Deep Learning-Driven Vector Acoustic Field Inversion

This monograph presents a deep learning framework for seabed characterization by fusing vector acoustic field physics with neural networks. It introduces Stokes parameters from vector hydrophones as robust features for geoacoustic inversion and develops specialized networks (BP MTL-TCN U-Net + ATT-BP) to estimate sediment parameters and extract dispersion curves. Validated in the Yellow Sea the method achieves core-comparable accuracy in minutes significantly outperforming traditional techniques in speed and robustness. The work highlights the synergy between physical principles and data-driven learning offering a scalable solution for real-time seabed mapping and advancing autonomous ocean sensing.