Real-time nowcasting of NO2 products from geostationary environment monitoring spectrometer using a conditional generative adversarial network

Abstract

In East Asia, megacities like Seoul, Tokyo, and Shanghai frequently recording high nitrogen dioxide (NO₂) concentrations due to traffic and industrial activity require urgent efforts to enhance short-term monitoring and forecasting systems. This research presents a deep-learning (DL) model for nowcasting atmospheric NO₂ concentration products derived from the geostationary environment monitoring spectrometer (GEMS) on the Geo-Kompsat-2B satellite from 1-h to 3-h. The DL model utilizes pairs of GEMS NO₂ products as input and output datasets. The nowcasting DL model was developed using a data-to-data (D2D) translation method incorporating conditional generative adversarial network techniques. The D2D-nowcast NO₂ model was trained and tested for 1, 2, and 3-h predictions. The test results of the D2D model demonstrated excellent statistical performance, including a correlation coefficient of 0.805, a root-mean-square error of 0.162 ⨉ 10¹⁶ molecules/cm², and a bias of 0.046 ⨉ 10¹⁶ molecules/cm² for the 3-h prediction. Furthermore, the D2D-nowcast NO₂ concentrations were validated using the Tropospheric Monitoring Instrument and Pandora NO₂ measurements, demonstrating high agreement. Consequently, this study aims to support real-time operational monitoring by supplementing temporal gaps in satellite observations without relying on numerical models and provides valuable supplements for decision-making by air quality forecasters.