Measuring the Emission Changes and Meteorological Dependence of Source‐Specific BC Aerosol Using Factor Analysis Coupled With Machine Learning

Abstract Reducing ambient black carbon (BC) relies on the targeted control of anthropogenic emissions. Measuring emission changes in source‐specific BC aerosol is essential to assess the effectiveness of regulatory policies but is difficult due to the presence of meteorology and multiple co‐existing emissions. Herein, we propose a data‐driven approach, combining dispersion‐normalized factor analysis (DN‐PMF) with a machine learning weather adjustment (deweathering) technique, to decompose ambient BC into source emissions and meteorological drivers. Six refined BC sources were extracted from the factor analysis of aethalometer multi‐wavelength BC and concurrent observational datasets. In addition to the widely reported dominant sources, such as vehicular emissions (VE) and coal/biomass burning (BB), a discernible port and shipping emission source were identified with potential impacts on coastal air quality. The source‐specific BC showed abrupt changes in response to interventions (e.g., holidays) after separating weather‐related confounders. Significant reductions in deweathered coal and BB, VE, and local dust verified the effectiveness of policies, such as clean winter‐heating and support for the Clean Air Actions. As revealed by a post‐hoc model explanation technique, the evolution of the boundary layer was the predominant meteorological driver exerting the opposite impact on local sources with respect to distant regional‐wide sources, that is, the port and shipping emissions.