Detecting Ultrasonic Harbor Porpoise Clicks with Hetrodyning and Teager–Kaiser Energy Operator

This study advances research in the design of autonomous machine listening agents to identify the presence of marine mammals. We compare two common deep learning architectures in the detection of ultrasonic harbor porpoise clicks in noisy coastal soundscapes. We reduce the high dimensionality of the ultrasonic underwater recordings with a heterodyned Teager–Kaiser energy operator approach, which shifts the ultrasonic signal down to lower frequencies. We cross-validate our deep learning models by day to increase our model’s ability to generalize between the ever-changing background noise environment. We demonstrate our approach over a large dataset of (n=8286) examples of harbor porpoise click steams recorded in the Bay of Fundy. We discuss our results in the context of emphasizing the need to overcome such shifts in noise to design a robust porpoise click classifier that is capable of real-time deployment and able to generalize to all real-world conditions.