Seabed type and source parameters predictions using ship spectrograms in convolutional neural networks | JoVE Visualize

Area of Science:

Ocean acoustics
Machine learning
Geophysics

Background:

Surface ship acoustic spectrograms offer rich information for environmental sensing.
Convolutional neural networks (CNNs) are powerful tools for analyzing complex signal data.
Accurate seabed characterization and vessel parameter estimation are crucial for marine operations.

Purpose of the Study:

To investigate the efficacy of CNNs in predicting seabed type, ship speed, and closest point of approach (CPA) range from broadband acoustic spectrograms.
To compare the performance of different CNN architectures and input data representations.
To assess the generalization capability of CNNs trained on synthetic data when applied to measured acoustic data.

Main Methods:

Trained three CNN architectures of varying sizes using multitask learning (classification for seabed type, regression for ship speed and CPA).
Utilized synthetic acoustic data generated with measured sound speed profiles and varied source parameters for training.
Evaluated model performance on both synthetic datasets (interpolation/extrapolation) and a measured dataset from the 2017 Seabed Characterization Experiment (SBCEX 2017).

Main Results:

The smallest CNN architecture generalized better to the measured SBCEX 2017 data compared to larger networks, despite slightly lower accuracy on synthetic test data.
Complex pressure spectral values yielded the most accurate and consistent ship speed and CPA predictions with the smallest network.
Absolute pressure values provided more accurate seabed type predictions than complex values when using the smallest network.

Conclusions:

CNNs, particularly smaller architectures, can effectively predict seabed characteristics and vessel parameters from ship-generated acoustic data.
Synthetic data generation is a viable approach for training deep learning models in the absence of extensive labeled field data.
The choice of input data representation (complex vs. absolute pressure values) impacts prediction accuracy depending on the target variable (seabed type vs. vessel parameters).