Whale Vocalization Classification Using Feature Extraction With Resonance Sparse Signal Decomposition and Ridge Extraction

Whales communicate using whistle vocalizations that are essentially underwater acoustic frequency-modulated tones. Inevitable environmental noise decreases recognition accuracy of these sounds during wide range detection. In this paper, we propose a robust time – frequency analysis method that combines resonance sparse signal decomposition (RSSD) and spectrogram ridge extraction. We apply RSSD to extract whistle components from the raw signal, and then we segment the ridge regions of the whistle spectrograms. By applying a partial derivative method, we extract the whistle spectrogram ridge representing an accurate trace of the whistle vocalization. From these results, we extract ridge features and use an SVM or a random forest to identify the whale species. We evaluated our method using experiments with samples for four whale species. Compared with direct ridge extraction directly without RSSD, our proposed method achieved better extraction of frequency characteristics of the vocalizations. Our proposed method achieved an accuracy rate of over 98% for sounds from four species when using five training samples.