Machine learning connects satellinte measurements to ocean velocity fields
Submesoscale currents, ocean flows that have a spatial scale of tens of kilometers and temporal scale of hours, is an intriguing subject because it is closely related to the ventilation of the ocean while at the same time still underexplored because it is ageostrophic. The availability of submesoscale observation is still quite limited and largely depends on detections from high resolution satellites like SWOT. It remains a challenge to convert high resolution data like sea surface heights from observation to unobserved data like velocities or vorticities. Recently, CAOS PhD Qiyu Xiao and Professor Shafer Smith and their collaborators have published a paper trying to explore the possibility of utilizing neural networks to solve this problem. They found that the performance of neural networks is highly related to the loss function used, as well as the property of the flows. While mean square error works well for flow with little waves, the result can be quite distorted when strong wave activities exist. They also found that transfer learning could be helpful when real data is less accessible. This work is highlighted in EOS and NASA ECCO.