Classification of dispersive calls using a convolutional neural network

Abstract

Low-frequency acoustic signals in shallow water are highly impacted by interactions with the sea surface and seabed. The acoustic field is then conveniently described by modal theory, and the received signal can be modeled by a set of modes that propagate dispersively. It is now well established that the time-frequency dispersion of normal modes, as measured with a single hydrophone, can be used to localize the source and/or estimate the propagation environment. This method has notably been used to range vocalizations from baleen whales in shallow water. However, this method requires at least two modes to be present in the recorded call. Here, we use a convolutional neural network (CNN) to detect and classify dispersed gunshots (impulse calls) from Southern right whales, using a dataset recorded in Baja de San Antonio, Argentina. The CNN outputs the confidence of an input belonging to the following three classes: at least two modes, less than two modes, and no call present. We show that the CNN can isolate multi-modal dispersive gunshots from large audio data with high precision. Such signals can then be further processed to localize the source and/or characterize the environment. [Work supported by the Office of Naval Research.]