Uncovering Benzene Pollution Patterns Using an Interpretable, Setting-Aware Artificial Intelligence Approach

We investigated benzene variability in an urban environment using an interpretable, setting-based artificial intelligence framework. A seven-year dataset (2017–2023) of hourly pollutant concentrations (benzene, NO2, SO2, CO, O3) measured in Zagreb (Croatia) was analyzed, as were meteorological variables. Multiple-ensemble decision tree models were developed, with hyperparameters optimized using metaheuristic algorithms. The best-performing model, Extra Trees optimized by the Sine Cosine Algorithm, achieved an R2 of 0.87. Model interpretation employed Shapley additive explanations (SHAP), followed by PaCMAP embedding and HDBSCAN clustering to identify coherent environmental settings. Seven settings (C0–C6) and one residual group were identified, representing pollution-enhancing, suppressing, and transitional regimes. Two settings dominated benzene extremes. C6 reflected winter stagnation, characterized by strong combustion influence (CO contribution of 11.9%), shallow boundary layers (~290 m), weak winds, and high humidity. C4 represented a synoptic stability regime with enhanced heat fluxes and diminished after the COVID-19 period, consistent with altered anthropogenic activity. Low-benzene settings (C0, C1, C3) were associated with stronger mixing and higher oxidizing capacity, while transitional settings (C2, C5) reflected moderate conditions. Overall, the results show that a small number of environmental settings governed the benzene extremes, providing a transferable and interpretable framework for air quality assessment and policy support.