The gap between column-integrated satellite retrievals and patchy ground-level monitoring makes it difficult to accurately characterize the spatial distribution of nitrogen dioxide (NO₂), which is crucial for environmental health. In order to close this observational gap in temperate climate regions, this study suggests an approach for integrating machine learning. We created a geographical predictive model utilizing a Random Forest Regressor by combining ground-truth data from OpenAQ with Sentinel-5P TROPOMI tropospheric column densities from January 2023 to January 2025. To estimate ground-level concentrations, the model combines satellite observations with geographic coordinates and category location contexts. With a Coefficient of Determination (R²) of 0.4337 and a Mean Absolute Error (MAE) of 8.25 µg/m³, the model effectively established a spatial transfer function despite the different physical characteristics of the two datasets. As a proof-of-concept for low-latency air quality assessment, these measurements show a strong capacity to resolve spatial variability in surface NO₂ using satellite inputs. This study provides a scalable framework for improving surveillance in regions without dense sensor infrastructure by validating the effectiveness of machine learning in downscaling satellite products for localized monitoring.
Sint et al. (Tue,) studied this question.