Emma Wilkinson, Flagler College

Class Year:

Mentor:

Abstract:

Passive acoustic monitoring (PAM) provides an opportunity to study the population ecology of bottlenose dolphins (Tursiops truncatus) on a large geographic scale, but as technology advances, the greatest challenge researchers face is developing efficient methods for processing large volumes of PAM data. The creation of an automatic identification algorithm using animal calls, such as signature whistles, is a solution to the costly and impractical manual processing of large data volumes. Signature whistles are a resource for the development of an auto-ID algorithm because they contain identifying information about the dolphins that are present. Bottlenose dolphins are protected under the Marine Mammal Protection Act, which makes developing effective methods of population monitoring a top priority for resource managers. We created a deep convolutional neural network model capable of recognizing the signature whistles of bottlenose dolphins within PAM data. We developed an algorithm to automate the processing of PAM data including loading, windowing, and Fourier transformation. The model was trained on a balanced data set of 170 mel-spectrograms with 55.7% prediction accuracy. When applied to field data, the model was 33.3% accurate in discriminating between dolphin signature whistles and other calls and noise. The prediction accuracy indicates a deep learning network as a promising method of signature whistle identification. Our future work will focus on improving the performance of the method and adapting it for monitoring dolphin presence and population size based solely on acoustic data. The auto-ID algorithm will enhance the understanding of dolphin spatio-temporal distribution, migration, and social structure, and help to inform effective management and conservation policies.