A hybrid land use regression/line-source dispersion model for predicting intra-urban NO2

Land use regression (LUR) has evolved as a standard tool for estimating intra-urban variation in long-term air pollution exposures. LUR, however, generally relies on observed spatial relationships between distributed sources and pollution measures, rather than meteorological or physio-chemical information. Incorporating traffic-meteorological interaction information into LURs via dispersion models may improve accuracy and physical interpretability. To examine whether deterministic dispersion output improves LUR predictions for nitrogen dioxide (NO2), we incorporated hourly Caline3QHCR dispersion information into existing winter-time regression models originally designed to disentangle effects of multiple sources (e.g., legacy industry, vehicle traffic) and concentration modifiers (e.g., elevation) across Pittsburgh, PA. Caline3QHCR output improved overall cross-validated R2 values by 0.10, 0.03, and 0.05 for the weekday, full-week, and merged years models, respectively. The addition of Caline3 output as an independent covariate effectively displaced two road-specific predictors, one built environment predictor and one meteorological predictor across two winter-season models. The incorporation of study-specific dispersion principles may corroborate overall LUR model interpretability and could be applied similarly to preexisting LUR models that lack strong source-concentration relationships. Model improvements were observed at sites with relatively higher NO2 concentrations that exhibited higher traffic volumes and were in closer proximity to roadways (< 300 m), which may have an important bearing towards better characterizing exposures in both near-roadway and high-traffic environments.