Abstract
Urban air pollution poses a major threat to public health and environmental sustainability. This study proposes a structured machine learning (ML)-based framework to examine how temporal and spatial resolution choices affect the accuracy of urban air pollution modeling. The research is conducted in two distinct phases. In the temporal phase, the impact of incorporating pollutant autocorrelation into ML models (Multilayer Perceptron (MLP) and Random Forest (RF)) is analyzed by comparing results with and without temporal lag features derived through autoregressive (AR) modeling. In the spatial phase, emission inventory data are aggregated at three spatial resolutions (500 m, 750 m, and 1000 m) to evaluate their effect on model performance in predicting PM and NOx concentrations. Results from the temporal modeling phase indicate that including lag features significantly improves PM predictions: RMSE for PM₁₀ is reduced by 25.9% (from 92.56 µg/m(3) to 68.59 µg/m(3)), and for PM(2.5) by 38.9% (from 61.10 µg/m(3) to 37.30 µg/m(3)). Conversely, for NOx, RMSE increases by 53.2% (from 7.90 µg/m(3) to 12.10 µg/m(3)), indicating pollutant-specific temporal behavior. In spatial modeling, a coarser resolution (1000 m) yields better performance for PM (RMSE = 13.51 kg/year), while a finer resolution (500 m) is more effective for NOx (RMSE = 307.50 kg/year). Among the evaluated algorithms, MLP consistently achieves the highest predictive accuracy across both temporal and spatial scenarios. These findings underscore the importance of selecting appropriate temporal and spatial resolutions tailored to each pollutant type. The proposed framework offers a flexible, resolution-aware modeling strategy that can support more effective urban air quality management policies.