Deep Learning-Based Eddy Segmentation With Vector Data for Biochemical Analysis in Ocean Simulations

Eddies are dynamic, swirling structures in ocean circulation that significantly influence the distribution of heat, nutrients, and plankton, thereby impacting marine biological processes. Accurate eddy segmentation from ocean simulation data is essential for enabling subsequent biological and physical analyses. However, leveraging vector-valued inputs, such as ocean velocity fields, in deep learning-based segmentation models poses unique challenges due to the complexity of representing the vector input in multiple combinations for training. In this article, we discuss such challenges and provide our solutions. In particular, we present a detailed study into multiple input encoding strategies, including raw velocity components, vector magnitude, and angular direction, and their impact on eddy segmentation performance. We introduce a two-branch attention U-Net architecture that separately encodes vector magnitude and direction. We evaluate seven different network configurations across four large-scale 3-D ocean simulation datasets, employing four different segmentation metrics. Our results demonstrate that the proposed two-branch architecture consistently outperforms single-branch variants.