Many fields within animal biology and notably studies in animal behaviour and movement ecology have benefited greatly from the development of advanced biologging techniques (1–3). The improvements in biologging since its inception in 1964 when G.L. Kooyman measured dive depth and duration of Weddell seals Leptonychotes weddelli (4) are manifold, most of which can be grouped into four categories. First, tracking of animal movements became more automated. While initially relying on hand-held receivers, the data logged by biologging devices has evolved to now being relayed automatically by satellites (e.g., 5), mobile networks (e.g., 6) or dedicated networks of receiver stations (e.g., 7). Second, while initially only large animals could be tracked, using e.g. 1.63 kg transmitters on caribou Rangifer tarandus (8), development of miniature tags now enables researchers to track increasingly smaller organisms such as songbirds (e.g., 9, 10). Third, the spatial resolution of tag fixes is becoming increasingly accurate. For example, the positioning accuracy of Global Positioning System (GPS) tags is typically under 10 m (11). In addition, nowadays, tags can not only log spatiotemporal information of animals, but also record animal physiology, behaviour and ambient environmental information by various on-board miniature sensors (12). Among the sensors, accelerometer (ACC) data has been mainly used to study animal behaviours and energy expenditure (e.g., 13). However, large amount of raw ACC data can be a burden to limited on-board storage or device battery capacity when data have to be transmitted remotely (14).
On-board data processing, to shrink raw data volume, is one way to solve the aforementioned constraints in remote behavioural data collection and transmission using ACC data. Popular ACC data processing procedures, be it used on-board of tracking devices or during post-processing, involve the calculation of activity indices over pre-defined time windows of ACC data such as ODBA (Overall Dynamic Body Accelration, e.g., 15), VeDBA (Vectorial Dynamic Body Acceleration, e.g., 16) and RMS (Root Mean Square, e.g., 17). These indices have been mostly used to represent activity levels of tracked animals (e.g., RMS; 18) or to study animals’ energy expenditure (e.g., VeDBA; 19). When these indices were calculated on-board of trackers, the research period could be extended considerable, even when using small trackers on small animals. As a prime example, Bäckman, Andersson (20) used data loggers that summarized single axis ACC data into 12 activity indices (by scoring relative intensity of ACC data) per hour to study migration and activity patterns of the red-backed shrike Lanius collurio for up to 14.5 months. Similar data processing procedures have also been used in biomechanics studies using ACC data. For example, indices such as ODBA have been used to evaluate fish swimming behaviour and effort (21). Also, tail beat frequencies derived from ACC data were used to approximate swimming speed in sailfish Istiophorus platypterus (22). Yet other on-board ACC data processing focussed on the animal’s biomechanics were used by Cox, Orgeret (23), summarizing ACC data on-board of satellite transmitters to study swimming effort and pitch angle (body angle relative to the horizontal plane) of southern elephant seals Mirounga leonine.
Accelerometers have also been widely used to study animal behaviours. ACC data is translated into behaviour using either unsupervised or supervised machine learning. Supervised behaviour classification relies on direct behavioural observations to assist in the translation process of the ACC data, whereas unsupervised behaviour classification relies on the ACC data itself and post-hoc behavioural interpretation of the ACC classes. Many supervised behaviour classification studies make use of machine learning where a range of summary statistics or “features” are calculated from the segmented ACC data. Next, using segments for which also direct behavioural observations are available, a behaviour classification model is “trained” using machine learning. In unsupervised behaviour classification studies, ACC data is clustered based on similarities in the ACC data after which expert opinion is used to allocate a likely behaviour to each cluster. Commonly, unsupervised behaviour classification consists of processing ACC data by wavelet transformation followed by k-means clustering (e.g., 24). Several studies applied on-board ACC data processing for animal behaviours and demonstrated their advantages. Nuijten et al. (2020) calculated raw ACC data into features on-board of trackers to study behaviours of Bewick's swans Cygnus columbianus bewickii. In this way, they were able to sample one 2 second bout of ACC data every 2 minutes. Korpela et al. (2020) used on-board behaviour classification through ACC data to control data sampling of the more energy consuming on-board camera to extend the runtimes of field experiments. A very recent developed tracking system (25) proved that on-board continuous behaviour classification from ACC data was energy, weight and cost efficient and allowed for continuous recording of behaviour that could be transmitted through the 3G mobile network on a daily basis.
We used this energy-efficient, solar-powered system on six free-ranging Pacific black ducks Anas superciliosa to evaluate some of the improvements this system could achieve over trackers recording intermittent or no behavioural data. The trackers continuously recorded eight different behaviours complemented with hourly GPS positions. Aside from providing general performance data for these six individuals, we provide in-depth analyses covering 72 days of continuous behavioural and hourly positional data for one individual. To illustrate how continuous on-board classification of animal behaviours can benefit animal-behaviour and movement-ecological studies, we used the data thus obtained to compare time-activity budgets using continuous versus intermittently sampled ACC data. We also compared energetically costly displacement behaviour estimates calculated from continuous behaviour records with calculations based on hourly positional data in isolation. Finally, we used the behaviour records for a more in-depth home range estimation, allowing focussing on which parts of their environment animals display specific (critical) behaviours and when.
Home range is a concept in animal ecology that has found wide use since the beginnings of animal telemetry (26). In most studies, home range is typically estimated based on animals’ geographical locations over time. Although these methods estimate the geographical distribution of animals and how that may vary over time, they may not be able to inform on how and why animals use their home range to satisfy their (daily, annual and life cycle) needs. Instead of using time spent in spatial locations for home range description, Powell and Mitchell (27) proposed that more metrics can be used to better describe animals’ home range. For this they suggested a range of potential metrics such as energy expenditure, energy gained, giving-up densities for resources, danger from predators, potential for competition, potential for mutualisms and access to mates. Following some of the suggestions of Powell and Mitchell (27) to better assess the environmental requirements of animals, we calculated and depicted an animal’s home range integrating not only spatial and temporal information from GPS fixes, but also using energy expenditure and detailed behaviour information from ACC data. Aside from illustrating these advancements, we also discuss additional ways in which on-board data processing of ACC data can assist in studying and conserving free-ranging animals in their natural environments.