A Neural Network-Based Atmospheric Correction Algorithm for GOCI Imagery Over Coastal Waters

The geostationary ocean color imager (GOCI) has provided eight observations per day since 2010 and has been widely used in coastal dynamics of bio-optical parameters. However, accurate atmospheric correction (AC) of GOCI data over coastal waters is still a challenge, hindering the quantitative retrieval of biogeochemical parameters. Here, we proposed a new AC method for coastal waters based on a neural network (denoted NN_3S). The NN_3S algorithm was designed to derive remote sensing reflectance ( R-rs ) from Rayleigh-corrected reflectance ( R-rc ) and the training of NN_3S used 0.85 million pairs of high-quality R-rs - R-rc for 2019 generated by the near-infrared (NIR) iterative algorithm (NIR_AC) in SeaDAS. The performance of NN_3S was evaluated with ground measurements from three aerosol robotic network-ocean color (AERONET-OC) stations. The results showed a notable reduction in the band-averaged mean absolute percentage difference (MAPD) for the 412-, 443-, 490-, 555-, and 667-nm R-rs retrievals when utilizing NN_3S, with decreases of 17.4%, 32.2%, and 16.59% observed in comparison to retrievals by NIR_AC, the Korea Ocean Satellite Center AC algorithm (KOSC) in GOCI data processing system version 2.0 (GDPS 2.0), and the ocean color-simultaneous marine and aerosol retrieval tool (OC-SMART), respectively. More importantly, the practical application of the NN_3S algorithm indicated successful retrievals over turbid waters. Furthermore, the daily percentage of valid observations (DPVOs) for NN_3S compared to NIR_AC and KOSC increased by 1.97% and 12.82% in June and by 4% and 10.83% in December, respectively. Smoother spatial patterns than NIR_AC were also found. These results indicate that NN_3S can be a reliable AC option for GOCI over coastal areas.