Elsevier

Ecological Informatics

Volume 60, November 2020, 101152
Ecological Informatics

Accounting for unknown behaviors of free-living animals in accelerometer-based classification models: Demonstration on a wide-ranging mesopredator

https://doi.org/10.1016/j.ecoinf.2020.101152Get rights and content

Highlights

  • Accelerometers are used to classify behaviors of unobservable animals.

  • If some behaviors are unknown, accelerometers classify incorrectly.

  • We use a probabilistic classifier to evaluate and address this problem.

  • We simultaneously develop a high-accuracy modeling framework to classify behaviors.

Abstract

Describing the behaviors of free-living animals is broadly useful for ecological and physiological research, but obtaining accurate records for difficult-to-observe species presents a considerable challenge. Tri-axial accelerometers are increasingly used for this purpose by exploiting behavioral observations from accelerometer-carrying animals to predict behaviors of unobserved conspecifics. We developed a modeling approach to predict behaviors of wolverines from collar-mounted accelerometers using Support Vector Machines. By applying a temporal smoothing function and setting a lower threshold for a-posteriori prediction probabilities, we improve the predictive performance of our model and simultaneously create a framework for explicitly accounting for behaviors unknown to the model, a problem that remains largely unaddressed in similar studies. We demonstrate that such an approach can achieve a model-averaged accuracy of 98.3%, with high predictive performance for the behaviors resting, running, scanning, tearing at food, and transferring items with the mouth, a behavior typically associated with caching food among captive wolverines. To illustrate the utility of this approach, we apply this model to a sample of seven free-living wolverines in Arctic Alaska.

Introduction

Describing the behaviors of free-living animals can provide important insights regarding a wide range of ecological processes. Taken alone, analysis of such behavioral records can be used to investigate temporal patterns in activity, including association among behaviors, yielding insights regarding circadian rhythms in the daily partitioning of behaviors or inter-individual differences in such temporal patterns (Garthe et al., 2003; Yoda and Ropert-coudert, 2007). When coupled with environmental and physiological information, behavioral analyses can address how such extrinsic and intrinsic factors influence behavioral decisions made by animals, including tradeoffs such as allocating time between foraging and antipredator behavior (Hamel and Côté, 2008; Studd et al., 2019; Switalski, 2003).

However, since documenting behavior has traditionally relied on direct observation, it is often a difficult or impossible task to assemble comprehensive records for remotely tracked free-living animals that have not been directly observed. Species that occupy areas that are remote or logistically difficult for human observers to access, such as under water, under snow, or in trees, present obvious challenges, as do species that range widely, travel quickly, or for which human observation alters behavior.

The rise of accelerometer-derived behavioral records promises to reduce these obstacles (Shepard et al., 2008). This process, whereby free-living animals are tagged with tri-axial accelerometers and the resulting data are used to predict the behaviors of the wearer, has been applied to a variety of marine (Battaile et al., 2015; Viviant et al., 2010; Whitney et al., 2010), and increasingly, terrestrial species (Hammond et al., 2016; McClune et al., 2014; Pagano et al., 2017; Wang et al., 2015). Resulting behavioral records have been used to investigate behaviors important to life history and fitness, including predation and mating events, and foraging strategies.

Using accelerometer data to classify behavior typically begins by building a predictive classification model based on observer-labeled accelerometer data. The labeled data are collected either by directly observing conspecific or surrogate species while wearing accelerometers (Campbell et al., 2013), or with the use of additional biologgers, such as video cameras, affixed to free-living individuals (Nakamura et al., 2015; Pagano et al., 2017; Watanabe and Takahashi, 2012). A classification model, such as a statistical learning classifier (Tatler et al., 2018) or decision tree analysis (Bellsolá, 2019) can then be applied to the labeled data to train and evaluate candidate models, after which the final model can be applied to free-living individuals where no direct observations or ancillary data for determining behaviors are available.

Here, we developed and evaluated the first predictive model that can be used to classify behaviors of free-living wolverines (Gulo gulo) using collar-mounted tri-axial accelerometers, based on visual observations of captive wolverines wearing similar collar-mounted accelerometers. Further, we used labeled accelerometer data from these captive conspecifics to create a framework by which behaviors not exhibited by the captive wolverines, and therefore unknown to the model, would be classified as “unknown,” rather than incorrectly classified to the best fitting known acceleration pattern. By developing such a model, we hoped to broaden the field of possible questions that can be addressed regarding the interactions of the environment, physiology, and ecology of wolverines, and provide a framework that other researchers can employ for other species to address similar questions while explicitly addressing the problem of incorrect attribution for behaviors unknown to the model. Finally, to demonstrate the utility of our modeling approach, we applied several candidate models to a small sample of free-living wolverines and assessed temporal trends of resting, running, vigilance behavior, and behaviors associated with handling food.

Section snippets

Methods

A schematic outlining the workflow is included in Fig. 1.

Model development and evaluation

Model tuning yielded optimum values of gamma = 0.01 and cost = 10. The base model had an overall predictive accuracy of 94.6% (95% CI: 93.2–95.8%) and correctly classified 326 (95% CI: 288–360) of the 433 observations in the 30% portion of the full training dataset used for testing (75.4%, 95% CI: 66.5–83.1%). Performance for individual behaviors ranged from a precision of 0 to 0.98, and recall ranged from 0 to 0.97 (Table 3). The “base + smoothing” model had an overall accuracy of 95.8% (95%

Discussion

Classifying behaviors from accelerometer data is an increasingly popular technique for addressing questions relating to the ecology and physiology of free-living animals. Considerable progress has been made in the field, particularly in evaluating the performance of different classification models (Nathan et al., 2012; Tatler et al., 2018) and the integration of multiple data sources, such as GPS and acoustic recorders, with acceleration to predict behavior (Shamoun-Baranes et al., 2012; Studd

Funding

This work was supported by the M.J. Murdock Charitable Trust, Wilburforce Foundation, the University of Alaska Fairbanks (UAF) Erich Follmann Memorial Student Research Fund, a UAF College of Natural Science and Mathematics Travel Grant, Wildlife Conservation Society, 69 generous individuals via a crowdfunding campaign, and a National Science Foundation Graduate Research Fellowship under Grant No. 1650114.

Declaration of Competing Interest

The authors declare no competing interests.

Acknowledgements

We thank the staff at Nordens Ark, in particular E. Nygren and E. Andersson, for providing crucial access to captive wolverines, logistical support for wolverine immobilizations, and accommodations for researchers. We thank the staff at Toolik Field Station, particularly J. Timm, as well as M. Kynoch, C. Haddad, and S. Andersen with Wildlife Conservation Society for logistical support in the field. We are grateful to J. McIntyre for feedback on early drafts of this manuscript and statistical

References (46)

  • O.R. Bidder et al.

    Love thy neighbour: automatic animal behavioural classification of acceleration data using the k-nearest neighbour algorithm

    PLoS One

    (2014)
  • E.O. Brigham et al.

    The fast fourier transform

    IEEE Spectr.

    (1967)
  • C.L. Buck et al.

    Annual cycle of body composition and hibernation in free-living arctic ground squirrels

    J. Mammal.

    (1999)
  • H.A. Campbell et al.

    Creating a behavioural classification module for acceleration data: using a captive surrogate for difficult to observe species

    J. Exp. Biol.

    (2013)
  • H. Cao et al.

    An integrated framework for human activity recognition

  • M. Chimienti et al.

    The use of an unsupervised learning approach for characterizing latent behaviors in accelerometer data

    Ecol. Evol.

    (2016)
  • H.E. Chmura et al.

    Biologging physiological and ecological responses to climatic variation: new tools for the climate change era

    Front. Ecol. Evol.

    (2018)
  • S. Garthe et al.

    Temporal patterns of foraging activities of northern gannets, Morus bassanus, in the northwest Atlantic Ocean

    Can. J. Zool.

    (2003)
  • A.C. Gleiss et al.

    Making overall dynamic body acceleration work: on the theory of acceleration as a proxy for energy expenditure

    Methods Ecol. Evol.

    (2011)
  • H.N. Golden et al.

    Immobilization of wolverines with Telazol® from a helicopter

    Wildl. Soc. Bull.

    (2002)
  • S. Grünewälder et al.

    Movement activity based classification of animal behaviour with an application to data from cheetah (Acinonyx jubatus)

    PLoS One

    (2012)
  • T.T. Hammond et al.

    Using accelerometers to remotely and automatically characterize behavior in small animals

    J. Exp. Biol.

    (2016)
  • C. Hsu et al.

    A Practical Guide to Support Vector Classification

    (2010)
  • Cited by (6)

    • Wolverines (Gulo gulo) in a changing landscape and warming climate: A decadal synthesis of global conservation ecology research

      2022, Global Ecology and Conservation
      Citation Excerpt :

      Early wolverine research employed radio telemetry, snow-tracking, and den-site studies; these revealed rich details on wolverine movement, resource selection, and natality from focal locales (Banci, 1987; Copeland et al., 2007; Hornocker and Hash, 1981; Krebs et al., 2004; Lofroth and Krebs, 2007; Whitman et al., 1986) but spatially and temporally limited by logistics. Global Positioning Satellite (GPS) collars provided greater opportunities for behavioral study, and new technologies such as accelerometers are emerging to measure activity and energetics in the wild (Glass et al., 2020). Advances in DNA analysis allowed scientists to ask increasingly sophisticated questions about wolverine genetics (Kyle and Strobeck, 2001; Sawaya et al., 2019; Schwartz et al., 2009) and evolution (Krejsa et al., 2021; Rico et al., 2015).

    • Trade-off between predation risk and behavioural thermoregulation drives resting behaviour in a cold-adapted mesocarnivore

      2021, Animal Behaviour
      Citation Excerpt :

      Since this method only classifies behaviours that are exhibited continuously for at least 10 s, grouping behaviours that occur at shorter time intervals as ‘unknown’, estimates for total time exhibiting any given behaviour are biased low. Nevertheless, the model performs quite well at predicting 10 s bouts of vigilance behaviour (94% precision, termed ‘scanning’ in Glass et al., 2020, p. 4), defined as ‘surveying [the] surroundings by moving [the] head while torso and legs remain stationary’. Therefore, we applied this model to the seven wolverines from whom 10 Hz accelerometer data were collected and calculated the number of 10 s periods spent vigilant in the 5 min preceding and following each resting period (hereafter ‘vigilance’).

    View full text