Mitigating bias induced by missing data in new-generation geostationary satellite monitoring of ground-level NO2 via machine learning.

The Geostationary Environment Monitoring Spectrometer (GEMS) has revolutionized air quality monitoring with hourly resolution from geostationary Earth orbit (GEO). However, satellite-derived air quality data often face limitations and biases due to missing data. Given the growing role of GEO environmental satellites, it is crucial to evaluate these limitations and correct biases in detail on an hourly basis. Based on GEMS measurements, this study assesses the potential for improving data availability and mitigating bias in monitoring ground-level nitrogen dioxide (NO2) concentrations in eastern China through a machine learning framework that integrates gap-filling and column-to-ground conversion processes. The results indicate that the gap-filling process significantly enhanced data availability from 10 to 50 % to full coverage across China. Furthermore, the seamless data substantially reduced bias in estimating the annual mean of ground-level NO2 concentrations, eliminating a significant underestimation of over 3.0 μg/m3 in 36.3 %, 47.2 %, and 63.6 % of the area for 8 a.m., 2 p.m., and 3 p.m., respectively. These findings enhance our understanding of the biases induced by missing data in new-generation GEO satellite measurements and highlight the need for seamless spatio-temporal mapping of air quality to address these limitations.