Surface-to-volume physics-informed model for 3D mesoscale eddy detection: Multi-source multi-scale fusion

Abstract

Mesoscale eddies play a crucial role in the ocean, and accurate detection of their three-dimensional (3D) structures is essential for understanding marine dynamics. However, 3D eddy reconstruction remains challenging: traditional methods provide limited pixel-level boundary delineation, while existing deep learning approaches suffer from insufficient physical constraints and unrealistic eddy geometries. To address this issue, we propose a surface-to-volume physics-informed model (StV-PI-EddyNet) for extracting 3D mesoscale eddies. We construct a high-quality 3D eddy dataset covering the North Pacific (100°E–100°W, 5°N–65°N) down to 1000 m for model training. The model takes sea surface height (SSH), sea surface temperature (SST), and surface current components (U/V) as inputs and consists of three core modules: 2D feature extraction, 3D feature inversion, and 3D semantic segmentation. It embeds a vorticity prediction branch for physical constraints and uses U-Net-like architectures with spatial and channel attention mechanisms. The loss function combines weighted Dice loss and a vorticity-driven physical loss (mean square error + gradient consistency loss) to ensure high segmentation accuracy and physical consistency. Experimental results show StV-PI-EddyNet model achieves 96.97% global accuracy, 84.80% weighted Dice coefficient, and 0.8955 macro-averaged F1-score. It effectively mitigates boundary distortion issues of existing methods, generating 3D eddy masks with high spatial precision. This model enjoys good scalability and can be further optimized by integrating earth observation advances in architecture design, data sources and application generalization, promoting its in-depth applications in marine and especially underwater dynamic research.