Machine learning in spatiotemporal PM2.5 hotspots analysis for better targeted mitigation strategies

Abstract

Rapid urbanization has driven significant air pollution, posing serious health risks, especially to marginalized communities. Despite ongoing control efforts, gaps remain in designing effective mitigation strategies. This study uses machine learning for spatiotemporal analysis of PM_2.5 hotspots in New York City, revealing a 10–25 % decline in concentrations from 2010 to 2019 but identifying 173 consistent, 62 emerging, and 27 declining hotspots. Among tested models, AutoGluon achieved 93 % accuracy and 88 % weighted F1 score, outperforming CatBoost, XGBoost, and Random Forest. Temporal features contributed 13.5 % of predictive power, while spatial features, like road networks (1.8 %), vehicle operations (0.7 %), and land use (0.7 %), collectively accounted for 3.2 %. Scenario analysis showed traffic volume restrictions and truck prohibitions effectively reduced pollution, with congestion pricing proving most impactful across eight hotspot regions. Equity-focused scenarios highlighted improvements in 15 areas under congestion pricing and 19 with odd–even license plate access. This framework offers actionable insights for targeted, equitable pollution control.