Large-Scale Whale Call Classification Using Deep Convolutional Neural Network Architectures

As the rapid development of deep learning techniques, extensive interest has been taken into the applications of deep learning methods on challenging problems of different domains. In view of the recent success of convolutional neural network (CNN) in various tasks of audio analysis, a comparative performance study of different the-state-of-the-art CNN architectures on a large-scale whale-call classification task is investigated in this paper. On the basis of deep neural network models, distinctive features of whale sub-populations are extracted to obtain higher level abstract representations for the accurate classification, which is significantly superior to the traditional classification approaches using manual features based on expert knowledge. In particular, a large open-source acoustic dataset recorded by audio sensors carried by whales in different locations is employed for performance comparison. Based on the experiments, it is found that the advancement of popular CNN architectures significantly improve the accuracy on the whale call classification task. The accuracy and computational efficiency varies with the change of the CNN architectures. Xception provides the best performance among all four CNN architectures while an ensemble of CNN models can produce even better results.