely measured. Exposure disparity patterns may vary by pollutants, suggesting a complex interplay between environmental and socioeconomic factors in environmental justice.

Evaluating low-cost monitoring designs for PM2.5 exposure assessment with a spatiotemporal modeling approach

Determining the most feasible and cost-effective approaches to improving PM2.5 exposure assessment with low-cost monitors (LCMs) can considerably enhance the quality of its epidemiological inferences. We investigated features of fixed-site LCM designs that most impact PM2.5 exposure estimates to be used in long-term epidemiological inference for the Adult Changes in Thought Air Pollution (ACT-AP) study. We used ACT-AP collected and calibrated LCM PM2.5 measurements at the two-week level from April 2017 to September 2020 (N of monitors [measurements] = 82 [502]). We also acquired reference-grade PM2.5 measurements from January 2010 to September 2020 (N = 78 [6186]). We used a spatiotemporal modeling approach to predict PM2.5 exposures with either all LCM measurements or varying subsets with reduced temporal or spatial coverage. We evaluated the models based on a combination of cross-validation and external validation at locations of LCMs included in the models (N = 82), and also based on an independent external validation with a set of LCMs not used for the modeling (N = 30). We found that the model's performance declined substantially when LCM measurements were entirely excluded (spatiotemporal validation R2 [RMSE] = 0.69 [1.2 μg/m3]) compared to the model with all LCM measurements (0.84 [0.9 μg/m3]). Temporally, using the farthest apart measurements (i.e., the first and last) from each LCM resulted in the closest model's performance (0.79 [1.0 μg/m3]) to the model with all LCM data. The models with only the first or last measurement had decreased performance (0.77 [1.1 μg/m3]). Spatially, the model's performance decreased linearly to 0.74 (1.1 μg/m3) when only 10% of LCMs were included. Our analysis also showed that LCMs located in densely populated, road-proximate areas improved the model more than those placed in moderately populated, road-distant areas.

Investigating Use of Low-Cost Sensors to Increase Accuracy and Equity of Real-Time Air Quality Information

U.S. Environmental Protection Agency (EPA) air quality (AQ) monitors, the "gold standard" for measuring air pollutants, are sparsely positioned across the U.S. Low-cost sensors (LCS) are increasingly being used by the public to fill in the gaps in AQ monitoring; however, LCS are not as accurate as EPA monitors. In this work, we investigate factors impacting the differences between an individual's true (unobserved) exposure to air pollution and the exposure reported by their nearest AQ instrument (which could be either an LCS or an EPA monitor). We use simulations based on California data to explore different combinations of hypothetical LCS placement strategies (e.g., at schools or near major roads), for different numbers of LCS, with varying plausible amounts of LCS device measurement errors. We illustrate how real-time AQ reporting could be improved (or, in some cases, worsened) by using LCS, both for the population overall and for marginalized communities specifically. This work has implications for the integration of LCS into real-time AQ reporting platforms.

Neighborhood Emission Mapping Operation (NEMO): A 1-km anthropogenic emission dataset in the United States

We present an unprecedented effort to map anthropogenic emissions of air pollutants at 1 km spatial resolution in the contiguous United States (CONUS). This new dataset, Neighborhood Emission Mapping Operation (NEMO), is produced at hourly intervals based on the United States Environmental Protection Agency (US EPA) National Emission Inventories 2017. Fine-scale spatial allocation was achieved through distributing the emission sources using 108 spatial surrogates, factors representing the portion of a source in each 1 km grid. Gaseous and particulate pollutants are speciated into model species for the Carbon Bond 6 chemical mechanism. All sources are grouped in 9 sectors and stored in NetCDF format for air quality models, and in shapefile format for GIS users and air quality managers. This dataset shows good consistency with the USEPA benchmark dataset, with a monthly difference in emissions less than 0.03% for any sector. NEMO provides the first 1 km mapping of air pollution over the CONUS, enabling new applications such as fine-scale air quality modeling, air pollution exposure assessment, and environmental justice studies.

Communities of color are disproportionately exposed to long-term and short-term PM2.5 in metropolitan America

We conducted a novel investigation of neighborhood-level racial/ethnic exposure disparities employing measures aligned with long-term and short-term PM_2.5 air pollution benchmarks across metropolitan contexts of the contiguous United States, 2012-2016. We used multivariable generalized estimating equations (GEE) to quantify PM_2.5 exposure disparities based on the census tract composition of people of color (POC) and POC groups (Hispanic/Latina/x/o, Black, Asian). We examined eight census tract-level measures of longer-to-shorter term exposures derived from data on modeled daily ambient PM_2.5 concentrations. We found associations between increased POC composition and greater exposure to all PM_2.5 measures, with associations strengthening across measures of longer-to-shorter term exposures. In a GEE with a negative binomial distribution, a standard deviation increase in POC composition predicted a 0.6% increase (incidence rate ratio (IRR): 1.006, 95% confidence interval (CI): 1.005-1.008) in the number of days PM_2.5 concentrations were ≥5 μg/m³ (longest-term benchmark). In a GEE with an inverse Gaussian distribution, a standard deviation increase in POC composition predicted a 0.110 μg/m³ (1.0%) increase (B: 0.110, 95% CI: 0.076-0.143) in mean PM_2.5 concentration. In GEEs with a negative binomial distribution, the effect of a standard deviation increase in POC composition on exposure strengthened to 2.6% (IRR:1.026, 95% CI:1.017-1.035), 3.4% (IRR:1.034, 95% CI:1.022-1.047), 4.2% (IRR:1.042, 95% CI:1.025-1.058), 16.2% (IRR:1.162, 95% CI:1.117-1.210), 22.7% (IRR:1.227, 95% CI:1.137-1.325) and 28.3% (IRR:1.283, 95% CI:1.144-1.439) with respect to the number of days PM_2.5 concentrations were ≥10, 12, 15, 25, 35 and 55.5 μg/m³. POC group models indicated exposure disparities based on greater Hispanic/Latina/x/o, Asian, and Black composition. Evidence for stronger POC associations with shorter-term (higher concentration) PM_2.5 exceedances suggests that reducing PM_2.5 would attenuate racial/ethnic exposure disparities.

ZIP Code-Level Estimation of Air Quality and Health Risk Due to Particulate Matter Pollution in New York City

Exposure to PM_2.5 is associated with hundreds of premature mortalities every year in New York City (NYC). Current air quality and health impact assessment tools provide county-wide estimates but are inadequate for assessing health benefits at neighborhood scales, especially for evaluating policy options related to energy efficiency or climate goals. We developed a new ZIP Code-Level Air Pollution Policy Assessment (ZAPPA) tool for NYC by integrating two reduced form models─Community Air Quality Tools (C-TOOLS) and the Co-Benefits Risk Assessment Health Impacts Screening and Mapping Tool (COBRA)─that propagate emissions changes to estimate air pollution exposures and health benefits. ZAPPA leverages custom higher resolution inputs for emissions, health incidences, and population. It, then, enables rapid policy evaluation with localized ZIP code tabulation area (ZCTA)-level analysis of potential health and monetary benefits stemming from air quality management decisions. We evaluated the modeled 2016 PM_2.5 values against observed values at EPA and NYCCAS monitors, finding good model performance (FAC2, 1; NMSE, 0.05). We, then, applied ZAPPA to assess PM_2.5 reduction-related health benefits from five illustrative policy scenarios in NYC focused on (1) commercial cooking, (2) residential and commercial building fuel regulations, (3) fleet electrification, (4) congestion pricing in Manhattan, and (5) these four combined as a "citywide sustainable policy implementation" scenario. The citywide scenario estimates an average reduction in PM_2.5 of 0.9 μg/m³. This change translates to avoiding 210-475 deaths, 340 asthma emergency department visits, and monetized health benefits worth $2B to $5B annually, with significant variation across NYC's 192 ZCTAs. ZCTA-level assessments can help prioritize interventions in neighborhoods that would see the most health benefits from air pollution reduction. ZAPPA can provide quantitative insights on health and monetary benefits for future sustainability policy development in NYC.

Selecting Data Analytic and Modeling Methods to Support Air Pollution and Environmental Justice Investigations: A Critical Review and Guidance Framework

Given the serious adverse health effects associated with many pollutants, and the inequitable distribution of these effects between socioeconomic groups, air pollution is often a focus of environmental justice (EJ) research. However, EJ analyses that aim to illuminate whether and how air pollution hazards are inequitably distributed may present a unique set of requirements for estimating pollutant concentrations compared to other air quality applications. Here, we perform a scoping review of the range of data analytic and modeling methods applied in past studies of air pollution and environmental injustice and develop a guidance framework for selecting between them given the purpose of analysis, users, and resources available. We include proxy, monitor-based, statistical, and process-based methods. Upon critically synthesizing the literature, we identify four main dimensions to inform method selection: accuracy, interpretability, spatiotemporal features of the method, and usability of the method. We illustrate the guidance framework with case studies from the literature. Future research in this area includes an exploration of increasing data availability, advanced statistical methods, and the importance of science-based policy.

Combining Machine Learning and Numerical Simulation for High-Resolution PM2.5 Concentration Forecast

Forecasting ambient PM_2.5 concentrations with spatiotemporal coverage is key to alerting decision makers of pollution episodes and preventing detrimental public exposure, especially in regions with limited ground air monitoring stations. The existing methods rely on either chemical transport models (CTMs) to forecast spatial distribution of PM_2.5 with nontrivial uncertainty or statistical algorithms to forecast PM_2.5 concentration time series at air monitoring locations without continuous spatial coverage. In this study, we developed a PM_2.5 forecast framework by combining the robust Random Forest algorithm with a publicly accessible global CTM forecast product, NASA's Goddard Earth Observing System "Composition Forecasting" (GEOS-CF), providing spatiotemporally continuous PM_2.5 concentration forecasts for the next 5 days at a 1 km spatial resolution. Our forecast experiment was conducted for a region in Central China including the populous and polluted Fenwei Plain. The forecast for the next 2 days had an overall validation R² of 0.76 and 0.64, respectively; the R² was around 0.5 for the following 3 forecast days. Spatial cross-validation showed similar validation metrics. Our forecast model, with a validation normalized mean bias close to 0, substantially reduced the large biases in GEOS-CF. The proposed framework requires minimal computational resources compared to running CTMs at urban scales, enabling near-real-time PM_2.5 forecast in resource-restricted environments.

Evaluating the Utility of High-Resolution Spatiotemporal Air Pollution Data in Estimating Local PM2.5 Exposures in California from 2015–2018

Air quality management is increasingly focused not only on across-the-board reductions in ambient pollution concentrations but also on identifying and remediating elevated exposures that often occur in traditionally disadvantaged communities. Remote sensing of ambient air pollution using data derived from satellites has the potential to better inform management decisions that address environmental disparities by providing increased spatial coverage, at high-spatial resolutions, compared to air pollution exposure estimates based on ground-based monitors alone. Daily PM2.5 estimates for 2015–2018 were estimated at a 1 km2 resolution, derived from NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS) satellite instrument and the Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm in order to assess the utility of highly refined spatiotemporal air pollution data in 92 California cities and in the 13 communities included in the California Community Air Protection Program. The identification of pollution hot-spots within a city is typically not possible relying solely on the regulatory monitoring networks; however, day-to-day temporal variability was shown to be generally well represented by nearby ground-based monitoring data even in communities with strong spatial gradients in pollutant concentrations. An assessment of within-ZIP Code variability in pollution estimates indicates that high-resolution pollution estimates (i.e., 1 km2) are not always needed to identify spatial differences in exposure but become increasingly important for larger geographic areas (approximately 50 km2). Taken together, these findings can help inform strategies for use of remote sensing data for air quality management including the screening of locations with air pollution exposures that are not well represented by existing ground-based air pollution monitors.

Publicly available low-cost sensor measurements for PM2.5 exposure modeling: Guidance for monitor deployment and data selection

High-resolution, high-quality exposure modeling is critical for assessing the health effects of ambient PM_2.5 in epidemiological studies. Using sparse regulatory PM_2.5 measurements as principal model inputs may result in two issues in exposure prediction: (1) they may affect the models' accuracy in predicting PM_2.5 spatial distribution; (2) the internal validation based on these measurements may not reliably reflect the model performance at locations of interest (e.g., a cohort's residential locations). In this study, we used the PM_2.5 measurements from a publicly available commercial low-cost PM_2.5 network, PurpleAir, with an external validation dataset at the residential locations of a representative sample of participants from the Adult Changes in Thought - Air Pollution (ACT-AP) study, to improve the accuracy of exposure prediction at the cohort participant locations. We also proposed a metric based on principal component analysis (PCA) - the PCA distance - to assess the similarity between monitor and cohort locations to guide monitor deployment and data selection. The analysis was based on a spatiotemporal modeling framework with 51 "gold-standard" monitors and 58 PurpleAir monitors for model development, as well as 105 home monitors at the cohort locations for model validation, in the Puget Sound region of Washington State from June 2017 to March 2019. After including calibrated PurpleAir measurements as part of the dependent variable, the external spatiotemporal validation R² and root-mean-square error, RMSE, for two-week concentration averages improved from 0.84 and 2.22 μg/m³ to 0.92 and 1.63 μg/m³, respectively. The external spatial validation R² and RMSE for long-term averages over the modeling period improved from 0.72 and 1.01 μg/m³ to 0.79 and 0.88 μg/m³, respectively. The exposure predictions incorporating PurpleAir measurements demonstrated sharper urban-suburban concentration gradients. The PurpleAir monitors with shorter PCA distances improved the model's prediction accuracy more substantially than the monitors with longer PCA distances, supporting the use of this similarity metric.

Integration of complete ensemble empirical mode decomposition with deep long short-term memory model for particulate matter concentration prediction

The atmospheric particulate matter (PM) with a diameter of 2.5 μm or less (PM2.5) is one of the key indicators of air pollutants. Accurate prediction of PM2.5 concentration is very important for air pollution monitoring and public health management. However, the presence of noise in PM2.5 data series is a major challenge of its accurate prediction. A novel hybrid PM2.5 concentration prediction model is proposed in this study by combining complete ensemble empirical mode decomposition (CEEMD) method, Pearson's correlation analysis, and a deep long short-term memory (LSTM) method. CEEMD was employed to decompose historical PM2.5 concentration data to different frequencies in order to enhance the timing characteristics of data. Pearson's correlation was used to screen the different frequency intrinsic-mode functions of decomposed data. Finally, the filtered enhancement data were inputted to a deep LSTM network with multiple hidden layers for training and prediction. The results evidenced the potential of the CEEMD-LSTM hybrid model with a prediction accuracy of approximately 80% and model convergence after 700 training epochs. The secondary screening of Pearson's correlation test improved the model (CEEMD-Pearson) accuracy up to 87% but model convergence after 800 epochs. The hybrid model combining CEEMD-Pearson with the deep LSTM neural network showed a prediction accuracy of nearly 90% and model convergence after 650 interactions. The results provide a clear indication of higher prediction accuracy of PM2.5 with less computation time through hybridization of CEEMD-Pearson with deep LSTM models and its potential to be employed for air pollution monitoring.

Excess Morbidity and Mortality Associated with Air Pollution above American Thoracic Society Recommended Standards, 2017-2019

Rationale: Over the past year, the American Thoracic Society (ATS), led by its Environmental Health Policy Committee, has reviewed the most current air quality scientific evidence and has revised their recommendations to 8 µg/m³ and 25 µg/m³ for long- and short-term fine particulate matter (PM_2.5) and reaffirmed the recommendation of 60 ppb for ozone to protect the American public from the known adverse health effects of air pollution. The current EPA standards, in contrast, expose the American public to pollution levels that are known to result in significant morbidity and mortality. Objectives: To provide county-level estimates of annual air pollution-related health outcomes across the United States using the most recent federal air quality data, and to support the ATS's recent update to the long-term PM_2.5 recommended standard. This study is presented as part of the annual ATS/Marron Institute "Health of the Air" report. Methods: Daily air pollution values were obtained from the U.S. Environmental Protection Agency's (EPA) Air Quality System for monitored counties in the United States from 2017-2019. Concentration-response functions used in the EPA's regulatory review process were applied to pollution increments corresponding to differences between the rolling 3-year design values and ATS-recommended levels for long-term PM_2.5 (8 µg/m³), short-term PM_2.5 (25 µg/m³), and ground-level ozone (O₃; 60 ppb). Health impacts were estimated at the county level in locations with valid monitoring data. Results: Meeting ATS recommendations throughout the country prevents an estimated 14,650 (95% CI: 8,660 - 22,610) deaths; 2,950 (95% CI: 1,530 - 4,330) lung cancer incidence events; 33,100 (95% CI: 7,300 - 71,000) morbidities, and 39.8 million (95% CI: 14.6 - 63.3 million) impacted days annually (see Table 1). This prevents 11,850 more deaths; 2,580 more lung cancer incidence events; 25,400 more morbidities; and 27.2 million more impacted days than meeting EPA standards alone. Conclusions: Significant health benefits to be gained by U.S. communities that work to meet ATS-recommended air quality standards have now been identified under scenarios meeting the new ATS recommendation for long-term PM_2.5 (8 µg/m³). The "Health of the Air" report presents an opportunity for air quality managers to quantify local health burdens and EPA officials to update their standards to reflect the latest science.

Beyond air pollution at home: Assessment of personal exposure to PM2.5 using activity-based travel demand model and low-cost air sensor network data

Assessing personal exposure to air pollution is challenging due to the limited availability of human movement data and the complexity of modeling air pollution at high spatiotemporal resolution. Most health studies rely on residential estimates of outdoor air pollution instead which introduces exposure measurement error. Personal exposure for 100,784 individuals in Los Angeles County was estimated by integrating human movement data simulated from the Southern California Association of Governments (SCAG) activity-based travel demand model with hourly PM_2.5 predictions from my 500 m gridded model incorporating low-cost sensor monitoring data. Individual exposures were assigned considering PM_2.5 levels at homes, workplaces, and other activity locations. These dynamic exposures were compared to the residence-based exposures, which do not consider human movement, to examine the degree of exposure estimation bias. The results suggest that exposures were underestimated by 13% (range 5-22%) on average when human movement was not considered, and much of the error was eliminated by accounting for work location. Exposure estimation bias increased for people who exhibited higher mobility levels, especially for workers with long commute distances. Overall, the personal exposures of workers were underestimated by 22% (5-61%) relative to their residence-based exposures. For workers who commute >20 miles, their exposure levels can be at most underestimated by 61%. Omitting mobility resulted in underestimating exposures for people who reside in areas with cleaner air but work in more polluted areas. Similarly, exposures were overestimated for people living in areas with poorer air quality and working in cleaner areas. These could lead to differential estimation biases across racial, ethnic and socioeconomic lines that typically correlate with where people live and work and lead to important exposure and health disparities. This study demonstrates that ignoring human movement and spatiotemporal variability of air pollution could lead to differential exposure misclassification potentially biasing health risk assessments. These improved dynamic approaches can help planners and policymakers identify disadvantaged populations for which exposures are typically misrepresented and might lead to targeted policy and planning implications.

Examining PM2.5 concentrations and exposure using multiple models

Epidemiologic studies have found associations between fine particulate matter (PM_2.5) exposure and adverse health effects using exposure models that incorporate monitoring data and other relevant information. Here, we use nine PM_2.5 concentration models (i.e., exposure models) that span a wide range of methods to investigate i) PM_2.5 concentrations in 2011, ii) potential changes in PM_2.5 concentrations between 2011 and 2028 due to on-the-books regulations, and iii) PM_2.5 exposure for the U.S. population and four racial/ethnic groups. The exposure models included two geophysical chemical transport models (CTMs), two interpolation methods, a satellite-derived aerosol optical depth-based method, a Bayesian statistical regression model, and three data-rich machine learning methods. We focused on annual predictions that were regridded to 12-km resolution over the conterminous U.S., but also considered 1-km predictions in sensitivity analyses. The exposure models predicted broadly consistent PM_2.5 concentrations, with relatively high concentrations on average over the eastern U.S. and greater variability in the western U.S. However, differences in national concentration distributions (median standard deviation: 1.00 μg m^-3) and spatial distributions over urban areas were evident. Further exploration of these differences and their implications for specific applications would be valuable. PM_2.5 concentrations were estimated to decrease by about 1 μg m^-3 on average due to modeled emission changes between 2011 and 2028, with decreases of more than 3 μg m^-3 in areas with relatively high 2011 concentrations that were projected to experience relatively large emission reductions. Agreement among models was closer for population-weighted than uniformly weighted averages across the domain. About 50% of the population was estimated to experience PM_2.5 concentrations less than 10 μg m^-3 in 2011 and PM_2.5 improvements of about 2 μg m^-3 due to modeled emission changes between 2011 and 2028. Two inequality metrics were used to characterize differences in exposure among the four racial/ethnic groups. The metrics generally yielded consistent information and suggest that the modeled emission reductions between 2011 and 2028 would reduce absolute exposure inequality on average.

Characterizing outdoor infiltration and indoor contribution of PM2.5 with citizen-based low-cost monitoring data

Epidemiological research on the adverse health outcomes due to PM_2.5 exposure frequently relies on measurements from regulatory air quality monitors to provide ambient exposure estimates, whereas personal PM_2.5 exposure may deviate from ambient concentrations due to outdoor infiltration and contributions from indoor sources. Research in quantifying infiltration factors (F_inf), the fraction of outdoor PM_2.5 that infiltrates indoors, has been historically limited in space and time due to the high costs of monitor deployment and maintenance. Recently, the growth of openly accessible, citizen-based PM_2.5 measurements provides an unprecedented opportunity to characterize F_inf at large spatiotemporal scales. In this analysis, 91 consumer-grade PurpleAir indoor/outdoor monitor pairs were identified in California (41 residential houses and 50 public/commercial buildings) during a 20-month period with around 650000 h of paired PM_2.5 measurements. An empirical method was developed based on local polynomial regression to estimate site-specific F_inf. The estimated site-specific F_inf had a mean of 0.26 (25^th, 75^th percentiles: [0.15, 0.34]) with a mean bootstrap standard deviation of 0.04. The F_inf estimates were toward the lower end of those reported previously. A threshold of ambient PM_2.5 concentration, approximately 30 μg/m³, below which indoor sources contributed substantially to personal exposures, was also identified. The quantified relationship between indoor source contributions and ambient PM_2.5 concentrations could serve as a metric of exposure errors when using outdoor monitors as an exposure proxy (without considering indoor-generated PM_2.5), which may be of interest to epidemiological research. The proposed method can be generalized to larger geographical areas to better quantify PM_2.5 outdoor infiltration and personal exposure.

Estimating hourly PM2.5 concentrations at the neighborhood scale using a low-cost air sensor network: A Los Angeles case study

Predicting PM_2.5 concentrations at a fine spatial and temporal resolution (i.e., neighborhood, hourly) is challenging. Recent growth in low cost sensor networks is providing increased spatial coverage of air quality data that can be used to supplement data provided by monitors of regulatory agencies. We developed an hourly, 500 × 500 m gridded PM_2.5 model that integrates PurpleAir low-cost air sensor network data for Los Angeles County. We developed a quality control scheme for PurpleAir data. We included spatially and temporally varying predictors in a random forest model with random oversampling of high concentrations to predict PM_2.5. The model achieved high prediction accuracy (10-fold cross-validation (CV) R² = 0.93, root mean squared error (RMSE) = 3.23 μg/m³; spatial CV R² = 0.88, spatial RMSE = 4.33 μg/m³; temporal CV R² = 0.90, temporal RMSE = 3.85 μg/m³). Our model was able to predict spatial and diurnal patterns in PM_2.5 on typical weekdays and weekends, as well as non-typical days, such as holidays and wildfire days. The model allows for far more precise estimates of PM_2.5 than existing methods based on few sensors. Taking advantage of low-cost PM_2.5 sensors, our hourly random forest model predictions can be combined with time-activity diaries in future studies, enabling geographically and temporally fine exposure estimation for specific population groups in studies of acute air pollution health effects and studies of environmental justice issues.

Estimating the Impact of COVID-19 on the PM2.5 Levels in China with a Satellite-Driven Machine Learning Model

China implemented an aggressive nationwide lockdown procedure immediately after the COVID-19 outbreak in January 2020. As China emerges from the impact of COVID-19 on national economic and industrial activities, it has become the site of a large-scale natural experiment to evaluate the impact of COVID-19 on regional air quality. However, ground measurements of fine particulate matters (PM2.5) concentrations do not offer comprehensive spatial coverage, especially in suburban and rural regions. In this study, we developed a machine learning method with satellite aerosol remote sensing data, meteorological fields and land use parameters as major predictor variables to estimate spatiotemporally resolved daily PM2.5 concentrations in China. Our study period consists of a reference semester (1 November 2018-30 April 2019) and a pandemic semester (1 November 2019-30 April 2020), with six modeling months in each semester. Each period was then divided into subperiod 1 (November and December), subperiod 2 (January and February) and subperiod 3 (March and April). The reference semester model obtained a 10-fold cross-validated R2 (RMSE) of 0.79 (17.55 μg/m3) and the pandemic semester model obtained a 10-fold cross-validated R2 (RMSE) of 0.83 (13.48 μg/m3) for daily PM2.5 predictions. Our prediction results showed high PM2.5 concentrations in the North China Plain, Yangtze River Delta, Sichuan Basin and Xinjiang Autonomous Region during the reference semester. PM2.5 levels were lowered by 4.8 μg/m3 during the pandemic semester compared to the reference semester and PM2.5 levels during subperiod 2 decreased most, by 18%. The southeast region was affected most by the COVID-19 outbreak with PM2.5 levels during subperiod 2 decreasing by 31%, followed by the Northern Yangtze River Delta (29%) and Pearl River Delta (24%).

A hybrid approach to predict daily NO2 concentrations at city block scale

Estimating the ambient concentration of nitrogen dioxide (NO2) is challenging because NO2 generated by local fossil fuel combustion varies greatly in concentration across space and time. This study demonstrates an integrated hybrid approach combining dispersion modeling and land use regression (LUR) to predict daily NO2 concentrations at a high spatial resolution (e.g., 50 m) in the New York tri-state area. The daily concentration of traffic-related NO2 was estimated at the Environmental Protection Agency's NO2 monitoring sites in the study area for the years 2015-2017, using the Research LINE source (R-LINE) model with inputs of traffic data provided by the Highway Performance and Management System and meteorological data provided by the NOAA Integrated Surface Database. We used the R-LINE-predicted daily concentrations of NO2 to build mixed-effects regression models, including additional variables representing land use features, geographic characteristics, weather, and other predictors. The mixed model was selected by the Elastic Net method. Each model's performance was evaluated using the out-of-sample coefficient of determination (R2) and the square root of mean squared error (RMSE) from ten-fold cross-validation (CV). The mixed model showed a good prediction performance (CV R2: 0.75-0.79, RMSE: 3.9-4.0 ppb). R-LINE outputs improved the overall, spatial, and temporal CV R2 by 10.0%, 18.9% and 7.7% respectively. Given the output of R-LINE is point-based and has a flexible spatial resolution, this hybrid approach allows prediction of daily NO2 at an extremely high spatial resolution such as city blocks.

Using Kriging incorporated with wind direction to investigate ground-level PM2.5 concentration

Conventional interpolation methods, such as spatial averaging, nearest neighbor, inverse distance weight and ordinary Kriging (OK); for estimating the spatial distribution of ground-level particulate matter (PM) data, do not account for the wind direction for estimating the spatial distribution of PM_2.5. In this work, an interpolation algorithm, Win-OK accounting for the wind direction, is developed. In contrast to ordinary Kriging where all locations (irrespective of the wind direction) in the vicinity of a site is considered, the new algorithm (Win-OK) predicts the value at a certain location based on the measured values at locations upwind as determined by the wind direction. This new methodology, Win-OK is validated by applying it to analyze the hourly spatial distribution of ground-level PM_2.5 concentrations during Chinese New Year and Chinese National Day in 2017 in Xinxiang city, China. The performance of OK and Win-OK are compared by using them to build PM_2.5 concentration heat-maps. A "leave-one-out" cross validation methodology is used to calculate the root-mean-square error (RMSE) and standard deviation for evaluating both algorithms. The results show that OK sometimes gives an extremely high RMSE value using a Gaussian semi-variance model, and the standard deviation significantly deviates from the measured values. Win-OK was found to more accurately predict the PM_2.5 spatial distribution in a specific sector. The performance of Win-OK is more stable than OK as established by comparing the calculated RMSE and standard deviation from predictions of both algorithms. Win-OK with a spherical semi-variance model is the most accurate method investigated here for deriving the spatial distribution of ground-level PM_2.5. The new algorithm developed here could improve the prediction accuracy of PM_2.5 spatial distribution by considering the effect of wind direction.

Imputing Satellite-Derived Aerosol Optical Depth Using a Multi-Resolution Spatial Model and Random Forest for PM2.5 Prediction

A task for environmental health research is to produce complete pollution exposure maps despite limited monitoring data. Satellite-derived aerosol optical depth (AOD) is frequently used as a predictor in various models to improve PM2.5 estimation, despite significant gaps in coverage. We analyze PM2.5 and AOD from July 2011 in the contiguous United States. We examine two methods to aid in gap-filling AOD: (1) lattice kriging, a spatial statistical method adapted to handle large amounts data, and (2) random forest, a tree-based machine learning method. First, we evaluate each model’s performance in the spatial prediction of AOD, and we additionally consider ensemble methods for combining the predictors. In order to accurately assess the predictive performance of these methods, we construct spatially clustered holdouts to mimic the observed patterns of missing data. Finally, we assess whether gap-filling AOD through one of the proposed ensemble methods can improve prediction of PM2.5 in a random forest model. Our results suggest that ensemble methods of combining lattice kriging and random forest can improve AOD gap-filling. Based on summary metrics of performance, PM2.5 predictions based on random forest models were largely similar regardless of the inclusion of gap-filled AOD, but there was some variability in daily model predictions.