Detection and classification of underwater transients with data driven methods based on time-frequency distributions and non-parametric classifiers

Due to the complexity of underwater transients and background interference, model based approaches to transient detection/classification are often not practical. This has motivated an interest for data-driven, model-free methods. One such method was presented by Jones and Sayeed (see Proceedings of the 1995 IEEE International Conference on Acoustics, Speech and Signal Processing CASSP 95, Detroit, MI, p.1033-1036) and modified by Oliveira and Barroso (see Proc. of MTS/IEEE Oceans 2000, August 2000), where it was applied to the detection of underwater transients. We extend that approach, to allow its use in the more demanding environment of a brown water environment, where background noise is constituted by a multitude of different interferences, non-white, and highly non-stationary. Also, the assumption of linear separability amongst the transients and the background noise in the time-frequency or related domains will be discarded, leading to the use of an additional classifier stage. A technique to minimize the number of prototypes on this classifier is presented. The developed methods are used to detect and classify real underwater transients, recorded off the Portuguese coast. Estimation of the overall error rate of the method is obtained using cross-validation with the available data set, showing that these methods can effectively be used in real environment situations.