Abstract
BackgroundMultiple methods have been developed to infer behavioral states from animal movement data, but rarely has their accuracy been assessed from independent evidence, especially for location data sampled with high temporal resolution. Here we evaluate the performance of behavioral segmentation methods using acoustic recordings that monitor prey capture attempts.MethodsWe recorded GPS locations and ultrasonic audio during the foraging trips of 11 Mexican fish-eating bats, Myotis vivesi, using miniature bio-loggers. We then applied five different segmentation algorithms (k-means clustering, expectation-maximization and binary clustering, first-passage time, hidden Markov models, and correlated velocity change point analysis) to infer two behavioral states, foraging and commuting, from the GPS data. To evaluate the inference, we independently identified characteristic patterns of biosonar calls (“feeding buzzes”) that occur during foraging in the audio recordings. We then compared segmentation methods on how well they correctly identified the two behaviors and if their estimates of foraging movement parameters matched those for locations with buzzes.ResultsWhile the five methods differed in the median percentage of buzzes occurring during predicted foraging events, or true positive rate (44–75%), a two-state hidden Markov model had the highest median balanced accuracy (67%). Hidden Markov models and first-passage time predicted foraging flight speeds and turn angles similar to those measured at locations with feeding buzzes and did not differ in the number or duration of predicted foraging events.ConclusionThe hidden Markov model method performed best at identifying fish-eating bat foraging segments; however, first-passage time was not significantly different and gave similar parameter estimates. This is the first attempt to evaluate segmentation methodologies in echolocating bats and provides an evaluation framework that can be used on other species.
Highlights
Multiple methods have been developed to infer behavioral states from animal movement data, but rarely has their accuracy been assessed from independent evidence, especially for location data sampled with high temporal resolution
The distribution of times between buzzes has a long right-tail towards higher time intervals, which has an exponentially decreasing shape, indicating that most buzzes occur in clusters, i.e., in foraging bouts (Additional file 1: Figure S11)
Hidden Markov models (HMM) had the highest median balanced accuracy (67.3%), followed by Correlated velocity change point (CVCP) (64.1%), Expectation-maximization and binary clustering (EMbC) (63.2%), k-means clustering (kmC) (62.7%), and First-passage time (FPT) (61.5%), though post-hoc tests indicated that FPT and CVCP did not significantly differ from HMM (Wilcoxon paired test with Bonferroni correction: FPT vs. HMM p = 1; CVCP vs. HMM, p = 0.15; Fig. 4b)
Summary
Multiple methods have been developed to infer behavioral states from animal movement data, but rarely has their accuracy been assessed from independent evidence, especially for location data sampled with high temporal resolution. Such area-restricted search (ARS) behavior has been predicted from optimal foraging theory and observed in several animals that exploit patchy resources [5, 6] This example illustrates the basic approach underlying path segmentation methods, which detect patterns in movement to provide insight into underlying behavioral states and partition tracks into segments of distinct behavioral states [7, 8]. These methods typically fall into three categories: pattern description, process identification, or change point detection [7]. Correlated velocity movement (CVM) models use continuous time to allow for irregularly sampled data and can be combined with change point detection (CVCP) [16]
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have