Synergistic Aerosol Optical Depth Fusion from GEMS, AMI, and GOCI-II Sensors on GEO-KOMPSAT: Statistical and Deep Learning Integration for Enhanced Atmospheric Monitoring

Aerosol optical depth (AOD) data fusion with aerosol datasets from the Geostationary Korea Multi-Purpose Satellites (GEO-KOMPSAT) has been performed using both statistical and deep neural network-based methods. The GEO-KOMPSAT mission includes an Advanced Meteorological Imager (AMI) onboard GK-2A, as well as a Geostationary Environment Monitoring Spectrometer (GEMS) and the successor to the Geostationary Ocean Color Imager (GOCI-II) onboard GK-2B. The statistical fusion method corrects the bias of each aerosol product by assuming a Gaussian error distribution. Subsequently, the Maximum Likelihood Estimation (MLE) fusion technique accounts for pixel-level uncertainty by weighting the Root Mean Square Error (RMSE) of each AOD product for every pixel. A deep neural network (DNN)-based fusion model has been trained to target AERONET AOD values, employing fully connected hidden layers. Both the statistical and DNN-based fusion results generally outperform the individual GEMS and AMI AOD datasets. Notably, when quantifying higher aerosol loading, DNN AOD outperforms MLE AOD. The selection of DNN AOD, which incorporates all aerosol products in the process, effectively addresses the rapid increase in uncertainty at higher aerosol loading levels. Overall, the fusion AOD, particularly the DNN AOD, exhibits performance closest to the MODIS dark target algorithm, showing slightly less variance and a negative bias. Both fusion algorithms stabilize diurnal error variations and provide better insights into hourly aerosol evolution. Applying aerosol fusion techniques to future geostationary satellite projects, such as TEMPO and GeoXO, could facilitate the production of high-quality global aerosol data.