Striking the right balance: Three-dimensional ocean sound speed field reconstruction using tensor neural networks

Siyuan Li, Lei Cheng, Ting Zhang, Hangfang Zhao, Jianlong Li

Accurately mapping the three-dimensional sound speed field (SSF) of the ocean is a cornerstone of underwater acoustics. Whether you are localizing a submarine, mapping the seafloor, or optimizing underwater communications, you need to know how fast sound travels through the water at any given point. However, the ocean is vast, and our sensors—like underwater gliders and floats—are few and far between.

In a recent paper, researchers from Zhejiang University propose a novel solution: Tensor Neural Networks (TNNs). By combining the structural efficiency of tensor mathematics with the non-linear power of deep learning, they’ve created a model that reconstructs the ocean's acoustic environment with unprecedented accuracy from sparse, noisy data.

TL;DR

This paper introduces a 3D SSF-tailored Tensor Neural Network that balances model expressiveness with parameter conciseness. It outperforms state-of-the-art methods by effectively capturing complex ocean features while naturally rejecting measurement noise.

What Problem Does This Paper Address?

The core challenge in 3D SSF reconstruction is that the problem is "highly under-determined." We have a massive volume of water to map but only a tiny handful of actual measurements.

The authors frame this as a balancing act between two types of error:

  1. Representation Error ($E_1$): The error that occurs when a model is too simple (like a basic linear interpolation) to capture the complex, non-linear "wiggles" in sound speed caused by ocean currents and eddies.
  2. Identification Error ($E_2$): The error that occurs when a model is too complex (like a massive standard neural network). With so few data points, a complex model can "hallucinate" patterns or simply fail to find the right configuration among millions of parameters.

Existing methods usually lean too far one way. Splines are concise but lack "expressiveness"; deep neural networks are expressive but often too "wordy" (over-parameterized) for sparse data.

Methods

The authors propose a Tensor Neural Network (TNN) built on Tensor Contraction Layers (TCLs).

  • The Architecture: Instead of standard "flat" layers, the TNN uses tensor-based operations (specifically based on the Tucker decomposition). This allows the network to exploit the multi-dimensional correlations inherent in 3D space (latitude, longitude, and depth) directly.
  • The "Secret Sauce": By stacking these TCLs with non-linear activation functions (like ReLU or tanh), the model gains the ability to represent complex oceanographic features. Crucially, it does this with a fraction of the parameters required by a standard Multi-Layer Perceptron (MLP).
  • Noise Rejection: The authors provide a theoretical proof (Proposition 1) showing that because the model is constrained by a low-rank tensor structure, it is mathematically "disinclined" to fit random Gaussian noise, effectively filtering the data as it reconstructs the field.

Datasets

The researchers evaluated their model using real-life 3D SSF data from the South China Sea (SCS).

  • Source: Data was derived from the Hybrid Coordinate Ocean Model (HYCOM), which uses temperature, salinity, and depth measurements.
  • Scale: A $20 \times 20 \times 20$ grid representing a $152\text{ km} \times 152\text{ km} \times 190\text{ m}$ region.
  • Simulation: They simulated sparse sampling by randomly selecting only 10% to 40% of the grid points and adding varying levels of Gaussian noise to mimic real-world sensor error.

Key Insights and Ideas

  • The "Aha!" Moment on Conciseness: A standard MLP for this task used ~8 million parameters and failed miserably. The proposed TNN used only 875 parameters and achieved the best results. This proves that for sparse physical fields, "smaller and smarter" beats "bigger and deeper."
  • Generalization of Classical Models: The authors show that if you remove the non-linear activations, the TNN simplifies back into a classical Tucker decomposition. This makes the TNN a more powerful, "generalized" version of existing tensor completion methods.
  • Built-in Denoising: Unlike other methods that require a separate denoising step, the TNN’s architecture acts as a natural low-pass filter for noise while preserving the "sharp" features of the sound speed field.

Key Results

The TNN was compared against several "State-of-the-Art" (SOTA) methods, including Gaussian Process Regression (GPR) and various tensor completion algorithms (LRTC, Tucker-ALS).

  • Accuracy: In almost every scenario, the TNN achieved the lowest Root Mean Squared Error (RMSE). For example, at a 20% sampling ratio with moderate noise, the TNN achieved an RMSE of 0.35, significantly beating GPR (0.52) and the matrix-based BMCG (0.72).
  • Visual Fidelity: Error surface maps show that the TNN produces much smoother and more physically realistic reconstructions, especially in the vertical dimension where other methods often struggle.
  • Regularization: Adding Total Variation (TV) regularization further boosted performance in extremely sparse conditions (10% sampling).

Limitations and Future Work

While the TNN is powerful, the authors note a few areas for growth:

  • Hyper-parameter Tuning: Finding the "sweet spot" for the number of layers and their dimensions (the "rank" of the tensors) currently requires manual trial and error.
  • Physical Priors: The current model is purely data-driven. The authors suggest that incorporating actual fluid dynamics or oceanographic equations into the loss function could make the reconstructions even more robust.
  • Uncertainty Quantification: Moving toward a Bayesian TNN framework would allow the model to not just give an estimate, but also tell the user how "confident" it is in that estimate.

Reference: Li, S., Cheng, L., Zhang, T., Zhao, H., & Li, J. (2023). Striking The Right Balance: Three-Dimensional Ocean Sound Speed Field Reconstruction Using Tensor Neural Networks. J. Acoust. Soc. Am. arXiv:2308.04930

Summary updated

Abstract

Accurately reconstructing a three-dimensional (3D) ocean sound speed field (SSF) is essential for various ocean acoustic applications, but the sparsity and uncertainty of sound speed samples across a vast ocean region make it a challenging task. To tackle this challenge, a large body of reconstruction methods has been developed, including spline interpolation, matrix/tensor-based completion, and deep neural networks (DNNs)-based reconstruction. However, a principled analysis of their effectiveness in 3D SSF reconstruction is still lacking. This paper performs a thorough analysis of the reconstruction error and highlights the need for a balanced representation model that integrates expressiveness and conciseness. To meet this requirement, a 3D SSF-tailored tensor DNN is proposed, which uses tensor computations and DNN architectures to achieve remarkable 3D SSF reconstruction. The proposed model not only includes the previous tensor-based SSF representation model as a special case but also has a natural ability to reject noise. The numerical results using the South China Sea 3D SSF data demonstrate that the proposed method outperforms state-of-the-art methods. The code is available at https://github.com/OceanSTARLab/Tensor-Neural-Network.
BibTeX
@article{li2023striking,
  author = {Siyuan Li and Lei Cheng and Ting Zhang and Hangfang Zhao and Jianlong Li},
  title = {{Striking the right balance: Three-dimensional ocean sound speed field reconstruction using tensor neural networks}},
  journal = {The Journal of the Acoustical Society of America},
  year = {2023},
  doi = {10.1121/10.0020670},
}

Loading related content...

Figure