Behavioural mapping of a pelagic seabird: combining multiple sensors and a hidden Markov model reveals the distribution of at-sea behaviour

Ben Dean, Robin Freeman, Holly Kirk, Kerry Leonard, Richard A. Phillips, Chris M. Perrins, Tim Guilford

Abstract

The use of miniature data loggers is rapidly increasing our understanding of the movements and habitat preferences of pelagic seabirds. However, objectively interpreting behavioural information from the large volumes of highly detailed data collected by such devices can be challenging. We combined three biologging technologies—global positioning system (GPS), saltwater immersion and time–depth recorders—to build a detailed picture of the at-sea behaviour of the Manx shearwater (Puffinus puffinus) during the breeding season. We used a hidden Markov model to explore discrete states within the combined GPS and immersion data, and found that behaviour could be organized into three principal activities representing (i) sustained direct flight, (ii) sitting on the sea surface, and (iii) foraging, comprising tortuous flight interspersed with periods of immersion. The additional logger data verified that the foraging activity corresponded well to the occurrence of diving. Applying this approach to a large tracking dataset revealed that birds from two different colonies foraged in local waters that were exclusive, but overlapped in one key area: the Irish Sea Front (ISF). We show that the allocation of time to each activity differed between colonies, with birds breeding furthest from the ISF spending the greatest proportion of time engaged in direct flight and the smallest proportion of time engaged in foraging activity. This type of analysis has considerable potential for application in future biologging studies and in other taxa.

1. Introduction

Understanding the at-sea behaviour of pelagic seabirds is both difficult, because of their elusive lifestyles, and important, because of their vulnerability to changes in the marine environment and their status as indicators of ocean health [1]. Advances in miniature telemetry and data-logging systems continue to revolutionize the remote observation of long-distance movements in these and other elusive species [27], but such systems can be expensive as well as impactful (especially on smaller species), and the resulting data are often under-used. Many tracking studies use recorded locations to examine movement and habitat use only in broad terms [36], leaving key questions about where and when animals engage in specific behaviours unanswered. However, behaviour may be inferred by characterizing patterns of movement based solely on the geometry or complexity of an animal's path using techniques such as tortuosity [8,9], positional entropy [10,11] or first-passage time [12]. In addition, modelling approaches such as Gaussian mixtures have been used to classify animal tracking data into discrete modes of movement [2,13], while state–space models including hidden Markov models (HMMs) [10,1419] have been used to identify different modes of movement and the dynamics of switching between them.

HMMs assume that a sequence of observations (e.g. tracking data) is generated by an underlying unobserved sequence of discrete states (modes of movement, or different behaviours) and that the state occupied at each step in the hidden sequence depends only on the state occupied in the preceding step [20]. This approach is well suited to uncovering hidden patterns of behaviour in animal tracking data because it is consistent with the fact that behaviour is never directly observed, makes use of the inherent temporal dynamics within the data and uses the non-independence of sequential data points. Where these methods have been applied solely to animal tracking data they have offered important insights into behaviour. However, enriching tracking data by concurrent deployment of additional devices that record activity [21] may provide both increased power to resolve behavioural states, and the information required to validate and interpret those states as specific behaviours.

The aims of this study were to identify the principal types of behaviour in which a volant, pelagic, diving seabird engages while at sea; to segment global positioning system (GPS) tracks based on those behaviours; and to explore their spatio-temporal distribution. We studied the Manx shearwater, a 400 g procellariiform seabird that breeds predominantly in colonies at remote locations off the west coasts of Britain and Ireland. Females lay a single egg in an underground nest and during the incubation period, both parents alternate stints of 5–7 days at the nest with long foraging trips [2] on which they recover body condition. During the chick-rearing period, both parents make repeated, shorter duration foraging trips to sea [2,22], returning at night to provision the chick. See Brooke [23] for a detailed description of the breeding behaviour. While marine surveys [24] and GPS tracking [2] have provided valuable insights into the routes and destinations of these foraging trips, little is known about individual and colony-based patterns of behaviour at sea.

Our approach applied an HMM to data from two independent but simultaneously deployed loggers (GPS and saltwater immersion) to categorize at-sea activity into discrete behavioural states. We subsequently interpreted and validated the HMM classification using path tortuosity and data from a third independent device (time–depth recorder, TDR) deployed concurrently. We were then able to segment individual GPS tracks into the most likely sequence of behavioural states. By training and validating the HMM on a detailed multi-sensor dataset and applying it to a much larger but less-detailed tracking dataset, we were able to map the spatial distribution of behaviours and explore where, when and for how long birds from different colonies engaged in each activity, during different stages of the breeding season, over 3 years.

2. Material and methods

2.1. Data loggers

Loggers were deployed on breeding birds from two colonies in the Irish Sea during the incubation (May–June) and chick-rearing (July–August) periods of the 2009, 2010 and 2011 breeding seasons: Skomer Island, Wales (Southern Irish Sea: 51°44′ N, 5°17′ W) and Lighthouse Island in the Copelands group, Northern Ireland (Northern Irish Sea: 54°41′ N, 5°31′ W). As far as practicable, deployments were carried out simultaneously on both islands to minimize any effects of temporal changes in environmental conditions. Breeding adults were taken from their burrows by hand through a purpose-built inspection hatch and weighed prior to device deployment. A total of 168 deployments were made on 117 individuals (table 1). All birds were fitted with at least two data-logging devices: a GPS logger (modified i-gotU GT-120: mobile action) and a British Antarctic Survey geolocator-immersion logger (Mk13, 14 or 18L). Twenty-five of the birds tracked from Skomer (table 1) were additionally fitted with a CEFAS G5 TDR.

View this table:
Table 1.

Number of birds tracked from each colony during the incubation (inc.) and chick-rearing (rear.) periods of each year. All birds were tracked using GPS and geolocator-immersion loggers. Birds marked with an asterisk also carried TDRs. Some individuals were tracked during the same period in 2 or more years: incubation Skomer five birds, Copeland eight birds; rearing Skomer five birds, Copeland nine birds.

GPS loggers were configured to record location once every 5 or 10 min and were attached dorsally using thin strips of Tesa marine cloth tape underlying a small number of contour feathers [2]. Geolocator-immersion loggers test for saltwater immersion every 3 s and record the proportion of samples immersed in each 10 min interval. Immersion loggers were attached by two lightweight cable ties to a metal leg ring, so as to be immersed when sitting on the sea surface. TDRs were configured to record pressure every 1 s and were attached to the central four tail feathers using thin strips of Tesa tape (similar to Wilson et al. [25]). GPS weighed 15 g in dual device deployments and 10.5–11.3 g (smaller battery) in triple device deployments. Geolocator-immersion loggers weighed 1.5–1.9 g and TDRs 2.7 g. The total mass of devices was approximately 18 g for all deployments. On return, birds were recaptured from the burrow, reweighed and devices removed. Capture and deployment took less than 15 min, and removal less than 10 min. During 2010 and 2011, we weighed the chicks of tracked adults, to determine the size of meals delivered during nest visits. The effect of carrying devices was tested on birds tracked from Skomer in 2009 when compared with hatching success, chick growth rate and fledging success with a group of unmanipulated control nests.

2.2. Hidden Markov model

HMM is a temporal pattern recognition method that identifies discrete states and time-dependent state changes within time-series data. For example, flight and rest might result in distinctly different patterns of movement (fast versus negligible speeds), each constituting a discrete state. The method is well suited to revealing the sequence of such states in animal tracking data where behaviour itself is never directly observed and sequential data points are usually non-independent. A full description of HMM methods is presented by Rabiner [20]. Here, we assumed that at-sea behaviour can be described as a Markov process such that it is composed of an unknown but discrete number of distinct states equating to the birds' principal activities. Foraging trips are segmented into discrete time steps (logger sampling intervals) and in each time step, a bird is engaged in one activity. Between time steps, the bird can switch from the current state (activity) to a new state, or remain in the same state according to a set of state transition probabilities. The sequence of states is hidden, because we cannot directly observe the bird, but at each time step, we make observations in the form of logger data, according to a set of emission probabilities, which might be given by Gaussian components for continuous variables or by conditional probability tables for discrete variables. By learning the state transition and emission probabilities, we can estimate the probability of the observations recorded at each time step belonging to a particular state.

2.3. Model data

Two data series were used to train the HMM: saltwater immersion and ground speed (figure 1). Because immersion was recorded at regular 10 min intervals, average speed values were calculated over the same intervals as follows: GPS locations were interpolated to 1 min frequency using piecewise cubic hermite polynomials as in Tremblay et al. [26]. Mean ground speed was then calculated between those interpolated locations falling within each 10 min interval. In validation tests (see the electronic supplementary material), this interpolation gave a better representation of speed within 10 min intervals than the original recorded locations, especially where acceleration or deceleration occurred between intervals. Speed data were log(+1)-transformed and immersion data (a proportion) were mapped to the whole real line (−∞, ∞) using a logit transform with a small value ɛ (the minimum non-zero proportion) added as in Warton & Hui [27]. Data collected within 1 km of the colony were omitted to exclude colony-based behaviours, such as sitting or moving on the ground.

Figure 1.

Example speed and saltwater-immersion data used to train the model. Proportion of 10 min intervals spent immersed (logit-transformed) is shown in (a); average speed (m s−1) over the same intervals (log-transformed) is shown in (b).

2.4. Model training

We trained the HMM on data from the 25 birds from Skomer fitted with all three logger types using the HMM toolbox [28] in Matlab (The MathWorks, Natick, MA, USA). In this implementation, we were constrained to the assumption that the emission probabilities were given by Gaussian components in the observed data. For the log-transformed speed data, this assumption was valid. The logit-transformed immersion data contained a peak of low values, a peak of high values and a range of values between. As such, the assumption of Gaussian emission distributions was not ideal for the immersion data, but provided a pragmatic solution to the partitioning of distinct groups within the data.

HMMs can be simple (few states), but have large error around them, or complicated to the extreme case of one state per observation and zero error. As a model selection exercise, we trained a series of candidate HMMs with an increasing number of putative states (2–10). The iterative addition of states (each explaining fewer observations with less error) increased log-likelihood monotonically, favouring over-fitting of complex (many state) models. However, the increase in log-likelihood was greatest between the two- and three-state models, resulting in a distinct knee point [29] (figure 2). We therefore chose a three-state HMM as a trade-off between model accuracy and complexity.

Figure 2.

Log-likelihood calculated for candidate models with an increasing number of states. The increase in log-likelihood was greatest between the two- and three-state models, resulting in a distinct knee point.

HMM parameters were initialized using k-means clustering [30] and improved through unsupervised training using the Baum–Welch algorithm [20] which, given a sequence of observations, iteratively adjusts the model parameters to maximize the probability of obtaining those observations. We then applied the trained final HMM to each individual track in the training data, using the Viterbi algorithm, which calculates the sequence of hidden states that best explains a sequence of observations [20]. We interpreted the nature of the behaviour that was characteristic of each of the emergent states, based on the typical values for immersion and speed, as well as track straightness [31] (calculated as the ratio between the straight-line distance and the track-line distance over a moving 100 min section of the interpolated GPS track) and time submerged (recorded by the TDRs).

2.5. Mapping behaviour

By linking the HMM classification to the corresponding GPS locations, we were able to plot the sequence of activities undertaken during the course of individual foraging trips, showing where, when and for how long each bird engaged in each of the activities. Finally, we applied the HMM trained and interpreted on the subset of deployments in which all three devices were used, to the whole dataset including a further 143 deployments on birds tracked only with GPS and geolocator-immersion loggers. We used the Viterbi algorithm to classify the most likely behaviour at each 10 min time-step along each GPS track. This enabled us to plot the sequence and spatial distribution of behaviours and to calculate the time spent engaged in each activity. To explore the spatial distribution of activities for birds from different colonies, during different periods of the breeding season, we calculated kernel density distributions for each activity over 1 km2 grid cells in ArcGIS, using a bandwidth (h) of 17 km, selected using least-squares cross validation, and which produced contiguous cores without over-smoothing.

3. Results

3.1. Trip metrics and impact of loggers

A total of 168 successful deployments were made on 117 individuals, simultaneously recording location and immersion data during 315 foraging trips mainly within the Irish and Celtic Sea (table 1). An additional four very long-range offshore trips into the Atlantic were excluded from the analysis as extreme outliers.

The median duration of deployments was 8 days (maximum 15) during incubation and 5 days (maximum 13) during chick-rearing. Overall, incubation deployments recorded 1.1 foraging trips and chick-rearing deployments recorded 2.4 trips. The departure weights of birds tracked during incubation and the minimum-recorded weights for birds tracked during chick-rearing are the best approximation of a bird's minimum body mass without food in the stomach, or large fat deposits, in each period. These were similar (407 and 412 g, respectively; electronic supplementary material, table S1), two-sample t-test: t167 = 1.06, p = 0.31, as were the weights of birds from different colonies: Copeland incubation 407 g, Skomer incubation 406 g, two-sample t-test: t60 = 0.12, p = 0.90; Copeland rearing 415 g, Skomer rearing 409 g, two-sample t-test: t105 = 1.01, p = 0.32. The overall mean minimum weight was 410 ± 30 g; so the combined device weight (18 g) comprised 4.1–4.7% of minimum body weight. During incubation trips, birds from Copeland increased their mass by slightly more (32 g) than birds from Skomer (24 g), two-sample t-test: t51 = 1.07, p = 0.29. Following chick-rearing trips, birds from Copeland and Skomer delivered similar size meals: 44 and 45 g, respectively, two-sample t-test: t142 = 0.31, p = 0.75.

Median trip duration was longer during incubation (8 days) than during chick-rearing (1 day; electronic supplementary material, table S1), Wilcoxon rank-sum: z283 = 10.09, p < 0.001. During incubation, median trip duration was slightly shorter for birds from Copeland (7 days) than from Skomer (8 days), Wilcoxon rank-sum: z47 = 0.99, p = 0.32. During chick-rearing, median trip duration was slightly longer from Copeland (2 days) than from Skomer (1 day), Wilcoxon rank-sum: z234 = 1.47, p = 0.142. During incubation trips, birds from Skomer covered greater distances overall and per day (24 h), and reached greater maximum distances from the colony than birds from Copeland, Wilcoxon rank-sum: z47 = 2.13, 2.89, 4.85; p = 0.033, 0.004, p < 0.001, respectively. During chick-rearing trips, birds from Skomer covered greater distances overall and per day, and reached greater maximum distances from the colony than birds from Copeland, Wilcoxon rank-sum: z234 = 1.89, 6.87, 5.50; p = 0.059, p < 0.001, p < 0.001, respectively.

For the nine nests from which adults were tracked on Skomer during incubation 2009, hatching success was 77.7 per cent compared with 82.4 per cent in the 17 control nests. For the 10 nests from which adults were tracked on Skomer during chick-rearing 2009, fledging success was 100 per cent compared with 93 per cent in the 15 remaining control nests. The mean growth rate of chicks whose parents carried tracking devices (8.7 ± 1.2 g day−1) was not significantly different to that of chicks in the control nests (9.3 ± 1.0 g day−1), two-sample t-test: t24 = 1.34, p = 0.19. There were no differences in trip duration between triple-instrumented birds (1.5 days, interquartile range (IQR) 1–2) and dual-instrumented birds tracked in the same period from the same colony (1 day IQR 1–2 days), Wilcoxon rank-sum, z59 = 0.76, p = 0.45.

3.2. Interpreting behavioural states

Of the candidate models trained with 2–10 states, the increase in log-likelihood was greatest between the two- and three-state models (figure 2), suggesting that the at-sea behaviour of Manx shearwaters was most parsimoniously described by three principal activities. Although information on diving and track straightness were not used to train the HMM, these differed significantly between the three states: proportion of time submerged Kruskal–Wallis χ22,72 = 58.37, p < 0.001; and track straightness ratio χ22,72 = 28.68, p < 0.001. This suggests that these states, identified purely on the basis of temporal patterns in combined immersion and ground speed data, represent three distinct activities.

The first activity was consistent with sustained direct flight. First, it was characterized by high-speed movement (median speed = 8.9 m s−1), within the theoretical flight speed range for the species [32]. Second, there was high or constant contact with the sea surface (median proportion of time immersed = 0). Third, track straightness was high (median straightness ratio = 0.83 compared with the recorded maximum of 0.87 in this study), indicating directed movement. Fourth, the HMM state transition probabilities (figure 3) indicated that this state was stable over time and likely to be followed by the same state (p = 0.68). Finally, only a very small proportion of the total time spent diving occurred during this state (median proportion = 0.01).

Figure 3.

State transition probability matrix learned for the final three-state HMM. Given that a particular state is occupied at time t, the matrix shows the probabilities of switching to each of the other states, or remaining in the same state at time t + 1.

The second activity was consistent with sitting and drifting on the sea surface. First, it was characterized by slow movement (median speed = 0.65 m s−1), slower than the 0.85 m s−1 previously calculated for drifting birds [2]. Second, there was high to constant contact with the sea surface (median proportion of time immersed = 1). Third, there was high track straightness (median ratio = 0.86). Fourth, the transition probabilities indicated that this state was highly stable and likely to be followed by the same state (p = 0.81). Only a small proportion of the total time spent diving occurred during this state (median proportion = 0.03).

We propose that the third activity is consistent with foraging, incorporating convoluted search flight and periods of immersion in pursuit of prey. First, speed was intermediate (median speed = 2.01 m s−1), slightly slower than the separation between drifting and flying speeds calculated by Guilford et al. [2]. Second, there was intermediate contact with the sea (median proportion of time immersed = 0.63). Third, track straightness was low (median straightness ratio = 0.62 compared with the recorded minimum of 0.49 in this study), suggesting tortuous movements within restricted areas. Fourth, the transition probabilities indicated that this state was stable over time and likely to be followed by further periods of foraging (p = 0.69). Finally, diving was very strongly associated with this activity, with a median proportion of 0.96 of total dive time occurring within this state.

Maps of individual bird tracks classified by the HMM (figure 4) reveal the spatial distribution and sequential patterns of activity, and further support our proposal that the discovered activities represent sustained direct flight, sitting and drifting, and foraging. In particular, by mapping the locations of diving activity along with the HMM classification, we visually demonstrate the high level of agreement between foraging activity and diving.

Figure 4.

(a) Three day and (b) 5 day foraging trips from Skomer (S), classified according to activity using the HMM. Ten-minute intervals classified as direct flight are mapped as green circles, foraging as red and sitting as blue. Intervals where diving occurred are mapped as grey circles, scaled in proportion to total time submerged.

3.3. Timing of activities

During both incubation and chick-rearing, the majority of sitting activity occurred during the evening and hours of darkness (figure 5), as birds roosted on the sea surface. Direct flight and foraging activities typically increased during the hour preceding sunrise, with a morning peak in flight activity just after sunrise, followed by a lull around midday coincident with an increase in sitting. Flight then peaked again, prior to sunset and decreased rapidly as darkness fell. Foraging activity was highly constrained to the hours of daylight and twilight (figure 5), which were shorter during chick-rearing than during incubation.

Figure 5.

Overall daily activity patterns for all birds during incubation and chick-rearing. The distributions are normalized to sum to unity. Timing (GMT) of sunrise and sunset at the centre of the Irish Sea (53°0′ N, 5°13′ W) are shown as dashed lines.

The percentage of daylight hours engaged in direct flight and foraging activity differed between incubation and chick-rearing, Wilcoxon rank-sum: (flight) z167 = 5.35, p < 0.001, (sitting) z167 = 0.25, p < 0.79, (foraging) z167 = 4.23, p < 0.001. On average, an incubation trip comprised 10 per cent direct flight, 28 per cent sitting and 63 per cent foraging, whereas a chick-rearing trip comprised 15 per cent direct flight, 28 per cent sitting and 57 per cent foraging (figure 6). During both periods, birds from Skomer spent a greater percentage of daylight hours engaged in direct flight and a smaller proportion engaged in foraging activity than birds from Copeland (figure 6). During incubation trips, birds from Skomer spent a median of 12 per cent of the day in direct flight and 60 per cent engaged in foraging, compared with 6 per cent and 64 per cent for birds from Copeland, Wilcoxon rank-sum: z61 = 4, p < 0.001 and z = 1.74, p < 0.08). Birds from both colonies spent 27 per cent of the day sitting, Wilcoxon rank-sum: z61 = 0.85, p < 0.39). During chick-rearing trips, birds from Skomer spent 19 per cent of the day in direct flight and 54 per cent engaged in foraging, compared with 12 per cent and 60 per cent for birds from Copeland, Wilcoxon rank-sum: z105 = 5.04, p < 0.001; and z = 2.74, p < 0.01. Birds from both colonies spent a similar percentage of the day sitting: Skomer 27 per cent and Copeland 28 per cent, Wilcoxon rank-sum: z105 = 0.64, p < 0.52.

Figure 6.

Ternary plots showing the percentage of daylight hours spent in each activity by birds from Copeland (red circles) and Skomer (blue circles) during (a) incubation and (b) chick-rearing. Large squares indicate corresponding median values.

3.4. Spatial distribution of activities

The distribution of each activity for birds tracked from both colonies, in both periods of the breeding season and in each year, is shown in (example) figure 7 and (all) electronic supplementary material, figures S1–S4. These highlight discrete areas where birds engaged in foraging activity, show the important routes between foraging areas, where birds engaged in direct flight activity, and indicate that sitting activity (most of which was nocturnal roosting) generally occurred within foraging areas or close to the colony. Those birds not visiting the nest usually spent the night roosting on the sea away from the colony, while birds that did visit the nest on a given night tended to roost within 20 km from the colony prior to landfall and resume roosting following their visit (figure 8). During both incubation (figure 9) and chick-rearing (figure 10), birds from Copeland typically foraged within relatively discrete, local areas, within 100–120 km of the colony. Birds from Skomer showed a more varied and dispersed foraging distribution. During incubation, few birds foraged locally, most commuting 200 km or more to the south Irish coast or to the area southwest of the Isle of Man, where Copeland birds were foraging locally (figure 9). During the chick-rearing period, birds from Skomer foraged locally much more frequently, but continued to commute 200 km or more to the area southwest of the Isle of Man (figure 10).

Figure 7.

Example kernel density distributions of each activity for birds from (a) Copeland and (b) Skomer during the 2009 chick-rearing period: (i) foraging (red), (ii) direct flight (green), (iii) sitting (blue). Densities are shaded from the lightest (95% occupancy) to the darkest (10% occupancy). Locations of Copeland (C) and Skomer (S) and the approximate positions of the Irish Sea front (curved line) and Celtic Sea Front (short line) are shown.

Figure 8.

Nocturnal locations classified as sitting (roosting) around Skomer (S): (a) prior to most birds visiting the colony and (b) after most birds have left the colony. Locations are coloured by hour (GMT).

Figure 9.

Kernel density foraging distributions of birds from (a) Copeland and (b) Skomer during the (i) 2009, (ii) 2010 and (iii) 2011 incubation periods. Densities are shaded from the lightest (95% occupancy) to the darkest (10% occupancy). Approximate positions of the Irish Sea Front (ISF; curved line) and Celtic Sea Front (CSF; short line) are shown.

Figure 10.

Kernel density foraging distributions of birds from (a) Copeland and (b) Skomer during the (i) 2009, (ii) 2010 and (iii) 2011 chick-rearing periods. Densities are shaded from the lightest (95% occupancy) to the darkest (10% occupancy). Approximate positions of the Irish Sea Front (ISF; curved line) and Celtic Sea Front (CSF; short line) are shown.

4. Discussion

4.1. Impacts of tracking

Combined deployment of GPS, immersion loggers and TDRs provided highly detailed data on the at-sea movements and behaviour of individual Manx shearwaters. A previous meta-analysis of procellariiform tracking studies [33] found negative impacts (nest desertion and increased trip duration) on a range of species carrying devices comprising 0.6–5.5% of body weight, although many of those deployments were of longer duration than here. Our control test suggests that our devices did not have detectable medium-term (breeding season) effects on reproductive success. We did not test for an effect of devices on trip duration, but the median incubation trip duration from Skomer in this study (8 days) was slightly longer than that previously reported for untagged birds from Skomer (5.4–6.9 days) [23], and mean trip duration from Skomer during chick-rearing (1.9 days) was similar to that previously reported for birds from Skomer carrying 2 g tags (1.8 days) [22]. It is possible that the devices had some effects on behaviour; however, our results suggest that these would have been minor, given the relatively short deployment durations.

4.2. At-sea behaviour

The combination of data from multiple loggers (GPS, geolocator-immersion and TDR) and the application of an HMM allowed us to identify, interpret and explore three discrete states in the at-sea behaviour of a pelagic seabird during foraging trips away from the colonies. These three activities were stable over time, generally persisting for many consecutive time-steps within individual tracks. Our use of an HMM offered an integrated approach to identifying hidden states within the immersion and speed data, without requiring prior categorization of the data into discrete groups of observations, while the deployment of multiple loggers provided additional data for interpretation and validation. The assumption of Gaussian emission distributions for logit-transformed immersion data may introduce a small bias. Each of the states roughly corresponded to a partitioning into low-immersion, high-immersion and intermediate values. Gaussians fitted to the high and low peaks were likely to be skewed slightly towards the centre of the distribution, increasing the probability that some intermediate values would be categorized as low immersion, or high immersion. This may slightly underestimate the occurrence of the intermediate (foraging) state consistently across all birds, but otherwise we do not believe that it significantly affects the findings.

Because of some irregularity in GPS signal acquisition, occasional missing locations and the need to calculate speed within discrete intervals, we found it necessary to interpolate regular locations prior to implementing the HMM. We used a curvilinear interpolation, which typically produces more realistic estimates of animal movements than linear interpolation [26] and is less prone to underestimating path length and speed. Our tests indicated that it performed well in recreating GPS data recorded at a higher resolution (see the electronic supplementary material). For some behaviours, interpolation may not represent a realistic model of complex fine-scale movement between sampled locations. An alternative approach would have been to estimate location and behavioural state simultaneously within an integrated state–space framework [14,19], incorporating models of individual movement such as random walks, correlated random walks or Lévy walks. While such approaches may offer some advantages, they can be computationally more complex, and their performance may depend on the choice of appropriate movement models and their parameters.

Sustained direct flight was characterized by little or no saltwater immersion, low path tortuosity and an average movement speed of 8.9 m s−1. Assuming that the movement speeds calculated here from the GPS (which include wind velocity) are reasonably representative of flight speeds (which exclude wind velocity), this is within the theoretical flight speed range previously calculated for Manx shearwaters [32], but slightly slower than that calculated from previous GPS tracking (11 m s−1), using devices of similar mass [2]. For long-distance movements, birds may be expected to fly close to their maximum range velocity (ca 14 m s−1; [32]). Here, it seems that birds were flying closer to their minimum power velocity (ca 7.5 m s−1; [32]), suggesting some exploitation of non-powered flight such as shear-soaring [34]. Flight activity typically persisted over long periods and occurred where birds travelled long distances, often via relatively direct routes (figure 4). The post-sunrise and pre-sunset peaks in direct flight activity (figure 5) are likely to represent the majority of outbound flights from the vicinity of the colony to foraging areas, and return flights to the vicinity of the colony prior to visiting the nest. These peaks are more pronounced during chick-rearing than incubation, as would be expected, given the much higher frequency of nest visits during this period. A small amount of diving (1%) occurred during a few of the intervals classified as flight. There is evidence in other pelagic seabirds that prey may be ingested even during extremely rapid (>100 km h−1) and directed flight [35]; hence, it may be impossible to classify behaviour unequivocally in every instance (especially short-lived ones) on the basis of 10 min intervals of movement and immersion data.

Sitting on the sea was characterized by very slow movement and high or continuous immersion, typically persisting over short periods during daylight hours and long periods at night. Sitting on the water at night was apparent as long series of consecutive time steps forming straight or curvilinear series of locations (figure 8). Movement speeds during sitting activity (0.65 m s−1) were consistent with tidal drifting speeds previously estimated for shearwaters from GPS data (0.85 m s−1) [2]. The virtual absence of diving (3%) during sitting suggests that this activity may largely represent resting behaviour. Much sitting on the surface occurred close to the colonies during the evening and early hours of darkness: as with many procellariids, Manx shearwaters form dense roosting flocks on the sea adjacent to colonies prior to nest visits [23] (figure 8a). Our analysis showed that nocturnal at-sea roosting adjacent to the colony continued after 01.00, when most birds had already visited their burrow (figure 8b). After departing the colony, birds resumed roosting on the sea until first light, whereas others spent nights adjacent to the colony without visiting their burrow. Wilson et al. [25] used the locations of radio-tracked birds, assumed to be in nocturnal roosting flocks, to estimate the extent of those flocks around several colonies as a tool to define marine protected areas. For Skomer, they estimated a 95 per cent kernel density core that extended 4 km from the colony, suggesting that this area would encompass most roosting birds. Our analysis suggests that nocturnal roosting may routinely occur at much greater distances from the colony.

Foraging activity was characterized by intermediate movement speeds around 2 m s−1 and by intermediate levels of immersion which, combined with the relatively high probability (0.69) of persisting over several consecutive 10 min intervals (figure 3), suggests repeated transitions between the air and the sea surface/submersion. The majority (96%) of diving occurred during this activity, reflecting the majority of attempts at prey capture. Foraging typically occurred within restricted areas (figure 4), reflected by a significantly higher track tortuosity than during either direct flight or sitting. This is consistent with a pattern of foraging involving attraction movements between sparse, locally dynamic prey patches, convoluted search flight (area-restricted search), frequent landings and take-offs from the surface, and dives in pursuit of prey. Many intervals classified as foraging did not involve diving, suggesting that our third activity class includes search behaviour as well as prey pursuit, and that birds may often engage in unsuccessful searches without locating prey. A strong visual component to searching and prey capture is suggested by the timing of foraging activity, which appeared to be constrained to the hours of daylight and twilight (figure 5). Thus, we tentatively infer that the primary foraging mode of Manx shearwaters involves local-scale area-restricted search involving frequent landings, perhaps to sample or view prey, and then, if conditions are suitable, attempted prey pursuit and capture.

However, the search behaviours associated with foraging are likely to be multi-scaled phenomena [3638]. Over larger scales, where the distance between prey patches exceeds the perceptual range of the predator, straight movement may be the most effective method of searching for prey [39]. In our analysis, large-scale straight-line search without significant periods of immersion would be categorized as sustained direct flight. Birds often made sustained flights more or less directly to a restricted area and then engaged in foraging (figure 4). The state transition probabilities (figure 3) support this: intervals classed as flight were most likely to be followed by further intervals of flight (p = 0.68), foraging was more likely to follow a period of sustained flight than of sitting (p = 0.31 versus 0.19) and flight was more likely to lead to foraging than sitting (p = 0.31 versus 0.002). It is possible, then, that shearwaters also searched for food over larger scales using long, relatively direct flights away from the colony, switching to area-restricted search only where conditions were perceived as suitable. Nevertheless, the direct nature of some track segments and the consistency of some distant foraging locations also suggest that many larger scale direct movements were informed by previous experience (or the following of other birds informed by previous experience), with shearwaters commuting directly to known profitable areas, and then engaging in local search and pursuit behaviour. This would imply an important role for learning and the building of a large-scale memorized map of foraging areas. The nature of the cues, or representations involved in searching and map learning have yet to be investigated, but almost certainly include a strong social element, with the location of prey patches signalled by the activities of conspecifics or other foragers [24,4042]. Shearwaters are typically recorded feeding in large flocks at sea [24,40] where there is likely to be a strong element of social enhancement of individual foraging success [43].

4.3. Colony differences

During both incubation and chick-rearing, birds breeding on Skomer made trips of longer range and covered greater distances per day than birds breeding on Copeland. The results of the HMM confirm that birds from Skomer spent a greater percentage of daylight hours engaged in direct flight and consequently less time engaged in foraging activity than birds from Copeland. Taken together, this suggests that it may be more difficult for birds breeding on Skomer to locate and reach the prey resources necessary during the breeding season.

Mapping these behaviours reveals that birds breeding on Skomer typically foraged in more distant and dispersed locations than birds from Copeland. The foraging distributions of birds from both colonies consistently overlapped in the area southwest of the Isle of Man, which corresponds with the location of the western ISF [44] and in particular the stratified waters to the north and west (figures 8 and 9). The aggregation of seabirds at frontal systems is well documented [24,4547], and the ISF is likely to provide profitable feeding opportunities as increased phytoplankton growth [48,49] attracts high densities of zooplankton and fish [50,51]. During chick-rearing trips, Skomer birds foraged locally (within 100 km of the colony) more frequently than during incubation trips, but continued to make longer range trips to the area around the ISF. The increase in local foraging is likely to reflect the requirement to return more frequently to the nest to provision the chick. This requirement may (to some extent) be facilitated by increasing opportunities to forage locally as the summer progresses and variable local fronts such as the Celtic Sea front [44] (figure 10) become established.

Differences in the percentage of time engaged in direct flight and foraging activities support the proposal [2] that Copeland breeders benefit from reduced travel costs because of the colony's proximity to the ISF. During incubation, the median trip duration from Copeland was slightly shorter (7 versus 8 days) and the mean increase in mass slightly (though not significantly) greater (32 versus 24 g) than for birds from Skomer. During chick-rearing, birds from both colonies delivered similar-sized meals to their chicks (45 versus 44 g), but birds from Skomer spent more time engaged in direct flight (covering greater distances per day and overall) and less time engaged in foraging activity than birds from Copeland. This suggests that birds breeding on Skomer must work harder to locate prey, presumably at a greater cost to their own body condition (minimum weights of birds from Copeland during chick-rearing were heavier, though not significantly, than those from Skomer: 415 versus 409 g).

5. Conclusions

Here, we have demonstrated an approach that makes full use of a large, new, multiple sensor biologging dataset, applying an HMM to reveal patterns of behaviour and segment foraging trips according to the activities undertaken. We used this technique to identify the principal at-sea activities of the Manx shearwater, to explore the spatio-temporal variation in those activities, and to map those areas that were important for different types of behaviour. We have shown important differences in foraging trip behaviour, which may have significant energetic consequences for birds breeding on different colonies and have highlighted the importance of one key foraging area in the Irish Sea, the ISF. This type of analysis has considerable potential for application in future biologging studies and in other taxa, to examine behaviour at a range of temporal and spatial scales. For example, analyses of the spatio-temporal structure of activity sequences and transitions between behavioural states in relation to the location of prey may improve our understanding of how animals respond to prey distributions and the decisions they make when searching. While at large scales, this type of behavioural mapping has high value for conservation planning, allowing better identification of the key areas and habitats where animals engage in specific types of behaviour. This type of information could greatly improve assessments of the degree of interaction between particular threats and those behaviours that increase risk: for example, foraging and fisheries, commuting flight and wind turbines, or roosting and surface pollution.

Acknowledgements

We thank the Skomer and Skokholm Islands Advisory Council and the Copeland Bird Observatory for permission to work on the Islands. We also thank the staff and volunteers from Skomer and Copeland, especially Jo and David Millborrow, Chris Taylor, Dave Boyle, Neville McKee, George Henderson and Annette Fayet. This work was funded by Microsoft Research, Cambridge, the UK Natural Environment Research Council and the Northern Ireland Environment Agency.

  • Received July 17, 2012.
  • Accepted September 12, 2012.

References

View Abstract