|Title||Clicks classification of sperm whale and long-finned pilot whale based on continuous wavelet transform and artificial neural network|
|Publication Type||Journal Article|
|Year of Publication||2018|
|Authors||Jiang, Jia-jia, Bu Ling-ran, Wang Xian-quan, Li Chun-yue, Sun Zhong-bo, Yan Han, Hua Bo, Duan Fa-jie, and Yang Jian|
|EndNote Rec Number||12071|
|Keywords||anthropogenic noise, bearded seal, bottlenose whale, cavitation in cetaceans, decibel, dolphin, marine mammals, prediction, snapping shrimp, sonar, strandings, thresholds, Underwater acoustics, whales|
Passive acoustic observation of whales is an increasingly important tool for whale research. Clicks are the predominant vocalizations of toothed whales, such as sperm whales and long-finned pilot whales. Classifying clicks of sperm whales and long-finned pilot whales is an essential task for the passive acoustic observation of the two whale species, especially in the case that both whale species vocalize in the same observed area. In this paper, we proposed a method performing the automated classification of clicks produced by sperm whales and long-finned pilot whales. First, the two types of whales’ original sounds were denoised using a wavelet denoising method. Then, a dual-threshold endpoint detection algorithm was utilized to detect and pick out all clicks from the denoised sounds. The continuous wavelet transform was applied to decompose the picked clicks, and a wavelet coefficient matrix can be obtained for each picked click. Focusing on the energy distribution and duration difference between the two types of whales’ clicks, we proposed a feature-vector extraction algorithm based on the wavelet coefficient matrix. For each picked click, scale (frequency) features and time feature were obtained respectively and they were used to form the feature vector. Finally, a back propagation (BP) neural network was designed as a classifier of feature-vector to output final classification result. The experiment results show the proposed method can obtain high classification performances. The effect of training dataset size, and the number of training features on the classification performance was also examined in the experiments.