The acoustic reflection coefficient of a layered seabed exhibits an oscillatory structure that varies with frequency and grazing angle. These oscillatory characteristics are linked to the seabed's stratification and its geoacoustic parameters. Through numerical simulation, the influence of seabed layering and geoacoustic parameters on the frequency-angle oscillatory structure of the bottom reflection coefficient (BRC) is investigated. Based on this, a deep learning geoacoustic inversion method is introduced for retrieving near-field seabed layering and its geoacoustic parameters from the frequency- and angle-dependent bottom reflection coefficient. A deep neural network model based on self-attention and cross-attention mechanism, Cross-ViT, is employed to learn features from the two-dimensional BRC matrix. The model is trained to perform the inversion of geoacoustic parameters using a multi-task learning strategy with gradient normalization (GradNorm). Simulation results indicate that, compared to convolutional neural network and transformer models, the presented model more effectively learns the mapping between seabed reflection characteristics and multiple geoacoustic parameters and possesses relatively strong noise robustness. The method's effectiveness is validated using near-field acoustic data from an acoustic inversion experiment conducted on the northern continental shelf of the South China Sea in 2022.
Ma et al. (Thu,) studied this question.