Abstract
Mesoscale eddies play a significant role in marine energy and matter transportation. Due to their huge impact on the ocean, mesoscale eddy detection has been studied for many years. However, existing methods mainly use single-modal data, such as the sea surface height (SSH), to detect mesoscale eddies, resulting in inaccurate detection results. In this paper, we propose an end-to-end mesoscale eddy detection method based upon multi-modal data fusion. Particularly, we don’t only use SSH, but also add data of other two modals, i.e., the sea surface temperature (SST) and the velocity of flow, which are closely related to mesoscale eddy detection. Moreover, we design a novel network named SymmetricNet, which is able to achieve multi-modal data fusion in mesoscale eddy detection. The proposed SymmetricNet mainly contains a downsampling pathway and an upsampling pathway, where the low-level feature maps from the downsampling pathway and the high-level feature maps from the upsampling pathway are merged through lateral connections. In addition, we apply dilated convolutions to the network structure to increase the receptive field without sacrificing resolution. Experiments on multi-modal mesoscale eddy dataset demonstrate the advantages of the proposed method over previous approaches for mesoscale eddy detection.
Introduction
Mesoscale eddies are the "weather systems" of the ocean—massive swirling vortices (100–300 km wide) that transport heat, nutrients, and salt across the globe. Detecting them accurately is vital for climate modeling and marine biology. However, traditional detection methods and early AI models have a major "blind spot": they rely almost exclusively on Sea Surface Height (SSH) data. Because eddies are physically complex and have irregular shapes, looking at SSH alone often leads to "fake" detections or missed eddies. To truly "see" an eddy, a model needs a multi-modal view that includes Sea Surface Temperature (SST) and flow velocity, which provide critical context that height data lacks.
Methods: The SymmetricNet Architecture
The authors propose SymmetricNet, an end-to-end deep learning framework designed for semantic segmentation. The model labels every pixel in an image as "cyclonic eddy," "anti-cyclonic eddy," or "background."
Multi-modal Fusion and Symmetric Design
The model processes a four-channel input (SSH, SST, zonal velocity, and meridional velocity) simultaneously, allowing it to learn correlations between different physical properties. SymmetricNet features a symmetric encoder-decoder architecture, consisting of a downsampling pathway to capture high-level semantics and an upsampling pathway to restore spatial resolution.
Lateral Connections and Dilated Convolutions
Instead of simple skip connections, the model uses lateral connections (LC) to merge low-level, high-resolution feature maps from the downsampling pathway with high-level semantic maps from the upsampling pathway. This ensures that fine-grained spatial details are preserved during the reconstruction phase.
Furthermore, dilated convolutions are applied within the network structure. These allow the model to expand its "receptive field"—understanding larger spatial contexts—without sacrificing the resolution of the output. This is particularly important for identifying the large-scale structure of eddies while maintaining sharp boundaries.
Optimization
The model optimizes training using a hybrid loss function that combines Dice Loss (to handle class imbalance where eddies are small relative to the ocean background) and Cross-Entropy Loss for stability.
Dataset and Experimental Setup
The researchers constructed a custom multi-modal dataset using data from the Copernicus Marine Environment Monitoring Service (CMEMS):
- Timeframe: 10 years of data (2000–2009).
- Resolution: 0.25-degree spatial resolution.
- Ground Truth: Experts manually labeled SSH images to create a reliable benchmark for training and testing.
- Scope: The final dataset consisted of 512 training samples and 256 test samples, formatted as 128x128 pixel patches.
Key Results
SymmetricNet demonstrated clear superiority over existing models like EddyNet and PSPNet:
- Global Accuracy: Achieved 97.06% on the multi-modal test set.
- Precision by Category: The model reached 91.51% precision for cyclonic eddies and 87.85% for anti-cyclonic eddies.
- Qualitative Success: Visual comparisons showed that SymmetricNet’s detections were the closest to expert-labeled ground truth, successfully ignoring the "noise" that caused other models to fail.
Conclusion and Future Work
The study confirms that while SSH is the most important single indicator, it is insufficient for robust detection. SST helps avoid missed eddies, while flow velocity acts as a filter to remove false positives. While SymmetricNet is a significant step forward, the authors identify several areas for future improvement:
- Labeling Bias: Future work could involve creating "fused" labels that incorporate temperature and flow data from the start, rather than relying solely on SSH-based labels.
- Dataset Expansion: Increasing the dataset size beyond the current 1,000 samples would improve generalization across different sea areas.
- Temporal Tracking: Moving from static snapshots to trajectory tracking would allow for following an eddy's entire lifecycle from formation to dissipation.
Summary updated