Generative Adversarial Network for Reconstructing Diurnal Sea Surface Temperature using Satellite Data over North-west Pacific

Estimating diurnal variations of Sea Surface Temperature (SST) is important for studying air-sea heat exchange. Existing operational diurnal SSTs are derived from numerical models incorporating satellite data, and assimilated with in-situ measurements, which are very accurate. However, numerical model-based methods incur significant computational costs for identifying the diurnal cycle of SST from various heat flux sources (i.e., sensible, latent heat). In this study, we first proposed a Generative Adversarial Network (GAN) method to reconstruct high-resolution diurnal SST using satellite observations as an actual diurnal signal from the ocean surface layer. A generator in the GAN model was trained using the diurnal variability-related variables, including the hourly SSTs and shortwave radiation measurements from Himawari-8 geostationary satellite observations, to estimate diurnal SSTs. The discriminator in the GAN model was learned to reduce the difference in spatiotemporal variability of diurnal SSTs between a satellite data-assimilated numerical model product (Global Ocean OSTIA Diurnal Skin Sea Surface Temperature; Copernicus marine service) and estimated SST from the generator. The results showed that the reconstructed SST had a better spatial distribution of ocean phenomena such as front and eddy than compared with the numerical model-derived SST. It implied that the GAN model could simulate a high spatial variability of SSTs using satellite-based data with a spatial resolution of 2km. The proposed GAN model produced high validation accuracy, resulting in the coefficient of determination of 0.99, bias of -0.2℃, and root mean square errors of 0.58℃ when compared with in situ SST Quality Monitor drifting buoy data. Since we use geostationary satellite data, the proposed model can capture real diurnal variability of SST more frequently than existing numerical model data using analysis data. In addition, the proposed deep learning model is much more computationally efficient than the numerical models.