A convolution neural network for dolphin species identification using echolocation clicks signal

The identification of echolocation clicks signals is of great important in protection and research in marine life. Indo-Pacific humpbacked dolphin is first class national protected animals, previous research has proved that the different region Indo-Pacific humpbacked dolphin has accent of their own species groups. An artificial neural network is proposed for dolphin species identification in this paper. The clicks signals were detected by Teager-Kaiser Energy Operator (TKEO), then cepstrum operation were applied to the clicks signals for better feature representation before feed in the neural network. Experiment were conducted based on signals of Indo-Pacific humpbacked dolphins from both Xiamen and Leizhou Bay. The accuracy of species identification was up to 99% and optimal choice of hyperparameter of neural network for this task is discussed.