Atmospheric correction under cloud edge effects for Geostationary Ocean Color Imager through deep learning

The Geostationary Ocean Color Imager (GOCI) is widely employed in tracking diurnal dynamics of oceanic conditions. However, the current atmospheric correction (AC) algorithms for GOCI often mask pixels around cloud edges to exclude pixels contaminated by cloud edge effects (CEEs; including stray light, cloud shadows, and cloud adjacent effects (AEs)), which results in massive data loss. In this paper, we propose a novel AC algorithm to correct these CEE-affected pixels based on deep learning (namely, DLACC) to achieve a comparable quality level to those pixels far away from cloud edges. We used the standard near-infrared iterative AC algorithm in SeaDAS (namely, NIR) to obtain Rayleigh-corrected reflectance (Rrc) and remote sensing reflectance (Rrs). Then, high-quality Rrs values were extracted using a quality assurance (QA) algorithm. Next, we obtained 2,821,668 matchups by matching CEE-free noontime Rrs values with CEE-affected Rrc values in a time window of 1 h. These matchups were used to develop DLACC. Validations using in situ data from Aerosol Robotic Network-Ocean Color (AERONET-OC) stations showed that more matchups were obtained with the DLACC model than with the NIR algorithm and Korea Ocean Satellite Center standard AC algorithm in GDPS 2.0 (KOSC), and the accuracies were similar. More importantly, the DLACC algorithm is more tolerant of AEs and reduces AEs in a 2-pixel distance from cloud edges by 10% in the blue bands and by 50% in the green band. As a result, more valid observations are obtained with the DLACC algorithm, with the daily percentage of valid observations (DPVOs) increasing by 71% and 62% over those with the NIR algorithm and the KOSC algorithm, respectively. These increases in valid observations could further result in more consistent patterns in terms of space and time. Our DLACC algorithm can not only be used to process GOCI images but also provide an open framework to develop corresponding deeplearning models for other geostationary satellites.