Abstract
Accurately reconstructing a three-dimensional ocean sound speed field (3D 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-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 both expressiveness and conciseness. To meet this requirement, a 3D SSF-tailored tensor deep neural network is proposed, which utilizes tensor computations and deep neural network 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.
Introduction
Accurately mapping the three-dimensional sound speed field (SSF) of the ocean is a cornerstone of underwater acoustics. Applications ranging from submarine localization and seafloor mapping to underwater communication optimization depend heavily on precise knowledge of sound speed at any given spatial coordinate. However, the ocean's vastness and the limited deployment of sensors—such as underwater gliders and floats—result in highly sparse and noisy datasets.
The core challenge in 3D SSF reconstruction is that the problem is highly under-determined. Researchers from Zhejiang University frame this as a balancing act between two primary types of error:
- Representation Error ($E_1$): Occurs when a model is too simple (e.g., basic linear interpolation) to capture complex, non-linear oceanographic features like currents and eddies.
- Identification Error ($E_2$): Occurs when a model is over-parameterized (e.g., standard deep neural networks). With sparse data, complex models may "hallucinate" patterns or fail to converge effectively.
Existing methods often fail to strike this balance; splines lack expressiveness, while deep neural networks are often too parameter-heavy for sparse physical fields.
Methodology: Tensor Neural Networks
To address these limitations, the authors propose a Tensor Neural Network (TNN) built on Tensor Contraction Layers (TCLs). This architecture combines the structural efficiency of tensor mathematics with the non-linear modeling power of deep learning.
Architecture and Design
Unlike standard "flat" layers in Multi-Layer Perceptrons (MLPs), the TNN utilizes tensor-based operations derived from Tucker decomposition. This allows the network to directly exploit the multi-dimensional correlations inherent in 3D space (latitude, longitude, and depth). By stacking TCLs with non-linear activation functions (e.g., ReLU or tanh), the model represents complex features with a fraction of the parameters required by standard networks.
Noise Rejection
A key theoretical contribution is the proof that the TNN's low-rank tensor structure provides a natural ability to reject noise. The model is mathematically "disinclined" to fit random Gaussian noise, effectively filtering the data during the reconstruction process without requiring a separate denoising step.
Experimental Evaluation
The model was evaluated using real-world 3D SSF data from the South China Sea (SCS), derived from the Hybrid Coordinate Ocean Model (HYCOM). The researchers simulated sparse sampling by randomly selecting only 10% to 40% of grid points from a $20 \times 20 \times 20$ region and adding varying levels of Gaussian noise.
Key Results
- Efficiency: While a standard MLP used approximately 8 million parameters and failed to reconstruct the field accurately, the proposed TNN achieved superior results using only 875 parameters. This demonstrates that "smaller and smarter" architectures are more effective for sparse physical fields.
- Accuracy: The TNN consistently outperformed state-of-the-art (SOTA) methods, including Gaussian Process Regression (GPR) and various tensor completion algorithms. For instance, at a 20% sampling ratio, the TNN achieved a Root Mean Squared Error (RMSE) of 0.35, compared to 0.52 for GPR.
- Generalization: The authors demonstrated that without non-linear activations, the TNN simplifies to a classical Tucker decomposition, making it a more powerful, generalized version of existing tensor completion methods.
Conclusion and Future Work
The TNN-aided 3D SSF reconstruction algorithm offers a robust solution for mapping the ocean's acoustic environment from limited data. Future research directions include automating hyper-parameter tuning (such as tensor rank), incorporating physical priors from fluid dynamics into the loss function, and developing a Bayesian framework for uncertainty quantification.
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