Energetic balance is a central driver of individual survival and population change, yet estimating energetic costs in free- and wide-ranging animals presents a significant challenge. Animal-borne activity monitors (using accelerometer technology) present a promising method of meeting this challenge and open new avenues for exploring energetics in natural settings.
To determine the behaviours and estimated energetic costs associated with a given activity level, three captive reindeer (Rangifer tarandus tarandus) at the Toronto Zoo were fitted with collars and observed for 53 h. Activity patterns were then measured over 13 months for 131 free-ranging woodland caribou (R. t. caribou) spanning 450 000 km2 in northern Ontario. The captive study revealed a positive but decelerating relationship between activity level and energetic costs inferred from previous behavioural studies.
Field-based measures of activity were modelled against individual displacement, vegetation abundance (Normalized Difference Vegetation Index), snow depth and temperature, and the best fit model included all parameters and explained over half of the variation in the data. Individual displacement was positively related to activity levels, suggesting that broad differences in energetic demands are influenced by variation in movement rates. After accounting for displacement, activity was highest at intermediate levels of vegetation abundance, presumably due to foraging behaviour. Snow depth, probably associated with digging for winter forage, moderately increased activity. Activity levels increased significantly at the coldest winter temperatures, suggesting the use of behavioural thermoregulation by caribou. These interpretations of proximate causal factors should be regarded as hypotheses subject to validation under normal field conditions.
These results illustrate the landscape characteristics that increase energetic demands for caribou and confirm the great potential for the use of accelerometry in studies of animal energetics.
An organism's balance between energy acquisition and expenditure affects its survival and reproduction (Brown et al. 2004); thus, ecologists are interested in quantifying patterns of energetic gains and losses. Recent work suggests animal movement trajectories can be strongly influenced by energetic tradeoffs (Shepard et al. 2013). Energetic constraints are expected to determine the extent of a species' range (Anderson & Jetz 2005; Buckley 2008), and energy deficits are likely associated with population decline. Therefore, examining variation in energy expenditure across environmental gradients, evocatively termed the energetic landscape (Wilson, Quintana & Hobson 2011; Shepard et al. 2013), is an important component of understanding and predicting patterns of space use, range shifts or retractions as well as population dynamics.
A relatively new, yet promising, method of estimating energetic costs is body accelerometry (Wilson et al. 2006; Halsey et al. 2009a; Gleiss, Wilson & Shepard 2011; Shepard et al. 2013). Accelerometers, typically integrated into tracking devices attached to an animal, measure body movements along three axes and approximate the animal's locomotion, a primary source of energy expenditure in animals (Alexander 2003). While laboratory research confirms the utility of this method, practical use in field studies remains in development (e.g. Wilson, Quintana & Hobson 2011; Nathan et al. 2012). In laboratory experiments for a range of species, these acceleration patterns have been calibrated against the rate of oxygen consumption, a commonly used measure of metabolic rate and energy expenditure (Elliott & Davison 1975). These calibrations have reliably yielded positive linear relationships between body acceleration and oxygen consumption (Halsey et al. 2009b). Accelerometry data, however, when measured at a subsecond frequency, present complications for use in field studies due to limitations of on-board data storage and the amount of information that can be transmitted via satellite. Manufacturers of wildlife tracking collars have therefore implemented the use of activity sensors derived from integrated accelerometry information (Body, Weladji & Holand 2012). Here, we evaluate the potential for the use of such activity sensors based on accelerometer scores as a proxy for metabolic costs applicable to free-ranging animals in remote settings.
To explore the use of this type of accelerometer technology, we paired a field-based study of woodland caribou (Rangifer tarandus caribou) with a short captive study of European reindeer (Rangifer tarandus tarandus). Prior to field deployment of accelerometer-equipped collars, we conducted an observational study of three captive reindeer fitted with the same collars, to verify the relationships among the simple accelerometer statistic, behaviour and activity level, as well as estimated energy expenditure. One hundred and sixty woodland caribou in northern Ontario were then fitted with collars for a multiyear study of caribou ranging ecology. The collar accelerometer measurements are reported as total active seconds within a 5- or 20-min time interval (see 'Materials and methods'), allowing for easy transmission via satellite and deployment of collars over multiple years. Here, we report on one full year of movement and accelerometry data for 131 female caribou ranging across 450 000 km2, a large fraction of the province of Ontario, Canada.
This study is among the first to report such data for a large number of free-ranging individuals inhabiting a large and varied area and to consider the environmental drivers of variation in energetic costs at a landscape scale (Wilson, Quintana & Hobson 2011). Our central aim, after verifying the utility of a simple activity measurement in a captive study, is to characterize the broad-scale patterns of variation in individual activity levels and to identify the environmental parameters underlying these patterns. Identifying the factors that increase activity, and thus energetic costs, will be a key component in determining whether energetic constraints are an important cause of decline in Ontario caribou populations (Schaefer 2003; Callaghan, Virc & Duffe 2011; Festa-Bianchet et al. 2011).
Materials and methods
Three individuals from a herd of seven female captive European reindeer at Metropolitan Toronto Zoo (ON, Canada) were fitted with collars and observed during a 12-day period from 15 December 2009 to 4 January 2010. Use of collars was preapproved by a zoo review committee, and animal handling was carried out by zoo staff. All observations took place roughly between 8:30 am and 4:30 pm. Although this subspecies is slightly smaller than that subsequently studied in the wild, behavioural patterns are otherwise known to be similar. The reindeer were 2, 13, 15 years of age; two of the three were substantially older than typically found in the wild. The observed reindeer had access to ground level grass and hay as well as pellets provided in a trough. Cratering to acquire food was unimportant for the zoo animals, although animals did dig through shallow snow occasionally to forage on grasses. As a result, foraging costs due to cratering are not included here.
The collars (Gen4 model number TGW-4680; Telonics Inc., Mesa, AZ, USA) use uGPSi-20 accelerometers to detect movement. Accelerometer measurements were reported as ‘active seconds’, where an active second was recorded if the change in acceleration from 1 s to the next in any one of three orthogonal dimensions exceeded a threshold as set by the manufacturer (about 0·3 g, or 2·87 m s−2). The sensor is also sensitive to rotation, as this movement also causes a change in the measured acceleration, and rotation angles of 17–45° thereby trigger an ‘active second’. This activity sensor falls somewhere between full dynamic acceleration and static acceleration. The virtue of such a measure is that it allows data to be collected and stored over appreciable periods, which could in principle enhance field applications in remote settings. For this captive study of short duration, a continuous stream of second-by-second data was recorded and stored on the device and then later downloaded to a computer.
Observations began in the morning before the reindeer were fed and continued during and after feeding until either the reindeer was recumbent or until the zoo closed for the evening. The following behavioural categories were recorded: lying down, standing, walking, grazing, rubbing antlers, trotting, social interaction and trough feeding. Data were logged with each change in behavioural category, and the total duration of each behavioural bout was recorded (to the second) along with a time stamp. Observations focussed on a single individual at a time and observation periods averaged approximately 20 min. We accumulated 21, 13 and 19 h of observation for the 15-, 13- and 2-year-old reindeer, respectively.
Accelerometer data were matched, second-by-second, with the behavioural observations. Energetic costs associated with each behaviour category (Boertje 1985a; Fancy & White 1986) were used to assign an energetic cost to each observed second. These data were then summarized over 5- and 20-min intervals, randomly selected with no overlap from the larger data set. This allowed comparison with the 5- and 20-min recording periods of the field-based study. Inverse negative exponential curves were then fitted to the data to describe the relationship between activity level and the estimated energetic expenditure during that interval.
During February to April of 2009 and 2010, 160 female woodland caribou across northern Ontario were net-gunned from a helicopter and fitted with GPS telemetry collars of the type described above. Capturing and handling were performed according to Ontario Ministry of Natural Resources animal care regulations. Telemetry collars were distributed in an effort to provide good geographic representation across the range of prominent vegetation communities inhabited by caribou in Ontario, including upland conifer forests in the west and lowland communities dominated by fens and bogs in the northern and eastern portions of range adjacent to Hudson Bay and James Bay. The study area, determined by the ranging patterns of all individuals, is roughly 450 000 km2 (Fig. 2).
Each collar recorded GPS coordinates and activity counts every 5 h for 13 months and transmitted the data via an Argos uplink every 5 days. Four activity intervals were recorded with each location fix, two before and two after the fix, and the total number of active seconds was reported for each. Activity intervals were 5 or 20 min, differing by collar, and thus, the maximum possible activity counts were 300 or 1200 active seconds. We accordingly normalized the data by dividing by the maximum possible activity level, allowing us to merge data from both collar types. Activity counts differed depending on the interval length, and this is treated appropriately in the statistical analyses (see below). A subset of data including only consecutive 5-h steps was used, and total displacement (in metres) was calculated for each step. Data from the first 24 h following collaring and the last 24 h before any mortality were excluded from the data set, as were any points that were likely due to device error (those indicating an improbable velocity >200 km per 5 h and a mean activity count of 0) or fixes that were not associated with at least three activity intervals. This resulted in a final data set based on 103 000 fixes of 155 000 possible.
Each GPS location was coupled with three environmental variables that are expected to be associated with foraging and thermoregulatory behaviour. We expected foraging behaviour to be associated with vegetation abundance and snow depth. Caribou often feed on lichen and ground level plants in the winter and must dig through the snow to reach them, a behaviour known as cratering. The amount of available forage was estimated using maps of Normalized Difference Vegetation Index (NDVI) at a 250-m resolution and 16-day intervals (Land Processes Distributed Active Archive Center, U.S. Geological Survey – Earth Resources Observation and Science Center, lpdaac.usgs.gov). NDVI values range from −2000 to 10 000, with higher values indicating more vegetation. A mean NDVI value was calculated from the area within a 250-m radius of each point, half the median 5-h displacement. Thermoregulatory behaviour was expected to be associated with temperature. Maps of snow depth (metres) and temperature (°C), based on real-time weather data and meteorological modelling, were obtained at a 40-km resolution for each 3-h period (North America Regional Reanalysis data set DSI-6175, NOAA Operational Model Archive and Distribution System, nomads.ncdc.noaa.gov). Each GPS point was associated with the nearest location and time for these two meteorological variables.
Location, activity count, displacement and environmental measures were averaged for each individual and month. Means were only included if the individual had at least 10 days of activity data within the month (leaving 131 individuals and 1231 individual months, and between 1 and 13 months of data per individual).
Mean monthly activity counts (transformed to percentage of possible maximum) were modelled against displacement, NDVI, snow depth and temperature, using a generalized least squares model (gls function, version 2.13.1, The R Foundation for Statistical Computing, Vienna, Austria), which included individual and x–y coordinates to account for nonindependence and spatial autocorrelation structure in the residuals. All models also included interval type (5 or 20 min) as a parameter, as the mean activity levels differed significantly, with a slightly higher mean for the 5-min activity intervals. Statistical models considered included a null model (intercept and interval type only), all linear and second-order combinations, and two-way interactions between environmental variables. For simplicity and ease of interpretation, second-order terms and two-way interactions were not considered within the same model. Candidate models were compared using Akaike's Information Criteria (AIC) with the best model having the highest model weight (ranging from 0 to 1) (Burnham & Anderson 2002). True R2 values cannot be calculated for these models, yet an estimate for the variance explained by the best model can be calculated by comparing the residual variance of the null model to that of the best model (Zheng 2000). These values are reported where appropriate.
Activity measures for the captive reindeer ranged from 0% to 46% of seconds active within an interval, with a mean and standard deviation of 7·6 ± 8·0%. Of all observations, 43% of time was spent lying down, 27% grazing, 15% standing, 8% feeding at trough, 5% walking, 0·4% rubbing antlers, 0·4% in social interactions and 0·1% trotting. Direct observations confirmed that intervals with low activity scores were associated with lower cost behaviours such as lying down, while intervals with higher activity scores were associated with higher cost behaviours such as grazing and walking (Fig. 1a). Energy expenditure, as estimated by the energetic costs of the behaviours observed within each interval, was significantly correlated with the activity measures (Fig. 1b). Nonlinear models fit to the data explained 64% of the variance for 5-min intervals and 67% for 20-min intervals, and all model coefficients were significant with P ≤ 0·0001.
For the study of free-ranging woodland caribou, measures of activity ranged from 2·9% to 94·8% of maximum possible activity level, with a mean and standard deviation of 17·4 ± 14·9%. That is, on average, individuals were active for 17·4% of the seconds within a 5- or 20-min activity interval. The mean for 5-min intervals was 18·0 ± 16·5%, whereas the mean for 20-min intervals was 17·1±13·8%.
Mean annual activity levels ranged from 14% to 21% and varied across the study range (Fig. 2), with a mean across all individuals of 17·6 ± 3·0%. Mean monthly activity levels ranged from 15·4% to 22·0%, with activity levels highest in December and January and variation greatest in the summer months. The best model for monthly mean activity (Table 1) included displacement, NDVI, snow depth and temperature, with a model weight of 0·49 and an estimated 51% of the variance explained (Table 2). This model included positive linear relationships with displacement and snow depth and second-order terms for NDVI and temperature (see Fig. 3, for the shapes of these curves). A model with x and y coordinates did not provide a more plausible model, indicating that displacement and local environmental variables captured a large fraction of observed spatial variation.
Table 1. Top ten models (ordered by strength of model fit) for analysis of mean monthly activity, including degrees of freedom (d.f.), ΔAIC relative to the best fit model and model weight (ω). Explanatory variables are displacement (disp), vegetation abundance (NDVI), snow depth (snow) and temperature (temp). When second-order terms were included in the model, the univariate terms were necessarily included as well. The best explanatory model, with a model weight of 0·49, explained an estimated 51% of the variance
AIC, Akaike's Information Criteria; NDVI, Normalized Difference Vegetation Index.
disp NDVI2 snow temp2
disp NDVI2 temp2
disp NDVI2 snow2 temp2
disp snow temp2
disp NDVI temp2
disp snow temp snow:temp
disp temp snow:temp
disp snow2 temp2
disp NDVI snow temp2
Table 2. Best model fit for mean monthly activity versus candidate model including second-order terms
Interval type (20 vs. 5 min)
Normalized Difference Vegetation Index (NDVI)
All top-ranked models include displacement and a displacement-only model explained 37% of the data variance. A plot of mean monthly activity versus displacement (Fig. 3a) illustrates the strong association between these variables. The scatterplot of monthly activity versus displacement (Fig. 3a) suggests there may be a lower bound to the data distribution, characteristic of so-called polygonal relationships (Scharf, Juanes & Sutherland 1998). Plotting the residuals after accounting for displacement (Fig. 3b,c) illustrates the remaining variance in activity patterns that are associated with the environmental variables.
The captive study suggested that activity scores derived from accelerometry were associated with differences in reindeer behaviour that have been linked in earlier laboratory studies with predictable variation in energy expenditure. There was, however, appreciable variance left unexplained, larger than that reported for other published correlates of metabolic rate, such as heart rate (Nilssen et al. 1984; Fancy & White 1986) or O2 consumption (Halsey et al. 2009b). Additional research on meaningful summary statistics for accelerometer data would be of great value. The use of ‘overall dynamic body acceleration’ (ODBA), for example, has proven to be highly accurate in laboratory studies (Halsey, Shepard & Wilson 2011). It would instructive in a future study to measure both the derived activity metric used in the current study and ODBA.
For the study of free-ranging caribou, displacement was associated with activity level. This result is not unexpected. Locomotion is a primary component of activity in a vagile animal (Boertje 1985a; Duquette & Klein 1987), is associated with elevated heart rates and likely metabolism in reindeer (Nilssen et al. 1984) and is a behaviour that is accurately captured by accelerometry (Gleiss, Wilson & Shepard 2011). After accounting for displacement, our best explanatory model suggested a parabolic relationship with vegetation abundance, where foraging activity is highest at intermediate levels of NDVI. Separate analyses for colder and warmer months also suggested that this reflects a seasonal difference in foraging patterns. At lower measures of vegetation abundance seen during the winter, we found the expected positive correlation and caribou are more active where there is more vegetation, presumably due to foraging behaviour. At higher levels of vegetation abundance seen during the summer months, we found the opposite pattern. One plausible explanation is that caribou living in these areas reach asymptotic levels of food intake when vegetation is plentiful and therefore actually spend less time foraging (Boertje 1985b). NDVI is significantly and positively correlated with percentage conifer tree cover, as estimated from a LANDSAT satellite land cover classification (Spectranalysis Inc. 2004), in both warm and cool months. Conifer cover is a preferred habitat for caribou and source of forage, particularly for lichens in winter months (Brown, Rettie & Mallory 2006; Brown et al. 2007).
Our best model also included a positive linear relationship between activity and snow depth, suggesting that activity levels are higher in deeper snow, presumably due to cratering behaviour. Note that higher snow depths depress movement across the landscape (Avgar et al. 2013) and yet increase overall activity levels. As this behaviour is potentially energetically costly (Fancy & White 1985), especially during the lean winter months, this aspect warrants further analysis at a finer scale.
The best model for monthly mean activity levels suggested an inverted polynomial relationship with temperature. Considerably higher activity levels were only found at the lowest temperatures. The lower critical temperature for reindeer is estimated to be about −30 °C (Tyler & Blix 1990), yet we see rises in activity at mean monthly temperatures much above this threshold. The temperatures actually experienced by the caribou during the coldest months were obviously colder than the mean and factors such as wind contribute considerably to heat loss as well. Caribou are well adapted to cold temperatures and are known to rely upon shivering and metabolic (or nonshivering) thermogenesis (Soppela, Nieminen & Timisjàrvi 1986), yet this pattern suggests that they may also use behavioural thermal substitution. Little is known, however, about activity as a form of thermoregulation in large mammals (McNab 2002; Humphries & Careau 2011), and we cannot say what form of behaviour this thermoregulatory activity may take. Thermoregulatory constraints are an important determinant of activity patterns (e.g. Armstrong & Robertson 2000) and range limits (e.g. Porter et al. 2002) for many species.
The modelled relationships summarize the primary components of energetic costs for woodland caribou. Landscapes that increase movement rates or displacement distances, of low to intermediate vegetation abundance, with greater snow depth, and lower temperatures are expected to increase energetic costs. These parameters explain over half the variation in the monthly data. The results of this study suggest that activity sensors based on accelerometry may be useful for field estimation of activity patterns and energetic costs in free-ranging animals. Our results also demonstrated that activity levels varied with both biological and environmental parameters. In sum, this work supports the growing recognition that accelerometry offers a new and important addition in remote biotelemetry (Cooke et al. 2004), allowing ecologists to probe otherwise unobtainable patterns.
A number of practical limitations would still limit the utility of accelerometry-based activity scores for precise estimation of absolute energetic costs in the field. The range of behaviours in captive animals, especially those in zoos or holding pens, no doubt differs from that seen in free-ranging caribou (Boertje 1985b; Fancy & White 1986). Our zoo animals exhibited much lower frequencies of walking trotting and (particularly) running than typically found in free-living ungulates, as well as foraging infrequently from ground level lichen, grasses or forbs. Moreover, the field estimates of energetic costs we have derived from the literature are possibly clouded by complexities associated with spatial variation in topographic relief, day-to-day variation in snow hardness, depth and surface ablation, as well as variation among individuals in their internal condition, motivational state and accumulated energy deficits. Although we feel confident in concluding from our study that the relative magnitude of energetic costs in free-living caribou varies considerably with temporal and spatial variation in environmental conditions as well as mobility levels, limited knowledge about caribou behavioural budgets as well as lack of a direct calibration of metabolic costs still hampers our ability to precisely predict absolute costs across the caribou energetic landscape (Shepard et al. 2013).
This work was funded by grants from the Forest Ecosystem Science Cooperative Inc, Ontario Ministry of Natural Resources (OMNR) and Natural Sciences and Engineering Research Council (NSERC) of Canada as well as a Vanier Fellowship to T.A. We thank the staff of the OMNR, notably Lyle Walton, for support in the field operations, the Toronto Zoo and staff, and Luba Broitman for software development. We also thank Douglas Morris, Daniel Fortin and two anonymous reviewers for comments on the manuscript. None of the authors has a conflict of interest.