The Surface Water and Ocean Topography (SWOT) mission provides unprecedented high-resolution observations of sea surface height (SSH), offering new opportunities to study mesoscale and submesoscale ocean dynamics. A key challenge, however, is that SSH contains signals from both eddy motions and internal waves. These signals become increasingly difficult to distinguish at smaller scales, where their spatial and temporal scales strongly overlap.
Here we develop a deep-learning framework to investigate the scale-dependent separability of eddy and wave SSH signals. Using a high-resolution, tide-resolving ocean simulation, we first construct reference wave and non-wave components through Lagrangian eddy–wave separation. A three-dimensional U-Net is then trained to predict the separated SSH components from different input configurations, including a single SSH snapshot, multi-hour SSH sequences, and SSH combined with sea surface temperature (SST).
Our results show that single-snapshot SSH can recover large-scale and mesoscale non-wave signals with high skill, but its performance decreases rapidly toward submesoscales. Adding temporal SSH information substantially improves the separation at smaller scales, suggesting that time evolution provides essential constraints for distinguishing propagating waves from eddy and frontal motions. We further find that SST provides complementary information that enhances submesoscale separation. These results highlight how AI-enabled SWOT SSH and SST fusion can help disentangle submesoscale eddies, fronts, and internal waves, opening new opportunities to study multiscale ocean dynamics in the high-resolution satellite era.