Vertical Velocity Diagnosed From Surface Data With Machine Learning

Submesoscale vertical velocities, w, are important for the oceanic transport of heat and biogeochemical properties, but observing w is challenging. New remote sensing technologies of horizontal surface velocity at O(1) km resolution can resolve surface submesoscale dynamics and offer promise for diagnosing w subsurface. Using machine learning models, we examine relationships between the three-dimensional w field and remotely observable surface variables such as horizontal velocity, density, and their horizontal gradients. We evaluate the machine learning models' sensitivities to different inputs, spatial resolution of surface fields, the addition of noise, and information about the subsurface density. We find that surface data is sufficient for reconstructing w, and having high resolution horizontal velocities with minimal errors is crucial for accurate w predictions. This highlights the importance of finer scale surface velocity measurements and suggests that data-driven methods may be effective tools for linking surface observations with vertical velocity and transport subsurface. Vertical velocities, w, are associated with ocean currents that move toward or away from the ocean surface and are important for connecting the surface and deep ocean. It is extremely difficult to measure w directly, but measurements of other variables that are related to w, such as horizontal currents that move along the surface in the north-south or east-west directions, can be exploited to predict w. Here, we investigate the feasibility of inferring w from other more easily measurable data. We compare three machine learning methods to see which is best at finding relationships between more easily measurable variables (the input data) and w at different depths. We test how using different input variables, adding noise to the input data, or changing the spatial resolution of the input data, impact the w predictions. Our results show that machine learning models are successful at reconstructing the 3D w field using high-resolution (similar to 1 km) surface data, and in particular, surface horizontal velocities are the most important to include. This study shows that data-driven methods are promising for relating remotely sensed surface measurements of the ocean to vertical velocities below the surface, which can help provide us with a better understanding of the 3D ocean. Machine learning models diagnose the oceanic 3D submesoscale vertical velocity field within and below the mixed layer from surface data High resolution horizontal surface velocity data is crucial for obtaining accurate vertical velocities subsurface Convolutional Neural Networks are relatively robust to noisy data and coarser data resolution