Derivation and Evaluation of Satellite-Based Surface Current

Observations of real-time ocean surface currents allow one to search and rescue at ocean disaster sites and investigate the surface transport and fate of ocean contaminants. Although real-time surface currents have been mapped by high-frequency (HF) radar, shipboard instruments, satellite altimetry, and surface drifters, geostationary satellites have proved their capability in satisfying both basin-scale coverage and high spatiotemporal resolutions not offered by other observational platforms. In this paper, we suggest a strategy for the production of operational surface currents using geostationary satellite data, the particle image velocimetry (PIV) method, and deep learning-based evaluation. We used the model scalar field and its gradient to calculate the corresponding surface current via PIV, and we estimated the error between the true velocity field and calculated velocity field by the combined magnitude and relevance index (CMRI) error. We used the model datasets to train a convolutional neural network, which can be used to filter out bad vectors in the surface current produced by arbitrary model scalar fields. We also applied the pretrained network to the surface current generated from real-time Himawari-8 skin sea surface temperature (SST) data. The results showed that the deep learning network successfully filtered out bad vectors in a surface current when it was applied to model SST and created stronger dynamic features when the network was applied to Himawari SST. This strategy can help to provide a quality flag in satellite data to inform data users about the reliability of PIV-derived surface currents.