Deep Learning Improves Reconstruction of Ocean Vertical Velocity

Abstract

AbstractOcean vertical velocity (w) plays a key role in regulating the exchanges of mass, heat and nutrients between the surface and deep ocean. However, direct observation remains difficult due to its small magnitude and large spatiotemporal variability. Therefore, w fields are generally diagnosed using dynamic‐based methods. In this study, we developed a deep neural network (DNN) to reconstruct three‐dimensional fields of ocean vertical velocity based on sea surface height (SSH) fields. Compared to dynamic‐based methods, the DNN shows improved performance in the w reconstruction within upper 500 m in terms of higher correlation and less error. Remarkably, the DNN requires only a ∼45 × 45 km size SSH image as input to estimate w at the center. This suggests that the DNN has great potential for w reconstruction in the future combined with high‐resolution observations such as the Surface Water and Ocean Topography mission.