Source-linked AI summary
EddyNet: A Deep Neural Network For Pixel-Wise Classification of Oceanic Eddies
Redouane Lguensat, Miao Sun, Ronan Fablet, Evan Mason, Pierre Tandeo, Ge Chen
TL;DR
Automated eddy detection and classification from SSH maps requires a pixel-wise approach suited to ocean remote sensing. The paper proposes EddyNet, a U-Net-like segmentation network using activation and overlap-loss alternatives, and evaluates its classification performance, including ghost eddies.
Problem
The paper addresses automated pixel-wise detection and classification of oceanic eddies from SSH maps.
Method
EddyNet uses a U-Net-like deep segmentation architecture with SELU or ReLU plus Batch Normalization and Dice-based overlap loss alternatives.
Results
EddyNet assigns ghost-eddy centers to the correct classes 55% of the time for anticyclonic eddies and 45% for cyclonic eddies.
Takeaways & Limitations
The approach transfers deep-learning image-segmentation methods to ocean eddy classification and provides an alternative to conventional detection methods for ghost eddies.
Abstract
from arXiv · showhide
This work presents EddyNet, a deep learning based architecture for automated eddy detection and classification from Sea Surface Height (SSH) maps provided by the Copernicus Marine and Environment Monitoring Service (CMEMS). EddyNet is a U-Net like network that consists of a convolutional encoder-decoder followed by a pixel-wise classification layer. The output is a map with the same size of the input where pixels have the following labels \{'0': Non eddy, '1': anticyclonic eddy, '2': cyclonic eddy\}. We investigate the use of SELU activation function instead of the classical ReLU+BN and we use an overlap based loss function instead of the cross entropy loss. Keras Python code, the training datasets and EddyNet weights files are open-source and freely available on https://github.com/redouanelg/EddyNet.
I. INTRODUCTION
The paper presents EddyNet, a deep neural network for automated eddy detection and classification from SSH maps. It investigates SELU activations, Dice-based loss, and dropout within this image-segmentation setting.
- EddyNet applies deep-learning image-segmentation methods to pixel-wise eddy detection and classification from AVISO-SSH maps.
- The architecture investigates SELU activations as an alternative to ReLU with Batch Normalization.
- Dice-based loss produces better scores for cyclonic and anticyclonic classes than categorical cross-entropy.
- Dropout layers are added to prevent EddyNet from overfitting.
- The authors describe the work as the first deep-learning architecture for pixel-wise eddy classification.
II. PROBLEM STATEMENT AND RELATED WORK
Oceanic eddies transport important physical quantities and influence ocean circulation, motivating automated detection. Existing approaches include physical-parameter, geometric-contour, hybrid, and machine-learning methods, while this work trains on PET14 outputs.
- Oceanic eddies transport carbon, heat, phytoplankton, and salt, making their detection relevant to circulation, weather, mixing, and climate models.
- Physical-parameter methods such as Okubo-Weiss are criticized for expert dependence, regional specificity, and sensitivity to noisy SSH maps.
- Geometric methods model eddies as elliptical shapes using closed contours, while hybrid methods combine physical and geometric ideas.
- EddyNet uses PET14 eddy detection and classification outputs as its training database.
III. DATA PREPARATION
The dataset uses 15 years of daily PET14 detections paired with CMEMS SSH maps, sampled into 128 × 128 Southern Atlantic patches for training and testing.
- The dataset contains 15 years of daily detected and classified eddies from 1998–2012, paired with AVISO-SSH maps from CMEMS.
- Training uses the first 14 years, while 2012 is reserved for testing.
- Memory constraints set the architecture input size to 128 × 128 pixels.
- Randomly sampling one patch per SSH map produces 5100 training samples from the Southern Atlantic region.
- Land and missing-data fill values are replaced by zero to avoid outliers without affecting eddy detection.
A. EddyNet architecture
EddyNet is a three-stage U-Net-like encoder-decoder with skip connections, contrasting SELU activation with ReLU plus Batch Normalization and using transposed convolutions for upsampling.
- The encoder has three stages, each using two 3 × 3 convolutions followed by activation and 2 × 2 max pooling.
- EddyNet compares SELU activation with the classical ReLU plus Batch Normalization configuration.
- The decoder uses transposed convolutions to restore the original resolution.
- Skip connections transfer information from the contracting path to the expanding path.
- The final design uses three stages with 32 filters throughout after the original U-Net overfit the small training set.
B. Loss metric
EddyNet uses a differentiable Dice-based overlap loss for multiclass segmentation, optimizing the mean of one-vs-all soft Dice coefficients across the three classes.
- B. Loss metric: The Dice coefficient measures twice the predicted–ground-truth intersection divided by the sum of their areas.A perfect segmentation scores 1, while a completely mistaken segmentation scores 0.
- B. Loss metric: EddyNet uses a soft, differentiable Dice coefficient that retains softmax probabilities without binarization.The probabilities are compared with one-hot ground-truth class indicators.
- B. Loss metric: The multiclass loss averages three one-vs-all soft Dice coefficients, one for each class.The network minimizes one minus this soft mean Dice coefficient.
A. Assessment of the performance
The evaluation compares activation functions and training losses on held-out SSH patches, finding that Dice loss improves relevant-class and overall Dice scores while ReLU+BN achieves better metrics at longer training cost.
- A. Assessment of the performance: Example segmentation outputs label anticyclonic eddies green, cyclonic eddies brown, and non-eddy regions blue.The examples compare outputs from EddyNet and EddyNet S.
- A. Assessment of the performance: The study compares EddyNet and EddyNet S, along with Dice Loss and categorical cross-entropy, using test-set metrics from 50 random patch sets.Each set contains 360 SSH 120 × 120 maps from 2012.
- A. Assessment of the performance: Dice loss yields higher Dice coefficients for cyclonic and anticyclonic classes in both architectures and a better overall mean Dice coefficient than CCE loss.The comparison uses the original, non-soft mean Dice coefficient.
- A. Assessment of the performance: EddyNet with ReLU+BN obtains better metrics than EddyNet S with SELU, but requires a longer training procedure.Training EddyNet S takes nearly half the time needed for training EddyNet.
B. Ghost eddies
EddyNet is evaluated for identifying ghost eddies, which conventional detection and tracking algorithms may miss between consecutive maps.
- B. Ghost eddies: EddyNet assigns ghost-eddy centers to the correct classes 55% of the time for anticyclonic eddies and 45% for cyclonic eddies.The authors present EddyNet as potentially relevant for detecting ghost eddies missed by conventional methods.
VI. CONCLUSION
EddyNet is presented as a transferable deep-learning tool for oceanic eddy detection and classification, with extensions proposed for temporal, geographic, and multisource applications. The work also emphasizes practical reuse through training, fine-tuning, and integration with existing methods.
- EddyNet applies deep-learning image-segmentation methods to eddy detection and classification from SSH maps.
- Future work includes temporal SSH volumes, a 3D architecture, additional surface information such as Sea Surface Temperature, and global deployment.
- The authors identify omitted postprocessing constraints and eddy tracking as further development opportunities.
- Users can train EddyNet from scratch with their own segmentation ground truth or fine-tune the provided weights on their datasets.
- Figure 5 compares marked ghost-eddy centers in SSH maps with PET14 and EddyNet segmentations using distinct class colors.