Estimating Ocean Surface Currents With Machine Learning

Global surface currents are usually inferred from directly observed quantities like sea-surface height, wind stress by applying diagnostic balance relations (like geostrophy and Ekman flow), which provide a good approximation of the dynamics slow, large-scale at large scales low Rossby numbers. However, newer generation satellite altimeters upcoming SWOT mission) will capture more high wavenumber variability associated with unbalanced components, but temporal sampling can potentially lead to aliasing. Applying these balances may an incorrect un-physical estimate flow. In this study we explore Machine Learning (ML) algorithms as alternate route infer observable quantities. We train our ML models SSH, SST, available primitive equation ocean GCM simulation outputs inputs make predictions (u,v), then compared against true output. As baseline example, demonstrate that linear regression model is ineffective predicting velocities accurately beyond localized regions. comparison, relatively simple neural network (NN) predict over most global ocean, lower mean squared errors than + Ekman. Using local stencil neighboring grid points additional input features, deep learning effectively “learn” spatial gradients physics currents. By passing stenciled variables through convolutional filters help learn much faster. Various training strategies explored using systematic feature hold out multiple combinations point data fed (2D/3D), understand effect each on NN's ability represent A sensitivity analysis reveals besides geographic information in some form essential ingredient required for making accurate models.