Inversion of Sediment Parameters in Shallow Water Waveguides Using Polarization Characteristics of Vector Acoustic Fields

This paper presents a method for inverting sediment layer parameters in shallow water waveguides based on the polarization characteristics of vector acoustic fields. By harnessing the robust nonlinear modeling capabilities of deep learning, this approach seamlessly integrates the polarization properties of the acoustic field with temporal convolutional networks (TCN) and multi-task learning (MTL), creating a customized neural network model specifically designed for inversion. Accurate inversion of sedimentary layer parameters was achieved by using the Stokes parameters, derived from particle velocity data received by a single vector hydrophone, as input features for the neural network. An analysis was conducted within a simulation environment to examine the impact of mismatches in source depth and range on the performance of the inversion method. A comparative assessment was undertaken against inversion results that used group speed (dispersion characteristics) as input. The findings revealed that the Stokes parameters demonstrated heightened sensitivity to variations in sediment layer parameters, leading to more accurate inversions with reduced errors, particularly outstanding in the inversion of sediment sound speed. Ultimately, the robustness and effectiveness of this inversion method are validated through experiment data in the South China Sea.