Development and validation of satellite-derived surface NO ₂ estimates using machine learning versus traditional approaches in North America

Abstract. Nitrogen dioxide (NO2) is one of the key pollutants with profound implications for air quality, and human health, and is needed to establish the air quality health index (AQHI). Currently, over 600 surface air monitoring stations are distributed across Canada and the United States measuring NO2, but many areas remain unmonitored leading to incomplete information for health risk assessments. This study leverages Tropospheric Monitoring Instrument (TROPOMI) satellite observations and machine learning models to derive high-resolution surface NO2 concentrations, provides enhanced spatial coverage and accuracy, revealing urban-rural NO2 gradients across North America. Existing traditional methods rely on scaling with modeled profiles to obtain NO2 surface concentrations from satellite observations. Here, we compare this traditional method to a machine learning approach that utilizes NO2 observations from TROPOMI, together with meteorological parameters, land cover type, topography, and emission inventories. Our results show that the machine learning (using random forest) yields less bias between the surface monitoring measurements and the “satellite-derived” surface concentrations, significantly improved the correlation coefficient (R2∼ 0.77–0.91) compared to the traditional method (R2∼ 0.39–0.57) and yields to significantly less bias.