Pitfalls in the categorization of behaviour: a comparison of dolphin whistle classification methods

The categorization of behaviour patterns into separate classes is crucial to the study of animal behaviour. Traditionally researchers have classified behaviour patterns through careful observation by eye. Recently this method has been increasingly replaced by computer methods. While the definition and fine scale analysis that can be achieved with computers is desirable, only a few studies have actually looked at how these methods perform in comparison with human observation. I compared the classification of bottlenose dolphin, Tursiops truncatus, whistles by human observers with the performance of three computer methods: (1) a method developed by McCowan (1995, Ethology, 100, 177-193); (2) a comparison of cross-correlation coefficients using hierarchical cluster analysis; and (3) a comparison of average difference in frequency along two whistle contours also using hierarchical cluster analysis. The whistle sample consisted of 104 randomly chosen whistles from a group of four captive bottlenose dolphins recorded both during periods when one was separate from the rest of the group and while they all swam in the same pool. The sample contained five individual-specific signature whistles and several nonsignature whistles. Five human observers, without knowledge of the recording context, were more likely than the computer methods to identify signature whistles that were used only while an animal was isolated from the rest of the group. I discuss the limitations of methods commonly used for pattern recognition in communication studies. The discrepancies between methods show how crucial it is to obtain an external validation of the behaviour classes used in studies of animal behaviour. Copyright 1999 The Association for the Study of Animal Behaviour.