Atmospheric-correction-free red tide quantification algorithm for GOCI based on machine learning combined with a radiative transfer simulation

Abstract

Satellite ocean color observation has been an effective way to detect red tide in a wide coastal area of many countries. However, red tide water, which has high red and infra-red reflectance, often causes failure in atmospheric correction, making quantification algorithms based on remote sensing reflectance produce large errors. This study proposes a new framework that tackles the difficulties in red tide quantification stemming from atmospheric variability, limited in-situ training data, and image artifacts, through the combined use of radiative simulation, machine learning, and in-situ measurements. The framework was applied to the geostationary ocean color imager (GOCI) to monitor Margalefidinium blooms that are frequent in Korean coasts. The estimation results were validated first quantitatively with independent ship survey data, and then qualitatively with other 3 red tide algorithms. Finally, the operational robustness of the proposed framework was analyzed based on the data acquired in the entire outbreak period in 2018. The results showed that the proposed algorithm produced a high correlation with field data (R2 ∼ 0.89) and high detection rate and low false alarms compared to the other red tide algorithms. The monitoring result for 2018 also demonstrated that the initiation, expansion, peak, and termination of red tide were successfully identified by the satellite data, which coincides with the field survey results provided by the national fishery agency.