Ultra-high-resolution mapping of ambient fine particulate matter to estimate human exposure in Beijing

The study addresses how declining regional PM2.5 levels and increasing contributions from local sources (50–80%) complicate exposure assessment that relies only on ambient concentrations. Traditional approaches (monitoring interpolation, CTMs at ~1 km) and earlier high-resolution LUR models inadequately capture micro-environment heterogeneity. The research aims to develop an ultra-high-resolution (30 m) PM2.5 field for Beijing, integrate it with fine-scale population activity, account for indoor/outdoor exposure differences to estimate personal internal dose, and assess implications for health burden estimation, thereby improving exposure characterization and revealing exposure inequalities unobservable with coarser ambient data.

Prior studies linked PM2.5 to major diseases and used exposure assessment models tying ambient levels to health effects. Earlier high ambient pollution made inter-city differences prominent, minimizing local heterogeneity. As ambient levels declined, local source contributions and micro-environment hot spots became more important. Fusion of auxiliary data (satellite, CTM outputs, meteorology) improved accuracy, but models often remained coarse, and high-resolution variables were resampled to ~1 km. Recent availability of ultra-high-resolution data (e.g., Landsat-8 TOA, Sentinel-derived products) and machine learning improved estimates versus linear LUR due to nonlinearity in PM2.5–predictor relationships. Existing cohort exposure assignments often used nearest monitors, coarse satellite retrievals, or CTMs at ~1 km; street-level or multi-scale integrated systems exist but can be data and computation intensive. This study builds on that literature by fusing multi-source ultra-high-resolution predictors using machine learning to produce a 30-m PM2.5 field and linking it to population activity and I/O exposure for internal dose and mortality assessment.

Study area: Beijing, China. Air quality modeling used WRF (v3.8.1) and CMAQ (v5.2) for January, April, July, and October of 2019 (3-day spin-up each), representing seasons. Four nested domains (36×36, 12×12, 4×4, and ~1.33×1.33 km²) with 32 vertical layers (lowest ~37 m) were configured; SLUCM was coupled in the innermost domain to improve urban meteorology. Emissions: MEIC (0.25°) for China (base year 2015) with Beijing gridded local emissions. Meteorological initial/boundary conditions from NCEP FNL; FDDA enabled with NCEP ADP surface/upper-air observations. CMAQ initialized via ICON with boundary nesting across domains. Ultra-high-resolution assimilation: Constructed an auxiliary dataset of 9 variable types, including: hourly PM2.5 and meteorology from WRF-CMAQ (e.g., PBLH, SFCTMP, WSPD10, RH), national monitoring PM2.5 (34 sites; 2018-11-01 to 2019-11-01), 30-m land use, Landsat-8 band-2 TOA reflectance at 30 m aggregated to monthly medians, POIs, road length/emissions (within 1000 m buffers), building location, population distribution, elevation, GDP, and other variables. Coarser-than-30 m inputs were split to finer meshes assuming homogeneous properties. Models: Compared stepwise LUR and a random forest (RF; optimized integrated-tree model). Validation via 10-fold CV and LOOCV using monitoring data; RF outperformed LUR (higher R2, lower RMSE), so selected for assimilation. The trained model was applied monthly, combining hourly CTM outputs with monthly/annual variables to produce 30-m PM2.5 maps. Population distribution: Hourly mobile signaling (Baidu Smart Eye) provided relative crowdedness; weekdays/weekends patterns assumed stable across the year. Gridded total population from WorldPop scaled the relative densities. Because raw mobile data resolution is hundreds of meters, inverse distance weighting interpolated to 30 m to align with building-scale analysis. Indoor PM2.5 and I/O exposure: Indoor concentrations were estimated by applying literature-based I/O ratios to ambient PM2.5 by land-use type: 0.9 in residential buildings and 0.7 in public buildings (without air cleaners). Sensitivity analyses varied I/O ratios (±0.1 to 0.2) and population distribution between residential/public buildings. Personal internal dose was computed by combining indoor/outdoor concentrations with exposure time patterns for micro-environments and inhalation rates, with age/gender standardization. Within downtown (inside the Fifth Ring), personal hourly/daily doses were mapped. Health impact assessment: Mortality attributable to long-term PM2.5 was assessed for ischaemic heart disease (IHD), stroke, COPD, and lung cancer (LC) using the Global Exposure Mortality Model (GEMM), applied to both the annual WRF-CMAQ ambient field and the ambient high-resolution (30 m) assimilation field to quantify resolution impacts on burden estimates. Statistical analyses used MATLAB R2021a; spatial processing used ESRI ArcGIS 10.3.

- 30-m PM2.5 mapping achieved substantially higher agreement with monitors (R2=0.78–0.82) compared to raw WRF-CMAQ (R2=0.31–0.64). Annual mean PM2.5 increased from 30.9 µg/m³ (WRF-CMAQ) to 34.3 µg/m³ (assimilated), correcting southeastern overestimation and northwestern underestimation and revealing local hotspots. - Population-weighted ambient PM2.5: 34.6 µg/m³ on weekdays and 34.5 µg/m³ on weekends for Beijing, about 15% higher than unweighted averages. Assimilation adjusted district-level estimates (elevated in NW districts; lowered in Tongzhou and Pinggu), with little weekday–weekend difference in the citywide population-weighted mean despite fine-scale shifts in micro-environment exposure locations. - Indoor concentrations: Average indoor PM2.5 was 26.5 µg/m³ (range 10.5–39.5). Within the Fifth Ring, mean indoor 26.9 µg/m³ and outdoor 41.6 µg/m³, higher in the southeast and lower in the northwest, consistent with ambient patterns. Indoor levels in Beijing have declined relative to past years but remain higher than many developed-country settings. - Personal internal dose: Within the Fifth Ring, mean hourly internal doses were ~24.6 µg/h (ambient outdoor), 22.6 µg/h (residential indoor), and 16.0 µg/h (public indoor). Age/gender-standardized daily internal dose per person was 568.2 µg/d using ambient WRF-CMAQ, 594.5 µg/d using ambient assimilation (+~5%), and 512.9 µg/d when accounting for I/O differences (−14% vs ambient assimilation). Males exceeded females by ~120–150 µg/d for ages >10 years (~20 µg/d among children). Children <10 years ~250 µg/d; ages 10–64 years 450–600 µg/d; ≥65 years ~350–450 µg/d. - Sensitivity: Varying residential/public I/O ratios by 0.2 independently or ±0.1 simultaneously changed total internal dose by ≤±10%; shifting the residential/public population distribution ratio by ±0.1 changed dose within ±2%. - Mortality burden: GEMM-based annual deaths (95% CI) from IHD, stroke, COPD, LC were 20,540 (16,908–24,086) with WRF-CMAQ vs 25,462 (20,901–29,881) with high-resolution assimilation. Stroke contributed nearly half of total attributable deaths. The ~25.9% higher ambient level in the assimilated field led to a 24.0% higher mortality estimate, highlighting health-assessment sensitivity to exposure resolution.

The ultra-high-resolution (30 m) assimilation reconciles CTM biases and captures local source-driven heterogeneity, improving exposure estimation beyond what is possible with ~1 km CTMs or monitor interpolation. While population-weighted ambient means show small weekday–weekend differences, fine-scale population activity reveals substantial shifts in micro-environment exposure that are invisible in aggregate metrics. Incorporating I/O exposure yields materially different internal dose estimates (−14% vs ambient-only), and age/gender differences are substantial due to inhalation rates. Health burden estimates increase appreciably when using high-resolution ambient fields, demonstrating that model resolution can influence policy-relevant outcomes. The work clarifies distinctions between individual and population exposure risks and suggests considering dynamic standards and targeted interventions where population density and exposure intersect. Overall, integrating 30-m PM2.5 with high-resolution population activity and I/O adjustments addresses the study aim of more accurately quantifying exposure and associated health impacts in a megacity.

This study develops and validates a 30-m PM2.5 mapping framework for Beijing by fusing multi-source ultra-high-resolution predictors with an RF model, substantially outperforming raw CTM outputs. By integrating 30-m population distributions and I/O exposure differences, it estimates personal internal doses across demographic groups and demonstrates pronounced exposure heterogeneity and its implications for health burden estimation, including a ~24% increase in attributable mortality when using high-resolution ambient fields versus coarse CTM outputs. The framework provides a basis for refined exposure assessments, inequality identification, and potentially more precise epidemiological analyses. Future work should: (1) extend assessments across multiple years to track urban-scale heterogeneity trends; (2) refine population activity characterizations and I/O exposure modeling (including indoor sources and air cleaners); (3) integrate meteorology-dependent I/O transport and street-canyon effects; and (4) develop exposure–response relationships based on internal dose to inform evolving air quality baselines and standards.

Uncertainties arise from WRF-CMAQ and RF modeling. Population activity was inferred from mobile signaling heat maps, likely underrepresenting non-permanent residents and lacking age/gender stratification. Indoor concentrations were estimated using simplified, literature-based I/O ratios without accounting for air cleaners or building-specific indoor sources; I/O ratios can vary widely (reported >180 in some venues) and are influenced by meteorology (humidity, wind speed/direction) and urban street-canyon effects, which were not explicitly modeled. Data scarcity limited dynamic I/O ratio modeling and detailed building ventilation characterization. Personal indoor dose estimates were provided only for downtown areas, with higher uncertainty in suburban regions with sparse buildings. These factors may affect the precision and generalizability of exposure and dose estimates.