A Spatiotemporal Prediction Model for Black Carbon in the Denver Metropolitan Area, 2009-2020

Abstract

Studies on health effects of air pollution from local sources require exposure assessments that capture spatial and temporal trends. To facilitate intraurban studies in Denver, Colorado, we developed a spatiotemporal prediction model for black carbon (BC). To inform our model, we collected more than 700 weekly BC samples using personal air samplers from 2018 to 2020. The model incorporated spatial and spatiotemporal predictors and smoothed time trends to generate point-level weekly predictions of BC concentrations for the years 2009-2020. Our results indicate that our model reliably predicted weekly BC concentrations across the region during the year in which we collected data. We achieved a 10-fold cross-validation R² of 0.83 and a root-mean-square error of 0.15 μg/m³ for weekly BC concentrations predicted at our sampling locations. Predicted concentrations displayed expected temporal trends, with the highest concentrations predicted during winter months. Thus, our prediction model improves on typical land use regression models that generally only capture spatial gradients. However, our model is limited by a lack of long-term BC monitoring data for full validation of historical predictions. BC predictions from the weekly spatiotemporal model will be used in traffic-related air pollution exposure-disease associations more precisely than previous models for the region have allowed.

Figure 1.

(A–E) Central and distributed site sampling locations for each campaign. Note: Locations are jittered to protect study participant privacy.

Figure 2.

Distributions of calibrated BC concentrations measured at distributed sites (residential and community) by campaign week. Note: Due to smaller sample sizes in Campaign 5, concentrations for individual filters are shown.

Figure 3.

Plots showing temporal trend functions used in the model: (A) smoothed temporal trend functions with four degrees of freedom per year for the entire modeling period (2009 to 2020), (B) weekly BC measurements and fitted trends at the central site BC monitor (2016–2020), and (C–E) weekly BC measurements and fitted trends at three representative distributed sites.

Figure 4.

Results from the 10-fold cross-validation: (A) observed and predicted weekly BC (μg/m³) and (B) long-term average BC concentrations (μg/m³) with 95% prediction intervals.

Figure 5.

Map of the long-term average BC predicted for 2018 near downtown Denver, CO. Concentrations are shown at a 250 m resolution.

Figure 6.

Boxplots of the weekly BC predicted at distributed sampling sites for each month, 2009–2020.