Deep learning-based super-resolution of GEMS hyperspectral data using GOCI-II fusion: Advancing high-resolution air quality monitoring

Hyperspectral observations from geostationary satellites provide detailed spectral information that is highly valuable for air quality monitoring, enabling improved characterization of aerosols and trace gases through their distinct spectral signatures. The Geostationary Environment Monitoring Spectrometer (GEMS) offers continuous hyperspectral measurements with high temporal resolution over the Asia–Pacific region, making it well suited for monitoring diurnal variations in atmospheric composition. However, the relatively coarse spatial resolution of hyperspectral geostationary sensors limits their ability to resolve fine-scale spatial heterogeneity in air pollution, especially in urban regions. This trade-off between spectral fidelity and spatial resolution remains a fundamental limitation of single-sensor satellite-based air quality monitoring. To address this challenge, this study develops a deep learning–based fusion framework that integrates hyperspectral radiance from GEMS with high-spatial-resolution multispectral observations from the Geostationary Ocean Color Imager-II (GOCI-II). A self-supervised learning strategy is used to improve the spatial resolution of GEMS Level-1C (L1C) radiance by using spatial patterns from GOCI-II. This makes hyperspectral super-resolution possible without needing high-resolution hyperspectral ground truth data. Validation against the original GEMS L1C data shows that the super-resolved radiance is very consistent in both space and time, with correlation coefficients (R) over 0.95 and normalized root mean square error (nRMSE) under 10%. The resulting super-resolved radiance preserves spectral information while providing substantially finer spatial detail than existing satellite products. Furthermore, the enhanced hyperspectral radiance is linked to surface-level air pollutant (e.g., PM₁₀, PM_2.5, and NO₂) concentrations through artificial intelligence-based models, demonstrating its applicability for high-resolution air quality monitoring. The proposed multi-satellite fusion framework highlights the value of integrating complementary satellite observations with data-driven approaches for urban-scale air quality analysis.