Development of roadway link screening model for regional-level near-road air quality analysis: A case study for particulate matter

Abstract Conducting high-resolution air quality analysis by applying microscale dispersion models at the regional scale poses a formidable computing challenge, because a huge number of receptors and the extensive network of roadway links (emission sources) must be processed. As a way to minimize computation cost without undermining estimation precision, this study proposes an innovative link screening methodology, using a supervised machine learning random forest (RF) classification algorithm, that eliminates links with zero or negligible concentration contributions from modeled link-receptor combinations. The study uses 79,328 receptor-link pairs randomly selected from the Atlanta Metropolitan area to train and test the model. The final link screening model employs six variables, including link attributes, urban variables, and meteorological conditions. The RF classifier successfully identifies the small portion of links that contribute more than 95% of concentrations that are estimated by the same model using every link-receptor pair. The efficiency and precision of the smaller dispersion model runs developed using the RF classifier (the ‘reduced-link’ model) are compared to the dispersion modeling without the link-screening process (the ‘whole-link’ model) for downtown Atlanta and northwest Atlanta. Results show that AERMOD run-times for reduced-link models are only 0.2%–1.1% of the times required for whole-link models, because far fewer links are handled during the AERMOD simulation (0.1%–0.6% of links in the whole-link model). The correlation between estimates of the two models ranges from 95% to 97%, depending upon the density of the road network, link activity, link emission rates, meteorology, etc.