Transient Detection in Impulsive Noise Using Low-Variance Spectrum Estimation

This paper considers the problem of detecting unknown bandpass acoustic transients in highly impulsive ambient noise such as that produced by snapping shrimp. Standard transient detection techniques apply the periodogram which is a classical direct nonparametric spectrum estimation method. The periodogram can be limited due to its poor bias properties resulting from substantial sidelobe leakage and is an inconsistent estimator in the sense that its variance does not decrease with the sample size. We propose the use of low-variance spectrum estimation techniques such as multitaper spectrum estimation whereby the power spectrum estimate is obtained as an average of a number of direct spectrum estimates computed using different windows (or tapers). In addition, methods for improving power spectrum estimates based on denoising by wavelet thresholding are considered. Detectors derived based on these approaches are tested using artificially generated bandpass transients inserted in impulsive ambient sea noise.