Clustering and classification of vertical movement profiles for ecological inference of behavior

Abstract

Vertical movements can expose individuals to rapid changes in physical and trophic environments—for aquatic fauna, dive profiles from biotelemetry data can be used to quantify and categorize vertical movements. Inferences on classes of vertical movement profiles typically rely on subjective summaries of parameters or statistical clustering techniques that utilize Euclidean matching of vertical movement profiles with vertical observation points. These approaches are prone to subjectivity, error, and bias. We used machine learning approaches on a large dataset of vertical time series (N = 28,217 dives) for 31 post‐nesting leatherback turtles (Dermochelys coriacea). We applied dynamic time warp (DTW) clustering to group vertical movement (dive) time series by their metrics (depth and duration) into an optimal number of clusters. We then identified environmental covariates associated with each cluster using a generalized additive mixed‐effects model (GAMM). A convolutional neural network (CNN) model, trained on standard dive shape types from the literature, was used to classify dives within each DTW cluster by their shape. Two clusters were identified with the DTW approach—these varied in their spatial and temporal distributions, with dependence on environmental covariates, sea surface temperature, bathymetry, sea surface height anomaly, and time‐lagged surface chlorophyll a concentrations. CNN classification accuracy of the five standard dive profiles was 95%. Subsequent analyses revealed that the two clusters differed in their composition of standard dive shapes, with each cluster dominated by shapes indicative of distinct behaviors (pelagic foraging and exploration, respectively). The use of these two machine learning approaches allowed for discrete behaviors to be identified from vertical time series data, first by clustering vertical movements by their movement metrics (DTW) and second by classifying dive profiles within each cluster by their shapes (CNN). Statistical inference for the identified clusters found distinct relationships with environmental covariates, supporting hypotheses of vertical niche switching and vertically structured foraging behavior. This approach could be similarly applied to the time series of other animals utilizing the vertical dimension in their movements, including aerial, arboreal, and other aquatic species, to efficiently identify different movement behaviors and inform habitat models.

Incorporating multidimensional behavior into a risk management tool for a critically endangered and migratory species

“…(a) For eastern Pacific leatherback turtles, individual means and standard deviations for hidden Markov model (HMM) parameters for move persistence velocity (μ) as a function of the parameters for the proportion of deep D1 (Barbour et al., 2023) dives ( P ) and (b) monthly proportions of HMM‐derived behavioral states ( S1, transiting behavior; S2, residential or foraging behavior; S3, deep diving and exploratory behavior) for 28 leatherback turtles in the eastern Pacific.…”

“…The results from our HMMs and monthly UDs showed EP leatherbacks have multidimensional behaviors varying in space and time and latitudinal differences in monthly space use and identified behaviors (S1, S2, S3) consistent with past findings. Previous dynamic time warp clustering analyses showed that dives for turtles in this population can be classified into unique categories (D1 [deeper dives] or D2 [shallower dives]), each representing different vertical behaviors (Barbour et al, 2023). The S3 behavioral state identified by the HMMs likely represents the use of deep dives (D1) to shed excess heat gained from traveling through warm equatorial waters (Okuyama et al, 2021;Shillinger et al, 2010Shillinger et al, , 2011Wallace & Jones, 2008) and prey search in the nutrient-poor SP Gyre, where zooplankton prey is dispersed at depth (Saba et al, 2008;Shillinger et al, 2011;Stromberg et al, 2009).…”

“…The S3 behavioral state identified by the HMMs likely represents the use of deep dives (D1) to shed excess heat gained from traveling through warm equatorial waters (Okuyama et al, 2021;Shillinger et al, 2010Shillinger et al, , 2011Wallace & Jones, 2008) and prey search in the nutrient-poor SP Gyre, where zooplankton prey is dispersed at depth (Saba et al, 2008;Shillinger et al, 2011;Stromberg et al, 2009). The S2 behavioral state is likely more of a traditional foraging or residential state, where turtles are using slower speeds and more tortuous movements to perform diurnal dives (deeper D1 dives during the day, shallower D2 dives at night) to forage on prey patches near more productive gyre areas (Bailey, Fossette, et al, 2012;Fernandez-Alamo & Farber-Lorda, 2006;Okuyama et al, 2021;Saba et al, 2008;Shillinger et al, 2011;Barbour et al, 2023).…”

“…Sea Mammal Research Unit Satellite Relay Data Logger tags transmitted surface locations via the ARGOS satellite system. These provided additional pressure‐sensor dive information (e.g., total dive duration, maximum dive depth, and dive profiles, including intermediate depth and duration points), with dives categorized as vertical movements when depth reached was >10 and <1500 m. Each track was processed to obtain daily locations with current‐corrected persistence velocities (Copernicus Marine Service [https://marine.copernicus.eu/]), total dive counts, and counts of deep versus shallow dive types (D1, deep; D2, shallow [Barbour et al, 2023]) (version 4.2.0 [R Core Team, 2022]) (further processing details in Appendix S2).…”

“…Slow, more tortuous movements can result in longer residence times and increased area usage that increase the risk of interaction with fishing vessels in the same area; dive depths can increase the risk if turtles are diving at the same depth and area as fishing gear. Each state was defined by a unique set of values for mean daily move persistence velocity (horizontal movement, the tendency of turtles to persist in a given direction, obtained by multiplying daily swimming speeds by the cosine of the turning angles), mean daily number of dives, and daily proportion of D1 (deeper) and D2 shallower) (Barbour et al, 2023) dives. The HMMs were fit with a Markov chain Monte Carlo algorithm and were implemented using the Bayesian programing language Stan and its R interface, rstan (Stan Development Team, 2022).…”