Feature visualizations in geoacoustic inversion using convolutional neural network

The acoustic signals propagating in different environments have distinct features which are related to the geoacoustic parameters. A convolutional neural network (CNN) is applied to extract features from signals in the frequency domain to estimate the geoacoustic parameters in shallow water. The outputs of the trained CNN layers with different depths are visualized to express the features extracted from the input data. The network input is the normalized sample covariance matrices (SCMs) of the broadband data received by a vertical line array. Simulated acoustic data generated by the acoustic propagation model are used as the training data, validation data, and test data. Simulation visualization results demonstrate that the trained CNN can extract features of geoacoustic parameters and have good robustness in geoacoustic inversion even on noisy test data.