Multivariate analysis of underwater sounds

This thesis considers the use of multivariate statistical methods in relation to a common signal processing problem, that of detecting features in sound recordings which contain interference, distortion and background noise.

Two separate but related areas of study are undertaken, first, the compression and noise reduction of sounds; second, the detection of intermittent departures ('signals') from the background sound environment ('noise'), where the latter may be evolving and changing over time.

Compression and noise reduction are two closely related areas that have been studied for a wide range of signals, both one dimensional (such as sound) and two dimensional (such as images). Many well known techniques used in this field are based on the Fourier transform. In this work, we show how the comparative recent wavelet transform is superior for sound data involving short duration signals (such as shrimp clicks) whilst being at least as good as the Fourier transform for longer duration signals (such as dolphin whistles). Various noise reduction techniques involving thresholding wavelet transforms are examined and compared. We show how none of the standard threshold methods cope well with underwater sounds to any reasonable degree and propose a new technique, known as RunsThresh to overcome the perceived problems. The performance of this new method is contrasted with that of various standard thresholds.

Signal identification for underwater sounds is an area that has been examined in detail in much previous work. Here, we build upon the results gleaned from noise reduction to develop a methodology for detecting signals. The underwater noise environment is dynamically modelled using recursive density estimation of certain summary features of its wavelet decomposition. Observations which are considered to be outliers from this distribution are flagged as 'signal'. The performance of our signal detection method is illustrated on artificial data, containing known signals, and on real data. This performance is compared with standard Fourier based methods for both cases.

Finally in this thesis, several ideas are presented and discussed which consider how the noise reduction and signal detection techniques examined in earlier chapters could be developed further, for example, in order to classify detected signals into different classes. These ideas are presented in outline only and are not followed up in detail, since they represent interesting directions for future study, rather than a primary focus of this thesis.