Discovering Prototypical Sound Speed Profiles with Deep Learning and Self-Organizing Maps

Sound speed profiles (SSPs), which measure the speed of sound underwater as a function of depth, are valuable for understanding acoustic characteristics such as transmission loss in particular regions of the ocean. As such, it can be advantageous to apply unsupervised learning techniques to SSPs in order to group regions of the ocean by their acoustic characteristics, called provincing. However, the variable length of SSPs (as ocean depth is not constant) makes tasks like this difficult. We demonstrate that a ResNet-based regression model that is trained to predict surface temperature from variable-length SSPs can be leveraged as a powerful fixed-length feature extractor for SSPs. We verify the effectiveness of this approach, referred to as ResNetSSP, by using self-organizing maps to find prototype SSPs in features extracted from a public CTD (conductivity, temperature, and depth) dataset.