Detection of internal waves from oceanic SHH field using neural-network based flow decomposition

Disentangling balanced eddy and unbalanced wave motions from the oceanic Sea Surface Height (SSH) fields, is a complex problem made even more challenging due to their non-linear interactions. The high resolution SSH snapshots from the recently launched Surface Water and Ocean Topography (SWOT) program will include scales down to O(10) km and resolve internal gravity waves (IGWs). At these scales the IGW signatures co-existing with those of the balanced motions pose two major challenges: (i) disentangling IGWs from balanced motions at the ocean surface, a regime with small scale separation and Rossby number near unity, and (ii) detecting IGW signals from two-dimensional snapshots. Here we address these challenges by using state-of-art flow decomposition methods combined with machine learning (ML) for a range of flow regimes.

The currently available flow decomposition methods rely on three-dimensional flow fields and methods to separate these motions from two-dimensional snapshots are currently non-existent. Here we develop a novel method using supervised ML to extract IGWs from snapshots of a flow field and apply it to SWOT SHH field. The initial training and testing is done using Convolution Neural Network algorithms, often used for instance in image classification and pattern recognition problems. The neural network (NN) is trained to detect IGWs in different dynamical regimes based on the decomposition outputs of velocities and model-derived SSH fields from a suite of idealised ocean models outputs of rotating stratified flows with different flow decomposition methods: Higher order decomposition (Eden et al., 2019; Chouksey et al., 2022), Optimal Balance (Masur et al., 2020; Chouksey et al., 2023), and Time Averaged Optimal Balance (Rosenau et al., 2023). The trained NN predicts the flow components from SSH fields generated by the SWOT-simulator and SWOT observations. Analysis using TensorFlow in a shallow water model shows promising results in the prediction of balanced and unbalanced motions by the trained NN. This novel ML-based flow decomposition method is the first of its kind and will provide support for the retrieval of IGW signatures to the SWOT community.