Assessing the utility and limitations of accelerometers and machine learning approaches in classifying behaviour during lactation in a phocid seal

Abstract

Background: Classifying behaviour with animal-borne accelerometers is quickly becoming a popular tool for remotely observing behavioural states in a variety of species. Most accelerometry work in pinnipeds has focused on classifying behaviour at sea often quantifying behavioural trade-offs associated with foraging and diving in income breeders. Very little work to date has been done to resolve behaviour during the critical period of lactation in a capital breeder. Capital breeding phocids possess finite reserves that they must allocate appropriately to maintain themselves and their new offspring during their brief nursing period. Within this short time, fine-scale behavioural trade-offs can have significant fitness consequences for mother and offspring and must be carefully managed. Here, we present a case study in extracting and classifying lactation behaviours in a wild, breeding pinniped, the grey seal (Halichoerus grypus). Results: Using random forest models, we were able to resolve 4 behavioural states that constitute the majority of a female grey seals' activity budget during lactation. Resting, alert, nursing, and a form of pup interaction were extracted and classified reliably. For the first time, we quantified the potential confounding variance associated with individual differences in a wild context as well as differences due to sampling location in a largely inactive model species. Conclusions: At this stage, the majority of a female grey seal's activity budget was classified well using accelerometers, but some rare and context-dependent behaviours were not well captured. While we did find significant variation between individuals in behavioural mechanics, individuals did not differ significantly within themselves; inter-individual variability should be an important consideration in future efforts. These methods can be extended to other efforts to study grey seals and other pinnipeds who exhibit a capital breeding system. Using accelerometers to classify behaviour during lactation allows for fine-scale assessments of time and energy trade-offs for species with fixed stores.

Coping styles in capital breeders modulate behavioural trade-offs in time allocation: assessing fine-scale activity budgets in lactating grey seals (Halichoerus grypus) using accelerometry and heart rate variability

“…It remains to be seen whether either strategy may impact energy leftover for lactation and how this may affect pup development. Locomotory behaviours were poorly identified in the current study design (see Shuert et al 2018) which prevents an accurate assessment of movement across the colony to and from freshwater sources, without the aid of additional sensors like GPS or magnetometers for dead-reckoning.…”

“…Grey seal activity was determined using head-mounted datalogging tri-axial accelerometers (AXY-Depth, Technosmart Europe, Italy) as described in Shuert et al (2018). Over two consecutive breeding seasons, 38 adult female grey seals were equipped with accelerometers for the core duration of lactation while on land.…”

“…Accelerometer data were transformed and summarized by 33 feature variables that characterize various elements of static and dynamic acceleration, corresponding to body position and movement, respectively, as well as elements of power spectrum density and frequency within the signal (Shuert et al 2018). These variables were used to classify the entirety of the accelerometry data for the duration of the deployed time for each individual into 6 behavioural states at the resolution of 1-s.…”

“…The behavioural states included were Resting, Alert, Flippering pup, Locomotion, Presenting/Nursing, and Comfort Movements. Video footage was used to label accelerometry data to train a random forest model that was then used to classify each female's activity budget as detailed in Shuert et al (2018). It was not possible to record data blind because our study involved focal animals in the field.…”

“…1). For the purposes of this analysis, extrinsic fixed effects included daily maximum temperature Table 1 Description of behaviour states for assessing time trade-offs in lactating female grey seals, a summary of static heave acceleration (stZ; median with 1st and 3rd quartile in brackets) found to be the most important variable for classifying behaviour in head-mounted accelerometers, as well as the mean proportion (± SD) of time spent in each behavioural state across deployments as classified through accelerometry (behaviour definitions and acceleration summary adapted from Shuert et al 2018 (DMT), daily mean wind speed (Mwind), and total daily precipitation (Precip). We included date, modelled as a polynomial of days since October 1st, to represent seasonal changes in colony density which roughly follows a negative second order polynomial pattern, with the highest density mid-way through the season.…”