Estimation of Sea Surface Salinity Around the Korean Peninsula Using Machine Learing

Salinity is one of the most important indicators of ocean circulation and affects the marine environment. Sea Surface Salinity (SSS) is highly related to various ocean-atmosphere phenomena and thus, the monitoring of SSS is crucial to investigate regional/global ocean environment and climate change. Field surveys, which are often used for SSS observation, are time-consuming and expensive, and do not cover vast areas with spatial continuity. On the other hand, satellite data (e.g., passive microwave) or numerical models can be used to quantify SSS with spatially continuous coverage. However, they have a relatively coarse resolution and often high regional uncertainty. In particular, existing satellite and model-derived SSS around the Korean Peninsula focusing on coastal areas has spatiotemporally varied uncertainty, which requires more regional SSS estimation models. The purpose of this study is to estimate SSS around the Korean Peninsula with higher resolution than existing satellite-derived SSS products using multi-sensor data fusion based on machine learning approaches such as random forest and neural network. In this study, Geostationary Ocean Color Imager (GOCI) and passive microwave satellites and Hybrid Coordinate Ocean Model (HYCOM) reanalysis data were used as main input data in this study. GOCI is the world first geostationary ocean color observation sensor, and it collects 8 images hourly per day at 500 m resolution. The reflectance data and basic products of GOCI with 500m resolution were used to improve the spatial resolution of SSS especially focusing on coastal regions. The results showed that SSS estimated using the proposed approach yielded a higher accuracy than the existing SSS data. This advanced model is expected to more accurately monitor SSS around the Korean peninsula.