The study of movement ecology has benefitted from the rapid improvements in telemetry devices, which are now used on a large number of aquatic and terrestrial organisms. As tags decrease in size, weight, and cost while increasing in bio-logging capacity and battery life, researchers have been able to collect ever-larger streams of data. However, statistical methods used to estimate latent behavioral states from these trajectories can sometimes be overly simplistic, rely on restrictive assumptions, or are difficult to fit. Our new model aims to improve upon the limitations of existing methods by implementing a non-parametric Bayesian framework to segment multivariate time series into relatively homogeneous units of behaviors and subsequently determine the likely number of behavioral states using a mixed-membership method. We also demonstrate that this model can successfully recover breakpoints and behavioral states from simulated trajectories that vary considerably in length (i.e., 1000, 5000, 10000, and 50000 observations). Additionally, we analyzed GPS tracks (step lengths and turning angles) from an endangered raptor species in Florida, the snail kite (Rostrhamus sociabilis), to estimate latent behavioral states and interpret these results with respect to their natural history.
Results/Conclusions
We simulated three-state trajectories from a correlated random walk, where step lengths were drawn from a gamma distribution and turning angles from a wrapped Cauchy distribution. Our model estimated breakpoints with 88.4% to 95.9% accuracy, where accuracy increased with track length. The model also properly identified the occurrence of three behavioral states and classified the dominant behaviors of time segments with 95.0% to 99.7% accuracy. Analysis of empirical snail kite movements using our Bayesian framework suggested the occurrence of three states, which comprised 94.1% of all behavior assignments. Distributions of step lengths and turning angles showed separation into distinct behaviors, labelled ‘resting’, ‘area-restricted search’, and ‘transit’. Temporal patterns of behavior corroborated the dispersal of fledgling snail kites from natal sites, as well as ‘transit’ behavior in adults immediately before and after peak breeding season. These results can be used further to evaluate time-matched environmental covariates with behavioral changes. Ultimately, this model generates unbiased estimates of behavioral states that can be used in additional analyses of animal movement.