A comparison of optimized methods for the detection of blue whale sounds

Abstract

Paul Thompson and colleagues published one of the first long-term studies of mysticete sounds [Thompson and Freidl, Cetology 45, 1–19 (1982)]. Thompson analyzed sounds manually, finding and tallying vocalizations to arrive at a view of seasonal occurrence. Today the detection and counting tasks are often done by computer, using various methods for pattern recognition. Here we examine and compare three such methods for detecting the sounds blue whales: matched filtering, which may work well because of the stereotypy of blue whale vocalizations; spectrogram correlation, which may work well for the same reason and also because of the noise removal that can be done with it; and a neural network, which has worked well in other contexts for detecting right whale calls. The methods are configured using optimization procedures specialized for each method, and the results are compared for vocalizations recorded at different signal-to-noise ratios. The optimized detectors are applied to SOSUS data to detect sounds characteristic of blues whales in the northeast Pacific.