Using accelerometers to develop time-energy budgets of wild fur seals from captive surrogates

Behaviour validation experiments with captive fur seals and sea lions

To validate the use of accelerometers for classifying behaviours of wild fur seals we used captive surrogates at three Australian marine facilities; Dolphin Marine Magic Coffs Harbour, Underwater World Sunshine Coast, and Taronga Zoo Sydney, from August to November 2014 and in August 2015. We used two adult Australian fur seals (Arctocephalus pusillus doriferus; one male, one female), three New Zealand fur seals (Arctocephalus forsteri; two male adults, one male juvenile), one subantarctic fur seal (Arctocephalus tropicalis; juvenile male), and six Australian sea lions (Neophoca cinerea; two adult males, three adult females, one juvenile female). Fur seals had accelerometers attached with tape to the fur, while sea lions wore a custom fitted harness with the accelerometer sewn into a pocket (for details of the animals used see Ladds et al., 2017b, Table 1). We pooled data from all four species for training machine learning models because fur seals and sea lions (otariids) are conservative in morphology and including species and/or attachment method as a factor only improves the accuracy of such models by ∼5% (Ladds et al., 2017b). This study was conducted under permits from Macquarie University ethics committee (ARA-2012_064) and Taronga ethics committee (4c/10/13).

Tri-axial accelerometers (CEFAS Technology Ltd, Lowestoft, UK) were set to record −8 g to +8 g at 1 Hz and 25 Hz simultaneously, with a wet/dry sensor active, and behaviours typical of wild fur seals were video-recorded during training sessions. We observed two types of sessions; feeding and behaviour. The feeding sessions aimed to provide seals with large food items that required some form of processing prior to eating. Behaviour sessions also incorporated some feeding events with small fish that did not require processing. Fish were thrown in the pool so that seals had to “capture” them mid-water as they sank. These two behaviours constituted foraging. During each behaviour session seals were instructed to perform a series of natural behaviours from their known behavioural repertoire, such as porpoising, swimming and grooming. While seals could not reach depths they would achieve in the wild, their pools provided adequate space to perform behaviours typical of wild fur seals. Captive fur seals had access to both land and water during trials, similar to a wild fur seal near their haul-out. In addition, seals were trained to swim consistently below the water for several minutes to replicate a deep dive for another project. The experimental set-up and training allowed us to record behaviours that lasted from less than a second (grabbing fish from the water column) to several minutes (swimming or resting).

Behaviours were manually matched to the accelerometry by two investigators. Where behaviours recorded by the investigators did not match, they reviewed the video such that both reached agreement. Twenty-six behaviours were grouped into four behavioural categories—foraging, travelling, grooming and resting, in three locations—land, water surface and underwater (for details of the behaviours and their groups see Ladds et al., 2016, Table 2). Foraging behaviours consisted of searching for prey and prey handling limited to dead fish. Grooming was any behaviour used in body maintenance or thermoregulation. To thermoregulate at sea, fur seals float with either their hind flippers (jughandling) or their fore flippers (sailing) in the air. Grooming involves the use of flippers to scratch or rub the body, including the face to clean whiskers. Resting was any period of stillness, while travelling was any period involving movement that was not foraging or grooming (Ladds et al., 2017b).

Estimate the energetic cost of each behaviour

Wild fur seal data collection

We tracked juvenile Australian fur seals from two colonies, Seal Rocks (Phillip Island, Victoria Australia, 38° 52′S–145°11′E, n = 6) during austral winter of 2013 and Lady Julia Percy (Victoria, Australia, 38°52′S–142°00′E, n = 8) during austral summer of 2014. These sites are the largest breeding colonies for Australian fur seals, with each site containing approximately 25% of the total population of the species (Kirkwood et al., 2010). In Australian fur seals, suckling ceases after 1 year, puberty occurs in females at approximately 3 years old and in males at 4–5 years old (Arnould & Warneke, 2002). We used animals between one and three years of age that were independently foraging (i.e., juveniles) for this study.

We identified juveniles by their mature pelage (i.e., lacking the lanugo of pre-moult pups) and size (<1.5 m and 40 kg) and captured individuals using a modified hoop-net and isofluorane gas sedation (Gales & Mattlin, 1998), then measured standard length (straight-line), girth and mass. Numbered tags were applied to the trailing edge of both fore-flippers (Super Tags^®, Dalton I.D. Systems Ltd, Henley-on-Thames, UK) to aid with identification and recapture. One of two types of location device (Kiwisat100, Sirtrack Ltd, New Zealand or Mk10; Wildlife Computers), a VHF transmitter (Sirtrack Ltd, 6 cm × 3  cm × 2 cm), and a tri-axial accelerometer G6A + (CEFAS technology Ltd, Lowestoft, UK) were glued directly to the fur on the dorsal midline of each fur seal (Fig. 1) using quick-setting epoxy (Araldite 2017; Aeropia Ltd, Crawley, UK or Araldite 268; Huntsman Advanced Materials, Victoria, Australia). A time-depth recorder (TDR) was also attached, either as part of the Mk10 device or a separate device (Mk9, Wildlife Computer; Fig. 1). The total instrument package mass (179 g–239 g) equated to <1% of the seal’s body mass and attached to maintain the lowest profile possible to minimise a drag effect. We observed animals until they had fully recovered from anaesthesia and released them at the site of capture. A minimum of 15 days lapsed before recapture (via hoop net and manual restraint), and devices were retrieved by cutting the hair beneath the glued instrument.

For the duration of the deployment, defined as from attachment until removal of a device, TDRs recorded depth (m) every second, and any drift in the depth sensors or error spikes were corrected prior to analyses using Zero-Offset Correction (Wildlife Computers, Redmond, WA, USA). Tri-axial accelerometers recorded acceleration on the X, Y, and Z axes at 1 sample per second (1 Hz) and temperature at 0.5 Hz. Accelerometers also recorded at a high resolution (20 Hz) when diving (depth > 1.5 m). This would sometimes continue after a diving event, giving high resolution data both at the surface and during diving.

To give an indication of time duration at sea, data from the TDRs were summarised into trips and dives. Trips started when a seal entered the water and ended when the seal hauled out, and excluded periods in water with minimal diving or dives <10 m (e.g., when seals rest at the water surface adjacent to colonies). Dives were defined as periods spent underwater below a minimum depth of 5 m to account for wave action at the surface. Fur seal physical parameters, trip duration, number of trips, average dive duration, mean maximum depth and maximum depth were derived from these parameters (Table S1).

Predicting behaviours of wild fur seals from accelerometers

As the high resolution data were only recorded for wild seals while at sea, the high resolution captive data were subsampled to include only those behaviours that occurred in the water. Low resolution (1 Hz) data were recorded continuously for both the wild and captive seals. The wet/dry sensor of the accelerometers from the wild fur seals (hereafter wild data) was used to indicate when individuals were in water or on land to improve the predictability of the models. We tested the accuracy of both high and low resolution accelerometry to classify behaviours.

To determine the behavioural state of surrogates using accelerometers, gradient boosting models (GBM) were trained in R using the package ‘xgboost’ (Chen, He & Benesty, 2016). GBM models are an extension of a random forest, whereby they build a classification tree on a subset of the data, then use a subsequent tree to learn from the errors of the previous trees. Trees are built successively until a stopping criteria is met, and the trees built are averaged together to provide an estimate of classification (Friedman, 2002). Captive data were pooled and split into one of three epochs (the number of samples on which summary data are calculated) for training the GBM. For high resolution (20 Hz) data epochs of 13, 25 and 75 samples were tested, which correspond to 0.65, 1.25 and 3.75 s of data, respectively. For low resolution (1 Hz) data epochs of 7.0, 15.0 and 21.0 samples were tested, which corresponded to 7.0, 15.0 and 21.0 s of data respectively. Training and testing longer epochs was not possible because there were too few events lasted for longer than a few seconds, meaning there were not enough examples to train a model. Down-sampling (randomly selecting behaviours from a pool until a specified number is met) was used to ensure that the behaviour categories had an even number of samples (Ladds et al., 2017b).

We coded 52 summary statistics and added five covariates describing some characteristic of the individual or the event to the second stage of model testing. These were included as they have previously been demonstrated to make a small improvement on prediction performance of the models (Ladds et al., 2017b). The covariates included were device attachment method (harness or tape), age, mass, sex and species of the individual. We included where the behaviour occurred (surface, underwater or land) in all models. Location was determined first by the wet/dry switch which indicated whether a seal was in the water or on land, then once a seal was more than 1 m under the surface (as determined by the depth device) they were classified as underwater. Summary statistics calculated included: mean, median, standard deviation, skewness, kurtosis, minimum, maximum, absolute value, inverse covariance, autocorrelation trend (the coefficient derived from a linear regression) for each of the three axes. We also calculated q as the square-root of the sum-of-squares of the three axis (Nathan et al., 2012), and included pair-wise correlations of the three axes (x–y, y–z, x–z) (Ravi et al., 2005). The inclination as azimuth were calculated as per Nathan et al. (2012). We calculated three measures of dynamic body acceleration (DBA) by first using a running mean of each axis over 3 s to create a value for static acceleration. We then subtracted the static acceleration at each point from the raw acceleration value to create a value for partial dynamic body acceleration (PDBA). The values of PDBA on each axis were used to calculate overall dynamic body acceleration (ODBA) (Shepard et al., 2008a; Wilson et al., 2006) and vectorial dynamic body acceleration (VeDBA) (Qasem et al., 2012). The integral of the start and end point of ODBA and VeDBA for each epoch ODBA and VeDBA was calculated using the package “MESS” in R (Ekstrom, 2014, R Core Development Team, 2015).

GBM models were run with the full suite of summary statistics derived from captive data and run over a grid of parameters (for details see Table 2 and additional file 2 in Ladds et al., 2017b). The combination of parameters that resulted in the highest accuracies was chosen for implementation on the high and low resolution wild data. Within each epoch, wild data were categorized using predictions from the GBM model built with captive data that produced the highest cross-validation accuracy and kappa values. Accuracy is a measure of the proportion of true positives identified by the model, while kappa is a performance measure that accounts for investigators’ observations agreeing or disagreeing by chance. Behaviour events were categorised for the duration of each deployment. Events were considered different when either the location or the behaviour category changed for an epoch, and the change occurred for longer than 15 s. A sensitivity analysis was conducted on the probability of each event being assigned to a behaviour group. For wild data, each event was assigned a probability of it being each behaviour category and then classified as the behaviour that had the highest probability. To evaluate how well our models classified behaviours, behaviours that were selected with less than 80% chance of occurring were examined, and the behaviour with the next highest probability was extracted. This allowed us to see when the model may have ‘confused’ two categories.

Apply the model to create time-energy budgets

To build time-energy budgets, we estimated DEE (MJ) which was assumed to be a function of the energy expenditure (EE) of a given behaviour event, its duration, the season it occurred in and where it occurred (land or water) summed over 24 h periods (Table 1, Eq (5)). An example of how this is calculated over a single dive is given in Fig. 2. Details of the calculations and assumptions made for the energetic models are in Table 1. The overall energetics model is defined by the sum of all of the behaviour events (e = e…E) that occur in a 24 h period (from midnight to midnight) for A (activity), G (grooming) and R (resting) as a function of season s ( s = winter, summer) and location l ( l =  land, water): (5)${\displaystyle \mathrm{DEE}\left(\mathrm{l}{\mathrm{O}}_2\right)=\sum _e^E{R}_{s,l}+\sum _e^E{A}_{s,l}+\sum _e^E{G}_{s,l}.}$

See Table 1 for definitions and details of each behaviour state.

For reporting and comparison with other energetics papers, the total oxygen used was converted into MJ. First, the total energy expended was converted to kilocalories using a factor of 5 kcal per l O₂, then converted to kilojoules using a conversion factor of 4.186  J  cal⁻¹ (Williams et al., 2007).

Statistical analysis

As it was not feasible to validate the captive model after applying it to the wild data, we conducted a sensitivity analysis on the probability of each event being assigned to a behaviour group. To do so, we adjusted the proportion of time wild fur seals spent in different behaviours (increasing some while decreasing others, such that the total proportion of time spent in the three behaviours always summed to 1) and investigated how it affected the overall DEE. We calculated 1,000 simulated proportions based on the actual range of time spent in each behavioural category, where each simulated proportion represented a day. We applied the energetics model to the simulated proportions and grouped the activities according to three behaviour categories: active (travelling and foraging) grooming and resting; then by two location categories: water and land. We then plotted the total DEE for the day against the simulated proportions for each of the five categories. As fur seals were of different sexes and were tagged in different seasons we tested for differences in DEE between sexes and seasons using post-hoc general linear hypothesis and a multiple comparison test via the Tukey method within the function glht from the package “multcomp” (Hothorn et al., 2013). Individual fur seal identification was included in models as a random factor and significance was set at p < 0.05. All analyses were completed in R (Version 3.1.3; R Core Development Team, 2015) and values reported as mean ± SEM. The datasets generated and analysed during the current study are available in the “Time-energy_budgets_from_accelerometers” repository: https://github.com/MoniqueLadds/Time-energy_budgets_from_accelerometers.git.

Behaviour validation experiments with captive fur seals and sea lions

Most epochs (99%) were assigned to a behaviour category with over 80% probability (Fig.  3). We investigated those behaviours that were assigned with a less than 80% (∼1% of total epochs) probability to understand where the model may have ‘confused’ behaviours. When the model was uncertain that an epoch was in the behaviour category “foraging” (less than chance), it generally predicted the epoch should be assigned as “travelling” and almost never “resting” or “grooming” (Fig. 3A). When there was uncertainty if an epoch should be categorised as “travelling”, with less than 50% chance, the behaviour category with the next highest probability was “foraging” (Fig. 3B). Grooming was rarely confused for other behaviours, but when there was uncertainty the model calculated “resting” with the next greatest probability (Fig. 3C). There was also little confusion with assigning an epoch to resting, but occasionally the model assigned a higher probability of foraging.

The models correctly classified surrogate behaviour (travelling, foraging, grooming, resting) with high accuracy (>68%), but the number of samples in an epoch used affected the results, where longer epochs (sampling time of behaviour) resulted in higher accuracies (Table 2). The best low resolution model (1 Hz) used 21 samples for an epoch, and the best high resolution model (20 Hz) used 75 samples for an epoch, both of which had the highest training, testing and kappa scores for their behaviour category. Given that the 1 Hz data classified behaviours with very high accuracy (90% out-of-sample using epochs of size 21, Table 2) and recorded for the duration of deployments on wild fur seals (20 Hz data only recorded while fur seals were diving and for a short time after), only the 1 Hz data for the activity budgets were analysed.

Wild fur seal data collection

Three fur seals from Seal Rocks and six fur seals from Lady Julia Percy were successfully recaptured (recapture rates of 50% and 67% respectively) and accelerometer data obtained. Fur seals made between two and 45 trips with durations of between 30 min and nine days. Diving parameters (Table S1) were very similar between individuals from both sites. Fur seals in winter (from Seal Rocks) made fewer and longer foraging trips than fur seals in summer (from Lady Julia Percy). All other diving parameters were very similar between winter and summer fur seals (Table S1). Similarly, there were few differences in diving parameters for male and female juvenile fur seals.

Predicting behaviours of wild fur seals from accelerometers

Figure 4 is an example of the output produced by the activity model for a wild fur seal, showing the end of a foraging bout, travelling back to land and then a short period of resting on land (hauled out). This figure demonstrates the strengths and weaknesses of the GBM built from surrogates. The model was very good at predicting when the wild individual was resting, as there was very little movement in the accelerometer. But this feature resulted in the dive ascent also being classified as resting as the seal rose slowly through the water column with limited body movement (Fig. 4B). Grooming was also classified accurately; it predominantly occurred immediately prior to or following a dive, or during the first hour or so after hauling out. Foraging and travelling were frequently misclassified by the model (Fig. 3); most commonly, the descent of a dive was classified as foraging when it most likely should have been travelling, and foraging appeared periodically during long trips returning to the haul out site.

Apply the model to create time-energy budgets

Overall, the time that fur seals spent in the three different locations was between 31–63% on land, between 3–25% underwater and between 28–47% at the surface of the water. The pressure sensor on the accelerometers on two of the wild fur seals (LJP_A10283 and LJP_A10284) failed for a portion of the deployment, which resulted in a significant underestimation of the time spent underwater. Each fur seal spent approximately half of their deployment resting (range 32–55%), predominantly on land (Fig. 5) and another 22% (range 17–33%) was used for grooming. Approximately 20% (range 13–25%) of fur seals’ time was foraging and approximately 12% travelling (range 8–22%).

There were no significant differences in the DEE for females (18.22 ± 5.91 MJ d⁻¹) and males (18.86 ± 6.01 MJ d⁻¹; post-hoc comparisons: Z =  − 0.35, p = 0.72) or for winter deployments (20.77 ± 7.00 MJ d⁻¹) and summer (17.55 ± 5.30 MJ d⁻¹; post-hoc comparisons: Z = 1.10, p = 0.27), which also represented site and year. Therefore, it was justifiable to pool the samples. The average DEE for wild individuals and locations pooled was 18.73 ± 5.73 MJ d⁻¹ (range: 8.24–32.04 MJ d⁻¹) and mass-specific DEE was 0.50 ±  0.14 MJ kg⁻¹ d⁻¹ (range: 0.08–0.81 MJ kg⁻¹ d⁻¹; Table 3). The maximum DEE was from a wild individual that spent 12 h continuously diving at sea (Fig. S1).

The most energetically expensive behaviour was foraging, making up over a third of the DEE (Table 3). Resting on land made up ∼15% of the overall energetic budget though this was the largest part of the activity budget (∼45%). Fur seals spent little time resting at sea (∼10%) and this behaviour represented ∼9% of the overall energetic budget. The least costly activities were underwater grooming (1.4%) and resting (1.5%) which were likely to be mistaken behaviour classification. Grooming only made up one seventh of DEE, less than resting, and most grooming activity was at the surface (12%). Travelling and foraging made up two-thirds of the energetic budget, though they only represent roughly a third of the activity budget.

The sensitivity analysis revealed that the proportion of time spent in different behaviours (active, grooming or resting) and locations (water or land) altered the expected DEE for a juvenile Australian fur seal. DEE increased with more time spent active where DEE increased up to 27 MJ d⁻¹ when a seal was active more that 50% of the time (Fig. 6A). DEE decreased to 13 MJ d⁻¹ as seals spent more time resting (Fig. 6C). There was no clear relationship between the time spent grooming (Fig. 6B) or time spent on land or in water (Figs. 6D–6E) and DEE.

Behaviour segmenting with accelerometers and machine learning

Supervised machine learning models trained with accelerometry data from captive animals reliably and accurately classified all four behaviour categories tested: foraging, grooming, resting and travelling. We expected that higher resolution data would enable the models to perform better at distinguishing the different behaviour types (Halsey et al., 2009), but were unable to compare this directly as we had too few long-duration behaviours recorded in water at high resolutions (>7 s). Despite this limitation, low resolution sampling produced very high out-of-sample (i.e., test) validation accuracies over a range of epoch sizes (72–90%; Table 2). The sample window size (epoch) influenced the result, with longer epochs tending to produce greater accuracies. The overall value of the summary statistic and its variation would be lower for long duration behaviours, i.e., low energy, repetitive behaviours are easier to distinguish (Diosdado et al., 2015). Shorter epochs are more likely to pick up irregular movements of the animal that arise from short duration, high energy activities (such as burst attacks on prey), that were rarely seen in the captive experiments (Bom et al., 2014). Instead, foraging was defined by handling of dead prey and actively searching the bottom of the pool for food hidden in the substrate (Ladds et al., 2017b). From this definition, foraging became the most difficult behaviour category for the model to classify.

The trained model based on captive animals predicted the behaviour of the wild fur seals, with minimal anomalies. Overall, the activity budgets matched expectations of wild fur seals (Battaile et al., 2015), where they spent most time resting (∼45%), and the rest of their time was split evenly among other activities—grooming (∼22%), foraging (∼20%) and travelling (∼12%). The model classified some behaviours incorrectly, primarily from two events—classifying the descent of the dive as foraging, when it was more likely travelling, and from identifying bouts of foraging during long bouts of surface swimming transiting back to the colony, when the fur seals were more likely grooming or simply changing direction. Explicitly defining some behaviours from accelerometers for other species has also been challenging. For example, using supervised machine learning models the foraging behaviour of plovers could not be classified (Bom et al., 2014), nor grooming of pumas (Wang et al., 2015). The total time that fur seals spent foraging and travelling is likely accurate because foraging and travelling were most often confused, thus combining the behaviours reduces the overall error, and is considered sufficient to summarise their activity budget. One major limitation of this study was the use of dead prey to induce foraging, so future studies should use live prey to help refine and improve the accuracy of models.

Time-energy budgets

Juvenile fur seals had an average DEE that was 4.2 times the predicted BMR, which was about the same as estimated mass-specific energy expenditure measured from adult female Antarctic fur seals and northern fur seals (4.7 times BMR; Jeanniard-du Dot et al., 2016) and slightly lower than adult female Californian sea lions (Zalophus californianus californianus) (5.2 times BMR; Ponganis et al., 1997). DEE did not differ for sex or for season (winter vs. summer) allowing us to pool the data. Sample size was too small and without adequate replication for further division. Even so, sex differences were not expected for juveniles, as any differences in diving abilities (Fowler et al., 2006), physiological parameters (Burns, Clark & Richmond, 2004) or survival (Beauplet et al., 2005) have been attributed to age or size rather than sex (Weise & Costa, 2007).