A marine mammal acoustic detection and localization algorithm using spectrogram image correlation

Abstract

Detection and localization of marine mammals at sea can be an important aspect of research and conservation. Detection can be accomplished on a single hydrophone using a simple pattern recognition technique that exploits the time-frequency complexity of the whale vocalization. In particular the frequency modulation of the signal can be an important feature by which vocalizations can be identified. The slope, duration and SNR of the frequency modulated signal in the time-frequency domain are important parameters which aid in the detection. When the call is detected on an array of hydrophones, time difference of arrival can be utilized for localization. This process can be complicated by multiple calls from a single animal, calls from multiple animals, transient noise sources and acoustics propagation effects. A technique which has had success involves the cross-correlation of the spectrogram image of the detected call with call on other hydrophones. This is a variation of the approach implemented by Mellinger and Clark (2000) to identify bowhead whale calls. They create a time-frequency kernel to correlate with received signals on a single phone. The technique presented in this paper uses one of the received calls as the template to match with other calls on other phones in the array. While Mellinger and Clark were required to search in time frequency, this algorithm need only search in time on detected signals. We present results for pilot and sperm whales on a bottom mounted array in deep water.