Deep Learning-based Eddy Segmentation with Vector-Data for Biochemical Analysis in Ocean Simulations.

Area of Science:

• Oceanography
• Marine Biology
• Computer Science

Background:

• Ocean eddies are vital for heat, nutrient, and plankton distribution, impacting marine ecosystems.
• Accurate segmentation of eddies from ocean simulation data is critical for physical and biological analyses.
• Deep learning models face challenges in processing vector-valued ocean velocity data for segmentation.

Purpose of the Study:

• To address challenges in deep learning-based eddy segmentation using vector-valued ocean velocity fields.
• To investigate the impact of different input encoding strategies on segmentation performance.
• To propose and evaluate an improved deep learning architecture for eddy segmentation.

Main Methods:

• Studied multiple input encoding strategies: raw velocity components, vector magnitude, and angular direction.
• Introduced a two-branch attention U-Net architecture to separately encode vector magnitude and direction.
• Evaluated seven network configurations on four large-scale 3D ocean simulation datasets using four segmentation metrics.

Main Results:

• The proposed two-branch attention U-Net architecture demonstrated superior performance compared to single-branch variants.
• Different input encoding strategies significantly impacted eddy segmentation accuracy.
• The architecture effectively captured complex vector-valued oceanographic data for segmentation.

Conclusions:

• The developed two-branch U-Net architecture offers a robust solution for deep learning-based ocean eddy segmentation.
• Effective encoding of vector-valued data is key to improving segmentation accuracy in oceanographic studies.
• This work advances the capability for analyzing dynamic ocean structures and their biological impacts.