Empirical analysis of latent space encodings for submerged small target acoustic backscattering data

With future sights on specialized classification methods, we work to generalize the acoustic backscattering data from sonar measurements of small targets submerged in water by learning a non-invertible mapping (encoding) to a low-dimensional vector space (ℝn). Finding the optimal dimensionality of this latent space is an important task. The encoding is accomplished by utilizing modality agnostic convolutional machine learning methods that have seen success in other signal and image processing domains. We have explored the autoencoder and its variants, the sparse autoencoder, and the variational autoencoder. Autoencoders encode input samples from a high-dimensional manifold to a lower latent vector space and then reverse the lossy mapping back to a high-dimensional manifold similar to the initial domain. The sparse autoencoder induces sparsity in the components of the latent vectors, and the variational autoencoder learns an encoding to n-dimensional gaussian distributions instead of n-dimensional vectors. The TREX13 data are the primary dataset used for training the networks used in experiments. We evaluate the change in models’ accuracies as the latent dimensionality is increased as well as the models’ ability to generalize to unseen data. Additionally, the PONDEX09 and PONDEX10 data are used to evaluate the models’ cross-domain efficacy.