Simulation-informed deep learning for enhanced SWOT observations of fine-scale ocean dynamics

TL;DR

This work tackles the difficulty of extracting fine-scale ocean dynamics from SWOT SSH data corrupted by KaRIn noise. It introduces SIMPGEN, an unsupervised adversarial framework that leverages simulated reference states to train a neural metric $M_\Phi$ and a neural inversion operator $G_\theta$, guiding reconstructions toward physically plausible fields without ground-truth labels. The method replaces a Gaussian prior with a neural latent-space prior $J_{prior}(x) = E_{x_s \in \mathcal{S}} ( M_\Phi(x) - M_\Phi(x_s) )^2$ and uses wavelet-based multi-scale analysis to capture scale- and direction-specific energy, while training with randomized observation error $R$ to reflect KaRIn uncertainty. Evaluations on synthetic and real SWOT data show SIMPGEN more faithfully preserves submesoscale features and spectral energy, yielding better RMSE, coherence, and temporal stability than supervised baselines and state-of-the-art CLS denoising, with potential for broader ocean state estimation and data-assimilation integration.

Abstract

Oceanic processes at fine scales are crucial yet difficult to observe accurately due to limitations in satellite and in-situ measurements. The Surface Water and Ocean Topography (SWOT) mission provides high-resolution Sea Surface Height (SSH) data, though noise patterns often obscure fine scale structures. Current methods struggle with noisy data or require extensive supervised training, limiting their effectiveness on real-world observations. We introduce SIMPGEN (Simulation-Informed Metric and Prior for Generative Ensemble Networks), an unsupervised adversarial learning framework combining real SWOT observations with simulated reference data. SIMPGEN leverages wavelet-informed neural metrics to distinguish noisy from clean fields, guiding realistic SSH reconstructions. Applied to SWOT data, SIMPGEN effectively removes noise, preserving fine-scale features better than existing neural methods. This robust, unsupervised approach not only improves SWOT SSH data interpretation but also demonstrates strong potential for broader oceanographic applications, including data assimilation and super-resolution.

Figures (7)

• Figure 1: Schematic of the proposed SIMPGEN architecture.
• Figure 2: (A–D) Example snapshot of SSH fields: (A) noiseless ground truth, (B) noisy field, (C) supervised U-Net denoised field, and (D) SIMPGEN denoised field. (E–L) Polar plots of the average wavelet coefficients normalized by the scale energy for two cases: across-track dominated snapshots (right column) and along-track dominated snapshots (left column). Panels (E,F) represent noiseless fields, (G, H) noisy fields, (I, J) supervised U-Net denoised fields, and (K, L) SIMPGEN denoised fields. (M) RMSE as a function of spatial scale, where the neural inversion operator outperforms the supervised model, particularly at smaller scales. (N) Power spectral density comparison across wavelengths.
• Figure 3: The figure presents the global (G1–G6) and regional analyses (Gulf Stream: A1–A7, Mediterranean Sea: B1–B7, Southern Ocean: C1–C7) in a structured panel layout. The first panel of the global analysis (G1) highlights the SWOT tracks and regions of interest, examples of SSH snapshots are provided in the regional rows. We use a single colorbar, but each region has its own numeric range, indicated by color coded ticks corresponding to the region box. For all analyses, the subsequent panels compare noisy fields (A1, B1, C1) with SIMPGEN denoised fields (A2, B2, C2) through the following elements: example snapshots of SSH fields, relative directional energy derived from wavelet coefficients normalized by the maximum scale energy (panels 3,4,5 and 6), and averaged across-track power spectral density (7). The wavelet energy analysis includes two datasets: along-track-dominated averages in panels 3 and 4, and across-track-dominated averages in panels 5 and 6. The panels 7 display the averaged power spectral density, where blue (red) curves represent the original (denoised) results, and shaded areas indicate 24-hour average variability for each scale. Dashed lines show constant spectral slopes, serving as references for typical dynamic regimes.
• Figure 4: Applications of directional wavelet analysis on synthetic plane-wave fields and anisotropic power-law correlated fields. The first column shows synthetic plane-wave fields at various orientations and wavelengths and the second column presents the corresponding wavelet coefficients, where the radial and angular coordinates represent scale and orientation, respectively. The third column displays power-law fields with an energy spectrum following $k^{-11/3}$, and the fourth column shows the wavelet analysis of these fields, revealing isotropic and anisotropic structures across scales.
• Figure 5: Panels A displays violin plot of the neural metric values for the different dataset. Panel B shows a polar plot of the neural metric sparse autoencoder activations, with positive (red) and negative (blue) values distributed across spatial directions and scales.