Exploring the plausibility and limitations of SSH interpolation with deep learning

Georgy Manucharyan - (University Of Washington, Seattle)

Co-Investigator(s):

Patrice Klein (California Institute of Technology)

Abstract:

Background. To obtain a more accurate estimate of sea surface height (SSH) it is imperative to extract as much information as possible from existing along-track altimetry data. A significant challenge here is that the altimetry data is inherently sparse and requires interpolation onto a regular grid in order to explore the mesoscale dynamics. Commonly used objective interpolation techniques for SSH are fast but do not conform to any dynamical models. Therefore such data products can result in the unphysical evolution of SSH anomalies and in unnecessarily-low spatial resolution because of their sub-optimal utilization of information contained in along-track data. This proposal revolves around a hypothesis that the quality of SSH interpolation could be improved if an accurate dynamical model predicting SSH evolution could be constructed. The recent research led by the PI demonstrated that in idealized models of eddying baroclinically unstable currents, the mesoscale variability exists in a dynamical state for which subsurface velocities can be approximately expressed in terms of SSH snapshots only via a highly non-linear mapping function - a deep convolutional neural network. It is the existence of this mapping that has been directly linked to improved SSH interpolation skills in idealized studies of mesoscale turbulence by the PI. This project will develop a new technique for SSH interpolation, assess the SSH predictability on mesoscale eddy timescales, and classify the dynamical regions in the ocean for which the non-linear dependencies withing the turbulent eddy field can be extracted with the goal of improving SSH interpolation and predictability.

Major Goals. The central objective of the proposal is to assess the plausibility and understand the dynamical limitations of a Deep Learning approach for SSH interpolation. The interpolation technique will be developed and its efficacy tested within the context of NASA's comprehensive eddy-resolving ocean models and a set of along-track altimetry data. Since the issue of SSH interpolation is directly linked to its predictability, as assessment of the utility of deep neural networks for short-term SSH predictability will be made. The connection will be made between the skill of SSH interpolation/predictability and the characteristics of mesoscale turbulence.

Methods and Datasets. Our key techniques will utilized Deep Neural Networks, including the Long-Short Term Memory Cells for analysis of temporally sparse along-track SSH information, Convolutional Networks for pattern analysis of SSH data and its dimensionality reduction, as well as deep Recurrent Neural Networks.The deep learning models will be trained and tested using a hierarchy of eddy resolving models ranging from idealized quasigeostrophic models of ocean mesoscale dynamics to comprehensive models with a key focus on the use of NASA's LLC4320/LLC2160 simulations. Transfer learning, a technique that allows to fine-tune the pre-trained neural networks on a much smaller number of real data samples will be used to interpolate the altimetry tracks.

Significance to the OSTST mission objectives. Our proposed research will contribute mainly to the priority Theme 1 as it is an inherently a physical oceanography study that attempts to reveal the highly-nonlinear connection between the nature of mesoscale turbulence and issue of SSH interpolation and prediction. We will also partially address Themes 2 as our study has a focus on strongly eddying regions of the ocean and Theme 3 as we propose to develop a new technique for interpolating along-track altimetry data onto a uniform grid.

Supported by NASA