Fusion of ocean data from multiple sources using deep learning: Utilizing sea temperature as an example

Mingqing Wang, Danni Wang, Yanfei Xiang, Yishuang Liang, Ruixue Xia, Jinkun Yang, Fanghua Xu, Xiaomeng HuangFrontiers in Marine Science 2023 · DOI

TL;DR

Researchers have developed ODF-Net, a deep learning framework that fuses fragmented satellite data, sparse underwater observations, and complex numerical models into a high-resolution 3D map of global sea temperatures. By integrating physics-based constraints and attention mechanisms, the model achieves significantly higher accuracy than traditional data assimilation methods while remaining computationally efficient.

What Problem Does This Paper Address?

Understanding the ocean is a "big data" problem with a "missing data" reality. We have three main sources of information, each with a fatal flaw:

  1. Satellite Remote Sensing: Provides excellent surface coverage but cannot see more than a few millimeters below the waves.
  2. In Situ Observations (Floats/Ships): Provide precise vertical profiles (like the Argo program) but are spatially sparse and "fragmented."
  3. Numerical Models/Reanalysis: Provide continuous 3D grids but often drift from reality due to simplified physics or initialization errors.

Traditional "Data Assimilation" (the math used in weather forecasting) tries to blend these, but it is computationally expensive and often struggles with non-linear ocean dynamics. This paper introduces a way to use deep learning to perform this fusion faster and more accurately, specifically targeting the creation of a continuous 3D sea temperature dataset.

Methods

The authors propose the Multi-Source Ocean Data Fusion Network (ODF-Net), which is built on a 1D Residual Network (1D-ResNet) architecture. The "secret sauce" of ODF-Net lies in three specific innovations:

  • Global Attention Mechanism (GAM): The model doesn't treat all data sources equally. It uses Channel Attention to weigh which input (e.g., a specific satellite vs. a specific model) is most reliable at a given time, and Spatial Attention to focus on the most critical depth layers.
  • Physics-Based Loss Function: Instead of just minimizing the mean squared error, the authors added "prior knowledge" constraints. The loss function penalizes the model if the vertical temperature gradients or the total heat content (integral) deviate from physical reality, suppressing "noise" in the vertical profiles.
  • Adaptive Activation (APReLU): Unlike standard ReLU functions, APReLU allows the network to learn the parameters of the activation function for each sample, providing better flexibility for complex non-linear ocean data.

Datasets

The study utilized a massive array of oceanographic data from 1994 to 2017:

  • Ground Truth: EN4-profiles (high-precision in situ temperature and salinity profiles).
  • Satellite Inputs: OISST, ERSST, and HadISST (Sea Surface Temperature); Aviso-SLA (Sea Level Anomaly); and CCMP (Sea Surface Wind).
  • 3D Gridded Models/Reanalysis: CESM2, BCC-CSM2-HR, EN4-analysis, SODA, ECCO, and HYCOM.
  • New Benchmark: The study produced the ODF-ST dataset, a global monthly 3D sea temperature dataset with a 0.25° × 0.25° spatial resolution.

Key Insights and Ideas

  • The "Aha" Moment in Attention: By analyzing the attention weights, the researchers found that the model naturally learned to trust the EN4-analysis and SODA datasets more than others. This provides a level of "interpretability" often missing in deep learning—the model's "logic" aligns with expert oceanographers' trust in those specific data sources.
  • Physics as a Regularizer: Simply throwing data at a neural network often results in "spiky" vertical profiles that are physically impossible. By forcing the model to respect the vertical temperature gradient, the researchers achieved a smoother, more realistic ocean structure.
  • Efficiency vs. Complexity: ODF-Net outperformed a state-of-the-art Transformer model while using only 60% of the trainable parameters, proving that domain-specific architectures (like 1D-ResNets for profiles) can be more efficient than "one-size-fits-all" models.

Key Results

  • Superior Accuracy: ODF-Net achieved an average RMSE of 0.74°C, significantly lower than the >0.9°C seen in other leading datasets like EN4-analysis or SODA.
  • High Correlation: The model showed a correlation coefficient ($R^2$) of 0.992 with observations.
  • Preserving the Thermocline: The model was particularly effective at the "thermocline" (the layer where temperature changes most rapidly with depth), where traditional models often fail.
  • ENSO Tracking: The resulting ODF-ST dataset accurately captured historical El Niño and La Niña events, matching the official indices provided by NOAA.

Limitations and Future Work

  • Polar Data Scarcity: The model’s performance was lower in the Arctic and Southern Oceans. This isn't necessarily a failure of the architecture, but a reflection of the lack of high-quality in situ observations in ice-covered or remote polar regions.
  • Single Variable Focus: This study focused primarily on sea temperature. The authors note that future work should expand this fusion technique to other critical variables like salinity and ocean currents to create a truly holistic 3D view of the ocean.
  • Real-time Potential: While the current study used historical data (1994–2017), the speed of the ODF-Net suggests it could eventually be used for near-real-time ocean monitoring.

Summary updated

Abstract

For investigating ocean activities and comprehending the role of the oceans in global climate change, it is essential to gather high-quality ocean data. However, existing ocean observation data have deficiencies such as inconsistent spatial and temporal distribution, severe fragmentation, and restricted observation depth layers. Data assimilation is computationally intensive, and other conventional data fusion techniques offer poor fusion precision. This research proposes a novel multi-source ocean data fusion network (ODF-Net) based on deep learning as a solution for these issues. The ODF-Net comprises a number of one-dimensional residual blocks that can rapidly fuse conventional observations, satellite observations, and three-dimensional model output and reanalysis data. The model utilizes vertical ocean profile data as target constraints, integrating physics-based prior knowledge to improve the precision of the fusion. The network structure contains channel and spatial attention mechanisms that guide the network model’s attention to the most crucial features, hence enhancing model performance and interpretability. Comparing multiple global sea temperature datasets reveals that the ODF-Net achieves the highest accuracy and correlation with observations. To evaluate the feasibility of the proposed method, a global monthly three-dimensional sea temperature dataset with a spatial resolution of 0.25°×0.25° is produced by fusing ocean data from multiple sources from 1994 to 2017. The rationality tests on the fusion dataset show that ODF-Net is reliable for integrating ocean data from various sources.
BibTeX
@inproceedings{wang2023fusion,
  author = {Mingqing Wang and Danni Wang and Yanfei Xiang and Yishuang Liang and Ruixue Xia and Jinkun Yang and Fanghua Xu and Xiaomeng Huang},
  title = {{Fusion of ocean data from multiple sources using deep learning: Utilizing sea temperature as an example}},
  booktitle = {Frontiers in Marine Science},
  year = {2023},
  doi = {10.3389/fmars.2023.1112065},
}
Figure