Improved Retrieval of AOD from Geostationary Environment Monitoring Spectrometer (GEMS) Based on Artificial Intelligence Techniques

Atmospheric aerosols are not only air pollutants that are harmful to human health but also play a crucial role in understanding climate change due to their influence on the Earth's radiation balance. As aerosol optical depth (AOD) provides a quantitative estimate of aerosol concentration in the atmosphere, it serves as a valuable proxy for air quality, offering insights into both short-term pollution events and long-term environmental trends. The primary method for obtaining satellite-based AOD involves using a radiative transfer model, which is highly computationally intensive and uncertain. An artificial intelligence (AI)-based model was proposed in this study to get AOD from the Geostationary Environment Monitoring Spectrometer (GEMS) over the Asia-Pacific regions (5°S–45°N, 75°E–145°E) in order to overcome these problems. The attentive Interpretable Tabular Learning Neural Network (TabNet) algorithm was developed using hyperspectral normalized radiance (NR) and observation angle data obtained by the GEMS satellite sensor, meteorological variables from the ERA5 (ECMWF Reanalysis v5), and other auxiliary data. Aerosol Robotic Network (AERONET) and the Sun-Sky Radiometer Observation Network (SONET)-measured AOD data were used as a dependent variable. The spatiotemporal robustness of the proposed model was evaluated using random, spatial, and temporal 10-fold cross-validation (CV) with in-situ data. The TabNet model showed stable and high accuracy of R² = 0.89 – 0.93 and within expected error (withinEE) = 75.1 – 84.6% across all validation methods (RDCV, SPCV, and TPCV). This meant that the results were more accurate than the operational GEMS AOD product, which is based on physics. The model demonstrates stable performance in simulating concentration changes during high aerosol events, such as dust storms, with high temporal resolution (hourly) and extensive coverage (Asia-Pacific region), suggesting its effective use for air quality monitoring.