Whistle Classification of Sympatric False Killer Whale Populations in Hawaiian Waters Yields Low Accuracy Rates

• Published in: Frontiers in Marine Science
• Main Authors: Yvonne Barkley, Erin M. Oleson, Julie N. Oswald, Erik C. Franklin
• Format: Article in Journal/Newspaper
• Language: English
• Published: Frontiers Media S.A. 2019
• Subjects: cetaceans; false killer whale; passive acoustic monitoring; population classification; Hawaiian archipelago; machine learning; Science; Q; General. Including nature conservation; geographical distribution; QH1-199.5; Killer Whale
• Online Access: https://doi.org/10.3389/fmars.2019.00645; https://doaj.org/article/b50a8c719f714a34a4e68b69f55358e4

Cetaceans are ecologically important marine predators, and designating individuals to distinct populations can be challenging. Passive acoustic monitoring provides an approach to classify cetaceans to populations using their vocalizations. In the Hawaiian Archipelago, three genetically distinct, sympatric false killer whale (Pseudorca crassidens) populations coexist: a broadly distributed pelagic population and two island-associated populations, an endangered main Hawaiian Islands (MHI) population and a Northwestern Hawaiian Islands (NWHI) population. The mechanisms that sustain the genetic separation between these overlapping populations are unknown but previous studies suggest that the acoustic diversity between populations may correspond to genetic differences. Here, we investigated whether false killer whale whistles could be correctly classified to population based on their characteristics to serve as a method of identifying populations when genetic or photographic-identification data are unavailable. Acoustic data were collected during line-transect surveys using towed hydrophone arrays. We measured 50 time and frequency parameters from whistles in 16 false killer whale encounters identified to population and used those measures to train and test random forest classification models. Random forest models that included three populations correctly classified 42% of individual whistles overall and resulted in a low kappa coefficient, κ = 0.15, indicating low agreement between models, and the true population. Whistles from the MHI population showed the highest correct classification rate (52%) compared to pelagic and NWHI whistles (42 and 36%, respectively). Pairwise random forest models classifying pelagic and MHI whistles proved slightly more accurate (62% accuracy, κ = 0.24), though a similar pelagic-NWHI model did not (56% accuracy, κ = 0.12). Results suggest that the time-frequency whistle characteristics are not suitable to confidently classify encounters to a specific false killer whale population, ...