A Case Study of Geoacoustic Inversion Based on Convolutional Neural Network Using Vertical Array Data

Geoacoustic inversion is regarded as a global or local optimization problem for the conventional matched-field inversion (MFI), while it is treated as a nonlinear regression problem in a machine learning (ML) framework in this paper. A case study is performed to evaluate the feasibility of these two inversion methods when the two-layer bottom is assumed to be a half space model in a shallow water environment. A convolutional neural network (CNN) with multi-task learning is used to estimate geoacoustic parameters simultaneously in shallow water. The network input is the normalized sample covariance matrices of the broadband data received by a vertical line array in frequency domain. The training data, validation data, and test data are generated by an acoustic propagation model. Localization and transmission loss (TL) for different typical bottom models are used to verify the performance of CNN and MFI. Simulation results demonstrate that the trained CNN is robust in geoacoustic inversion even on noisy test data with a moderate Signal-to-Noise Ratio (SNR) and achieve inversion performance comparable to the MFI.