Transferable Dual-Stream Representations for Mesoscale-Preserving Sea Surface Temperature Downscaling

2026-08-04

AI Summary

Existing sea surface temperature (SST) downscaling methods often fail to preserve mesoscale eddies—features critical to ocean dynamics—resulting in overly smoothed outputs, spectral distortions, and poor generalization. To address this, this work proposes EddyFlow, a novel framework that introduces, for the first time, a physics-informed dual-stream representation learning mechanism to jointly optimize prediction accuracy, structural fidelity, and cross-regional transferability at kilometer-scale resolution. By integrating deep representation learning, multiscale spectral analysis, physics-constrained loss functions, and cross-domain strategies, EddyFlow achieves substantial performance gains under zero-shot and few-shot settings: it reduces zero-shot RMSE by 21% over unseen ocean regions, attains a relative persistence skill of 85.6%, and yields a power spectral density ratio approaching the ideal value of 1.00.

Abstract

Deep learning models for scientific spatio-temporal downscaling often minimize reconstruction error while failing to preserve physically meaningful multi-scale structure. For sea surface temperature prediction, this can yield outputs that are numerically plausible yet overly smooth, missing mesoscale variability critical to regional ocean dynamics. Existing methods often focus on pixel-wise objectives or single-context conditioning, which limits their ability to preserve spectral fidelity and generalize across regions. To address this, we propose EddyFlow, a representation learning framework for kilometer-scale sea surface temperature downscaling that balances predictive accuracy, scale-dependent structure, and regional generalization. EddyFlow is trained on the Gulf of St.~Lawrence and evaluated in zero-shot and few-shot settings on the Bay of Fundy and the Gulf of Mexico. EddyFlow demonstrates that physics-informed representation learning reduces zero-shot RMSE by 21%, achieves up to 85.6% skill relative to persistence on unseen domains, and maintains near-ideal spectral fidelity with a PSD ratio of $\approx 1.00$.

Problem

• sea surface temperature downscaling
• mesoscale variability
• spectral fidelity
• regional generalization
• spatio-temporal downscaling

Innovation

• dual-stream representations
• mesoscale-preserving downscaling
• physics-informed representation learning
• zero-shot generalization
• spectral fidelity