Uncertainty-aware deep learning for downscaling air quality concentrations

Air pollution is a leading environmental risk factor to human health worldwide. Fine particulate matter (PM2.5) is of particular concern due to its ability to penetrate the lungs and enter the cardiovascular system, contributing to respiratory disease, cancer and early mortality. Accurate, high-resolution exposure estimates are therefore essential for assessing health risk and informing effective intervention strategies. However, regulated ground monitoring networks offer limited spatial coverage, and remote sensing and chemical transport models (CTMs) often fail to resolve local-scale variation in PM2.5 concentrations. A further limitation is that most machine learning approaches to this problem are deterministic, providing no indication of predictive confidence. This study addresses both gaps with a Bayesian deep learning framework that downscales coarse CAMS CTM output to 1 km resolution by leveraging fine-scale land use, population and meteorological covariates and trained against ground observations. Predictive uncertainty was quantified via Monte Carlo dropout, a Bayesian approximation method that adds minimal computational cost. The model achieved strong deterministic accuracy (R2 = 0.814) with well-calibrated, if slightly conservative, predictive intervals (95% coverage = 0.969) on held-out test data, offering a scalable approach for high-resolution exposure mapping with quantified uncertainty.