Deep learning-based eddy detection

==============================================================

Using 2D deep neural networks to segment and visualize 3D features  in scientific simulations

Volumetric (3D) datasets are commonly generated by scientific studies and analyzed by the scientific community. Identifying coherent features in volumetric datasets, such as eddies in ocean simulations, is a key task that can be achieved using segmentation. A classical volumetric data segmentation and visualization pipeline typically segments features from the dataset first and then visualizes those segmented features using techniques such as volume rendering. Many recent studies focus on employing deep learning-based solutions for analyzing volumetric simulations. These approaches generally use 3D datasets to train deep learning-based 3D segmentation networks. Unfortunately, training neural networks on 3D datasets is more demanding, requiring both annotated 3D datasets and substantial computing resources compared to training on 2D datasets. In this paper, we propose an alternative to classical 3D visualization pipelines. Specifically, our proposed pipeline visualizes eddies in 3D ocean simulations by applying volume rendering first, then using 2D segmentation networks trained on 2D annotations. Recognizing that volume rendering functions similarly to feature embedding, we introduce a more resource-efficient pipeline wherein we first map the 3D data into a 2D embedding space via volume rendering before training our deep learning-based segmentation algorithm in this 2D space. Then, we overlay the segmentation results onto the 2D-rendered volumetric data visualization. We demonstrate the effectiveness of our proposed eddy visualization pipeline across four different ocean simulations. Our experiments show that the proposed pipeline achieves competitive performance while significantly reducing training time compared to the classical 3D visualization pipeline. This approach is particularly beneficial for scientists using visualization systems with limited computational resources.

==============================================================

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

Eddy structures play important roles in nutrient transition in the biological ecosystem of oceans. Scientists look for accurate ways to identify and analyze eddy-like structures in ocean simulations as an important step to understanding how the nutrient transition affects the biological ecosystem. In eddy simulations, the velocity components are key attributes to identify and characterize an eddy structure. Training a deep network using such velocity vector-based inputs raises challenges regarding effectively handling this data for eddy segmentation. In this paper, we address these challenges by focusing on cases where the input consists of velocity data, represented as a vector with components along the x and y directions in 2D space. The vector-based input can be characterized by its magnitude and direction (angle) or individual axis-based components. However, the optimal way to provide this data to a deep segmentation network for training remains unclear in scientific simulations. Thus, we explore various input representations for eddy segmentation with deep learning and report our findings in this paper.