Morphology-constrained GAN for high-resolution proxy estimation of urban air pollution: A rapid assessment tool for urban planning

Fine-scale estimation of near-surface air pollution is important for health-oriented urban planning, yet observations at design-relevant spatial scales remain limited. Macro-scale statistical models often lack the spatial detail required for localized design evaluation, whereas high-fidelity physics-based simulations are difficult to apply repeatedly during early-stage design because of their scenario-specific setup and computational cost. To address this gap, we developed a rapid, morphology-constrained generative adversarial network (GAN) framework for estimating 30-m-grid proxy concentration surfaces for PM 2.5 , black carbon (BC), NO, NO 2 , and summer-mean O 3 . Using New York City as a case study, we integrated LiDAR-derived urban-form indicators, land-use zoning, and building footprints with proxy target surfaces generated by applying ordinary kriging to publicly available 300-m NYCCAS raster products. The framework was evaluated on a geographically held-out test set against Ridge regression, LightGBM, and U-Net. Relative to U-Net, the proposed GAN reduced mean absolute error by 17.5%–42.9% across the five pollutants. For detecting hotspots defined by the highest 20% of reference concentrations, it achieved F1 scores above 0.71 and Intersection over Union scores of 0.58–0.69. Applications to multiple design scenarios further revealed pollutant- and location-specific responses to morphological interventions. The framework provides a rapid decision-support tool for comparing urban-form alternatives during early-stage planning.