Abstract's details
From SWOT Swaths to Global Ocean: A Machine Learning Proof of Concept for Estimating Incoherent Tidal Variability
Event: 2025 SWOT Science Team Meeting
Session: Oceanography: Inversion/Assimilation
Presentation type: Poster
The unprecedented high-resolution, two-dimensional mapping of Sea Surface Height (SSH) provided by SWOT comes with trade-offs: spatial gaps between swaths during the 1-day repeat Cal/Val phase, and temporal gaps between revisits during the 21-day repeat Science phase. One of the key challenges for SWOT-based studies is therefore to propagate observed structures in space and time to produce complete global maps. In this study, we investigate the potential of machine learning techniques to extrapolate information collected during the SWOT Cal/Val phase—covering roughly 9% of the ocean—to the global ocean.
Incoherent internal tides emerge from the interaction of coherent baroclinic tides with mesoscale eddies. Unlike coherent tides, which have predictable amplitudes and phases at fixed locations, incoherent tides are temporally and spatially variable and therefore characterized by unpredictability. The high temporal sampling of the SWOT Cal/Val phase has provided valuable insights into both coherent and incoherent tidal signals, in regions such as the Amazon Shelf and New Caledonia. In this study, we explore the potential to extrapolate these signals globally.
Coherent and incoherent tidal variances have been estimated from ORCA12 model simulations with tidal forcing. We present a proof of concept to assess the ability of machine learning algorithms to predict incoherent tides variance beyond SWOT swaths. Assuming that coherent internal tide variance is globally known and that the variance from incoherent tides has been estimated along SWOT tracks, we test whether machine learning models can reconstruct these signals elsewhere. We evaluate several regression models—including linear regression, support vector regression (SVR), random forests, neural networks, XGBoost, and k-nearest neighbors—and combinations of input features such as bathymetry and mesoscale variability, to identify the key drivers of accurate extrapolation.
Back to the list of abstractIncoherent internal tides emerge from the interaction of coherent baroclinic tides with mesoscale eddies. Unlike coherent tides, which have predictable amplitudes and phases at fixed locations, incoherent tides are temporally and spatially variable and therefore characterized by unpredictability. The high temporal sampling of the SWOT Cal/Val phase has provided valuable insights into both coherent and incoherent tidal signals, in regions such as the Amazon Shelf and New Caledonia. In this study, we explore the potential to extrapolate these signals globally.
Coherent and incoherent tidal variances have been estimated from ORCA12 model simulations with tidal forcing. We present a proof of concept to assess the ability of machine learning algorithms to predict incoherent tides variance beyond SWOT swaths. Assuming that coherent internal tide variance is globally known and that the variance from incoherent tides has been estimated along SWOT tracks, we test whether machine learning models can reconstruct these signals elsewhere. We evaluate several regression models—including linear regression, support vector regression (SVR), random forests, neural networks, XGBoost, and k-nearest neighbors—and combinations of input features such as bathymetry and mesoscale variability, to identify the key drivers of accurate extrapolation.