Signal detection in underwater sound using wavelets

This article considers the use of wavelet methods in relation to a common signal processing problem, that of detecting transient features in sound recordings that contain interference or distortion. In this particular case, the data are various types of underwater sounds, and the objective is to detect intermittent departures (potential "signals") from the background sound environment in the data ("noise"), where the latter may itself be evolving and changing over time. We develop an adaptive model of the background interference, using recursive density estimation of the joint distribution of certain summary features of its wavelet decomposition. Observations considered to be outliers from this density estimate at any time are then flagged as potential "signals." The performance of our method is illustrated on artificial data, where a known "signal" is contaminated with simulated underwater "noise" using a range of different signal-to-noise ratios, and a "baseline" comparison is made with results obtained from a relatively unsophisticated, but commonly used, time-frequency approach. A similar comparison is then reported in relation to the more significant problem of detecting various types of dolphin sound in real conditions.