Spatiotemporal Patterns and Random Forest Prediction of Tropospheric NO₂ in Ankara, Türkiye

Urban air pollution remains a critical challenge in rapidly urbanizing metropolitan re-gions, where complex topography and uneven monitoring infrastructure limit accurate exposure assessment. Nitrogen dioxide (NO₂), primarily emitted from traffic and combus-tion sources, exhibits marked spatial heterogeneity that is often underrepresented by sparse ground-based stations. This study examines the spatiotemporal variability of tropospheric NO₂ over Ankara Province, Türkiye, for 2025 and develops and implements a machine learning-based downscaling framework integrating Sentinel-5P TROPOMI ob-servations with Sentinel-2 multispectral surface reflectance data, without relying on an-cillary meteorological or emission datasets. After rigorous quality filtering and temporal aggregation, a Random Forest regression model was used to generate annual NO₂ maps at 500 m resolution based solely on spectral predictors. Results indicate strong seasonal variability, with winter monthly means reaching 8.93 × 10⁻⁵ mol/m² and peak values ex-ceeding 30 × 10⁻⁵ mol/m², alongside a persistent urban–rural gradient radiating from the metropolitan core. The optimized model achieved consistent predictive performance (R² = 0.30; RMSE = 2.72 × 10⁻⁵ mol/m²), with SWIR and Red Edge bands contributing most strongly. These findings demonstrate that high-resolution urban NO₂ patterns can be re-liably inferred from optical satellite data alone, providing a transferable and scalable framework for air quality assessment in data-limited metropolitan environments.