Linking animal-borne video to accelerometers reveals prey capture variability

Abstract

Understanding foraging is important in ecology, as it determines the energy gains and, ultimately, the fitness of animals. However, monitoring prey captures of individual animals is difficult. Direct observations using animal-borne videos have short recording periods, and indirect signals (e.g., stomach temperature) are never validated in the field. We took an integrated approach to monitor prey captures by a predator by deploying a video camera (lasting for 85 min) and two accelerometers (on the head and back, lasting for 50 h) on free-swimming Adélie penguins. The movies showed that penguins moved the heads rapidly to capture krill in midwater and fish (Pagothenia borchgrevinki) underneath the sea ice. Captures were remarkably fast (two krill per second in swarms) and efficient (244 krill or 33 P. borchgrevinki in 78–89 min). Prey captures were detected by the signal of head acceleration relative to body acceleration with high sensitivity and specificity (0.83–0.90), as shown by receiver-operating characteristic analysis. Extension of signal analysis to the entire behavioral records showed that krill captures were spatially and temporally more variable than P. borchgrevinki captures. Notably, the frequency distribution of krill capture rate closely followed a power-law model, indicating that the foraging success of penguins depends on a small number of very successful dives. The three steps illustrated here (i.e., video observations, linking video to behavioral signals, and extension of signal analysis) are unique approaches to understanding the spatial and temporal variability of ecologically important events such as foraging.

Foraging is of central importance in ecology because it determines the energy gain, and ultimately, the fitness of animals (1, 2). However, monitoring foraging is difficult, especially in marine ecosystems where direct observations are rarely possible. Although the development of data-logging tags has allowed recording detailed behavior of marine predators (3), monitoring prey captures remains a challenge. The most straightforward (and perhaps ultimate) method is to equip animals with video cameras and observe their foraging behavior directly from their perspectives. This method was previously applied to several species of marine animals to show what, where, and how they capture in the natural habitats (4–9). However, the recording periods of those studies were short (<6 h), and thus quantitative analyses, such as those of the spatial and temporal variability in prey captures, were rarely possible. Generally, video cameras are power-consuming (requiring large batteries) and the movie data obtained are large (requiring large memories). Consequently, building a tiny animal-borne video camera that works for long periods is currently unrealistic.

For this reason, several indirect parameters are frequently used as proxy for prey captures. These parameters include: (i) stomach, esophageal, or visceral temperature (10–13); (ii) mouth or beak opening as monitored by magnet sensors (14–16); and (iii) head or jaw movement monitored by accelerometers (17, 18). When changes in these parameters exceed a threshold—that is, when a signal is detected—then prey capture is considered to have occurred. The underlying assumptions of this approach are that a signal appears at a sufficiently high probability when prey is actually captured (i.e., the hit rate is high), and that signals mistakenly appear at a sufficiently low probability when no prey are captured (i.e., the false-alarm rate is low). Crucially, no previous studies have validated these assumptions under natural conditions. In most previous studies, the indirect parameters have been calibrated in tanks or aquariums, where captive animals were equipped with the sensors and fed with dead (sometimes live) prey (19–22). Although these studies concluded that the hit rate was high, predator–prey interactions are different in the wild, and the actual hit rate may be low. More importantly, the false-alarm rate cannot be examined using captive animals. For example, a drop in the stomach and esophageal temperature and the opening of a mouth may represent the gulping of water rather than prey capture. Similarly, head movements detected by an accelerometer may represent a searching motion rather than prey capture. As a different approach to the calibration of indirect signals, some studies (17, 23) recorded two signals simultaneously in free-ranging marine predators and showed that the information from the two signals were fairly consistent. However, both signals are still indirect, and the actual “answer” is unknown.

In this study, we took an integrated approach for monitoring prey capture by a predator. We attached a small video camera (33 g, lasting for about 85 min) to the back, and two tiny accelerometers (9 g, lasting for about 50 h) to the head and back of individual free-swimming Adélie penguins Pygoscelis adeliae. First, we observed from the movies what, where, and how the penguins captured prey. Second, we calculated the accelerations of a penguin’s head relative to its body by subtracting the acceleration recorded on the body from that recorded on the head. This “head-only acceleration” was assumed to provide prey-capture signals, because penguins presumably move their heads relative to their bodies when they catch prey, such as krill and fishes. The signals were validated using the movies, which provided the true “answer” for a fraction of the acceleration records. By measuring the hit rates and false-alarm rates for a range of thresholds in the head-only acceleration, we determined the optimal threshold (which considered the trade-off between hit rates and false-alarm rates) and evaluated the signal performance objectively. Third, keeping in mind the utility and limitation of the signals, we extended the signal analysis to each bird’s complete behavioral dataset (which covered an entire foraging trip for most birds) to examine the temporal variability of prey capture in the penguins.

Temporal variability in the prey captures of Adélie penguins is of particular interest, because the diets of this species include krill (Euphausia superba and Euphausia crystallorophias) and fishes (mainly Pagothenia borchgrevinki) (24), which are distinctly different in their distributions. Whereas krill form swarms in the water column and appear in a highly variable density [<4–804 individuals per cubic meter (25)], P. borchgrevinki may only form loose, ill-defined schools underneath the sea ice (26). Given that the foraging success of a predator depends on the distribution of prey (e.g., ref. 27), we expect that the temporal foraging success of the penguins is highly variable for krill, but rather stable for P. borchgrevinki. In theory, variation in foraging success is a risk that could affect fitness, and foraging animals are generally sensitive to this risk (1). Nevertheless, the temporal variability of foraging success of marine predators was rarely quantified, except in some studies that used indirect feeding sensors (12, 13). The frequency distribution of krill density, when examined by horizontal scanning of the water using echo sounders, follows a power-law model (28). This observation led us to hypothesize that the frequency distribution of krill capture rate by penguins also follows a power-law model. A support of this hypothesis could provide significant insights into the energy budget of penguins, because power-law frequency distribution is a characteristic distribution in which variation is so large that a typical or average value is not informative (29).

Adélie penguins breeding on the coast of Lützow-Holm Bay, Antarctica, are unusual for this species in that their foraging area is covered by fast ice. They dive (mostly <30 m) at more shallow depths than Adélie penguins foraging in the open sea (30) and at more shallow depths than many other breath-hold divers (31). Therefore, their feeding behavior could be filmed without using artificial light (which not only requires large equipment but may also affect the animals’ vision), making them an excellent model for the study of underwater foraging behavior.

Results

Data were recorded for 14 birds (mean body mass, 3.92 kg). Movies of 78–89 min (mean, 84 min) duration were recorded for 11 of the 14 birds. For three birds, the video cameras became faulty because of water leakage. See Movie S1 for an example of the footage taken from the penguin’s perspective. Simultaneous records of head acceleration, body acceleration, depth, and temperature of 25.3 ± 12.1 h duration (mean ± SD, from the initial dive to the last recorded dive) were obtained for all 13 birds to which two accelerometers were attached. The entire foraging trip was covered for 10 of the 13 birds. For the bird (52_44a) that lacked the accelerometer on the head, 49.5 h of body acceleration, depth, and temperature data were recorded.

Video Observations.

The movies showed that the penguins searched in the water column and moved the heads rapidly to capture krill individually (Fig. 1A and Movie S2). We filmed 367 definitive capture events, in which krill was observed as cleared by the darting movement of the penguin’s head. Krill was often seen as coming close toward the camera until being captured, which meant that the penguin made a head-on approach at its prey (example 1 in Movie S2). In some cases, krill was observed displaying an escape behavior by bending its body repeatedly to swim backward (example 2 in Movie S2). These observations suggest that krill is easy to catch after being detected. In addition, 96 probable capture events were filmed, in which the same type of head movement was observed during krill-feeding dives, but the krill could not be seen (probably because it was either out of the camera view or behind the penguin’s head). Three penguins (52_20b, 52_44a, and 52_51a) encountered a dense swarm of krill in five dives, in which they captured krill individually as fast as two krill per second (example 3 in Movie S2). Penguins often approached a krill by swimming upward, and then changing the swimming direction at the time of capture and swimming downward. This characteristic behavior resulted in many inflections on the depth profiles when penguins were feeding on krill (Fig. 2A).

Fig. 1.

Images from penguin-borne video cameras showing the timing of the capture of (A) a krill in midwater (bird 52_44a, depth 14.9 m) and (B) a P. borchgrevinki underneath the fast ice (bird 52_49a, depth 1.3 m). Note the accelerometer on the penguin’s head in B.

Fig. 2.

Examples of time-series depth and head-only acceleration (i.e., the vector sum of triaxial acceleration of the body subtracted from that of the head) during (A) a krill-feeding dive in midwater, shown by bird 52_51a and (B) a P. borchgrevinki-feeding dive underneath the fast ice, shown by bird 52_49a. Timing of prey capture inferred from head-only acceleration (“Signal” at the bottom) is comparable with that observed in the simultaneously recorded movie (“Answer” at the bottom), resulting in a hit, miss, or false alarm (denoted by h, m, and f, respectively, at the bottom). Gray bars represent surfacing periods, when high acceleration values are not related to feeding events.

Furthermore, the movies showed that penguins searched underneath the sea ice and captured P. borchgrevinki by extending their neck toward the ice (Fig. 1B and Movie S3). Thirty definitive capture events were filmed, in which we observed that the head, body, or tail of P. borchgrevinki was grabbed by penguins with their beaks or at least P. borchgrevinki was cleared by the head movement of the penguin. Although P. borchgrevinki is morphologically well camouflaged when seen from below against the underside of sea ice (32), penguins in the movies repeatedly captured P. borchgrevinki from below. Escape behavior of the fish was not evident in most cases, suggesting an excellent stealth approach by penguins (examples 1 and 2 in Movie S3). However, in two cases, they chased a P. borchgrevinki toward the underside of the sea ice and caught it there, as if they used the ice surface as a barrier (example 3 in Movie S3). In addition, 54 probable capture events were filmed, in which the same type of head movement toward the ice was observed, but the prey could not be seen (probably because it was either out of the camera view or behind the penguin’s head). Depth profile was characterized by a relatively flat bottom at about 2 m and a few inflections pointing toward the sea surface, which corresponded to the timing of captures (Fig. 2B).

Although these two types of prey captures (i.e., krill and P. borchgrevinki) were dominant (96% of all prey capture events observed), some other minor types of prey captures were also observed. Two birds (52_43b and 52_64a) searched along the seabed, and one of the two (52_43b) captured at least four fish (species unknown) there (Movie S4); however, the movies were darker during bottom-feeding dives, precluding our complete evaluation when the birds caught their prey. Bird 52_43b once encountered a school of fish, apparently Antarctic silver fish Pleuragramma antarcticum, of which the bird captured 14 individual fish in just 20 s (Movie S5). Bird 52_44a captured four amphipods (species unknown) individually near the seabed (Movie S6).

No unsuccessful attempt of prey capture was observed in any individual birds.

Signal Validations.

The vector sums of triaxial accelerations were noisy for both the head and body, because of various body motions, such as stroking activities. However, these noises were cancelled out in the head-only acceleration, and clear peaks were observed. Generally, peaks coincided well with the timing of prey captures, regardless of whether the prey was krill or P. borchgrevinki (Fig. 2). Bird 52_51a dived continuously in the water column during its 89-min movie record (total dive duration, 48 min) and displayed 244 krill-capture events, composed of 208 definitive captures (in which both darting movement of the penguin’s head and the presence of krill were seen) and 36 probable captures (in which only darting movement was seen). Bird 52_49a dived underneath the fast ice repeatedly during its 78-min movie record (total dive duration, 40 min) and displayed 33 P. borchgrevinki capture events, composed of 20 definitive captures (in which both the penguin’s neck extension toward the ice and the presence of a fish were seen) and 13 probable captures (in which only the penguin’s neck extension was seen). These two datasets allowed us to analyze the receiver-operating characteristics (33) of the acceleration signals for krill capture and P. borchgrevinki capture separately. The analysis showed highly convex curves for both krill and P. borchgrevinki, demonstrating an excellent performance of the head-only acceleration signals (Fig. 3A). The sum of sensitivity (= hit rate) and specificity (= 1 – false-alarm rate), a criterion to determine the optimal threshold (33), showed a sharp peak at the threshold of 0.25 g (1 g = 9.81 m s⁻²) for krill and a broad peak between threshold values of 0.25–0.45 g for P. borchgrevinki (Fig. 3B). Thus, the optimal threshold was determined to be 0.25 g, at which sensitivity was 0.83 and specificity was 0.85 for krill, and 0.88 and 0.90, respectively, for P. borchgrevinki. Because of the similar values of sensitivity and specificity at the optimal threshold, the total number of prey captures during the 78- to 89-min record of the two birds (52_51a and 52_49a), as inferred from the acceleration signal (237 krill and 31 fish, respectively), were close to the actual “answers” (244 krill and 33 fish, respectively). The threshold of 0.25 g appeared to be a good choice for other individual birds as well, although a small sample size of capture events precluded us from analyzing their receiver-operating characteristics.

Fig. 3.

(A) The receiver-operating characteristic plot for the signal of krill capture (open circle) and P. borchgrevinki capture (closed circle), based on 244 capture events recorded for bird 52_51a and 33 events recorded for bird 52_49a, respectively. Highly convex curves toward the upper left corner in the plot (which represents the ideal point, with 100% hit rate and 0% false-alarm rate) indicate an excellent performance of the signals. A random guess with no predictive power would give a point along the diagonal line. (B) The sum of sensitivity (= hit rate) and specificity (= 1 − false-alarm rate), potentially ranging from 1 (a random guess) to 2 (the ideal signal), plotted against the threshold in the head-only acceleration (g) for the signal of krill capture (open circle) and P. borchgrevinki capture (closed circle). Arrow represents the optimal threshold (0.25 g) determined.

Many signals appeared when bird 52_43b searched on the seabed (Fig. S1). Although the movie taken during these bottom-feeding dives was often too dark to determine the timing of prey captures, most signals were apparently false alarms. Signals also appeared when bird 52_43b captured 14 individuals of P. antarcticum repeatedly, with a satisfactorily high hit rate (12 of 14) and low false-alarm rate (2 of 14). However, sample size was too small to determine the characteristics of diving behavior that focused on P. antarcticum. Therefore, we were not able to detect the capture events of P. antarcticum from the behavioral records that lacked simultaneous movie records. Head-only acceleration data were not available for amphipod capture events, which were shown by the bird (52_44a) that lacked the accelerometer on the head.

Extension of Signal Analysis.

A total of 1,017 krill-feeding dives, with a mean (± SD) dive depth and duration of 29.3 ± 19.7 m and 97 ± 38 s, respectively, and 1,161 P. borchgrevinki-feeding dives, with a mean dive depth and duration of 6.3 ± 5.0 m and 78 ± 27 s, respectively, were extracted from the entire behavioral records of 13 birds. At the optimal threshold determined above (0.25 g), a total of 11,028 krill capture events and 1,049 P. borchgrevinki capture events were detected. Krill were captured at broad depth ranges (<80 m) with a peak at 10 m, whereas P. borchgrevinki were captured at narrow depth ranges (<5 m), apparently affected by the fast-ice thickness (Fig. 4). The number of krill captured in a dive varied greatly (range, 0–61) and that in a hypothetical 100-s dive varied to an even greater degree (range, 0–98) (Fig. 5A). In contrast, the number of P. borchgrevinki captured in a dive fell within a narrow range (0–6); similarly, the number captured in a hypothetical 100-s dive fell within a narrow range (0–9) (Fig. 5B). A sample of krill-feeding dives large enough to test the power-law frequency distribution hypothesis was collected for five birds (Table 1). For four (52_20b, 52_30a, 52_35a, and 52_51a) of the five birds, the truncated power-law model fit better than the exponential model based on Akaike weights (Fig. 5A and Table 1). For the fifth bird (52_44b), neither model fit well.

Fig. 4.

Frequency distribution of the depth at which krill (Left) and P. borchgrevinki (Right) were captured, as inferred from the acceleration signal. Data from 13 birds were pooled.

Fig. 5.

Frequency distribution of the number of prey captured in a dive (standardized for a hypothetical 100 s dive) for (A) bird 52_30a feeding on krill, and (B) bird 52_28a feeding on P. borchgrevinki, as inferred from the acceleration signal. (Inset in A The log-log rank frequency plot based on the same data, with black circles representing values ≥x_min and gray circles representing values <x_min (x_min denotes the start of tail in the data). The model fits are truncated power-law (red) and exponential (blue) distributions.

Table 1.

Summary of model fitting

				Exponent		wAIC		Model
Bird ID	n*	xmin	xmax	TP	EX	TP	EX	Supported
52_20b	51	9	58	0.73	0.051	0.999	0.001	TP
52_30a	116	4	42	2.21	0.168	1.000	0.000	TP
52_35a	143	6	98	1.22	0.046	0.995	0.005	TP
52_44b	82	8	69	—	—	—	—	—
52_51a	88	7	35	2.00	0.144	0.712	0.288	TP

EX, exponential model; TP, truncated power-law model; wAIC, Akaike weight; x_max, tail end; x_min, tail start.

*Number of dives with prey capture values between x_min and x_max.

Stomach Contents.

Stomach contents (mean, 328 g; range 23–468 g by wet mass) sampled from 10 birds showed that their main prey were krill (E. superba and E. crystallorophias; mean, 38%; range, 0–100% by wet mass) and P. borchgrevinki (mean, 62%; range, 0–100% by wet mass). The degree of digestion varied from specimen to specimen for both krill and P. borchgrevinki. Based on relatively fresh 15 specimen of P. borchgrevinki, the mean fork length was 6 cm (range, 5–12 cm). A small number of amphipods (mean, 0.2%; range, 0–0.8% by wet mass) were also found.

Discussion

Although the large potential of animal-borne videos in ecological studies is well acknowledged (34, 35), video cameras are not only power- and memory-consuming, but also need to be small enough to be deployed on animals, thus resulting in severely limited recording durations (typically <6 h) that are often insufficient to give ecologically meaningful datasets. We demonstrate that this problem can be overcome by linking video to simultaneously recorded behavioral parameters. Our approach can be summarized in the following three steps: (i) Observations: directly observing the behavior in question from movies; (ii) Link: linking the observed behavior to particular behavioral signals; and (iii) Extension: extending the signal analysis to the complete behavioral records. We suggest that this three-step approach can be broadly applied in many animal species to examine the spatial and temporal variability of ecologically important events such as foraging.

Observations: The First Step.

Our movies showed that Adélie penguins feed on (i) krill in the water column, (ii) P. borchgrevinki underneath the fast ice, (iii) some fishes on the seabed, (iv) P. antarcticum in the water column, and (v) amphipods near the seabed, listed in the order of their importance. These variations agree with the diet analysis in the present and previous study (24) conducted in the same study area. The forging behavior shown in (ii) was previously reported for the emperor penguins Aptenodytes forsteri that dived through a man-made isolated ice hole with a video camera attached (5).

In addition, our movies showed that the foraging behavior of Adélie penguins is remarkably fast and efficient. Two krill were captured within a second in dense swarms; 244 krill and 33 P. borchgrevinki were, respectively, captured in an 89-min and 78-min recording period (total dive duration, 48 min and 40 min). Thirty-six of the 244 krill and 13 of the 33 P. borchgrevinki were not observed in the movies but were considered as captured because of the characteristic body motions of the penguins. A part of these events may represent unsuccessful attempts; however, unsuccessful attempts were not observed throughout the movies, indicating that they are rare. Considering the average size of krill previously collected from Adélie penguin stomach contents [41-mm body length (36), corresponding to 0.4 g body mass (37)], 244 krill might weigh 98 g. Considering the average P. borchgrevinki size from the stomach contents collected in this study [6-cm fork length, corresponding to 5-g body mass when extrapolated from the data of larger individuals (38)], 33 P. borchgrevinki might weigh 165 g. Collecting 98–165 g of food in 78–89 min (or in 40–48 min total dive duration) is reasonable for Adélie penguins breeding chicks, given that they spend 3.5–5.1 h diving per day on average (39) and, as suggested by an energetic model (40), require 820 g of krill per day.

Link: The Second Step.

We quantitatively link the foraging behavior observed in the movies to a behavioral parameter, that is, the head-only acceleration. This study is unique in that it validated a presumed signal of prey captures under natural conditions. Our receiver-operating characteristic analysis showed that the head-only acceleration is an excellent detector of capture events of krill and P. borchgrevinki, with a sensitivity and specificity of 0.83–0.90. This finding is not surprising, considering that penguins move their head rapidly when they catch krill or P. borchgrevinki, as observed from the movies. Compared with krill capture, P. borchgrevinki capture showed a more highly convex curve on the receiver-operating characteristic plot (Fig. 3A), higher values of sensitivity and specificity (Fig. 3B), and lower sensitivity to threshold (Fig. 3B). All these results indicate that P. borchgrevinki capture is easier to detect than krill capture. This result is also not surprising, considering that penguins show distinct body motions when feeding on P. borchgrevinki than when feeding on krill, as observed from the movies.

Head-only acceleration allowed us to detect individual krill-capture events, which were as fast as two events per second when a patch was encountered. Such high encounter rate cannot be detected by stomach, esophagus, or visceral temperature sensors, which are the current most widely used methods (10–13). The beak-opening method appears to have enough temporal resolution to detect individual krill-capture events (15, 16); however, this method has disadvantages, such as complex procedures for deployment and frequent failure because of breakage of the cable between the sensor on the beak and the logger on the back (23). High temporal resolutions, ease of attachment, and absence of a cable are major advantages of our accelerometer method over the temperature and beak-opening methods.

However, we found that the acceleration method did not work for bottom-feeding dives, although such dives formed a minor component of foraging behavior. Because penguins moved their heads when searching on the seabed, many false alarms appeared during those periods. This result highlights the importance of validating a feeding sensor for a range of prey types under natural conditions. In many previous studies (11–13, 16–18, 23), a single set of criteria was used to extract all possible prey captures from the records; however, such an approach might be risky. Foraging behaviors can be diverse and should be well understood before a feeding sensor is applied.

Extension: The Last Step.

We extended the signal analysis to the complete behavioral records and showed that the spatial and temporal variability in the foraging success of penguins are distinctly different when feeding on krill and P. borchgrevinki. Regarding spatial variability (although we are looking at the vertical dimension only), krill were captured at wide depth ranges (<80 m), but P. borchgrevinki were always captured at depths of <5 m (i.e., underneath the fast ice). Regarding temporal variability, the number of krill captured in a dive was highly variable (range, 0–61 individuals), but the number of P. borchgrevinki captured in a dive was more stable (range, 0–6 individuals).

Notably, we showed that the number of krill captured in a 100-s dive closely followed a power-law distribution, a model that was observed for the frequency distribution of krill density that occurs in nature (28). The range of our krill-capture data was relatively narrow for power-law model fitting. Nevertheless, this result is important because optimal foraging models for diving animals frequently assume prey encounter rates derived from the Poisson or normal distribution (41, 42). As recently suggested for fishes (43), feeding opportunities of diving animals might be more heterogeneous than most ecologists assume. Power-law distribution is fundamentally different from normal distribution in that the variation is so large that a typical or average value is not informative (29). This difference indicates that the success of penguins feeding on krill during a foraging trip depends on a small number of very successful dives, rather than a number of typical dives. Our results might explain a paradoxical previous report that time spent diving by breeding Adélie penguins does not affect brood growth rates (39), emphasizing the importance of recording foraging success rather than foraging duration.

Adélie penguins in our study area are subjected to high variability in foraging success or a high risk (1) when they feed on krill in the water column, but are subjected to a relatively low variability when they feed on P. borchgrevinki underneath the fast ice. A laboratory experiment showed that birds prefer highly variable resources when they are in a negative energy budget, but prefer less variable resources when they are in a positive energy budget (44). When predators make foraging decisions such as these based on their energy states, it is thought to help them avoid starvation, and hence increase fitness (1). Further investigations are required to confirm whether the penguins in the present study follow this scenario. Having access to both types of prey is unique for the Adélie penguins breeding in ice-covered areas, and apparently advantageous. This advantage might be offset by the disadvantages incurred by the presence of fast ice, including difficulty in finding cracks and long time and high energy expended for walking on the ice.

In conclusion, using Adélie penguins as a model, we show: (i) the direct observations of underwater prey captures, (ii) the validation of a detector of prey-capture events in the field, and (iii) the temporal variability of prey captures for two different prey types. The three-step approach described in this study (i.e., video observations, linking video to behavioral signals, and extension of signal analysis) can be used to examine the spatial and temporal variability in prey captures and possibly other events over a long period, and thus, has a large potential for future use in ecology.

Materials and Methods

Fieldwork.

We made a fieldwork at the Hukuro Cove colony (69°13′S, 39°38′E) in Lützow-Holm Bay, Antarctica, from late December in 2010 to early February in 2011. All of the experimental procedures were approved by Ministry of the Environment, Japan. Fifteen Adélie penguins, rearing one or two chicks, were equipped with a video camera (on the back) and two accelerometers (one on the back and the other on the head), except for one individual (52_44a) that lacked the accelerometer on the head. However, the instruments on one individual could not be recovered, and the final sample size was 14. In addition, stomach contents were collected from 10 birds using the standard stomach-flushing method (45), and prey species was determined. Two birds equipped with accelerometers and videos (52_35a and 52_44b) were included in these 10 birds. See SI Materials and Methods for details of the experimental procedures and the instruments equipped.

Data Analysis.

For both head and body accelerations, the values along triaxis were summed vectorially [√(x² + y² + z²)] and smoothed using a low-pass filter to remove high-frequency noises. The smoothed vector sum of the body acceleration was then subtracted from that of the head acceleration to obtain the head-only acceleration (g), a possible signal of prey capture events. To determine the optimal threshold of head-only acceleration and to quantitatively evaluate the performance of the signal, the concept of receiver-operating characteristic (33) was applied to our datasets with actual “answers” seen in the movies. With the optimal threshold determined, signal analysis was extended to the complete behavioral record for each bird, which overlapped the movies only partially. The acceleration signals were found to correlate well for the capture events of krill and P. borchgrevinki, but not for other minor types of prey captures (Results). Therefore, using acceleration signals to extract all possible prey captures would lead to inaccurate results; hence, we focused on the capture events of krill and P. borchgrevinki. See SI Materials and Methods for details of the receiver-operating characteristic analysis and how we extended the signal analysis.

Model Fitting.

The hypothesis that the frequency distribution of krill capture rate follows a power-law model was tested by using maximum-likelihood estimation and Akaike Information Criteria (46–49). As a parameter for model fitting, we used the number of krill captured in a dive. However, to remove the effect of variable dive duration, the number of krill captured in a dive was standardized to the number of krill captured in a hypothetical 100-s dive, by dividing the number of krill captured in a dive by the duration of the dive (s) and multiplying by 100; 100 s was chosen as it was close to the mean dive duration of krill-feeding dives (97 s). Note that the number of krill, a discrete number, was converted into a continuous number by this manipulation, thus allowing us to avoid the difficulty in dealing with a discrete number in power-law model fitting (29, 47). See SI Materials and Methods for details of the model fitting.