Assessment of contributors to airborne PAHs and heavy metals in PM10 using temporal, spatial, traffic and heating data in explainable machine learning models

Air pollution in urban areas originates from multiple interacting sources and is strongly influenced by meteorology, yet direct emission data are often incomplete. This study quantifies how meteorological conditions, station location, and proxy indicators of traffic and residential heating affect PM 10 -bound polycyclic aromatic hydrocarbons (PAHs) and metals in Zagreb, Croatia. Daily concentrations of PM 10 , selected PAHs, metals and NO 2 from four monitoring stations (2017–2020) were combined with local and ERA5 meteorology, highway traffic counts and gas consumption as emission proxies. Non-negative Matrix Factorization (NMF) was applied separately to PAHs and metals to identify dominant source-related patterns, while Random Forest regression and SHapley Additive Explanations (SHAP) were used to evaluate the influence of temporal, spatial, meteorological, traffic and heating predictors. NMF separated a heating-related PAH component dominated by Pyr and Flu from a traffic-related component characterised by BaA, Chry and BkF, and indicated enrichment of As and Pb at traffic- and industry-affected stations. Random Forest models showed higher predictive skill for PAHs (R 2 ≈ 0.60–0.68) than for metals (R 2 ≈ 0.24–0.42). Temperature and solar radiation were the main predictors for PAHs, whereas PM 10 , NO 2 and station indicators dominated the prediction of metals. These results demonstrate that integrating proxy emission indicators with explainable machine learning provides an efficient framework for characterising sources and supports season- and location-specific air quality management in data-limited urban environments.