EddyNet: A Deep Neural Network For Pixel-Wise Classification of Oceanic Eddies (1711.03954v1)

Abstract: 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.

• The paper presents EddyNet, a U-Net inspired network for pixel-wise classification of oceanic eddies in Sea Surface Height data.
• It leverages SELU activation and a Dice coefficient loss to expedite training and improve segmentation accuracy over traditional methods.
• Performance tests on Southern Atlantic SSH maps show promising detection of ghost eddies, supporting further advances in oceanographic research.

EddyNet: A Deep Neural Network for Pixel-Wise Classification of Oceanic Eddies

The paper "EddyNet: A Deep Neural Network For Pixel-Wise Classification of Oceanic Eddies" introduces EddyNet, a deep learning architecture tailored for the classification of oceanic eddies. This work is pertinent to researchers focused on the application of machine learning techniques to oceanography, specifically for detecting and categorizing eddies in Sea Surface Height (SSH) data obtained from the Copernicus Marine and Environment Monitoring Service (CMEMS).

EddyNet employs a convolutional encoder-decoder architecture that delivers pixel-wise classifications into three categories: non-eddy, anticyclonic eddies, and cyclonic eddies. The choice of network architecture takes inspiration primarily from U-shaped networks, notably the U-Net model, which is renowned for its efficacy in image segmentation tasks. This paper explores the impact of advanced activation functions, such as Scaled Exponential Linear Units (SELU), in enhancing the training process over the traditional ReLU and Batch Normalization (BN) configuration.

Key Contributions and Methodology

1. Data Preparation: The research utilizes Sea Surface Height maps from 1998 to 2012, focusing on the Southern Atlantic Ocean. For every SSH map, eddy segmentation masks are created using an existing detection algorithm, thereby generating a robust dataset for training EddyNet.
2. Network Architecture: EddyNet features three downsampling stages in its encoding path, with each stage comprised of two convolutional layers augmented by activation functions. In this context, the paper contrasts the use of SELU, which optimizes training duration, with the classical ReLU + BN approach. The inclusion of dropout layers, situated prior to maximum pooling and transposed convolution layers, is instrumental in mitigating overfitting phenomena.
3. Loss Function: Instead of conventional categorical cross-entropy, EddyNet utilizes a loss function based on the Dice coefficient, also known as the F1 measure. This choice is substantiated by improved segmentation performance metrics for cyclonic and anticyclonic eddy categories.
4. Performance Evaluation: The evaluation traces EddyNet's efficacy in comparison to conventional detection methodologies by conducting thorough experiments across SSH maps. The resulting analysis reveals that SELU markedly accelerates training phases while maintaining comparable classification performance to ReLU + BN setups.
5. Addressing Ghost Eddies: The paper discusses EddyNet's potential in detecting ghost eddies, which are typically elusive to traditional methods. Here, EddyNet demonstrates notable success in identifying ghost eddies' centers, establishing itself as a complementary tool alongside traditional techniques.

Implications and Future Research

EddyNet constitutes a significant stride in applying deep learning to oceanographic data analysis by automating eddy identification and enhancing classification preciseness. Its implementation condenses the complexities associated with conventional methods reliant on region-specific parameters or contour techniques, offering a generalized approach adaptable to various datasets.

The authors envision future advancements encompassing the integration of multi-temporal SSH data to harness temporal features, proposing the development of a 3D extension of the current 2D convolutional framework. Additionally, the fusion of diverse surface indicators such as Sea Surface Temperature alongside SSH may pioneer deeper insights into oceanic eddy dynamics. The simplification for global applicability and further refinement of post-processing techniques stand as promising trajectories for prospective research.

Conclusion

The paper culminates in open-sourcing the EddyNet model, testifying to a conviction in promoting reproducibility and facilitating broader research endeavors. By making the model, training data, and code accessible, the authors underscore their commitment to fostering collaboration within the oceanographic and remote sensing communities. Researchers and practitioners are encouraged to optimize EddyNet for their specific applications, potentially converging multiple methodologies to bolster eddy detection and tracking efficacy.