Passive acoustic monitoring (PAM) is an effective method for monitoring marine fauna. Automatic detection, identification, and quantification of echolocation clicks by toothed whales is commonly computationally inefficient. We present an autoregression and correlation-based algorithm for detection and classification of odontocete clicks. The detection process consists of three steps: (1) detection and extraction of impulsive signals from PAM recordings, (2) cross-correlation of the extracted sounds with template signals, and (3) binary classification of detected clicks. The extraction of impulsive signals from background noise is based on automatic detection of outliers in an autoregression model of noise. The model settings can be adjusted to detect signals of certain length. The cross-correlation analysis involves the use of template signals of high signal-to-noise ratio, which aremanually identified and extracted from PAM data. Using the cross-correlationcoefficient as a criterion allows distinguishing echolocation clicks by sperm or beaked whales from other impulsive sounds. Finally, binary classification utilizes thresholds determined through the receiver operating characteristic analysis of an independent subset of manually labeled, extracted signals. The approach was tested for detection of sperm and beaked whales in PAM datasets recorded on the Northwest Shelf of Australia and demonstrated low false positive rates of 3%–5%.
Read full abstract