Classification of behavioral state using hidden Markov model analysis of animal-attached tag data: Applications and future prospects

Abstract

Data from high-resolution animal-attached tags enable quantification of behavioral responses to anthropogenic noise. However, the duration of such detailed tag records on marine divers are typically too short to allow evaluation of the biological significance of such effects. To explore whether and how sperm whale behavior changed during exposure to sonar, we developed a discrete-time hidden activity state model that describes how observed parameters derived from measured Dtag data (depth, pitch, and clicking behavior) arise from five behavioral modes (surfacing, descent, bottom phase, ascent, resting, and silent active). Although the model assumed simple Markovian state-transitions, the state classification matched well with expert judgment of dive state. During experimental exposures to 1–2 kHz sonar, all four sperm whales tested reduced foraging time, increased silent active behavior, and buzz rates during foraging states decreased. None of those effects were found during 6–7 kHz experimental exposures of the same four whales, nor for three other whales exposed to distant sonar. Hidden classification of behavioral state using quantitative analysis of data collected by the animal attached tag is a procedure that has the potential to be processed autonomously on-board tags. This would enable collection and satellite telemetry of longer-term behavioral data sets with biologically significant interpretations.