A Hybrid ViT-CNN Architecture for Volumetric Target Classification with Sub-Bottom SAS Data

Deep learning convolutional neural networks (CNNs) have become the state-of-the-art (SOTA) method for automatically classifying targets captured above the seafloor for many 2D synthetic aperture sonar (SAS) imaging systems [1], [2] and recently for targets buried within the sub-bottom [3]–[5]. More specifically, Williams and Brown [3] designed the first 3D CNN architecture to classify objects buried in the sediment. Trung et al. [4] designed neural networks to account for the resonant scattering with volumetric SAS data in shallow water environments. Vetaw et al. [5] introduced a learning-based tone mapping approach to enhance the salient features of buried targets in the sediment while jointly the SOTA 3D CNN model training Williams and Brown [3].