Deep Machine Learning Techniques for the Detection and Classification of Sperm Whale Bioacoustics

Abstract

We implemented Machine Learning (ML) techniques to advance the study of sperm whale ( Physeter macrocephalus ) bioacoustics. This entailed employing Convolutional Neural Networks (CNNs) to construct an echolocation click detector designed to classify spectrograms generated from sperm whale acoustic data according to the presence or absence of a click. The click detector achieved 99.5% accuracy in classifying 650 spectrograms. The successful application of CNNs to clicks reveals the potential of future studies to train CNN-based architectures to extract finer-scale details from cetacean spectrograms. Long short-term memory and gated recurrent unit recurrent neural networks were trained to perform classification tasks, including (1) “coda type classification” where we obtained 97.5% accuracy in categorizing 23 coda types from a Dominica dataset containing 8,719 codas and 93.6% accuracy in categorizing 43 coda types from an Eastern Tropical Pacific (ETP) dataset with 16,995 codas; (2) “vocal clan classification” where we obtained 95.3% accuracy for two clan classes from Dominica and 93.1% for four ETP clan types; and (3) “individual whale identification” where we obtained 99.4% accuracy using two Dominica sperm whales. These results demonstrate the feasibility of applying ML to sperm whale bioacoustics and establish the validity of constructing neural networks to learn meaningful representations of whale vocalizations.

Review of Automatic Detection and Classification Techniques for Cetacean Vocalization

“…Bermant et.al [33] achieved a 99.5% detection accuracy of sperm whales echolocation clicks from annotated 650 spectrogram images of the click data using a CNN based approach to build an echolocation click detector. The technique design the detector to label annotated spectrogram image as 'click' or 'no click'.…”

“…The detector was reported to produce lower magnitude of false positive rate while exhibiting significantly increasing true positive rate. Deep learning approach to developing algorithms for detection and classification of cetacean signals is gaining more prominence as seen in recent works [25], [33], [41], [96]. Deep learning being an illustrative learning method where machine automatically learns the representations that are needed from the input raw data makes it a promising area to improving on existing techniques for detection and classification of cetacean species.…”

“…The cetacean sounds are generally categorized as clicks, whistles, songs and (burst) pulsed calls [30]- [32]. The click sounds which are produced by all cetaceans are used for echolocation and foraging [30], [33]. They are impulsive short pulses of substantial strength which last for micro seconds and extremely directional.…”

“…Application of a robust automatic detection and classification model to cetacean vocalizations can give understanding into their repertoire variation, individual vocal changeability, social setting relationships and some other crucial cetacean behavioral questions [38]. In the last few years, machine learning approaches and other automatic techniques have been used for the detection and classification of cetacean sounds [33], [39]- [46] with varying performance outputs. The basic block diagram of the detection and classification stages is shown in Fig.…”

Automated detection of Hainan gibbon calls for passive acoustic monitoring

“…Case studies reporting successful applications play an important role in developing and disseminating best practices, and in discriminating between those tasks that current deep learning methods are able to automate and those they cannot. Previous applications have used convolutional neural networks (CNNs; LeCun, Bengio, and Hinton (2015)) to identify various bird (Grill & Schlüter, 2017;Kahl et al, 2017;Stowell, Wood, et al, 2019) and whale species (Bergler et al, 2019;Bermant, Bronstein, Wood, Gero, & Gruber, 2019;Jiang et al, 2019;Shiu et al, 2020), bees (Kulyukin, Mukherjee, & Amlathe, 2018;Nolasco et al, 2019), as well as anomalous acoustic events in soundscapes (Sethi et al, 2020). These have shown, for example, that a generally good approach is to represent data as spectrograms and treat the problem as an image classification one, as well as providing specialised approaches for data augmentation on spectrogram inputs, such as pitch and time shifting and introducing background noise (Bergler et al, 2019;Sprengel, Jaggi, Kilcher, & Hofmann, 2016).…”