A Deep Learning Method for Inversing 3D Temperature Fields Using Sea Surface Data in Offshore China and the Northwest Pacific Ocean

Xiangyu Wu, Mengqi Zhang, Qingchang Wang, Xidong Wang, Jian Chen, Yinghao QinJournal of Marine Science and Engineering 2024 · DOI

TL;DR

This paper presents a novel deep learning approach using the U-Net architecture to reconstruct high-resolution 3D ocean temperature fields in the Northwest Pacific and offshore China, leveraging readily available sea surface temperature and height data. The method effectively captures mesoscale eddy dynamics, offering a valuable tool for oceanographic research where direct subsurface measurements are scarce.

What Problem Does This Paper Address?

Accurate, high-resolution 3D ocean temperature data is crucial for understanding ocean dynamics, air-sea interactions, and climate change. However, obtaining such data is challenging. In-situ measurements (like Argo floats) are sparse and discrete, while numerical simulations are computationally intensive and often region-specific. Satellite remote sensing provides abundant sea surface data, but it cannot penetrate the ocean to reveal subsurface temperature structures. This gap limits our understanding of vital oceanographic processes, especially mesoscale phenomena like eddies, which play a significant role in heat and material transport. This paper aims to bridge this gap by developing a method to infer 3D temperature fields from easily accessible sea surface data.

Methods

The core of this research is a deep learning model based on the U-Net architecture. U-Net is a convolutional neural network originally designed for image segmentation, known for its ability to preserve fine details through skip connections between its encoder and decoder paths. This architecture is well-suited for image-to-image translation tasks, making it ideal for reconstructing a 3D temperature field from 2D surface data.

The authors experimented with different input data formats and loss functions to optimize the reconstruction:

  • Input Variables: They compared using raw sea surface temperature (SST) and sea surface height (SSH) data against their anomalies (SSTA and SLA). Anomalies, representing deviations from the mean, proved to be more effective inputs.
  • Label Data: They also tested using normalized subsurface temperature anomalies (STA) versus raw STA as the target for reconstruction.
  • Loss Functions: Two common loss functions, Mean Absolute Error (MAE) and Mean Squared Error (MSE), were evaluated to measure the discrepancy between predicted and actual temperature values during training.

Data preprocessing involved interpolating all data to a common spatial resolution of 0.25° and normalizing the input variables to a similar range.

Datasets

The study utilized the Copernicus Marine Environment Monitoring Service (CMEMS) Global Ocean Reanalysis and Simulation (GLORYS12V1) dataset. This dataset provides daily average data for temperature, salinity, sea level, and mixed layer depth with a spatial resolution of 1/12° and 50 vertical layers.

  • Training and Validation: Data from January 2010 to December 2017 (2992 samples) were used for training and optimizing the model.
  • Testing: Data from 2018 (3357 samples) were used to evaluate the model's performance.

The study did not introduce a new benchmark dataset but rather used the GLORYS reanalysis data as both the source for training/testing and the ground truth for evaluating reconstruction accuracy.

Key Insights and Ideas

  • Anomalies are Key: Using anomalies of sea surface temperature and height (SSTA and SLA) as input significantly improves reconstruction accuracy compared to using raw SST and SSH. This suggests that the deviations from the norm hold more predictive power for subsurface temperature than the absolute values themselves.
  • U-Net's Strength: The U-Net architecture's ability to capture both high-level features and fine details through its skip connections is crucial for accurately reconstructing the complex 3D temperature structure.
  • Loss Function Nuances: While both MSE and MAE performed similarly, MSE showed a slight advantage in the upper ocean, likely due to its sensitivity to larger errors. However, the difference was minimal, suggesting that for this task, either function is a reasonable choice.
  • Eddy Reconstruction Power: The model successfully reconstructs the spatial and temporal characteristics of mesoscale eddies, demonstrating its potential for studying these dynamic ocean features.

Key Results

  • The U-Net model successfully reconstructed a 3D temperature field for water depths up to 1700 m in the Northwest Pacific and offshore China with a 0.25° spatial resolution.
  • Reconstruction accuracy was highest when using SSTA and SLA as input variables, with normalized STA as the label.
  • The Root Mean Square Error (RMSE) generally decreased with increasing depth, stabilizing below 200m.
  • Errors were more pronounced in areas with strong temperature gradients, such as near the thermocline and coastal regions.
  • The reconstructed mesoscale eddies showed high consistency with the original GLORYS data, indicating the model's capability in capturing these phenomena.

Limitations and Future Work

  • Limited Input Variables: The study primarily used SST and SSH. The authors suggest that incorporating other sea surface variables (e.g., salinity, wind data) could further enhance reconstruction accuracy.
  • Upper Ocean Challenges: While generally accurate, the model showed some significant errors in the upper water body, particularly around the thermocline (0-300m), where temperature gradients are steep.
  • Loss Function Exploration: The study focused on MSE and MAE. Exploring other, potentially more specialized, loss functions could lead to further improvements.
  • Model Optimization: The authors suggest that investigating ways to optimize and modify the U-Net model could lead to higher output accuracy in future work.

Summary updated

Abstract

Three-dimensional ocean temperature field data with high temporal-spatial resolution bears a significant impact on ocean dynamic processes such as mesoscale eddies. In recent years, with the rapid development of remote sensing data, deep learning methods have provided new ideas for the reconstruction of ocean information. In the present study, based on sea surface data, a deep learning model is constructed using the U-net method to reconstruct the three-dimensional temperature structure of the Northwest Pacific and offshore China. Next, the correlation between surface data and underwater temperature structure is established, achieving the construction of a three-dimensional ocean temperature field based on sea surface height and sea surface temperature. A three-dimensional temperature field for the water layers within the depth of 1700 m in the Northwest Pacific and offshore China is reconstructed, featuring a spatial resolution of 0.25°. Control experiments are conducted to explore the impact of different input variables, labels, and loss functions on the reconstruction results. This study’s results show that the reconstruction accuracy of the model is higher when the input variables are anomalies of sea surface temperature and sea surface height. The reconstruction results using the mean square error (MSE) and mean absolute error (MAE) loss functions are highly similar, indicating that these two loss functions have no significant impact on the results, and only in the upper ocean does the MSE value slightly outperform MAE. Overall, the results show a rather good spatial distribution, with relatively large errors only occurring in areas where the temperature gradient is strong. The reconstruction error remains quite stable over time. Furthermore, an analysis is conducted on the temporal-spatial characteristics of some mesoscale eddies in the inversed temperature field. It is shown that the U-net network can effectively reconstruct the temporal-spatial distribution characteristics of eddies at different times and in different regions, providing a good fit for the eddy conditions in offshore China and the Northwest Pacific. The inversed eddy features are in high agreement with the eddies in the original data.
BibTeX
@article{wu2024deep,
  author = {Xiangyu Wu and Mengqi Zhang and Qingchang Wang and Xidong Wang and Jian Chen and Yinghao Qin},
  title = {{A Deep Learning Method for Inversing 3D Temperature Fields Using Sea Surface Data in Offshore China and the Northwest Pacific Ocean}},
  journal = {Journal of Marine Science and Engineering},
  year = {2024},
  doi = {10.3390/jmse12122337},
}
Figure