Three-Dimensional Spatio-Temporal Slim Weighted Generative Adversarial Imputation Network: Spatio-Temporal Silm Weighted Generative Adversarial Imputation Net to Repair Missing Ocean Current Data

Ya Jie Yue, Juan Li, Y. Zhang, Meiqi Ji, Jingyao Zhang, Rui MaJournal of Marine Science and Engineering 2025 · DOI

TL;DR

This paper introduces a novel deep learning model, 3D-STA-SWGAIN, designed to fill in missing data in three-dimensional ocean current fields. By integrating spatio-temporal attention mechanisms and Wasserstein constraints, the model significantly improves data imputation accuracy, outperforming traditional methods and existing GAN-based approaches, particularly in scenarios with high data loss.

What Problem Does This Paper Address?

Accurate three-dimensional ocean current data is crucial for understanding ocean dynamics and improving climate models. However, obtaining this data is challenging due to the difficulty, cost, and susceptibility to component failures and environmental noise associated with sensor arrays. This often leads to discontinuous and incomplete datasets. Traditional interpolation methods struggle with non-uniform missing data and capturing complex spatio-temporal correlations, while existing deep learning models often fail to adequately address the unique challenges of marine data, such as high sparsity and complex missing patterns. This paper aims to bridge this gap by developing a robust method for reconstructing missing 3D ocean current data.

Methods

The core of this work is the 3D-STA-SWGAIN (Three-Dimensional Spatio-Temporal Slim Weighted Generative Adversarial Imputation Network). This unsupervised model builds upon the Generative Adversarial Network (GAN) framework, specifically incorporating several key innovations:

  • Spatio-Temporal Attention Mechanism: This module allows the model to dynamically focus on the most relevant spatio-temporal information within the data, effectively capturing complex relationships across both time and space. This is crucial for understanding the dynamic nature of ocean currents.
  • Wasserstein Constraints with Gradient Penalty (GP): The model utilizes Wasserstein distance for training, which is known to provide more stable training and better gradient flow compared to traditional GAN loss functions. The addition of a gradient penalty further enhances training stability and ensures Lipschitz continuity, preventing issues like abnormal vorticity.
  • 3D Input/Output: Unlike many previous methods that focus on 2D data, 3D-STA-SWGAIN is designed to handle four-dimensional data (time × depth × latitude × longitude), enabling it to capture the full 3D structure of ocean currents.
  • Slim-Weighted Generative Adversarial Imputation Network (SGAIN) Backbone: The model leverages an improved version of GAIN, which simplifies the architecture for better computational efficiency while maintaining effectiveness.

The generator learns to produce realistic ocean current data, while the discriminator is trained to distinguish between real and generated data. This adversarial process, guided by the spatio-temporal attention and Wasserstein-GP loss, leads to high-quality imputation.

Datasets

The study utilizes real-time analysis flow field data of the Northwest Pacific, reconstructed from sea surface velocity and sea surface height data. This includes:

  • SST dataset from the National Oceanic and Atmospheric Administration.
  • AVISO satellite altimetry dataset.
  • CCMP dataset from the National Centers for Environmental Prediction.

The data was assimilated and corrected using various international public sources and observational monitoring data. For experiments, data from the South China Sea region was extracted, covering a specific latitude and longitude range with a horizontal resolution of 0.125° and 25 depth layers.

Key Insights and Ideas

  • The critical role of spatio-temporal attention: The paper strongly demonstrates that explicitly modeling spatio-temporal dependencies is paramount for accurate ocean current imputation, especially in complex and dynamic environments.
  • 3D is essential: Moving beyond 2D imputation is not just an incremental improvement but a necessity for capturing the true dynamics of ocean currents, which are inherently three-dimensional.
  • GANs are powerful for imputation: The success of the 3D-STA-SWGAIN model highlights the potential of generative adversarial networks for tackling challenging data imputation tasks in scientific domains.
  • Robustness to diverse missing patterns: The model shows effectiveness across various missing data scenarios, including random, layer-wise (cloud cover), and block-shaped (sensor failure) patterns, indicating its broad applicability.

Key Results

  • Under a 50% random missing rate on the South China Sea dataset, 3D-STA-SWGAIN reduced the error by 37.2% compared to methods like GAIN.
  • In experiments involving missing whole layers (simulating cloud cover), the model achieved significantly lower errors, with MAE values as low as 0.002 at a 10% missing rate, outperforming traditional methods and other GAN-based approaches.
  • For block-shaped missing data, the model demonstrated superior performance, with MAE values as low as 0.0007 at a 10% missing rate, significantly outperforming other methods.
  • Ablation studies confirmed the substantial contribution of the spatio-temporal attention mechanism and the Wasserstein-GP loss function to the model's overall performance.

Limitations and Future Work

  • Computational Efficiency: The paper acknowledges that the model's computational complexity, particularly due to the cascaded 3D convolution and spatio-temporal attention modules, leads to long training cycles for large datasets.
  • Scalability: The parameter scale of the dynamic prompt generator grows linearly with input dimension, which can limit its application in higher-resolution scenarios.
  • Extreme Missing Data: The model's robustness to extremely high missing rates (e.g., over 70%) needs further improvement.
  • Physical Constraints: For extreme marine disaster events with strong nonlinear features, the current physical constraint loss term might be too smooth, potentially limiting its ability to capture sudden dynamic processes.

Future work suggested includes:

  • Designing a lightweight spatio-temporal attention mechanism module (e.g., using local window attention like Swin Transformer) to reduce computational complexity.
  • Developing real-time online adversarial training functionality.
  • Incorporating physical prior constraints (e.g., Navier-Stokes equations) to enhance dynamic rationality.
  • Investigating the risk of physical consistency and error propagation in generated data.

Summary updated

Abstract

Three-dimensional ocean observation is the foundation for accurately predicting ocean information. Although ocean observation sensor arrays can obtain internal data, their deployment is difficult, costly, and prone to component failures and environmental noise, resulting in discontinuous data. To address the severe missing data problem in three-dimensional ocean flow fields, this paper proposes an unsupervised model: Three-dimensional Spatio-Temporal Slim Weighted Generative Adversarial Imputation Network (3D-STA-SWGAIN). This method integrates spatio-temporal attention mechanisms and Wasserstein constraints. The generator captures the three-dimensional spatial distribution and vertical profile dynamic patterns through the spatio-temporal attention module, while the discriminator introduces gradient penalty constraints to prevent gradient vanishing. The generator strives to generate data that conforms to the real ocean flow field, and the discriminator attempts to identify pseudo-ocean current data samples. Through the adversarial training of the generator and the discriminator, high-quality completed data are generated. Additionally, a spatio-temporal continuity loss function is designed to ensure the physical rationality of the data. Experiments show that on the three-dimensional flow field dataset of the South China Sea, compared with methods such as GAIN, under a 50% random missing rate, this method reduces the error by 37.2%. It effectively solves the problem that traditional interpolation methods have difficulty handling non-uniform missing and spatio-temporal correlations and maintains the spatio-temporal continuity of the current field’s three-dimensional structure.
BibTeX
@article{yue2025three,
  author = {Ya Jie Yue and Juan Li and Y. Zhang and Meiqi Ji and Jingyao Zhang and Rui Ma},
  title = {{Three-Dimensional Spatio-Temporal Slim Weighted Generative Adversarial Imputation Network: Spatio-Temporal Silm Weighted Generative Adversarial Imputation Net to Repair Missing Ocean Current Data}},
  journal = {Journal of Marine Science and Engineering},
  year = {2025},
  doi = {10.3390/jmse13050911},
}
Figure