Matched-field geoacoustic inversion based on radial basis function neural network.

Area of Science:

• Ocean acoustics
• Geophysical signal processing
• Machine learning applications

Background:

• Matched-field inversion (MFI) is a standard technique for estimating geoacoustic parameters.
• Traditional MFI methods can be computationally intensive and sensitive to environmental variations.
• Neural networks (NNs) offer potential for improved parameter estimation through advanced data processing.

Purpose of the Study:

• To develop and evaluate a novel method combining multi-layer neural networks (NNs) with matched-field inversion (MFI) objective functions.
• To enhance the estimation accuracy and efficiency of geoacoustic parameters.
• To demonstrate the efficacy of the proposed method using numerical simulations and experimental data.

Main Methods:

• A radial basis function neural network (RBFNN) was extended to incorporate MFI objective functions.
• Shallow NN layers extracted frequency features from hydrophone data; deep layers performed inverse function approximation.
• A hybrid backpropagation and pseudo-inverse scheme with batch processing was used for efficient weight updates.

Main Results:

• The proposed NN-MFI method demonstrated inversion performance comparable to conventional MFI.
• Training the NN with a large dataset covering the parameter interval was crucial for performance.
• Successful application to both numerical simulations and the SWellEx-96 experimental dataset.

Conclusions:

• The integration of NNs with MFI objective functions provides a powerful approach for geoacoustic parameter estimation.
• The method effectively utilizes big data and MFI principles for robust inversion.
• This NN-based technique offers a competitive alternative to traditional MFI methods in underwater acoustics.