Whale-Call Classification Based on Transfer Learning and Ensemble Method

Abstract

Whale calls contain rich acoustic characteristics, which can fully perceive the underwater environment. The purpose of whale-call classification in complex underwater environment is not limited to the study of whale activity characteristics, but also to the study of large-scale underwater radiation noise. As an automatic feature extractor and classifier, convolutional neural networks (CNN) have achieved remarkable success in the fields of computer vision, speech processing and natural language processing. However, the deep convolutional neural networks have not been fully utilized in underwater acoustic feature classification tasks. In the classification task of supervised learning, robust and efficient models need a large number of label data, but the acquisition of underwater acoustic radiation data is difficult, and the process of labeling data is boring and complex. In this paper, we use the method of transfer learning which can transfer the structure from the field of image classification to complete the classification of large-scale whale calls. We use five models to classify the calls from different whale populations and regions, and compare the performance of these models. Finally, an ensemble method is adopted to obtain the optimal classification effect.