Unleashing the Potential of Geostationary Satellite Observations in Air Quality Forecasting Through Artificial Intelligence Techniques
Abstract. Air quality forecasting plays a critical role in mitigating air pollution. However, current physics-based air pollution predictions encounter challenges in accuracy and spatiotemporal resolution due to limitations in the understanding of atmospheric physical mechanisms, observational constraints, and computational capacity. The world’s first geostationary satellite UV-Vis spectrometer, i.e., the Geostationary Environment Monitoring Spectrometer (GEMS), offers hourly measurements of atmospheric trace gas pollutants at high spatial resolution over East Asia. In this study, we successfully incorporate Geostationary satellite observations into a neural network model (GeoNet) to forecast full-coverage surface nitrogen dioxide (NO2) concentrations over eastern China at 4-hour intervals for the next 24 hours. GeoNet leverages spatiotemporal series of satellite NO2 observations to capture the intricate relationships among air quality, meteorology, and emissions in both temporal and spatial domains. Evaluation against ground-based measurements demonstrates that GeoNet accurately predicts diurnal variations and spatial distribution details of next-day NO2 pollution, yielding the coefficient of determination of 0.68 and root mean square of error of 12.31 μg/m3, significantly surpassing traditional air quality model forecasts. The model’s interpretability reveals that geostationary satellite observations notably improve NO2 forecast capability more than other input features, especially over polluted regions. Our findings demonstrate the significant potential of geostationary satellite observations in artificial intelligence-based air quality forecasting, with implications for early warning of air pollution events and human health exposure.