Using a deep neural network to classify echolocation clicks and identify biogeographic patterns of Pacific white-sided dolphins

Abstract

Pacific white-sided dolphins are endemic to the Pacific Ocean; and two genetically distinct populations overlap along the west coast of North America. However, they are visually indistinguishable and the degree of spatial overlap remains unknown. Here, we use a deep neural network to show that the populations are acoustically distinguishable. Previous studies described two distinct echolocation click types associated with pacific white-sided dolphins and hypothesized that they were population-specific. Our neural network was trained to classify the click types based on spectral and temporal properties described previously. The neural network enabled us to analyze passive acoustic recordings from sites between the Gulf of California and the Gulf of Alaska over multiple years to investigate possible population-specific trends. The latitudinal occurrence pattern of the two click types supports the population-specific hypothesis: type A clicks continue to associate with the northern population distribution, and type B clicks with the southern population distribution. At long-term monitoring sites in Southern California, type B clicks were dominant during periods of warm water anomalies. This pattern may be an early indicator of future biogeographic shifts in the distribution of pacific white-sided dolphins and demonstrates the utility of long-term passive acoustic monitoring.