Open-source machine learning BANTER acoustic classification of beaked whale echolocation pulses

Machine learning with taxonomic family delimitation aids in the classification of ephemeral beaked whale events in passive acoustic monitoring

“…However, relying too much on context when constructing automatic classifiers can negatively affect their performance on rare signal events [33,34]. Recent studies have used both deep learning [33][34][35] and random forest [29] methods for automated classification of beaked whale echolocation clicks. While both approaches are effective, deep learning demonstrates higher and more robust classification performance [29,[33][34][35].…”

“…Recent studies have used both deep learning [33][34][35] and random forest [29] methods for automated classification of beaked whale echolocation clicks. While both approaches are effective, deep learning demonstrates higher and more robust classification performance [29,[33][34][35]. Our study supports these findings, showing high classification performance comparable with results of a deep learning method previously applied to the western North Atlantic [34], with both classifiers achieving precision and recall rates above 98% for beaked whale species in the test datasets.…”

“…The generalized classifier implemented in the case study example was implemented to be comparable with prior deep-learning-based methods [33][34][35] as a baseline model, to facilitate PLOS ONE comparison with our proposed targeted species classification pipeline. A random forest approach utilizing a hard negative filter to eliminate non-beaked whale species based on frequency sweep rate of detected clicks [29] has also been proposed, with slightly poorer performance than the baseline model. Our comparison results shows that when non-beaked whale dominant clicks were not removed prior to clustering, a higher incidence of partial misclassification of time bins occurred.…”

“…Automatic detectors have been developed to identify beaked whale acoustic events that meet certain heuristics [13,[21][22][23], but classification to species is often still manual and time-consuming [13,24], or inaccurate with high false positive and missed detections if automated [25,26]. Several machine learning techniques have emerged as effective tools for beaked whale species-level classification, by learning patterns from manually developed training sets [26][27][28][29] or groups of similar click types identified using unsupervised clustering algorithms [30][31][32]. By deriving clusters of similar click types based on the similarities of selected click features, unsupervised learning has automatically generated large training sets needed to teach supervised learning algorithms for species classification [29,33].…”

“…Several machine learning techniques have emerged as effective tools for beaked whale species-level classification, by learning patterns from manually developed training sets [26][27][28][29] or groups of similar click types identified using unsupervised clustering algorithms [30][31][32]. By deriving clusters of similar click types based on the similarities of selected click features, unsupervised learning has automatically generated large training sets needed to teach supervised learning algorithms for species classification [29,33]. This approach has helped discover and learn known and novel signals, and the acoustic identification of species previously not classifiable [31,34,35], allowing for realistic analysis with evolving knowledge of species acoustics recognition.…”