Multi-label classification of frog species via deep learning

Acoustic classification of frogs has received increasing attention for its promising application in ecological studies. Various studies have been proposed for classifying frog species, but most recordings are assumed to have only a single species. In this study, a method to classify multiple frog species in an audio clip is presented. To be specific, continuous frog recordings are first cropped into audio clips (10 seconds). Then, various time-frequency representations are generated for each 10-s recording. Next, instead of using traditional hand-crafted features, various features are extracted using pre-trained networks using three time-frequency representations: Fast-Fourier spectrogram, Constant-Q transform spectrogram, and Gammatone-like spectrogram. Finally, a binary relevance based multi-label classification approach is proposed to classify simultaneously vocalizing frog species with our proposed features. Our proposed method is verified using eight frog species widely distributed in Queensland, Australia. The results show that the proposed features extracted via pre-trained networks can achieve better classification performance when compared to hand-crafted features for classifying multiple simultaneously vocalizing species.