Advanced Classification Tools for The Identification and Classification of Underwater Sounds

Sounds in the ocean come from a wide variety of sources, the balance of which can vary greatly with geographical location. These source of sounds include a wide variety of civilian vessels (merchantmen, fishing boats and leisure craft), biological sounds (primarily marine mammals, but including fish and some invertebrates) and geophysical sounds (rain and seismic activity for example).

Additional complexity is added by the fact that the character of these sounds not only depends on the source, but can be strongly affected by the environment in which they are encountered.

The goal of this PhD will be to develop tools, based on machine learning, which assist with the task of identifying the sounds. This is a challenge which has been considered over a significant number of years with limited success. The traditional approach has been to utilise signal processing tools such as a spectrogram to provide input to various forms of classification methods.

ML and artificial intelligence methods are largely been based on the supposition that a large existing labelled dataset is available, an assumption which does not hold for underwater acoustic data. Collecting a set of (labelled) recordings which capture the full richness of acoustic encounters in the marine environment is an unrealistic challenge. Whilst these datasets tend to be large in terms of the number of Tb they occupy, the number of labelled events is small and tends to be very imbalanced (a large collection of common sounds and few examples of the interesting sounds).