Synthetic Aperture Sonar Micronavigation with Variational Inference of a State-Space Model

Synthetic aperture sonar (SAS) utilizes the motion of the carrying platform to combine coherently the backscattered signal from several pings, rendering seafloor maps with distinctively high resolution. Micronavigation aims to estimate the platform motion with sub-wavelength accuracy from the spatiotemporal coherence of diffuse backscatter on redundant measurements between successive pings and is essential for coherent SAS processing, when positioning information from navigational instruments is absent or inadequately accurate. Representation learning with variational autoencoders (VAE) offers an unsupervised data-driven micronavigation solution. This study incorporates the motion dynamics through a state-space model to self-supervise the training of the VAE and improve the accuracy of the variational inference scheme.