Augmentation of Oceanographic Data with Deep Learning

Satellite observations are limited by sampling rates and only measure the oceans surface, while ocean models are limited by their finite resolution and high momentum viscosity coefficients. Therefore both satellite altimetry data and ocean models lack information at small-scales, with satellite altimetry data also missing sub-surface information. Here we use machine learning to leverage observations and model data by predicting unresolved turbulent processes and sub-surface flow fields. As a proof-of-concept, we train convolutional neural networks on degraded-data from a high-resolution quasi-geostrophic ocean model. We show that convolutional neural networks successfully replicate the spatio-temporal variability of the `missing' eddy momentum forcing, and that the neural networks can generalise to different regions and model configurations. Initially the neural networks do not conserve momentum globally, but we show how to respect conservation laws without a significant reduction in accuracy. By training a new neural network, the sub-surface flow field can be predicted using only information at the surface. Our study shows that convoutional neural networks can exploit existing datasets, even when data is limited to a particular region, to a particular external forcing, or to the surface. Our results show that these data-driven approaches can exploited while respecting physical principles.