Capturing heterogeneity in urban air quality forecasting using a station-specific machine learning model

Research on predicting urban air quality using machine learning techniques typically employs global models that assume spatial homogeneity. This study aims to comprehensively evaluate the predictive limitations of global and station-based XGBoost methods and to examine the assumption of spatial homogeneity in complex and data-constrained environments such as Istanbul. To examine the link between traffic and air pollution, data from a single month were analysed, minimising seasonal variability. Global models showed satisfactory performance for coarse particulate matter (PM10; R2 = 0.92), but their effectiveness for nitrogen dioxide (NO2) was limited due to their inability to distinguish factors affecting local distribution (R2 = 0.35). However, the best accuracy was achieved with the proposed station-based methodology (R2 = 0.93 for nitrogen dioxide), resulting in an 85% improvement in root mean squared error. Furthermore, feature significance analysis and pure prediction trials show that traffic data, when considered as a single variable, can explain approximately 7–10% of the prediction variance in the models. This finding demonstrates that, contrary to popular belief, the impact of road traffic on urban air quality is quite limited, while region-specific spatial and meteorological factors have a more significant effect. Consequently, it is concluded that standard and traffic-focused interventions are insufficient for effective environmental management, and the adoption of region-specific methods is necessary.