Exposure models for particulate matter elemental concentrations in Southern California

• Published in: Environment international 2022-07, Vol.165, p.107247-107247, Article 107247
• Main Authors: Yin, Xiaozhe, Franklin, Meredith, Fallah-Shorshani, Masoud, Shafer, Martin, McConnell, Rob, Fruin, Scott
• Format: Article
• Language: English
• Subjects: Land use regression; Particulate matter; Particulate matter speciation; Spatial resolution
• ISSN: 0160-4120; 1873-6750
• DOI: 10.1016/j.envint.2022.107247

[Display omitted] Due to a scarcity of routine monitoring of speciated particulate matter (PM), there has been limited capability to develop exposure models that robustly estimate component-specific concentrations. This paper presents the largest such study conducted in a single urban area. Using samples that were collected at 220 locations over two seasons, quasi-ultrafine (PM0.2), accumulation mode fine (PM0.2-2.5), and coarse (PM2.5-10) particulate matter concentrations were used to develop spatiotemporal regression, machine learning models that enabled predictions of 24 elemental components in eight Southern California communities. We used supervised variable selection of over 150 variables, largely from publicly available sources, including meteorological, roadway and traffic characteristics, land use, and dispersion model estimates of traffic emissions. PM components that have high oxidative potential (and potentially large health effects) or are otherwise important markers for major PM sources were the primary focus. We present results for copper, iron, and zinc (as non-tailpipe vehicle emissions); elemental carbon (diesel emissions); vanadium (ship emissions); calcium (soil dust); and sodium (sea salt). Spatiotemporal linear regression models with 17 to 36 predictor variables including meteorology; distance to different classifications of roads; intersections and off ramps within a given buffer distance; truck and vehicle traffic volumes; and near-roadway dispersion model estimates produced superior predictions over the machine learning approaches (cross validation R-squares ranged from 0.76 to 0.92). Our models are easily interpretable and appear to have more effectively captured spatial gradients in the metallic portion of PM than other comparably large studies, particularly near roadways for the non-tailpipe emissions. Furthermore, we demonstrated the importance of including spatiotemporally resolved meteorology in our models as it helped to provide key insights into spatial patterns and allowed us to make temporal predictions.