Aerosol optical depth retrieval from Geostationary Environment Monitoring Spectrometer (GEMS): Advancing the first hyperspectral geostationary air quality mission using deep learning

Abstract

Aerosol optical depth (AOD) quantifies atmospheric aerosol loading and is fundamental for air quality monitoring, aerosol-climate interaction studies, and climate change assessments. Traditional satellite-based AOD retrieval methods using radiative transfer models are computationally intensive and prone to uncertainties stemming from assumptions about aerosol properties and surface reflectance. To mitigate these limitations, we proposed a deep learning-based AOD retrieval method using Attentive Interpretable Tabular Learning (TabNet). This approach enables accurate and efficient AOD retrieval from the Geostationary Environment Monitoring Spectrometer (GEMS), the world's first geostationary hyperspectral environmental satellite, over the Asia–Pacific region. The TabNet model integrates GEMS hyperspectral normalized radiance with meteorological and auxiliary variables, leveraging its high spectral resolution to improve accuracy. Model performance was evaluated through random, spatial, and temporal 10-fold cross-validation, demonstrating strong predictive accuracy with coefficient of determination (R2) values between 0.90 and 0.93 and 76.4–83.7 % of predictions falling within the expected error (EE) envelope of ±(0.05 + 0.15 × AOD). This significantly outperformed the physics-based GEMS Level-2 product (R2 = 0.57; within EE = 48.3 %), demonstrating enhanced accuracy and robustness of the TabNet approach. The TabNet-derived AOD effectively captured seasonal and regional aerosol distributions across the Asia–Pacific region and resolved hourly aerosol variability during extreme aerosol events such as biomass burning and dust storms. Moreover, it surpassed other satellite-based products, including Moderate Resolution Imaging Spectroradiometer Multi-Angle Implementation of Atmospheric Correction and Advanced Meteorological Imager AOD. These findings highlight the potential of the TabNet-based AOD retrieval framework as a powerful tool for air quality and climate applications, providing valuable data to support public health management. © 2025 Elsevier B.V.